Applied Econometrics Using Matlab

Applied Econometrics using MATLAB
James P. LeSage
Department of Economics
University of Toledo
October, 1999
.::UdecomBooks::.
Preface
This text describes a set of MATLAB functions that implement a host of

econometric estimation methods. Toolboxes are the name given by the
MathWorks to related sets of MATLAB functions aimed at solving a par-
ticular class of problems. Toolboxes of functions useful in signal processing,
optimization, statistics, finance and a host of other areas are available from
the MathWorks as add-ons to the standard MATLAB software distribution.
I use the term Econometrics Toolbox to refer to the collection of function
libraries described in this book.
The intended audience is faculty and students using statistical methods,
whether they are engaged in econometric analysis or more general regression
modeling. The MATLAB functions described in this book have been used
in my own research as well as teaching both undergraduate and graduate
econometrics courses. Researchers currently using Gauss, RATS, TSP, or
SAS/IML for econometric programming might find switching to MATLAB
advantageous. MATLAB software has always had excellent numerical algo-
rithms, and has recently been extended to include: sparse matrix algorithms,
very good graphical capabilities, and a complete set of object oriented and
graphical user-interface programming tools. MATLAB software is available
on a wide variety of computing platforms including mainframe, Intel, Apple,
and Linux or Unix workstations.
When contemplating a change in software, there is always the initial
investment in developing a set of basic routines and functions to support
econometric analysis. It is my hope that the routines in the Econometrics
Toolbox provide a relatively complete set of basic econometric analysis tools.
The toolbox also includes a number of functions to mimic those available
in Gauss, which should make converting existing Gauss functions and ap-
plications easier. For those involved in vector autoregressive modeling, a
complete set of estimation and forecasting routines is available that imple-
ment a wider variety of these estimation methods than RATS software. For
example, Bayesian Markov Chain Monte Carlo (MCMC) estimation of VAR
.::UdecomBooks::.
ii
models that robustify against outliers and accommodate heteroscedastic dis-

turbances have been implemented. In addition, the estimation functions for
error correction models (ECM) carry out Johansen’s tests to determine the
number of cointegrating relations, which are automatically incorporated in
the model. In the area of vector autoregressive forecasting, routines are
available for VAR and ECM methods that automatically handle data trans-
formations (e.g. differencing, seasonal differences, growth rates). This allows
users to work with variables in raw levels form. The forecasting functions
carry out needed transformations for estimation and return forecasted values
in level form. Comparison of forecast accuracy from a wide variety of vector
autoregressive, error correction and other methods is quite simple. Users
can avoid the difficult task of unraveling transformed forecasted values from
alternative estimation methods and proceed directly to forecast accuracy
comparisons.
The collection of around 300 functions and demonstration programs are
organized into libraries that are described in each chapter of the book. Many
faculty use MATLAB or Gauss software for research in econometric analysis,
but the functions written to support research are often suitable for only a
single problem. This is because time and energy (both of which are in short
supply) are involved in writing more generally applicable functions. The
functions described in this book are intended to be re-usable in any number
of applications. Some of the functions implement relatively new Markov
Chain Monte Carlo (MCMC) estimation methods, making these accessible
to undergraduate and graduate students with absolutely no programming
involved on the students part. Many of the automated features available in
the vector autoregressive, error correction, and forecasting functions arose
from my own experience in dealing with students using these functions. It
seemed a shame to waste valuable class time on implementation details when
these can be handled by well-written functions that take care of the details.
A consistent design was implemented that provides documentation, ex-
ample programs, and functions to produce printed as well as graphical pre-
sentation of estimation results for all of the econometric functions. This
was accomplished using the “structure variables” introduced in MATLAB
Version 5. Information from econometric estimation is encapsulated into a
single variable that contains “fields” for individual parameters and statistics
related to the econometric results. A thoughtful design by the MathWorks
allows these structure variables to contain scalar, vector, matrix, string,
and even multi-dimensional matrices as fields. This allows the econometric
functions to return a single structure that contains all estimation results.
These structures can be passed to other functions that can intelligently de-
.::UdecomBooks::.
iii
cipher the information and provide a printed or graphical presentation of

the results.
The Econometrics Toolbox should allow faculty to use MATLAB in un-
dergraduate and graduate level econometrics courses with absolutely no pro-
gramming on the part of students or faculty. An added benefit to using
MATLAB and the Econometrics Toolbox is that faculty have the option of
implementing methods that best reflect the material in their courses as well
as their own research interests. It should be easy to implement a host of ideas
and methods by: drawing on existing functions in the toolbox, extending
these functions, or operating on the results from these functions. As there is
an expectation that users are likely to extend the toolbox, examples of how
to accomplish this are provided at the outset in the first chapter. Another
way to extend the toolbox is to download MATLAB functions that are avail-
able on Internet sites. (In fact, some of the routines in the toolbox originally
came from the Internet.) I would urge you to re-write the documentation
for these functions in a format consistent with the other functions in the
toolbox and return the results from the function in a “structure variable”.
A detailed example of how to do this is provided in the first chapter.
In addition to providing a set of econometric estimation routines and doc-
umentation, the book has another goal. Programming approaches as well as
design decisions are discussed in the book. This discussion should make it
easier to use the toolbox functions intelligently, and facilitate creating new
functions that fit into the overall design, and work well with existing toolbox
routines. This text can be read as a manual for simply using the existing
functions in the toolbox, which is how students tend to approach the book.
It can also be seen as providing programming and design approaches that
will help implement extensions for research and teaching of econometrics.
This is how I would think faculty would approach the text. Some faculty
in Ph.D. programs expect their graduate students to engage in econometric
problem solving that requires programming, and certainly this text would
eliminate the burden of spending valuable course time on computer pro-
gramming and implementation details. Students in Ph.D. programs receive
the added benefit that functions implemented for dissertation work can be
easily transported to another institution, since MATLAB is available for
almost any conceivable hardware/operating system environment.
Finally, there are obviously omissions, bugs and perhaps programming
errors in the Econometrics Toolbox. This would likely be the case with any
such endeavor. I would be grateful if users would notify me when they en-
counter problems. It would also be helpful if users who produce generally
useful functions that extend the toolbox would submit them for inclusion.
.::UdecomBooks::.
iv
Much of the econometric code I encounter on the internet is simply too

specific to a single research problem to be generally useful in other appli-
cations. If econometric researchers are serious about their newly proposed
estimation methods, they should take the time to craft a generally useful
MATLAB function that others could use in applied research. Inclusion in
the Econometrics Toolbox would also have the benefit of introducing the
method to faculty teaching econometrics and their students.
The latest version of the Econometrics Toolbox functions can be found on
the Internet at: http://www.econ.utoledo.edu under the MATLAB gallery
icon. Instructions for installing these functions are in an Appendix to this
text along with a listing of the functions in the library and a brief description
of each.
.::UdecomBooks::.
Acknowledgements
First I would like to acknowledge my students who suffered through previous

buggy and error filled versions of the toolbox functions. I owe a great deal to
their patience and persistence in completing projects under such conditions.
I would like to thank Michael R. Dowd would was brave enough to use
early versions of the vector autoregressive library routines in his research.
In addition to providing feedback regarding the functions and reading early
drafts of the text, Mike also provided LaTeX assistance to produce this text.
David C. Black is another faculty colleague who tested many of the functions
and read early drafts of the text. Michael Magura who has always been a
great editor and helpful critic also read early drafts of the text.
Some of the functions in the Econometrics Toolbox represent algorithms
crafted by others who unselfishly placed them in the public domain on the
Internet. The authors names and affiliations have been preserved in these
functions. In all cases, the documentation was modified to match the format
of the Econometrics Toolbox and in many cases, the arguments returned
by the function were converted to MATLAB structure variables. In other
cases, functions written in other languages were converted to the MATLAB
language. I am indebted to those who implemented these routines as well.
The following persons have generously contributed their functions for
inclusion in the Econometrics Toolbox. C. F. Baum from Boston College,
Mike Cliff from UNC, Matt Hergott a Duke B-school graduate and Tony
E. Smith from University of Pennsylvania. Tony also spent a great deal of
time looking at the spatial econometrics functions and reporting problems
and bugs.
Many others have sent e-mail comments regarding errors and I am grate-
ful to all of you.
.::UdecomBooks::.
Contents
1 Introduction 1
2 Regression using MATLAB 5

2.1 Design of the regression library . . . . . . . . . . . . . . . . . 6
2.2 The ols function . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Selecting a least-squares algorithm . . . . . . . . . . . . . . . 12
2.4 Using the results structure . . . . . . . . . . . . . . . . . . . . 17
2.5 Performance profiling the regression toolbox . . . . . . . . . . 28
2.6 Using the regression library . . . . . . . . . . . . . . . . . . . 30
2.6.1 A Monte Carlo experiment . . . . . . . . . . . . . . . 31
2.6.2 Dealing with serial correlation . . . . . . . . . . . . . 32
2.6.3 Implementing statistical tests . . . . . . . . . . . . . . 38
2.7 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 41
Chapter 2 Appendix 42
3 Utility Functions 45
3.1 Calendar function utilities . . . . . . . . . . . . . . . . . . . . 45
3.2 Printing and plotting matrices . . . . . . . . . . . . . . . . . 49
3.3 Data transformation utilities . . . . . . . . . . . . . . . . . . 65
3.4 Gauss functions . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5 Wrapper functions . . . . . . . . . . . . . . . . . . . . . . . . 73
3.6 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 76
4 Regression Diagnostics 80
4.1 Collinearity diagnostics and procedures . . . . . . . . . . . . 80
4.2 Outlier diagnostics and procedures . . . . . . . . . . . . . . . 94
4.3 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 100
vi
.::UdecomBooks::.
CONTENTS vii
5 VAR and Error Correction Models 103

5.1 VAR models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.2 Error correction models . . . . . . . . . . . . . . . . . . . . . 113
5.3 Bayesian variants . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.3.1 Theil-Goldberger estimation of these models . . . . . 138
5.4 Forecasting the models . . . . . . . . . . . . . . . . . . . . . . 139
5.5 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 145
6 Markov Chain Monte Carlo Models 151

6.1 The Bayesian Regression Model . . . . . . . . . . . . . . . . . 154
6.2 The Gibbs Sampler . . . . . . . . . . . . . . . . . . . . . . . . 156
6.2.1 Monitoring convergence of the sampler . . . . . . . . . 159
6.2.2 Autocorrelation estimates . . . . . . . . . . . . . . . . 163
6.2.3 Raftery-Lewis diagnostics . . . . . . . . . . . . . . . . 163
6.2.4 Geweke diagnostics . . . . . . . . . . . . . . . . . . . . 165
6.3 A heteroscedastic linear model . . . . . . . . . . . . . . . . . 169
6.4 Gibbs sampling functions . . . . . . . . . . . . . . . . . . . . 175
6.5 Metropolis sampling . . . . . . . . . . . . . . . . . . . . . . . 184
6.6 Functions in the Gibbs sampling library . . . . . . . . . . . . 190
6.7 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 197
7 Limited Dependent Variable Models 204

7.1 Logit and probit regressions . . . . . . . . . . . . . . . . . . . 206
7.2 Gibbs sampling logit/probit models . . . . . . . . . . . . . . . 211
7.2.1 The probit g function . . . . . . . . . . . . . . . . . . 218
7.3 Tobit models . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
7.4 Gibbs sampling Tobit models . . . . . . . . . . . . . . . . . . 224
7.5 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 227
8 Simultaneous Equation Models 230

8.1 Two-stage least-squares models . . . . . . . . . . . . . . . . . 230
8.2 Three-stage least-squares models . . . . . . . . . . . . . . . . 235
8.3 Seemingly unrelated regression models . . . . . . . . . . . . . 240
.::UdecomBooks::.
CONTENTS viii
8.4 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 244
9 Distribution functions library 247

9.1 The pdf, cdf, inv and rnd functions . . . . . . . . . . . . . . . 248
9.2 The specialized functions . . . . . . . . . . . . . . . . . . . . 249
9.3 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 256
10 Optimization functions library 260

10.1 Simplex optimization . . . . . . . . . . . . . . . . . . . . . . . 261
10.1.1 Univariate simplex optimization . . . . . . . . . . . . 261
10.1.2 Multivariate simplex optimization . . . . . . . . . . . 268
10.2 EM algorithms for optimization . . . . . . . . . . . . . . . . . 269
10.3 Multivariate gradient optimization . . . . . . . . . . . . . . . 278
10.4 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 287
11 Handling sparse matrices 289

11.1 Computational savings with sparse matrices . . . . . . . . . . 289
11.2 Estimation using sparse matrix algorithms . . . . . . . . . . . 297
11.3 Gibbs sampling and sparse matrices . . . . . . . . . . . . . . 304
11.4 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 309
References 313
Appendix: Toolbox functions 320
.::UdecomBooks::.
List of Examples
1.1 Demonstrate regression using the ols() function . . . . . . . . 8

2.2 Least-squares timing information . . . . . . . . . . . . . . . . 12
2.3 Profiling the ols() function . . . . . . . . . . . . . . . . . . . . 28
2.4 Using the ols() function for Monte Carlo . . . . . . . . . . . . 31
2.5 Generate a model with serial correlation . . . . . . . . . . . . 33
2.6 Cochrane-Orcutt iteration . . . . . . . . . . . . . . . . . . . . 34
2.7 Maximum likelihood estimation . . . . . . . . . . . . . . . . . 37
2.8 Wald’s F-test . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.9 LM specification test . . . . . . . . . . . . . . . . . . . . . . . 40
3.1 Using the cal() function . . . . . . . . . . . . . . . . . . . . . 47
3.2 Using the tsdate() function . . . . . . . . . . . . . . . . . . . 48
3.3 Using cal() and tsdates() functions . . . . . . . . . . . . . . . 49
3.4 Reading and printing time-series . . . . . . . . . . . . . . . . 49
3.5 Using the lprint() function . . . . . . . . . . . . . . . . . . . . 51
3.6 Using the tsprint() function . . . . . . . . . . . . . . . . . . . 56
3.7 Various tsprint() formats . . . . . . . . . . . . . . . . . . . . . 58
3.8 Truncating time-series and the cal() function . . . . . . . . . 60
3.9 Common errors using the cal() function . . . . . . . . . . . . 60
3.10 Using the tsplot() function . . . . . . . . . . . . . . . . . . . 61
3.11 Finding time-series turning points with fturns() . . . . . . . . 64
3.12 Seasonal differencing with sdiff() function . . . . . . . . . . . 67
3.13 Annual growth rates using growthr() function . . . . . . . . 68
3.14 Seasonal dummy variables using sdummy() function . . . . . 68
3.15 Least-absolute deviations using lad() function . . . . . . . . 71
4.1 Collinearity experiment . . . . . . . . . . . . . . . . . . . . . 82
4.2 Using the bkw() function . . . . . . . . . . . . . . . . . . . . 86
4.3 Using the ridge() function . . . . . . . . . . . . . . . . . . . . 88
4.4 Using the rtrace() function . . . . . . . . . . . . . . . . . . . 90
4.5 Using the theil() function . . . . . . . . . . . . . . . . . . . . 92
4.6 Using the dfbeta() function . . . . . . . . . . . . . . . . . . . 94
ix
.::UdecomBooks::.
LIST OF EXAMPLES x
4.7 Using the robust() function . . . . . . . . . . . . . . . . . . . 97

4.8 Using the pairs() function . . . . . . . . . . . . . . . . . . . . 99
5.1 Using the var() function . . . . . . . . . . . . . . . . . . . . . 105
5.2 Using the pgranger() function . . . . . . . . . . . . . . . . . . 107
5.3 VAR with deterministic variables . . . . . . . . . . . . . . . . 110
5.4 Using the lrratio() function . . . . . . . . . . . . . . . . . . . 111
5.5 Impulse response functions . . . . . . . . . . . . . . . . . . . . 112
5.6 Using the adf() and cadf() functions . . . . . . . . . . . . . . 116
5.7 Using the johansen() function . . . . . . . . . . . . . . . . . . 119
5.8 Estimating error correction models . . . . . . . . . . . . . . . 123
5.9 Estimating BVAR models . . . . . . . . . . . . . . . . . . . . 128
5.10 Using bvar() with general weights . . . . . . . . . . . . . . . 130
5.11 Estimating RECM models . . . . . . . . . . . . . . . . . . . . 137
5.12 Forecasting VAR models . . . . . . . . . . . . . . . . . . . . . 140
5.13 Forecasting multiple related models . . . . . . . . . . . . . . 141
5.14 A forecast accuracy experiment . . . . . . . . . . . . . . . . . 143
6.1 A simple Gibbs sampler . . . . . . . . . . . . . . . . . . . . . 157
6.2 Using the coda() function . . . . . . . . . . . . . . . . . . . . 160
6.3 Using the raftery() function . . . . . . . . . . . . . . . . . . . 165
6.4 Geweke’s convergence diagnostics . . . . . . . . . . . . . . . . 166
6.5 Using the momentg() function . . . . . . . . . . . . . . . . . . 168
6.6 Heteroscedastic Gibbs sampler . . . . . . . . . . . . . . . . . 173
6.7 Using the ar g() function . . . . . . . . . . . . . . . . . . . . . 182
6.8 Metropolis within Gibbs sampling . . . . . . . . . . . . . . . 188
6.9 Bayesian model averaging with the bma g() function . . . . . 192
7.1 Logit and probit regression functions . . . . . . . . . . . . . . 206
7.2 Demonstrate Albert-Chib latent variable . . . . . . . . . . . . 212
7.3 Gibbs vs. maximum likelihood logit . . . . . . . . . . . . . . 214
7.4 Gibbs vs. maximum likelihood probit . . . . . . . . . . . . . . 215
7.5 Heteroscedastic probit model . . . . . . . . . . . . . . . . . . 217
7.6 Tobit regression function . . . . . . . . . . . . . . . . . . . . . 222
7.7 Gibbs sampling tobit estimation . . . . . . . . . . . . . . . . 225
8.1 Two-stage least-squares . . . . . . . . . . . . . . . . . . . . . 231
8.2 Monte Carlo study of ols() vs. tsls() . . . . . . . . . . . . . . 234
8.3 Three-stage least-squares . . . . . . . . . . . . . . . . . . . . 237
8.4 Using the sur() function . . . . . . . . . . . . . . . . . . . . . 242
9.1 Beta distribution function example . . . . . . . . . . . . . . . 248
9.2 Random Wishart draws . . . . . . . . . . . . . . . . . . . . . 252
9.3 Left- and right-truncated normal draws . . . . . . . . . . . . 253
9.4 Rejection sampling of truncated normal draws . . . . . . . . . 254
.::UdecomBooks::.
LIST OF EXAMPLES xi
10.1 Simplex maximum likelihood for Box-Cox model . . . . . . . 265

10.2 EM estimation of switching regime model . . . . . . . . . . . 275
10.3 Maximum likelihood estimation of the Tobit model . . . . . 281
10.4 Using the optsolv() function . . . . . . . . . . . . . . . . . . 286
11.1 Using sparse matrix functions . . . . . . . . . . . . . . . . . 291
11.2 Solving a sparse matrix system . . . . . . . . . . . . . . . . 292
11.3 Symmetric minimum degree ordering operations . . . . . . . 295
11.4 Solving for rho using the far() function . . . . . . . . . . . . 303
11.5 Gibbs sampling with sparse matrices . . . . . . . . . . . . . 307
.::UdecomBooks::.
List of Figures
2.1 Histogram of β̂ outcomes . . . . . . . . . . . . . . . . . . . . 33
3.1 Output from tsplot() function . . . . . . . . . . . . . . . . . . 62

3.2 Graph of turning point events . . . . . . . . . . . . . . . . . . 66
4.1 Ridge trace plot . . . . . . . . . . . . . . . . . . . . . . . . . . 91

4.2 Dfbeta plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3 Dffits plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4 Pairwise scatter plots . . . . . . . . . . . . . . . . . . . . . . . 99
5.1 Impulse response graphs . . . . . . . . . . . . . . . . . . . . . 113

5.2 Prior means and precision for important variables . . . . . . . 136
5.3 Prior means and precision for unimportant variables . . . . . 137
6.1 Prior Vi distributions for various values of r . . . . . . . . . . 171

6.2 Mean of Vi draws . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.3 Distribution of φ1 + φ2 . . . . . . . . . . . . . . . . . . . . . . 184
6.4 |In − ρW | as a function of ρ . . . . . . . . . . . . . . . . . . . 187
7.1 Cumulative distribution functions compared . . . . . . . . . . 206

7.2 Actual y vs. mean of latent y-draws . . . . . . . . . . . . . . 213
7.3 Posterior mean of vi draws with outliers . . . . . . . . . . . . 219
9.1 Beta distribution demonstration program plots . . . . . . . . 250

9.2 Histograms of truncated normal distributions . . . . . . . . . 253
9.3 Contaminated normal versus a standard normal distribution . 255
10.1 Plots from switching regime regression . . . . . . . . . . . . . 277
11.1 Sparsity structure of W from Pace and Berry . . . . . . . . . 291

11.2 An illustration of fill-in from matrix multiplication . . . . . . 294
xii
.::UdecomBooks::.
LIST OF FIGURES xiii
11.3 Minimum degree ordering versus unordered Pace and Barry

matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
.::UdecomBooks::.
List of Tables
2.1 Timing (in seconds) for Cholesky and QR Least-squares . . . 13

2.2 Digits of accuracy for Cholesky vs. QR decomposition . . . . 16
4.1 Variance-decomposition proportions table . . . . . . . . . . . 84

4.2 BKW collinearity diagnostics example . . . . . . . . . . . . . 85
4.3 Ridge Regression for our Monte Carlo example . . . . . . . . 88
6.1 A Comparison of FAR Estimators . . . . . . . . . . . . . . . 190
xiv
.::UdecomBooks::.
Chapter 1
Introduction
The Econometrics Toolbox contains around 50 functions that implement

econometric estimation procedures, 20 functions to carry out diagnostic
and statistical testing procedures, and 150 support and miscellaneous util-
ity functions. In addition, there are around 100 demonstration functions
covering all of the econometric estimation methods as well as the diagnostic
and testing procedures and many of the utility functions. Any attempt to
describe this amount of code must be organized.
Chapter 2 describes the design philosophy and mechanics of implementa-
tion using least-squares regression. Because regression is widely-understood
by econometricians and others working with statistics, the reader should be
free to concentrate on the ‘big picture’. Once you understand the way that
the Econometric Toolbox functions encapsulate estimation and other results
in the new MATLAB Version 5 ‘structure variables’, you’re on the way to
successfully using the toolbox.
Despite the large (and always growing) number of functions in the tool-
box, you may not find exactly what you’re looking for. From the outset,
examples are provided that illustrate how to incorporate your own functions
in the toolbox in a well-documented, consistent manner. Your functions
should be capable of using the existing printing and graphing facilities to
provide printed and graphical display of your results.
Chapters 3 through 10 focus more directly on description of functions
according to their econometric purpose. These chapters can be read as
merely a software documentation manual, and many beginning students
approach the text in this manner. Another approach to the text is for
those interested in MATLAB programming to accomplish research tasks.
Chapters 3 through 10 describe various design challenges regarding the task
.::UdecomBooks::.
CHAPTER 1. INTRODUCTION 2
of passing information to functions and returning results. These are used

to illustrate alternative design and programming approaches. Additionally,
some estimation procedures provide an opportunity to demonstrate coding
that solves problems likely to arise in other estimation tasks. Approaching
the text from this viewpoint, you should gain some familiarity with a host
of alternative coding tricks, and the text should serve as a reference to the
functions that contain these code fragments. When you encounter a similar
situation, simply examine (or re-use) the code fragments modified to suit
your particular problem.
Chapter 3 presents a library of utility functions that are used by many
other functions in the Econometrics Toolbox. For example, all printing of
results from econometric estimation and testing procedures is carried out by
a single function that prints matrices in a specified decimal or integer format
with optional column and row labels. Many of the examples throughout the
text also rely on this function named mprint.
Regression diagnostic procedures are the topic of Chapter 4. Diagnostics
for collinearity and influential observations from texts like Belsley, Kuh and
Welsch (1980) and Cook and Weisberg (1982) are discussed and illustrated.
Chapter 5 turns attention to vector autoregressive and error correction
models, as well as forecasting. Because we can craft our own functions in
MATLAB, we’re not restricted to a limited set of Bayesian priors as in the
case of RATS software. It is also possible to ease the tasks involved with
cointegration testing for error correction models. These tests and model for-
mation can be carried out in a single function with virtually no intervention
on the part of users. A similar situation exists in the area of forecasting.
Users can be spared the complications that arise from data transformations
prior to estimation that require reverse transformations to the forecasted
values.
A recent method that has received a great deal of attention in the statis-
tics literature, Markov Chain Monte Carlo, or MCMC is covered in Chap-
ter 6. Econometric estimation procedures are also beginning to draw on this
approach, and functions are crafted to implement these methods. Additional
functions were devised to provide convergence diagnostics (that are an in-
tegral part of the method) as well as presentation of printed and graphical
results. These functions communicate with each other via the MATLAB
structure variables.
Chapter 7 takes up logit, probit and tobit estimation from both a max-
imum likelihood as well as MCMC perspective. Recent MCMC approaches
to limited dependent variable models hold promise for dealing with non-
constant variance and outliers, and these are demonstrated in this chapter.
.::UdecomBooks::.
Simultaneous equation systems are the subject of Chapter 8, where we

face the challenge of encapsulating input variables for a system of equations
when calling our toolbox functions. Although MATLAB allows for variables
that are global in scope, no such variables are used in any Econometric Tool-
box functions. We solve the challenges using MATLAB structure variables.
Chapter 9 describes a host of functions for calculating probability den-
sities, cumulative densities, quantiles, and random deviates from twelve fre-
quently used statistical distributions. Other more special purpose functions
dealing with statistical distributions are also described.
The subject of optimization is taken up in Chapter 10 where we demon-
strate maximum likelihood estimation. Alternative approaches and func-
tions are described that provide a consistent interface for solving these types
of problems.
The final chapter discusses and illustrates the use of MATLAB sparse
matrix functions. These are useful for solving problems involving large ma-
trices that contain a large proportion of zeros. MATLAB has a host of
algorithms that can be used to operate on this type of matrix in an intelli-
gent way the conserves on both time and computer memory.
Readers who have used RATS, TSP or SAS should feel very comfortable
using the Econometrics Toolbox functions and MATLAB. The procedure
for producing econometric estimates is very similar to these other software
programs. First, the data files are “loaded” and any needed data trans-
formations to create variables for the model are implemented. Next, an
estimation procedure is called to operate on the model variables and a com-
mand is issued to print or graph the results.
The text assumes the reader is familiar with basic MATLAB commands
introduced in the software manual, Using MATLAB Version 5 or The Stu-
dent Edition of MATLAB, Version 5 User’s Guide by Hanselmann and Lit-
tlefield (1997). Gauss users should have little trouble understanding this
text, perhaps without reading the MATLAB introductory manuals, as the
syntax of MATLAB and Gauss is very similar.
All of the functions in the Econometrics Toolbox have been tested us-
ing MATLAB Version 5.2 on Apple, Intel/Windows and Sun Microsystems
computing platforms. The functions also work with Versions 5.0 and 5.1,
but some printed output that relies on a new string justification option in
Version 5.2 may not appear as nicely as it does in Version 5.2.
The text contains 71 example programs that come with the Econometric
Toolbox files. Many of these examples generate random data samples, so the
results you see will not exactly match those presented in the text. In addition
to the example files, there are demonstration files for all of the econometric
.::UdecomBooks::.
estimation and testing functions and most of the utility functions.

To conserve on space in the text, the printed output was often edited to
eliminate ‘white space’ that appears in the MATLAB command window. In
some cases output that was too wide for the text pages was also altered to
eliminate white space.
As you study the code in the Econometric Toolbox functions, you will see
a large amount of repeated code. There is a trade-off between writing func-
tions to eliminate re-use of identical code fragments and clarity. Functions
after all hide code, making it more difficult to understand the operations
that are actually taking place without keeping a mental note of what the
sub-functions are doing. A decision was made to simply repeat code frag-
ments for clarity. This also has the virtue of making the functions more
self-contained since they do not rely on a host of small functions.
Another issue is error checking inside the functions. An almost endless
amount of error checking code could be written to test for a host of possible
mistakes in the user input arguments to the functions. The Econometric
Toolbox takes a diligent approach to error checking for functions like least-
squares that are apt to be the first used by students. All functions check for
the correct number of input arguments, and in most cases test for structure
variables where they are needed as inputs. My experience has been that
after students master the basic usage format, there is less need for extensive
error checking. Users can easily interpret error messages from the function to
mean that input arguments are not correct. A check of the input arguments
in the function documentation against the user’s MATLAB command file
will usually suffice to correct the mistake and eliminate the error message. If
your students (or you) find that certain errors are typical, it should be fairly
straightforward to add error checking code and a message that is specific
to this type of usage problem. Another point in this regard is that my
experience shows certain types of errors are unlikely. For example, it is
seldom the case that users will enter matrix and vector arguments that have
a different number of observations (or rows). This is partly due to the way
that MATLAB works. Given this, many of the functions do not check for
this type of error, despite the fact that we could add code to carry out this
type of check.
.::UdecomBooks::.
Chapter 2
Regression using MATLAB
This chapter describes the design and implementation of a regression func-

tion library. Toolboxes are the name given by the MathWorks to related
sets of MATLAB functions aimed at solving a particular class of problems.
Toolboxes of functions useful in signal processing, optimization, statistics, fi-
nance and a host of other areas are available from the MathWorks as add-ons
to the standard MATLAB distribution. We will reserve the term Economet-
rics Toolbox to refer to the collection of function libraries discussed in each
chapter of the text. Many of the function libraries rely on a common util-
ity function library and on other function libraries. Taken together, these
constitute the Econometrics Toolbox described in this book.
All econometric estimation methods were designed to provide a consis-
tent user-interface in terms of the MATLAB help information, and related
routines to print and plot results from various types of regressions. Sec-
tion 2.1 discusses design issues and the use of MATLAB structures as a
way to pass results from the various regression estimation functions to as-
sociated routines. Implementation of a least-squares function is discussed
in Section 2.2 and the choice of least-squares algorithm is discussed in Sec-
tion 2.3 from a speed and numerical accuracy standpoint. Section 2.4 takes
up the design of related functions for printing and plotting regression re-
sults. The performance profiling capabilities of MATLAB are illustrated
in Section 2.5 and Section 2.6 demonstrates the use of alternative regres-
sion functions from the library in an applied setting. All of the regression
functions implemented in the library are documented in an appendix to
the chapter. Some of these functions are discussed and illustrated in later
chapters.
.::UdecomBooks::.
CHAPTER 2. REGRESSION USING MATLAB 6
2.1 Design of the regression library

In designing a regression library we need to think about organizing our func-
tions to present a consistent user-interface that packages all of our MATLAB
regression functions in a unified way. The advent of ‘structures’ in MAT-
LAB Version 5 allows us to create a host of alternative regression functions
that all return ‘results structures’.
A structure in MATLAB allows the programmer to create a variable
containing what MATLAB calls ‘fields’ that can be accessed by referencing
the structure name plus a period and the field name. For example, suppose
we have a MATLAB function to perform ordinary least-squares estimation
named ols that returns a structure. The user can call the function with
input arguments (a dependent variable vector y and explanatory variables
matrix x) and provide a variable name for the structure that the ols function
will return using:
result = ols(y,x);
The structure variable ‘result’ returned by our ols function might have
fields named ‘rsqr’, ‘tstat’, ‘beta’, etc. These fields might contain the R-
squared statistic, t−statistics for the β̂ estimates and the least-squares es-
timates β̂. One virtue of using the structure to return regression results is
that the user can access individual fields of interest as follows:
bhat = result.beta;
disp(‘The R-squared is:’);
result.rsqr
disp(‘The 2nd t-statistic is:’);
result.tstat(2,1)
There is nothing sacred about the name ‘result’ used for the returned
structure in the above example, we could have used:
bill_clinton = ols(y,x);
result2 = ols(y,x);
restricted = ols(y,x);
unrestricted = ols(y,x);
That is, the name of the structure to which the ols function returns its
information is assigned by the user when calling the function.
To examine the nature of the structure in the variable ‘result’, we can
simply type the structure name without a semi-colon and MATLAB will
present information about the structure variable as follows:
.::UdecomBooks::.
result =
meth: ’ols’
y: [100x1 double]
nobs: 100.00
nvar: 3.00
beta: [ 3x1 double]
yhat: [100x1 double]
resid: [100x1 double]
sige: 1.01
tstat: [ 3x1 double]
rsqr: 0.74
rbar: 0.73
dw: 1.89
Each field of the structure is indicated, and for scalar components the
value of the field is displayed. In the example above, ‘nobs’, ‘nvar’, ‘sige’,
‘rsqr’, ‘rbar’, and ‘dw’ are scalar fields, so their values are displayed. Matrix
or vector fields are not displayed, but the size and type of the matrix or
vector field is indicated. Scalar string arguments are displayed as illustrated
by the ‘meth’ field which contains the string ‘ols’ indicating the regression
method that was used to produce the structure. The contents of vector or
matrix strings would not be displayed, just their size and type. Matrix and
vector fields of the structure can be displayed or accessed using the MATLAB
conventions of typing the matrix or vector name without a semi-colon. For
example,
result.resid
result.y
would display the residual vector and the dependent variable vector y in the
MATLAB command window.
Another virtue of using ‘structures’ to return results from our regression
functions is that we can pass these structures to another related function
that would print or plot the regression results. These related functions can
query the structure they receive and intelligently decipher the ‘meth’ field
to determine what type of regression results are being printed or plotted.
For example, we could have a function prt that prints regression results and
another plt that plots actual versus fitted and/or residuals. Both these func-
tions take a regression structure as input arguments. Example 2.1 provides
a concrete illustration of these ideas.
The example assumes the existence of functions ols, prt, plt and data
matrices y, x in files ‘y.data’ and ‘x.data’. Given these, we carry out a regres-
sion, print results and plot the actual versus predicted as well as residuals
.::UdecomBooks::.
with the MATLAB code shown in example 2.1. We will discuss the prt and
plt functions in Section 2.4.
% ----- Example 2.1 Demonstrate regression using the ols() function
load y.data;
load x.data;
result = ols(y,x);
prt(result);
plt(result);
2.2 The ols function

Now to put these ideas into practice, consider implementing an ols function.
The function code would be stored in a file ‘ols.m’ whose first line is:
function results=ols(y,x)
The keyword ‘function’ instructs MATLAB that the code in the file ‘ols.m’
represents a callable MATLAB function.
The help portion of the MATLAB ‘ols’ function is presented below and
follows immediately after the first line as shown. All lines containing the
MATLAB comment symbol ‘%’ will be displayed in the MATLAB command
window when the user types ‘help ols’.
function results=ols(y,x)
% PURPOSE: least-squares regression
%---------------------------------------------------
% USAGE: results = ols(y,x)
% where: y = dependent variable vector (nobs x 1)
% x = independent variables matrix (nobs x nvar)
%---------------------------------------------------
% RETURNS: a structure
% results.meth = ’ols’
% results.beta = bhat
% results.tstat = t-stats
% results.yhat = yhat
% results.resid = residuals
% results.sige = e’*e/(n-k)
% results.rsqr = rsquared
% results.rbar = rbar-squared
% results.dw = Durbin-Watson Statistic
% results.nobs = nobs
% results.nvar = nvars
% results.y = y data vector
.::UdecomBooks::.
% --------------------------------------------------
% SEE ALSO: prt(results), plt(results)
%---------------------------------------------------
All functions in the econometrics toolbox present a unified documenta-

tion format for the MATLAB ‘help’ command by adhering to the convention
of sections entitled, ‘PURPOSE’, ‘USAGE’, ‘RETURNS’, ‘SEE ALSO’, and
perhaps a ‘REFERENCES’ section, delineated by dashed lines.
The ‘USAGE’ section describes how the function is used, with each input
argument enumerated along with any default values. A ‘RETURNS’ section
portrays the structure that is returned by the function and each of its fields.
To keep the help information uncluttered, we assume some knowledge on
the part of the user. For example, we assume the user realizes that the
‘.residuals’ field would be an (nobs x 1) vector and the ‘.beta’ field would
consist of an (nvar x 1) vector.
The ‘SEE ALSO’ section points the user to related routines that may
be useful. In the case of our ols function, the user might what to rely on
the printing or plotting routines prt and plt, so these are indicated. The
‘REFERENCES’ section would be used to provide a literature reference
(for the case of more exotic regression procedures) where the user could
read about the details of the estimation methodology.
To illustrate what we mean by consistency in the user documentation,
the following shows the results of typing ‘help ridge’, that provides user
documentation for the ridge regression function in the regression library.
PURPOSE: computes Hoerl-Kennard Ridge Regression

---------------------------------------------------
USAGE: results = ridge(y,x,theta)
where: y = dependent variable vector
x = independent variables matrix
theta = an optional ridge parameter
(default: best theta value ala Hoerl-Kennard)
---------------------------------------------------
RETURNS: a structure
results.meth = ’ridge’
results.beta = bhat
results.theta = theta (input or HK determined value)
results.tstat = t-stats
results.yhat = yhat
results.resid = residuals
results.sige = e’*e/(n-k)
results.rsqr = rsquared
results.rbar = rbar-squared
results.dw = Durbin-Watson Statistic
.::UdecomBooks::.
results.nobs = nobs
results.nvar = nvars
results.y = y data vector
--------------------------------------------------
SEE ALSO: rtrace, prt, plt
---------------------------------------------------
REFERENCES: David Birkes, Yadolah Dodge, 1993, Alternative Methods of
Regression and Hoerl, Kennard, Baldwin, 1975 ‘Ridge Regression: Some
Simulations’, Communcations in Statistics
Now, turning attention to the actual MATLAB code for estimating the
ordinary least-squares model, we begin processing the input arguments to
carry out least-squares estimation based on a model involving y and x. We
first check for the correct number of input arguments using the MATLAB
‘nargin’ variable.
if (nargin ~= 2); error(’Wrong # of arguments to ols’);
else
[nobs nvar] = size(x); [nobs2 junk] = size(y);
if (nobs ~= nobs2); error(’x and y must have same # obs in ols’);
end;
end;
If we don’t have two input arguments, the user has made an error which
we indicate using the MATLAB error function. The ols function will return
without processing any of the input arguments in this case. Another error
check involves the number of rows in the y vector and x matrix which should
be equal. We use the MATLAB size function to implement this check in
the code above.
Assuming that the user provided two input arguments, and the number
of rows in x and y are the same, we can pass on to using the input information
to carry out a regression.
The ‘nobs’ and ‘nvar’ returned by the MATLAB size function are pieces
of information that we promised to return in our results structure, so we
construct these fields using a ‘.nobs’ and ‘.nvar’ appended to the ‘results’
variable specified in the function declaration. We also fill in the ‘meth’ field
and the ‘y’ vector fields.
results.meth = ’ols’;
results.y = y;
results.nobs = nobs;
results.nvar = nvar;
The decision to return the actual y data vector was made to facilitate
the plt function that will plot the actual versus predicted values from the
.::UdecomBooks::.
regression along with the residuals. Having the y data vector in the structure
makes it easy to call the plt function with only the structure returned by a
regression function.
We can proceed to estimate least-squares coefficients β̂ = (X 0 X)−1 X 0 y,
but we have to choose a solution method for the least-squares problem.
The two most commonly used approaches are based on the Cholesky and
qr matrix decompositions. The regression library ols function uses the qr
matrix decomposition method for reasons that will be made clear in the next
section. A first point to note is that we require more than a simple solution
for β̂, because we need to calculate t−statistics for the β̂ estimates. This
requires that we compute (X 0 X)−1 which is done using the MATLAB ‘slash’
operator to invert the (X 0 X) matrix. We represent (X 0 X) using (r0 r), where
r is an upper triangular matrix returned by the qr decomposition.
[q r] = qr(x,0);
xpxi = (r’*r)\eye(nvar);
results.beta = r\(q’*y);
An alternative solution based on the Cholesky decomposition is faster,

but less accurate for ill-conditioned X data matrices. The Cholesky solution
could be implemented as:
xpxi = (x’*x)\eye(k);
results.beta = xpxi*(x’*y);
Given either of these solutions, we are in a position to use our estimates

β̂ to compute the remaining elements of the ols function results structure.
We add these elements to the structure in the ‘.yhat, .resid, .sige,’ etc.,
fields.
results.yhat = x*results.beta;
results.resid = y - results.yhat;
sigu = results.resid’*results.resid;
results.sige = sigu/(nobs-nvar);
tmp = (results.sige)*(diag(xpxi));
results.tstat = results.beta./(sqrt(tmp));
ym = y - mean(y);
rsqr1 = sigu; rsqr2 = ym’*ym;
results.rsqr = 1.0 - rsqr1/rsqr2; % r-squared
rsqr1 = rsqr1/(nobs-nvar);
rsqr2 = rsqr2/(nobs-1.0);
results.rbar = 1 - (rsqr1/rsqr2); % rbar-squared
ediff = results.resid(2:nobs) - results.resid(1:nobs-1);
results.dw = (ediff’*ediff)/sigu; % durbin-watson
.::UdecomBooks::.
2.3 Selecting a least-squares algorithm

In this section, we explore the speed and accuracy of the Cholesky and qr
approaches to solving the least-squares problem. Using a program like that
shown in example 2.2 we can examine execution times for the Cholesky and
qr solution methods. Execution times are found using the MATLAB tic
and toc functions. MATLAB begins timing execution when it encounters
the tic command and stops timing when the toc command is encountered,
printing out the execution time in the MATLAB command window.
% ----- Example 2.2 Least-squares timing information

n = 10000; k = 10;
e = randn(n,1); x = randn(n,k); b = ones(k,1);
y = x*b + e;
disp(’time needed for Cholesky solution’);
tic;
xpxi = (x’*x)\eye(k); % solve using the Cholesky decomposition
bhatc = xpxi*(x’*y);
toc;
disp(’time needed for QR solution’);
tic;
[q r] = qr(x,0); % solve using the qr decomposition
xpxi = (r’*r)\eye(k);
bhatq = r\(q’*y);
toc;
Using this type of program, we explored the timing differences between

the Cholesky and qr solution methods for different numbers of observations
and explanatory variables. The timing results (presented in seconds of time
required) from this exploration are shown in Table 2.1. These results were
produced on a MacIntosh G3 computer that contains a 266 Mhz PowerPC
750 chip. From the third part of the table that shows the ratio of times
for the qr versus Cholesky algorithms, we seen that for smaller problems
(100 to 1000 observations and 10 or 50 explanatory variables) the qr routine
takes 2 to 3.5 times as long as the Cholesky approach. As the number of
observations increases to 10000 or the number of variables equals 100, the
qr takes 1.5 to 2 times as long as the Cholesky method.
As indicated above, speed is not our only concern. Accuracy in the
face of ill-conditioning is also an important consideration in designing a
least-squares algorithm. A question that arises is — how sensitive are the
computer solutions (the estimates) to small perturbations in the data con-
tained in the X matrix? We would like to believe that small changes in the
elements of the X matrix would not lead to large changes in the solution
.::UdecomBooks::.
vector β̂. If small changes in the data elements lead to large changes in the
solution to an estimation problem, we say that the problem is ill-conditioned.
Table 2.1: Timing (in seconds) for Cholesky and QR Least-squares

Cholesky
nobs/nvar 10 50 100
100 0.0009 0.0128 0.0603
1000 0.0072 0.1112 0.6572
10000 0.2118 2.6344 8.8462
QR
nobs/nvar 10 50 100
100 0.0026 0.0250 0.0919
1000 0.0261 0.2628 1.1174
10000 0.4289 4.9914 17.0524
QR/Cholesky
nobs/nvars 10 50 100
100 2.8618 1.9535 1.5244
1000 3.5064 2.2806 1.7004
10000 2.0249 1.8947 1.9276
We can quantify the conditioning of a problem by calculating a condition

number, which in the case of the least-squares problem is the ratio of the
largest to the smallest eigenvalue of the data matrix X. The larger this
ratio, the more ill-conditioned is the least-squares problem. Belsley, Kuh,
and Welsch (1980, p. 114) use the condition number to state the following:
If the data are known to d significant digits and the condition number is on
the order of magnitude of 10r , then a small change in the last place of the
data can (but need not) affect the solution for β̂ in the (d − r)th place. A
proof of this statement along with a more detailed discussion can be found
in Belsley, Kuh, and Welsch (1980).
To see how this information is useful, consider an example where the
explanatory variables data are trusted to 4 digits, that is we have no faith
in our ability to measure these variables beyond 4 decimal digits. In this
case, a shift in the fifth decimal place of the X matrix data would produce
an observationally equivalent data set — one that cannot be distinguished
from the original data set given our measurement limitations. It would be
highly desirable that two observationally equivalent data sets produce the
same least-squares estimates of the β parameters. Continuing this example,
suppose the condition number of the X matrix is 1, 000 which can be written
as 103 . Then a shift in the fifth decimal place of the X matrix data could
affect the least-squares parameter estimates of β in its (5 − 3) = 2nd digit.
.::UdecomBooks::.
The implication is that only the first digit is numerically accurate, meaning
that an estimate of β̂ = .83 only informs us that the parameter is between
.80 and .90!
These statements are based on some theoretical bounds calculated in
Belsley, Kuh, and Welsch (1980), leading us to conclude that the results
could be affected in the ways stated. The theoretical bounds are upper
bounds on the potential problems, reflecting the worst that could happen.
To examine the numerical accuracy of the Cholesky and qr approaches to
solving the least-squares problem, we rely on a “benchmark” data set for
which we know the true parameter values. We can then test the two regres-
sion algorithms to see how accurately they compute estimates of the true
parameter values. Such a benchmark is shown in (2.1).
   
1 1 + γ 1 + γ ... 1 + γ (n − 1) + (n − 2)γ + ²
 1 γ+² γ ... γ   (n − 2)γ + ² 
   
   
 1 γ γ + ² ... γ   (n − 2)γ + ² 
X=
 .. .. .. .. .. y = 
  .. 

 . . . . .   . 
   
 1 γ γ ... γ + ²   (n − 2)γ + ² 
1 γ γ ... γ (n − 1) + (n − 2)γ − ²
(2.1)
This is a modification of a benchmark originally proposed by Wampler
(1980) set forth in Simon and LeSage (1988a, 1988b). The n by (n − 1)
matrix X and the nx1 vector y in (2.1) represent the Wampler benchmark
with the parameter γ added to the last (n − 2) columns of the X matrix,
and with (n − 2)γ added to the y vector. When γ = 0 this benchmark is
equivalent to the original Wampler benchmark. The modified benchmark
shares the Wampler property that its solution is a column of ones for all
values of ² > 0, and for all values of γ, so the coefficient estimates are unity
irrespective of ill-conditioning. This property makes it easy to judge how
accurate the least-squares computational solutions for the estimates are. We
simply need to compare the estimates to the true value of one.
The parameters γ and ² in (2.1) control the severity of two types of
near-linear relationships in the data matrix X. The parameter ² controls
the amount of collinearity between the last (n−2) columns of the data matrix
X. As the parameter ² is decreased towards zero, the last (n − 2) columns
move closer to becoming perfect linear combinations with each other. The
implication is that the parameter ² acts to control the amount of collinearity,
or the severity of the near linear combinations among the last (n−2) columns
of the data matrix. As we make the value of ² smaller we produce an
.::UdecomBooks::.
increasingly ill-conditioned least-squares problem.

The γ parameter serves to control a near linear relationship between
the intercept column and the last (n − 2) columns of the data matrix. As
the parameter γ is increased, the last (n − 2) columns of the matrix X
become more collinear with the intercept column producing a near linear
combination between the intercept term and the last (n − 2) columns. This
type of collinear relationship reflects a situation where each independent
variable becomes more nearly constant, exhibiting a near linear relationship
with the constant term vector.
Using this benchmark data set, we can examine the numerical accu-
racy of regression algorithms while continuously varying the amount of ill-
conditioning. In addition, we can examine the impact of two different types
of collinear relationships that one might encounter in practice. The first
type of collinearity — that controlled by the ² parameter — involves the last
(n − 2) columns of the data matrix representing the explanatory variables
in our least-squares model. The second type of collinearity — controlled by
the γ parameter — represents cases where the intercept term is involved in
a near linear relationship with the entire set of explanatory variables, the
last (n − 2) columns.
We will measure the numerical accuracy of the coefficient estimates pro-
duced by the Cholesky and qr regression algorithms using a formula from
Wampler (1980). This formula, can be thought of as representing the num-
ber of digits of accuracy and is shown in (2.2).
accj = −log10 (|1.0 − β̂ j |) (2.2)

In (2.2) accj represents the accuracy of the estimate β̂ j in estimating the
true value which is known to be 1.0, log10 represents the base 10 logarithm,
and the symbol | denotes that we are taking the absolute value of the quan-
tity 1.0 − β̂ j . As an example, consider a case where β̂ j = 1.001, according
to (2.2) we have three digits of accuracy.
Table 2.2 presents the digits accuracy results from the Cholesky and qr
algorithms over a range of values for the control parameters γ and ². Values
of n = 10, 25 and 50 were used to produce numerous estimates. The symbol
‘**’ is used in the table to designate negative digits of accuracy or inability
to solve the least-squares problem due to ill-conditioning. Accuracy for the
intercept term estimate is presented along with results for the ‘most’ and
‘least’ accurate slope estimates.
From Table 2.2 we conclude that the qr algorithm was capable of solving
the least-squares problem for all values of γ and ², producing at least 6
.::UdecomBooks::.
Table 2.2: Digits of accuracy for Cholesky vs. QR decomposition

Least accurate slope Cholesky
γ\² .01 .001 .0001 .00005
10 7.34 5.95 3.87 3.36
100 5.54 3.69 1.59 0.77
1000 3.56 1.48 ** **
5000 2.39 0.56 ** **
Least accurate slope qr
γ\² .01 .001 .0001 .00005
10 11.61 9.70 8.04 7.17
100 11.09 9.73 7.58 6.87
1000 10.11 8.74 7.61 6.95
5000 9.35 8.13 7.04 7.05
Most accuracy slope Cholesky
γ\² .01 .001 .0001 .00005
10 8.29 7.85 6.38 4.46
100 6.44 5.20 2.91 2.51
1000 4.68 2.76 ** 0.17
5000 3.48 2.71 ** **
Most accurate slope qr
γ\² .01 .001 .0001 .00005
10 12.70 11.12 8.92 8.02
100 12.19 10.89 8.50 7.79
1000 10.95 10.56 8.25 8.60
5000 10.16 8.71 7.94 8.68
Cholesky intercept accuracy
γ\² .01 .001 .0001 .00005
10 6.43 4.90 2.54 2.53
100 3.85 2.22 0.44 **
1000 0.72 ** ** **
5000 ** ** ** **
qr intercept accuracy
γ\² .01 .001 .0001 .00005
10 12.17 12.39 11.91 12.33
100 10.67 10.06 9.86 10.42
1000 8.63 8.29 8.72 8.16
5000 6.95 7.93 6.66 7.28
.::UdecomBooks::.
decimal digits of accuracy. In contrast, the Cholesky produced negative

decimal digits of accuracy (or was incapable of solving the least-squares
problem) in 4 of the 16 least accurate slope estimate cases, 3 of 16 cases
for the most accurate slope estimate, and 8 of the 16 cases for the intercept
estimates. The conclusion is that if we care about numerical accuracy in the
face of ill-conditioned data sets, we should rely on the qr algorithm to solve
the least-squares problem. With the advent of faster computers, the timing
differences do not appear as dramatic as the numerical accuracy differences.
2.4 Using the results structure

To illustrate the use of the ‘results’ structure returned by our ols function,
consider the associated function plt reg which plots actual versus predicted
values along with the residuals. The results structure contains everything
needed by the plt reg function to carry out its task. Earlier, we referred
to functions plt and prt rather than plt reg, but we will see in Chapter 3
that prt and plt are “wrapper” functions that call the functions prt reg
and plt reg where the real work of printing and plotting regression results
is carried out. The motivation for taking this approach is that separate
smaller functions can be devised to print and plot results from all of the
econometric procedures in the toolbox facilitating development. The wrap-
per functions eliminate the need for the user to learn the names of different
printing and plotting functions associated with each group of econometric
procedures, all results structures can be printed and plotted by simply in-
voking the prt and plt functions. The Econometrics Toolbox contains a
host of printing and plotting functions for producing formatted output from
vector autoregressions, cointegration tests, Gibbs sampling estimation, si-
multaneous equations estimators, etc. All of these functions can be accessed
using the wrapper functions prt and plt, or by directly calling the individual
function names such as plt reg.
The plt reg function is shown below. A check that the user supplied a
regression results structure can be carried out using the MATLAB isstruct
function that is true if the argument represents a structure. After this error
check, we rely on a MATLAB programming construct called the ‘switch-
case’ to provide the remaining error checking for the function.
function plt_reg(results);
% PURPOSE: plots regression actual vs predicted and residuals
%---------------------------------------------------
% USAGE: plt_reg(results);
% where: results is a structure returned by a regression function
.::UdecomBooks::.
%---------------------------------------------------
% RETURNS: nothing, just plots regression results
% --------------------------------------------------
% NOTE: user must supply pause commands, none are in plt_reg function
% e.g. plt_reg(results);
% pause;
% plt_reg(results2);
% --------------------------------------------------
% SEE ALSO: prt_reg(results), prt, plt
%---------------------------------------------------
if ~isstruct(results), error(’plt_reg requires a structure’); end;

nobs = results.nobs; tt=1:nobs; clf;
switch results.meth
case {’arma’,’boxcox’,’boxcox2’,’logit’,’ols’,’olsc’,’probit’,’ridge’, ...
’theil’,’tobit’,’hwhite’,’tsls’,’nwest’}
subplot(211), plot(tt,results.y,’-’,tt,results.yhat,’--’);
title([upper(results.meth), ’ Actual vs. Predicted’]);
subplot(212), plot(tt,results.resid); title(’Residuals’);
case {’robust’,’olst’,’lad’}
subplot(311), plot(tt,results.y,’-’,tt,results.yhat,’--’);
title([upper(results.meth), ’ Actual vs. Predicted’]);
subplot(312), plot(tt,results.resid); title(’Residuals’);
subplot(313), plot(tt,results.weight); title(’Estimated weights’);
otherwise
error(’method not recognized by plt_reg’);
end;
subplot(111);
The ‘switch’ statement examines the ‘meth’ field of the results structure
passed to the plt reg function as an argument and executes the plotting
commands if the ‘meth’ field is one of the regression methods implemented
in our function library. In the event that the user passed a result structure
from a function other than one of our regression functions, the ‘otherwise’
statement is executed which prints an error message.
The switch statement also helps us to distinguish special cases of robust,
olst, lad regressions where the estimated weights are plotted along with the
actual versus predicted and residuals. These weights allow the user to detect
the presence of outliers in the regression relationship. A similar approach
could be used to extend the plt reg function to accommodate other special
regression functions where additional or specialized plots are desired.
A decision was made not to place the ‘pause’ command in the plt reg
function, but rather let the user place this statement in the calling program
or function. An implication of this is that the user controls viewing regres-
.::UdecomBooks::.
sion plots in ‘for loops’ or in the case of multiple invocations of the plt reg
function. For example, only the second ‘plot’ will be shown in the following
code.
result1 = ols(y,x1);
plt_reg(result1);
plt_reg(result2);
If the user wishes to see the regression plots associated with the first
regression, the code would need to be modified as follows:
plt_reg(result1);
pause;
plt_reg(result2);
The ‘pause’ statement would force a plot of the results from the first
regression and wait for the user to strike any key before proceeding with the
second regression and accompanying plot of these results.
Our plt reg function would work with new regression functions that we
add to the library provided that the regression returns a structure containing
the fields ‘.y’, ‘.yhat’, ‘.resid’, ‘.nobs’ and ‘.meth’. We need simply add this
method to the switch-case statement.
A more detailed example of using the results structure is the prt reg
function from the regression library. This function provides a printout of
regression results similar to those provided by many statistical packages.
The function relies on the ‘meth’ field to determine what type of regression
results are being printed, and uses the ‘switch-case’ statement to implement
specialized methods for different types of regressions.
A small fragment of the prt reg function showing the specialized print-
ing for the ols and ridge regression methods is presented below:
function prt_reg(results,vnames,fid)
% PURPOSE: Prints output using regression results structures
%---------------------------------------------------
% USAGE: prt_reg(results,vnames,fid)
% Where: results = a structure returned by a regression
% vnames = an optional vector of variable names
% fid = optional file-id for printing results to a file
% (defaults to the MATLAB command window)
%---------------------------------------------------
% NOTES: e.g. vnames = strvcat(’y’,’const’,’x1’,’x2’);
.::UdecomBooks::.
% e.g. fid = fopen(’ols.out’,’wr’);

% use prt_reg(results,[],fid) to print to a file with no vnames
% --------------------------------------------------
% RETURNS: nothing, just prints the regression results
% --------------------------------------------------
% SEE ALSO: prt, plt
%---------------------------------------------------
if ~isstruct(results) % error checking on inputs

error(’prt_reg requires structure argument’);
elseif nargin == 1, nflag = 0; fid = 1;
elseif nargin == 2, fid = 1; nflag = 1;
elseif nargin == 3, nflag = 0;
[vsize junk] = size(vnames); % user may supply a blank argument
if vsize > 0, nflag = 1; end;
else, error(’Wrong # of arguments to prt_reg’); end;
nobs = results.nobs; nvar = results.nvar;
% make up generic variable names
Vname = ’Variable’;
for i=1:nvar;
tmp = [’variable ’,num2str(i)]; Vname = strvcat(Vname,tmp);
end;
if (nflag == 1) % the user supplied variable names
[tst_n nsize] = size(vnames);
if tst_n ~= nvar+1
warning(’Wrong # of variable names in prt_reg -- check vnames argument’);
fprintf(fid,’will use generic variable names \n’);
nflag = 0;
else,
for i=1:nvar; Vname = strvcat(Vname,vnames(i+1,:)); end;
end; % end of nflag issue
switch results.meth
case {’ols’,’hwhite’,’nwest’} % <=== ols,white,nwest regressions
if strcmp(results.meth,’ols’)
fprintf(fid,’Ordinary Least-squares Estimates \n’);
elseif strcmp(results.meth,’hwhite’)
fprintf(fid,’White Heteroscedastic Consistent Estimates \n’);
elseif strcmp(results.meth,’nwest’)
fprintf(fid,’Newey-West hetero/serial Consistent Estimates \n’);
end;
if (nflag == 1)
fprintf(fid,’Dependent Variable = %16s \n’,vnames(1,:));
end;
fprintf(fid,’R-squared = %9.4f \n’,results.rsqr);
fprintf(fid,’Rbar-squared = %9.4f \n’,results.rbar);
fprintf(fid,’sigma^2 = %9.4f \n’,results.sige);
fprintf(fid,’Durbin-Watson = %9.4f \n’,results.dw);
fprintf(fid,’Nobs, Nvars = %6d,%6d \n’,results.nobs,results.nvar);
.::UdecomBooks::.
fprintf(fid,’*******************************************************\n’);
% <=================== end of ols,white, newey-west case
case {’ridge’} % <=================== ridge regressions
fprintf(fid,’Ridge Regression Estimates \n’);
if (nflag == 1)
end;
fprintf(fid,’Ridge theta = %16.8g \n’,results.theta);
fprintf(fid,’*******************************************************\n’);
% <=================== end of ridge regression case
otherwise
error(’method unknown to the prt_reg function’);
end;
tout = tdis_prb(results.tstat,nobs-nvar); % find t-stat probabilities
tmp = [results.beta results.tstat tout]; % matrix to be printed
% column labels for printing results
bstring = ’Coefficient’; tstring = ’t-statistic’; pstring = ’t-probability’;
cnames = strvcat(bstring,tstring,pstring);
in.cnames = cnames; in.rnames = Vname; in.fmt = ’%16.6f’; in.fid = fid;
mprint(tmp,in); % print estimates, t-statistics and probabilities
The function mprint is a utility function to produce formatted printing

of a matrix with column and row-labels and is discussed in detail in Chap-
ter 3. All printing of matrix results for the Econometric Toolbox functions
is done using the mprint function.
The prt reg function allows the user an option of providing a vector of
fixed width variable name strings that will be used when printing the regres-
sion coefficients. These can be created using the MATLAB strvcat function
that produces a vertical concatenated list of strings with fixed width equal
to the longest string in the list. We give the user a break here, if the wrong
number of names is supplied, a warning is issued to the user but results are
still printed using generic variable names. I found this avoids some annoying
situations where generic names and printed results are preferable to an error
message with no printed output.
We can also print results to an indicated file rather than the MATLAB
command window. Three alternative invocations of the prt reg function
illustrating these options for usage are shown below:
vnames = strvcat(’y variable’,’constant’,’population’,’income’);

res = ols(y,x);
.::UdecomBooks::.
prt_reg(res); % print with generic variable names

prt_reg(res,vnames); % print with user-supplied variable names
fid = fopen(’ols.out’,’wr’); % open a file for printing
prt_reg(res,vnames,fid); % print results to file ‘ols.out’
The first use of prt reg produces a printout of results to the MATLAB
command window that uses ‘generic’ variable names:
Ordinary Least-squares Estimates
R-squared = 0.8525
Rbar-squared = 0.8494
sigma^2 = 0.6466
Durbin-Watson = 1.8791
Nobs, Nvars = 100, 3
***************************************************************
Variable Coefficient t-statistic t-probability
variable 1 1.208077 16.142388 0.000000
variable 2 0.979668 11.313323 0.000000
variable 3 1.041908 13.176289 0.000000
The second use of prt reg uses the user-supplied variable names. The
MATLAB function strvcat carries out a vertical concatenation of strings
and pads the shorter strings in the ‘vnames’ vector to have a fixed width
based on the longer strings. A fixed width string containing the variable
names is required by the prt reg function. Note that we could have used:
vnames = [’y variable’,
’constant ’,
’population’,
’income ’];
but, this takes up more space and is slightly less convenient as we have
to provide the padding of strings ourselves. Using the ‘vnames’ input in
the prt reg function would result in the following printed to the MATLAB
command window.
Dependent Variable = y variable
R-squared = 0.8525
sigma^2 = 0.6466
***************************************************************
constant 1.208077 16.142388 0.000000
population 0.979668 11.313323 0.000000
income 1.041908 13.176289 0.000000
.::UdecomBooks::.
The third case specifies an output file opened with the command:
fid = fopen(’ols.out’,’wr’);
The file ‘ols.out’ would contain output identical to that from the second use
of prt reg. It is the user’s responsibility to close the file that was opened
using the MATLAB command:
fclose(fid);
Next, we turn to details concerning implementation of the prt reg func-

tion. The initial code does error checking on the number of input arguments,
determines if the user has supplied a structure argument, and checks for
variable names and/or an output file id. We allow the user to provide a file
id argument with no variable names using the call: prt reg(result,[],fid)
where a blank argument is supplied for the variable names. We check for
this case by examining the size of the vnames input argument under the
case of nargin == 3 in the code shown below.

error(’prt_reg requires structure argument’);
elseif nargin == 1, nflag = 0; fid = 1;
elseif nargin == 2, fid = 1; nflag = 1;
elseif nargin == 3, nflag = 0;
[vsize junk] = size(vnames); % user may supply a blank argument
if vsize > 0, nflag = 1; end;
else
error(’Wrong # of arguments to prt_reg’);
end;
Variable names are constructed if the user did not supply a vector of
variable names and placed in a MATLAB fixed-width string-array named
‘Vname’, with the first name in the array being the row-label heading ‘Vari-
able’ which is used by the function mprint. For the case where the user
supplied variable names, we simply transfer these to a MATLAB ‘string-
array’ named ‘Vname’, again with the first element ‘Variable’ that will be
used by mprint. We do error checking on the number of variable names
supplied which should equal the number of explanatory variables plus the
dependent variable (nvar+1). In the event that the user supplies the wrong
number of variable names, we issue a warning and print output results using
the generic variable names.
nobs = results.nobs; nvar = results.nvar;

% make up generic variable names
.::UdecomBooks::.
for i=1:nvar;
tmp = [’variable ’,num2str(i)]; Vname = strvcat(Vname,tmp);
end;
if (nflag == 1) % the user supplied variable names
[tst_n nsize] = size(vnames);
if tst_n ~= nvar+1
warning(’Wrong # of variable names in prt_reg -- check vnames argument’);
fprintf(fid,’will use generic variable names \n’);
nflag = 0;
else,
for i=1:nvar; Vname = strvcat(Vname,vnames(i+1,:)); end;
end; % end of nflag issue
After constructing or transferring variable names, the ‘switch-case’ takes

over sending our function to the appropriate customized segment of code for
printing part of the regression results depending on the ‘meth’ field of the
results structure.
In the case of ‘ols’, ‘hwhite’ or ‘nwest’ (ordinary least-squares, White’s
heteroscedastic consistent estimator or Newey and West’s heteroscedastic-
serial correlation consistent estimator), we rely on a great deal of common
code. The title for the regression results printout will differ depending on
which of these methods was used to produce the results structure, with
everything else identical.
switch results.meth
case {’ols’,’hwhite’,’nwest’} % <====== ols,white,nwest regressions
if strcmp(results.meth,’ols’)
fprintf(fid,’Ordinary Least-squares Estimates \n’);
elseif strcmp(results.meth,’hwhite’)
fprintf(fid,’White Heteroscedastic Consistent Estimates \n’);
elseif strcmp(results.meth,’nwest’)
fprintf(fid,’Newey-West hetero/serial Consistent Estimates \n’);
end;
if (nflag == 1)
end;
fprintf(fid,’*******************************************************\n’);
% <=================== end of ols,white, newey-west case
A point to note is that use of the MATLAB ‘fprintf’ command with an
.::UdecomBooks::.
input argument ‘fid’ makes it easy to handle both the case where the user
wishes output printed to the MATLAB command window or to an output
file. The ‘fid’ argument takes on a value of ‘1’ to print to the command
window and a user-supplied file name value for output printed to a file.
Finally, after printing the specialized output, the coefficient estimates, t-
statistics and marginal probabilities that are in common to all regressions are
printed. The marginal probabilities are calculated using a function tdis prb
from the distributions library discussed in Chapter 9. This function returns
the marginal probabilities given a vector of t−distributed random variates
along with a degrees of freedom parameter. The code to print coefficient
estimates, t-statistics and marginal probabilities is common to all regression
printing procedures, so it makes sense to move it to the end of the ‘switch-
case’ code and execute it once as shown below. We rely on the function
mprint discussed in Chapter 3 to do the actual printing of the matrix of
regression results with row and column labels specified as fields of a structure
variable ‘in’. Use of structure variables with fields as input arguments to
functions is a convenient way to pass a large number of optional arguments
to MATLAB functions, a subject taken up in Chapter 3.
tout = tdis_prb(results.tstat,nobs-nvar); % find t-stat probabilities
tmp = [results.beta results.tstat tout]; % matrix to be printed
% column labels for printing results
bstring = ’Coefficient’; tstring = ’t-statistic’; pstring = ’t-probability’;
cnames = strvcat(bstring,tstring,pstring);
in.cnames = cnames; in.rnames = Vname; in.fmt = ’%16.6f’; in.fid = fid;
mprint(tmp,in); % print estimates, t-statistics and probabilities
The case of a ridge regression illustrates the need for customized code
to print results for different types of regressions. This regression produces
a ridge parameter estimate based on a suggested formula from Hoerl and
Kennard (1970), or allows for a user-supplied value. In either case, the
regression output should display this important parameter that was used to
produce the coefficient estimates.
case {’ridge’} % <=================== ridge regressions
fprintf(fid,’Ridge Regression Estimates \n’);
if (nflag == 1)
end;
fprintf(fid,’Ridge theta = %16.8g \n’,results.theta);
.::UdecomBooks::.

fprintf(fid,’*******************************************************\n’);
% <=================== end of ridge case
A design consideration here is whether to attempt a conditional state-

ment that would print the ridge parameter based on an ‘if’ statement that
checks for the ridge method inside the code for the ols, white, nwest case.
Taking this approach, we could incorporate the same code used for print-
ing the ols, white, nwest functions with the addition of a single additional
statement to print the ridge parameter when we detect a ridge regression in
the ‘meth’ field. The prt reg function was designed to allow easy modifica-
tion for printing results from new methods that are added to the regression
library. There is a trade-off between repeating code and complicating the
function so it becomes difficult to modify in the future.
An example of a more convincing case for writing separate code for differ-
ent regression procedures is the Theil-Goldberger regression. Here we want
to print prior means and standard deviations specified as input arguments
by the user in the output information. This specialized information can be
printed by the code segment handling the ’theil method as shown below:
case {’theil’} % <=================== theil-goldberger regressions

fprintf(fid,’\n’);
fprintf(fid,’Theil-Goldberger Regression Estimates \n’);
if (nflag == 1)
end;
fprintf(fid,’*******************************************************\n’);
vstring = ’Variable’; bstring = ’Prior Mean’; tstring = ’Std Deviation’;
tmp = [results.pmean results.pstd];
cnames = strvcat(bstring,tstring);
pin.cnames = cnames;
pin.rnames = Vname;
pin.fmt = strvcat(’%16.6f’,’%16.6f’);
pin.fid = fid;
mprint(tmp,pin);
fprintf(fid,’*******************************************************\n’);
fprintf(fid,’ Posterior Estimates \n’);
As in the case of all other regression methods, the common code at

the end of the prt reg function would print out the posterior estimates,
.::UdecomBooks::.
t-statistics and probabilities that should be in the ‘.beta, .tstat’ fields of the
structure returned by the theil function.
As an example of adding a code segment to handle a new regression
method, consider how we would alter the prt reg function to add a Box-
Jenkins ARMA method. First, we need to add a ‘case’ based on the Box-
Jenkins ‘meth’ field which is ‘arma’. The specialized code for the ‘arma’
method handles variable names in a way specific to a Box-Jenkins arma
model. It also presents output information regarding the number of AR and
MA parameters in the model, log-likelihood function value and number of
iterations required to find a solution to the nonlinear optimization problem
used to find the estimates, as shown below.
case {’arma’} % <=============== box-jenkins regressions

p = length(results.ar);
q = length(results.ma);
fprintf(1,’\n’);
fprintf(1,’Box-Jenkins ARMA Estimates \n’);
fprintf(1,’R-squared = %9.4f \n’,results.rsqr);
fprintf(1,’Rbar-squared = %9.4f \n’,results.rbar);
fprintf(1,’sigma^2 = %9.4f \n’,results.sige);
fprintf(1,’log-likelihood = %16.8g \n’,results.like);
fprintf(1,’Nobs = %6d \n’,results.nobs);
fprintf(1,’AR,MA orders = %6d,%6d \n’,p,q);
fprintf(1,’Iterations = %6d \n’,results.iter);
fprintf(fid,’*******************************************************\n’);
% create special variable names for box-jenkins
if results.cterm == 0;
for i=1:nvar
if i <= p, tmp = str2mat([’AR ’,num2str(i)]);
Vname = strvcat(Vname,tmp);
elseif i <= p+q, tmp = str2mat([’MA ’,num2str(i-p)]);
end;
end;
else
for i=1:nvar
if i <= p, tmp = str2mat([’AR ’,num2str(i)]);
elseif i == p+1, Vname = strvcat(Vname,’Const’);
else, tmp = str2mat([’MA ’,num2str(i-p-1)]);
end;
end;
end;
% <=================== end of boxjenkins case
.::UdecomBooks::.
Provided that we placed the regression estimates and t-statistics from the
arma regression routine into structure fields ‘results.beta’ and ‘results.tstat’,
the common code (that already exists) for printing regression results would
work with this new function.
2.5 Performance profiling the regression toolbox

This section demonstrates how to use the MATLAB performance profiling
function profile to examine regression functions and enhance their speed.
The profile command is provided with a ‘function’ argument which we
then execute. After executing the function, calling the profile function
again with the argument ‘report’ produces a printed summary indicating
how much time the function spent on each line of code . Example 2.3
illustrates profiling the ols function.
% ----- Example 2.3 Profiling the ols() function

nobs = 1000; nvar = 15; beta = ones(nvar,1);
xmat = randn(nobs,nvar-1); x = [ones(nobs,1) xmat];
evec = randn(nobs,1); y = x*beta + evec;
% profile the ols function
profile ols; result = ols(y,x); profile report;
% profile the prt_reg function
profile prt_reg; prt_reg(result); profile report;
This produced the following profile output for the ols function:
Total time in "ols.m": 0.1 seconds

100% of the total time was spent on lines:
[54 44 47]
43:
0.04s, 40% 44: [q r] = qr(x,0);
45: xpxi = (r’*r)\eye(nvar);
46:
0.01s, 10% 47: results.beta = r\(q’*y);
48: results.yhat = x*results.beta;
53: results.tstat = results.beta./(sqrt(tmp));
0.05s, 50% 54: ym = y - mean(y);
55: rsqr1 = sigu;
The total time spent to carry out the regression involving 1000 obser-
vations and 15 explanatory variables was 0.1 seconds. Three lines of code
accounted for 100% of the time and these are listed in order as: [54 44 47].
Line #54 accounted for 50% of the total time, whereas the qr decomposi-
tion on line #44 only accounted for 40% of the time. Line #54 computes
.::UdecomBooks::.
the mean of the y-vector used to determine the R-squared statistic. The
third slowest computation involved line #47 where the backsolution for the
coefficients took place, requiring 10% of the total time.
These results shed additional light on the speed differences between the
Cholesky and qr decomposition methods discussed earlier in Section 2.3.
Using either the Cholesky or qr method, we would still be spending the
majority of time computing the mean of the y-vector and backsolving for
the coefficients, and these computations account for 60% of the total time
required by the ols function. Saving a few hundredths of a second using the
Cholesky in place of the qr decomposition would not noticeably improve the
performance of the ols function from the user’s viewpoint.
The profile report on the prt reg function was as follows:
Total time in "prt_reg.m": 0.47 seconds

100% of the total time was spent on lines:
[367 354 36 364 141 140 139 137 135 125]
0.03s, 6% 36: if ~isstruct(results)
37: error(’prt_reg requires structure argument’);
0.01s, 2% 125: fprintf(fid,’\n’);
126: if strcmp(results.meth,’ols’)
134: if (nflag == 1)
0.01s, 2% 135: fprintf(fid,’Dependent Variable = %16s \n’,vnames(1,:));
136: end;
0.01s, 2% 137: fprintf(fid,’R-squared = %9.4f \n’,results.rsqr);
138: fprintf(fid,’Rbar-squared = %9.4f \n’,results.rbar);
0.01s, 2% 139: fprintf(fid,’sigma^2 = %9.4f \n’,results.sige);
0.01s, 2% 140: fprintf(fid,’Durbin-Watson = %9.4f \n’,results.dw);
0.01s, 2% 141: fprintf(fid,’Nobs, Nvars = %6d,%6d \n’,results.nob
142: fprintf(fid,’***************************************\n’);
353: tstat = results.tstat;
0.15s, 32% 354: tout = tdis_prb(tstat,nobs-nvar);
0.02s, 4% 364: fprintf(fid,’%16s %16s %16s %16s \n’,vstring,bstring,tst
366: for i=1:nvar;
0.21s, 45% 367: fprintf(fid,’%16s %16.6f %16.6f %16.6f\n’,Vname{i},tmp(i,
368: end;
Here we see that printing the results took 0.47 seconds, almost five times
the 0.10 seconds needed to compute the regression results. It is not surpris-
ing that computation of the marginal probabilities for the t-statistics on line
#354 took 32% of the total time. These computations require use of the in-
complete beta function which in turn draws on the log gamma function, both
of which are computationally intensive routines. Most of the time (45%) was
spent actually printing the output to the MATLAB command window which
is done in the ‘for-loop’ at line #367. (Note that we replaced the call to the
.::UdecomBooks::.
mprint function with the ‘for loop’ and explicit fprintf statements to make
it clear that printing activity actually takes time.)
One conclusion we should draw from these profiling results is that the
design decision to place computation of the marginal probabilities for the t-
statistics in the prt reg function instead of in the ols function makes sense.
Users who wish to carry out Monte Carlo experiments involving a large num-
ber of least-squares regressions and save the coefficient estimates would be
hampered by the slow-down associated with evaluating the marginal prob-
abilities in the ols function.
A second conclusion is that if we are interested in least-squares estimates
for β alone (as in the case of a two-stage least-squares estimation procedure),
we might implement a separate function named olsb. Computing β̂ coeffi-
cients using a specialized function would save time by avoiding computation
of information such as the R−squared statistic, that is not needed.
The comments concerning speed differences between Cholesky and qr
solutions to the least-squares problem are amplified by these profiling results.
It would take the same amount of time to print results from either solution
method, and as indicated, the time needed to print the results is five times
that required to compute the results!
We could delve into the time-profile for the tdis prb function which is
part of the distributions library discussed in Chapter 9. This turns out to
be un-enlightening as the routine spends 100% of its time in the MATLAB
incomplete beta function (betainc) which we cannot enhance.
A final point regarding use of the MATLAB profile command is that
you need to position MATLAB in the directory that contains the source file
for the function you are attempting to profile. For example, to profile the
tdis prb function which is in the ‘distrib’ directory, we need to move to
this directory before executing the profile tdis prb and profile report
commands.
2.6 Using the regression library

This section presents examples showing how some of the various regression
functions in the library might be used to solve econometric estimation prob-
lems. It seems a good idea when creating a MATLAB function to provide
an example program that illustrates use of the function. This has been
done for most of the functions in the Econometrics Toolbox. Demonstration
programs have an ‘underscore d’ attached to the function name. The file
‘ols d.m’ would contain a demonstration program for the ols function, and
.::UdecomBooks::.
the file ‘ridge d.m’ a demo of the ridge function.

Basic use of the regression functions follows a canned format involving:
1. read in the sample data
2. perform any transformations or calculations necessary to form the set

of explanatory variables and the dependent variable.
3. send the dependent and independent variables to the regression func-

tion for processing.
4. send the structure returned by the regression function to the prt or

plt function to print or plot results.
For specific examples of this canned format you can examine the demon-
stration files in the regression function library.
In this section, we wish to go beyond simple demonstrations of the var-
ious estimation procedures to illustrate how the results structures can be
useful in computing various econometric statistics and performing hypothe-
sis tests based on regression results. This section contains sub-sections that
illustrate various uses of the regression function library and ways to produce
new functions that extend the library. Later chapters illustrate additional
regression functions not discussed here.
2.6.1 A Monte Carlo experiment

As an initial demonstration of the ols function, consider a Monte Carlo
experiment where we generate 100 different data sets based on the same
explanatory variable matrix X and 100 different disturbance vectors ε that
are used to produce 100 sample y vectors.
We wish to carry out 100 regressions and save the estimates β̂ from
each regression. We will use the 100 sets of estimates to compute a mean
over the 100 trials, that should be close to the true values of β used to
generate the 100 different sample y vectors. This experiment illustrates to
introductory econometrics students the nature of the unbiasedness property
associated with the least-squares estimates. Although the means over 100
different samples are close to the true values for β, the amount of dispersion
in individual sample outcomes is usually a surprise to students.
% ----- Example 2.4 Using the ols() function for Monte Carlo
nobs = 100; nvar = 5; ntrials = 100;
b = ones(nvar,1); % true betas = 1
.::UdecomBooks::.
x = [ones(nobs,1) randn(nobs,nvar-1); % fixed x-matrix

bout = zeros(ntrials,nvar); % storage for estimates
for i=1:ntrials; % do ols in a for-loop
evec = randn(nobs,1); y = x*beta + evec;
out = ols(y,x); bout(i,:) = out.beta’; % save bhat’s
end;
bm = mean(bout); bs = std(bout); % find mean and std of bhats
fprintf(1,’Mean of the bhats \n’);
for i=1:nvar; fprintf(1,’%8.4f \n’,bmean(1,i)); end;
fprintf(1,’Std deviation of the bhats \n’);
for i=1:nvar; fprintf(1,’%8.4f \n’,bstd(1,i)); end;
% provide a histogram for each bhat
hist(bout); ylabel(’frequency of \beta outcomes’);
xlabel(’Estimated \beta values’);
legend(’\beta_1’,’\beta_2’,’\beta_3’,’\beta_4’,’\beta_5’);
We recover the estimates β̂ from the ‘results’ structure each time through
the loop, transpose and place them in the ‘ith’ row of the matrix ‘bsave’.
After the loop completes, we compute the mean and standard deviations of
the estimates and print these out for each of the 5 coefficients. MATLAB
mean and std functions work on the columns of matrices, motivating our
storage scheme for the ‘bsave’ matrix.
To provide a graphical depiction of our results, we use the MATLAB
hist function to produce a histogram of the distribution of estimates. The
hist function works on each column when given a matrix input argument,
producing 5 separate histograms (one for each column) that are color-coded.
We used the LaTeX notation, ‘backslash beta’ in the ‘ylabel’, ‘xlabel’ and
‘legend’ commands to produce a Greek symbol, β in the y and x labels and
the LaTeX underscore to create the subscripted βi , i = 1, . . . , 5 symbols in
the legend. Figure 2.1 shows the resulting 5-way histogram.
It is clear from Figure 2.1 that although the estimates are centered on
the true value of unity, the distribution extends down to 0.75 and up to 1.25.
The implication is that particular data samples may produce estimates far
from truth despite the use of an unbiased estimation procedure.
2.6.2 Dealing with serial correlation

This discussion illustrates how one would construct Cochrane-Orcutt and
maximum likelihood estimates in the face of first-order serial correlation in
the disturbances using MATLAB functions and the regression library.
In example 2.5, we generate a regression model with serially correlated
disturbances as in (2.3).
.::UdecomBooks::.
30
β
1
β
2
β3
25
β4
β5
20
frequency of β outcomes
15
10
0
0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4
Estimated β values
Figure 2.1: Histogram of β̂ outcomes
yt = X t β + u t
ut = ρut−1 + εt (2.3)
Example 2.5 carries out least-squares estimation, prints and plots the results
for the generated data set.
% ----- Example 2.5 Generate a model with serial correlation

n = 200; k = 3; evec = randn(n,1);
xmat = [ones(n,1) randn(n,k)]; y = zeros(n,1); u = zeros(n,1);
beta = ones(k,1); beta(1,1) = 10.0; % constant term
for i=2:n; % generate a model with 1st order serial correlation
u(i,1) = 0.4*u(i-1,1) + evec(i,1);
y(i,1) = xmat(i,:)*beta + u(i,1);
end;
% truncate 1st 100 observations for startup
.::UdecomBooks::.
yt = y(101:n,1); xt = xmat(101:n,:);
n = n-100; % reset n to reflect truncation
Vnames = strvcat(’y’,’cterm’,’x2’,’x3’);
result = ols(yt,xt); prt(result,Vnames); plt(result);
Cochrane-Orcutt estimation involves iteration using the least-squares

residuals in a regression on the residuals lagged one period to estimate new
values for ρ conditional on the least-squares estimates of β. The new value
of ρ is used to produce an updated estimate of β̂ and this process is iterated.
When the values for ρ converge, we stop the iteration process. The iteration
is carried out using a ‘while loop’ that checks for convergence every time
through the loop. Example 2.6 relies on the variables xt,yt generated in
the previous example to construct Cochrane-Orcutt estimates.
% ----- Example 2.6 Cochrane-Orcutt iteration
% NOTE: assumes xt,yt exist (from example 2.5)
converg = 1.0; rho = 0.0; iter = 1; xl = lag(xt,1); yl = lag(yt,1);
xlag = xl(2:n,:); ylag = yl(2:n,1); % truncate to feed the lag
y = yt(2:n,1); x = xt(2:n,:); n = n-1; % adjust n for truncation
Vnames = strvcat(’ystar’,’istar’,’x2star’,’x3star’);
disp(’Cochrane-Orcutt Estimates’);
while (converg > 0.0001),
% step 1, using initial rho = 0, do OLS to get bhat
ystar = y - rho*ylag; xstar = x - rho*xlag; res = ols(ystar,xstar);
% compute residuals based on beta ols estimates
e = y - x*res.beta; elag = lag(e);
% truncate 1st observation to account for the lag
et = e(2:n,1); elagt = elag(2:n,1);
% step 2, update estimate of rho using residuals from step 1
res_rho = ols(et,elagt); rho_last = rho; rho = res_rho.beta(1);
converg = abs(rho - rho_last);
fprintf(1,’rho = %8.5f, converg = %8.5f, iter = %2d \n’,rho,converg,iter);
iter = iter + 1;
end; % end of while loop
% after convergence produce a final set of estimates using rho-value
ystar = y - rho*ylag; xstar = x - rho*xlag;
res = ols(ystar,xstar); prt(res,Vnames);
The output from the Cochrane-Orcutt estimation in example 2.6 looks

as follows:
Cochrane-Orcutt Estimates
rho = 0.38148, converg = 0.38148, iter = 1
rho = 0.39341, converg = 0.01193, iter = 2
rho = 0.39363, converg = 0.00022, iter = 3
rho = 0.39364, converg = 0.00000, iter = 4
.::UdecomBooks::.
Dependent Variable = ystar

R-squared = 0.711
sigma^2 = 0.904
Nobs, Nvars = 99, 3
*********************************************************************
istar 9.80742294 62.19721521 0.00000000
x2star 0.97002428 9.99808285 0.00000000
x3star 1.05493208 12.06190233 0.00000000
As an exercise, consider packaging this procedure as a MATLAB function

that we could add to our regression library. This would involve:
1. Transferring the above code to a function and providing documenta-

tion for use of the function.
2. returning a regression function ‘results structure’ that is compatible

with the prt reg function from the library.
3. modifying the prt reg function by adding specific code for printing
the output. Your printing code should probably present the results
from iteration which can be passed to the prt reg function in the
results structure.
You might compare your code to that in olsc.m from the regression
function library.
Maximum likelihood estimation of the least-squares model containing
serial correlation requires that we simultaneously minimize the negative of
the log-likelihood function with respect to the parameters ρ, β and σε in the
problem. This can be done using a simplex minimization procedure that
exists as part of the MATLAB toolbox. Other more powerful multidimen-
sional optimization procedures are discussed in Chapter 10 which takes up
the topic of general maximum likelihood estimation of econometric models.
We will use a MATLAB function that minimizes a function of several
variables using a simplex algorithm. The function is named fmins and it
has the following input format for our application:
[xout, ooptions] = fmins(‘ar1_like’,xin,ioptions,[],y,x);
The string input argument ‘ar1 like’ is the name of a MATLAB function
we must write to evaluate the negative of the log-likelihood function. The
.::UdecomBooks::.
argument ‘xin’ is a vector of starting values for the parameters, ‘ioptions’

is a 4x1 vector of optimization input options, and the ‘y,x’ after the empty
matrix input argument ‘[]’ are the data vectors that will be passed to our
function ar1 like. The fmins function produces an output vector ‘xout’
containing the vector of parameter values that minimize the negative of the
log-likelihood function and a vector ‘ooptions’ containing information about
the optimization problem. For example, the function value at the minimum,
the number of iterations taken to find a minimum, and so on.
Turning attention to the function ar1 like, we have a few options here.
Some authors refer to the likelihood function for this model conditional on
the first observation as:
X
n
L(ρ, β) = (et − ρet−1 )2 (2.4)
t=2
But, this likelihood ignores the first observation and would only be appro-
priate if we do not view the serial correlation process as having been in
operation during the past. The function we use is not conditional on the
first observation and takes the form (see Green, 1997 page 600):
L(ρ, β) = −(n/2)ln(y ∗ − X ∗ β)0 (y ∗ − X ∗ β) + (1/2)ln(1 − ρ2 ) (2.5)

y ∗ = P y, X∗ = P X
 p 
(1 − ρ2 0 0 . . . 0 0
 −ρ 1 0 ... 0 0 
 
 
 0 −ρ 1 . . . 0 0 
P = 
 .. .. .. . . . ..

 . . . . .. .
 
 0 0 0 ... 1 0 
0 0 0 . . . −ρ 1
where σε2 has been concentrated out of the likelihood function. We compute
this parameter estimate using e∗0 e∗ /(n − k), where e∗ = y ∗ − X ∗ β̂.
The MATLAB function ar1 like to compute this log-likelihood for var-
ious values of the parameters ρ and β is shown below:
function llike = ar1_like(param,y,xmat)

% PURPOSE: evaluate log-likelihood for ols model with AR1 errors
%-----------------------------------------------------
% USAGE: like = ar1_like(b,y,x)
% where: b = parameter vector (k x 1)
% y = dependent variable vector (n x 1)
.::UdecomBooks::.
% x = explanatory variables matrix (n x k-1)

%-----------------------------------------------------
% NOTE: this function returns a scalar equal to -log(likelihood)
% b(1,1) contains rho parameter
% sige is concentrated out
%-----------------------------------------------------
% REFERENCES: Green, 1997 page 600
%-----------------------------------------------------
[n k] = size(x); rho = param(1,1); beta = param(2:2+k-1,1);
ys = y - rho*lag(y); xs = x - rho*lag(x);
ys(1,1) = sqrt(1-rho*rho)*y(1,1); xs(1,:) = sqrt(1-rho*rho)*x(1,:);
term1 = -(n/2)*log((ys - xs*beta)’*(ys - xs*beta));
term2 = 0.5*log(1-rho*rho);
like = term1+term2;
llike = -like;
Note that we do not strictly follow the mathematical formulas presented

in (2.6) when coding ar1 like to evaluate the negative of the log likelihood
function. Use of the large matrix P would be inefficient and require addi-
tional computer memory for storage of the sparse matrix P .
The code shown in example 2.7 to implement maximum likelihood es-
timation sets initial values for the parameters ρ, β based on the Cochrane-
Orcutt estimates produced by the code described previously. (They are
assumed to exist in this example.) We then set some options for use in the
optimization routine, and make a call to the fmins function using the data
vector y and matrix X.
% ----- Example 2.7 Maximum likelihood estimation
% NOTE: assumes Cochrane-Orcutt estimates exist (from example 2.6)
param = zeros(k+2,1); % initial values from Cochrane-Orcutt
param(1,1) = rho; % initial rho
param(2:2+k-1,1) = res.beta; % initial bhat’s
options(1,1) = 0; % no print of intermediate results
options(2,1) = 0.0001; % simplex convergence criteria
options(3,1) = 0.0001; % convergence criteria for function
options(4,1) = 1000; % set # of function evaluations
[mrho ooptions] = fmins(’ar1_like’,param,options,[],y,x);
niter = ooptions(1,10); % recover some info on optimization
llike = ooptions(1,8);
% find sig estimate based on ml rho and bhat’s
ystar = y - rho*ylag; xstar = x - rho*xlag; bhat = mrho(2:2+k-1,1);
sig = (ystar - xstar*bhat)’*(ystar - xstar*bhat); sig = sig/(n-k);
fprintf(1,’Maximum likelihood estimate of rho = %8.4f \n’,mrho(1,1));
fprintf(1,’Maximum likelihood estimate of sigma = %8.4f \n’,sige);
fprintf(1,’Maximum likelihood estimates of bhat = %8.4f \n’,bhat);
fprintf(1,’negative of Log-likelihood value = %8.4f \n\n’,llike);
fprintf(1,’number of iterations taken = %4d \n\n’,niter);
.::UdecomBooks::.
The resulting parameter estimates for ρ, β are returned from the fmins
function and used to compute an estimate of σε2 . These results are then
printed. For an example of a function in the regression library that imple-
ments this approach, see olsar1.
2.6.3 Implementing statistical tests

In this example, we demonstrate how the ‘results’ structure returned by the
ols function can be used in a function named waldf that carries out the
Wald F-test for a restricted versus unrestricted model. A joint F-test is
used to test whether the zero restrictions produced by omitting variables
from a model are consistent with the sample data.
We generate a data set containing ten explanatory variables and carry
out a regression. A second model based on only the first five explanatory
variables represents a restricted model. We wish to carry out a joint F-test
for the significance of variables 6 through 10 in the model. This involves
computing an F-statistic based on the residuals from the restricted (e0 er )
and unrestricted (e0 eu ) models:
F = (e0 er − e0 eu )/m
(2.6)
e0 eu /(n − k)
where m is the number of restrictions (five in our example) and n, k are the
number of observations and number of explanatory variables in the unre-
stricted model respectively.
We use ols for the two regressions and send the ‘results’ structures to a
function waldf that will carry out the joint F-test and return the results.
Assuming the existence of a function waldf to implement the joint F-
test, the MATLAB code to carry out the test would be as shown in example
2.8.
% ----- Example 2.8 Wald’s F-test

nobs = 200; nvar = 10; b = ones(nvar,1); b(7:nvar,1) = 0.0;
xmat1 = randn(nobs,nvar-1); evec = randn(nobs,1)*5;
x = [ones(nobs,1) xmat1]; % unrestricted data set
y = x*b + evec; x2 = x(:,1:5); % restricted data set
resultu = ols(y,x); prt(resultu); % do unrestricted ols regression
resultr = ols(y,x2); prt(resultr); % do restricted ols regression
% test the restrictions
[fstat fprob] = waldf(resultr,resultu);
disp(’Wald F-test results’); [fstat fprob]
.::UdecomBooks::.
All of the information necessary to carry out the joint F-test resides in
the two results structures, so the function waldf needs only these as input
arguments. Below is the waldf function.
function [fstat, fprb] = waldf(resultr,resultu)
% PURPOSE: computes Wald F-test for two regressions
%---------------------------------------------------
% USAGE: [fstat fprob] = waldf(resultr,resultu)
% Where: resultr = results structure from ols() restricted regression
% resultu = results structure from ols() unrestricted regression
%---------------------------------------------------
% RETURNS: fstat = {(essr - essu)/#restrict}/{essu/(nobs-nvar)}
% fprb = marginal probability for fstat
% NOTE: large fstat => reject as inconsistent with the data
%---------------------------------------------------
if nargin ~= 2 % flag incorrect arguments
error(’waldf: Wrong # of input arguments’);
elseif isstruct(resultu) == 0
error(’waldf requires an ols results structure as input’);
elseif isstruct(resultr) == 0
error(’waldf requires an ols results structure as input’);
end;
% get nobs, nvar from unrestricted and restricted regressions
nu = resultu.nobs; nr = resultr.nobs;
ku = resultu.nvar; kr = resultr.nvar;
if nu ~= nr
error(’waldf: the # of obs in the results structures are different’);
end;
if (ku - kr) < 0 % flag reversed input arguments
error(’waldf: negative dof, check for reversed input arguments’);
end;
% recover residual sum of squares from .sige field of the result structure
epeu = resultu.sige*(nu-ku); eper = resultr.sige*(nr-kr);
numr = ku - kr; % find # of restrictions
ddof = nu-ku; % find denominator dof
fstat1 = (eper - epeu)/numr; % numerator
fstat2 = epeu/(nu-ku); % denominator
fstat = fstat1/fstat2; fprb = fdis_prb(fstat,numr,ddof);
The only point to note is the use of the function fdis prb which re-
turns the marginal probability for the F-statistic based on numerator and
denominator degrees of freedom parameters. This function is part of the
distribution functions library discussed in Chapter 9.
As another example, consider carrying out an LM specification test for
the same type of model based on restricted and unrestricted explanatory
variables matrices. This test involves a regression of the residuals from the
restricted model on the explanatory variables matrix from the unrestricted
.::UdecomBooks::.
model. The test is then based on the statistic computed using n ∗ R2 from
this regression, which is chi-squared (χ2 ) distributed with degrees of freedom
equal to the number of restrictions.
We can implement a function lm test that takes as input arguments the
‘results’ structure from the restricted regression and an explanatory variables
matrix for the unrestricted model. The lm test function will: carry out the
regression of the residuals from the restricted model on the variables in
the unrestricted model, compute the chi-squared statistic, and evaluate the
marginal probability for this statistic. We will return a regression ‘results’
structure from the regression performed in the lm test function so users
can print out the regression if they so desire. Example 2.9 shows a program
that calls the lm test function to carry out the test.
% ----- Example 2.9 LM specification test

nobs = 200; nvar = 7; beta = ones(nvar,1);
beta(6:nvar,1) = 0.0; evec = randn(nobs,1)*5;
xmat1 = randn(nobs,nvar-1);
x = [ones(nobs,1) xmat1]; % unrestricted data set
xr = x(:,1:5); % restricted data set
y = x*beta + evec;
resultr = ols(y,xr); prt(resultr); % restricted ols
[lmstat lmprob result_lm] = lm_test(resultr,x); % do LM - test
disp(’LM-test regression results’); prt(result_lm);
disp(’LM test results’); [lmstat lmprob]
Note that one of the arguments returned by the lm test function is a re-
gression ‘results’ structure that can be printed using the prt function. These
regression results reflect the regression of the residuals from the restricted
model on the unrestricted data matrix. The user might be interested in see-
ing which particular explanatory variables in the unrestricted model matrix
exhibited a significant influence on the residuals from the restricted model.
The results structure is passed back from the lm test function just as if it
were any scalar or matrix argument being returned from the function.
The function lm test is shown below.
function [lmstat, lmprob, reslm] = lm_test(resr,xu)

% PURPOSE: computes LM-test for two regressions
%---------------------------------------------------
% USAGE: [lmstat lmprob, result] = lm_test(resultr,xmatrixu)
% Where: resultr = matrix returned by ols() for restricted regression
% xmatrixu = explanatory variables matrix from unrestricted model
%---------------------------------------------------
% RETURNS: lmstat = calculated chi-squared statistic
% lmprob = marginal probability for lmstat
.::UdecomBooks::.
% result = ols() results structure

%---------------------------------------------------
% NOTE: Expected value of (lmstat) = #restrictions
%---------------------------------------------------
if nargin ~= 2, error(’lm_test: Wrong # of inputs’);
elseif isstruct(resr) == 0, error(’lm_test: ols structure needed’);
end;
nobs1 = resultr.nobs; [nobs rsize] = size(xu);
if nobs1 ~= nobs
error(’lm_test: both inputs should have same # of obs’);
end;
er = resultr.resid; % recover residuals from restricted model
reslm = ols(er,xu); % step 1) regress er on all the x’s
rsqr = reslm.rsqr; % step 2) recover R^2
lmstat = rsqr*nobs; % compute chi-squared statistic
% determine the # of restrictions
[junk ku] = size(xu); kr = resultr.nvar; nrestrict = ku - kr;
lmprob = 1.0 - chis_prb(lmstat,nrestrict);
2.7 Chapter summary

We designed a general framework for regression functions that share common
functions for printing and plotting results. The use of MATLAB Version 5
structure variables allowed us to pass regression estimation results between
these functions using a single structure variable. This design also allows
us to implement auxiliary functions to carry out traditional econometric
statistical tests.
In Chapter 3 we will see how a single function prt and another plt can be
constructed to print and plot results from all functions in the Econometrics
Toolbox, including those in the regression library. This facility also derives
from the powerful possibilities associated with MATLAB structure variables.
.::UdecomBooks::.
Chapter 2 Appendix
The regression function library is in a subdirectory regress.
regression function library
------- regression program functions -----------
ar_g - Gibbs sampling Bayesian autoregressive model

bma_g - Gibbs sampling Bayesian model averaging
boxcox - Box-Cox regression with 1 parameter
boxcox2 - Box-Cox regression with 2 parameters
egarchm - EGARCH(p,q)-in-Mean regression model
emhergott - EM estimates of Hamilton’s markov switching model
garchs - garchs(1,1,1) model using non-central t-distribution
hmarkov_em - Hamilton’s markov switching model
hwhite - Halbert White’s heteroscedastic consistent estimates
lad - least-absolute deviations regression
lm_test - LM-test for two regression models
logit - logit regression
mlogit - multinomial logit regression
nwest - Newey-West hetero/serial consistent estimates
ols - ordinary least-squares
ols_g - Gibbs sampling Bayesian linear model
olsar1 - Maximum Likelihood for AR(1) errors ols model
olsc - Cochrane-Orcutt AR(1) errors ols model
olst - regression with t-distributed errors
probit - probit regression
probit_g - Gibbs sampling Bayesian probit model
ridge - ridge regression
robust - iteratively reweighted least-squares
rtrace - ridge estimates vs parameters (plot)
sur - seemingly unrelated regressions
switch_em - switching regime regression using EM-algorithm
theil - Theil-Goldberger mixed estimation
thsls - three-stage least-squares
tobit - tobit regression
tobit_g - Gibbs sampling Bayesian tobit model
tsls - two-stage least-squares
42
.::UdecomBooks::.
waldf - Wald F-test
-------- demonstration programs -----------------
ar_gd - demonstration of Gibbs sampling ar_g

bma_gd - demonstrates Bayesian model averaging
boxcox_d - demonstrates Box-Cox 1-parameter model
boxcox2_d - demonstrates Box-Cox 2-parmaeter model
demo_all - demos most regression functions
egarchm_d - demos egarchm function
garchs_d - demos garchs function
hmarkov_emd - demos Hamilton’s model
hmarkov_emd2 - another demo of Hamilton’s model
hwhite_d - H. White’s hetero consistent estimates demo
lad_d - demos lad regression
lm_test_d - demos lm_test
logit_d - demonstrates logit regression
mlogit_d - demonstrates multinomial logit
nwest_d - demonstrates Newey-West estimates
ols_d - demonstrates ols regression
ols_d2 - Monte Carlo demo using ols regression
ols_gd - demo of Gibbs sampling ols_g
olsar1_d - Max Like AR(1) errors model demo
olsc_d - Cochrane-Orcutt demo
olst_d - olst demo
probit_d - probit regression demo
probit_gd - demo of Gibbs sampling Bayesian probit model
ridge_d - ridge regression demo
robust_d - demonstrates robust regression
sur_d - demonstrates sur using Grunfeld’s data
switch_emd - demonstrates switching regression
theil_d - demonstrates theil-goldberger estimation
thsls_d - three-stage least-squares demo
tobit_d - tobit regression demo
tobit_d2 - tobit right-censoring demo
tobit_gd - demo of Gibbs sampling Bayesian tobit model
tobit_gd2 - Bayesian tobit right-censoring demo
tsls_d - two-stage least-squares demo
waldf_d - demo of using wald F-test function
-------- Support routines ------------------------
ar1_like - used by olsar1 (likelihood)

bmapost - used by bma_g
box_lik - used by box_cox (likelihood)
box_lik2 - used by box_cox2 (likelihood)
chis_prb - computes chi-squared probabilities
dmult - used by mlogit
egarchm_lik - likelihood used by egarchm
.::UdecomBooks::.
garchs_llf - likelihood used by garchs

herg_llf - likelihood used by emhergott
herg_llf2 - likelihood used by emhergott
hmarkov_llf - likelihood used by hmarkov_em
hmarkov_llf2 - likelihood used by hmarkov_em
fdis_prb - computes F-statistic probabilities
find_new - used by bma_g
grun.dat - Grunfeld’s data used by sur_d
grun.doc - documents Grunfeld’s data set
hessian - used by tobit to determine numerical hessian
lo_like - used by logit (likelihood)
mcov - used by hwhite
mderivs - used by mlogit
mlogit_lik - used by mlogit
nmlt_rnd - used by probit_g
nmrt_rnd - used by probit_g, tobit_g
norm_cdf - used by probit, pr_like
norm_pdf - used by prt_reg, probit
olse - ols returning only residuals (used by sur)
plt_eqs - plots equation systems
plt_reg - plots regressions
pr_like - used by probit (likelihood)
prt_eqs - prints equation systems
prt_gibbs - prints Gibbs sampling models
prt_reg - prints regressions
prt_swm - prints switching regression results
sample - used by bma_g
stdn_cdf - used by norm_cdf
stdn_pdf - used by norm_pdf
tdis_prb - computes t-statistic probabilities
to_like - used by tobit (likelihood)
.::UdecomBooks::.
Chapter 3
Utility Functions
This chapter presents utility functions that we will use throughout the re-
mainder of the text. We can categorize the various functions described here
into functions for:
1. Working with time-series that contain dates. MATLAB is relatively

weak in this regard when compared to standard econometric programs
for working with economic time-series like RATS and TSP.
2. General functions for printing and plotting matrices as well as produc-

ing LaTeX formatted output of matrices for tables.
3. Econometric data transformations.
4. Some functions that mimic Gauss functions, allowing us to more eas-

ily re-code the many econometric procedures available for Gauss to
MATLAB functions.
The material in this chapter is divided into four sections that correspond
to the types of utility functions enumerated above. A final section discusses
development of “wrapper” functions that call other functions to make print-
ing and plotting econometric procedure results simpler for the user.
3.1 Calendar function utilities

Working with time-series data often requires that we associate observations
with calendar dates, where the dates vary depending on the frequency of
the data. We construct three functions that help in this task.
45
.::UdecomBooks::.
CHAPTER 3. UTILITY FUNCTIONS 46
The first function cal stores information regarding the calendar dates
covered by our time-series data in a MATLAB structure variable. We would
use the function at the outset of our analysis to package information regard-
ing the starting year, period and frequency of our data series. This packaged
structure of calendar information can then be passed to other time series
estimation and printing functions.
For example, consider a case where we are dealing with monthly data
that begins in January, 1982, we would store this information using the call:
cal_struct = cal(1982,1,12);
which returns a structure, ‘cal struct’ with the following fields for the be-
ginning year, period and frequency.
cal_struct.beg_yr
cal_struc.beg_per
cal_struc.freq
The field ‘beg yr’ would equal 1982, ‘beg per’ would equal 1 for January,
and ‘freq’ would equal 12 to designate monthly data. Beyond setting up
calendar information regarding the starting dates for our time-series data,
we might want to know the calendar date associated with a particular ob-
servation. The cal function can provide this information as well, returning
it as part of the structure. The documentation for the function is:
PURPOSE: create a time-series calendar structure variable that
associates a date with an observation #
-----------------------------------------------------
USAGE: result = cal(begin_yr,begin_per,freq,obs)
or: result = cal(cstruc,obs)
where: begin_yr = beginning year, e.g., 1982
begin_per = beginning period, e.g., 3
freq = frequency, 1=annual,4=quarterly,12=monthly
obs = optional argument for an observation #
cstruc = a structure returned by a previous call to cal()
-----------------------------------------------------
RETURNS: a structure:
result.beg_yr = begin_yr
result.beg_per = begin_period
result.freq = frequency
result.obs = obs (if input)
result.year = year for obs (if input)
result.period = period for obs (if input)
Calling cal with an observation number argument, will produce a return

structure that associates that particular observation with a year and period,
.::UdecomBooks::.
recorded in the ‘.year’ and ‘.period’ fields. We also include the observation
number in the field ‘.obs’ of the returned structure. To illustrate use of this
function, consider that our data series start in January, 1982 and we wish to
determine the calendar year and period associated with observation number
83, we would make the call:
cal_struct = cal(1982,1,12,83)
ans =
beg_yr: 1982
beg_per: 1
freq: 12
obs: 83
year: 1988
period: 11
which informs us that observation number 83 is associated with November,

1988.
Another option allows you to call cal with a structure returned by a
previous call to the function itself. This might be useful if you set up a
calendar structure at the outset of your program and later wish to find
observation numbers associated with dates. Here is an example of this type
of usage.
% ----- Example 3.1 Using the datesf() function

load test.dat; % monthly mining employment for il,in,ky,mi,oh,pa,tn,wv
dates = cal(1982,1,12); % data covers 1982,1 to 1996,5
begf = ical(1995,6,dates); % beginning forecast period
nfor = 12; % number of forecast periods
datesf = cal(dates,begf+nfor-1); % add end of forecast period
Note that we could use: datesf = cal(1982,1,12,begf+nfor-1), so

the ability to use the previously returned structure ‘dates’ simply saves us
some typing.
The function ical used in the example above, serves as an inverse to the
cal function, returning an observation number associated with a particular
calendar date. It utilizes the information stored in the structure returned by
cal to determine this information. Forecasting provides an example where
we might use this function. Suppose our data series begins in January 1980
and we wish to produce a forecast beginning in June, 1991 we could utilize
the following calls to cal and ical to accomplish this task.
cstr = cal(1980,1,12);
begf = ical(1991,6,cstr);
.::UdecomBooks::.
This would set begf=138, which we can utilize when calling our fore-
casting function, or simply printing the data series observations for this
particular date.
The third function, tsdate allows us to display and print date strings
associated with annual, quarterly or monthly time-series, which is helpful
for printing results and forecasted values with date labels. Documentation
for this function is:
PURPOSE: produce a time-series date string for an observation #
given beginning year, beginning period and frequency of the data
---------------------------------------------------
USAGE: out = tsdate(beg_yr,beg_period,freq,obsn);
or: tsdate(beg_yr,beg_period,freq,obsn);
or: tsdate(cal_struct,obsn);
where: beg_yr = beginning year, e.g., 1974
beg_period = beginning period, e.g., 4 for april
freq = 1 for annual, 4 for quarterly 12 for monthly
obsn = the observation #
cal_struct = a structure returned by cal()
e.g., tsdate(1974,1,12,13) would print: Jan75

tsdate(1974,1,4,13) would print: Q1-77
tsdate(1974,1,1,13) would print 1986
out = tsdate(1974,1,12,13); would return a string ‘Jan75’
cstr = cal(1974,1,12);
tsdate(cstr,13); would print Jan75
---------------------------------------------------
RETURNS: a string, or: simply displays the date associated
with obsn (observation #) in the command window
The function uses MATLAB ‘nargout’ to determine whether the user has
requested an output argument or not. In the event that no output argument
is requested, the function calls the MATLAB datestr function with no semi-
colon after the call to produce a printed string in the MATLAB command
window. For cases where the user desires more control over the printing
format of the date string, or wishes to include the date string in an output
file, we can call the function with an output argument. As an illustration of
the first case, consider the following code:
% ----- Example 3.2 Using the tsdate() function
cstruct = cal(1982,1,4);
for i=1:2;
tsdate(cstruct,i);
end;
which would produce the following output in the MATLAB command win-
dow:
.::UdecomBooks::.
Q1-82
Q2-82
On the other hand, we could make a series of calls providing a MATLAB

cell-variable array as an output argument:
% ----- Example 3.3 Using cal() and tsdates() functions
cstruct = cal(1982,1,12);
for i=1:6;
fdates{i} = tsdate(cstruct,i);
end;
that would place each date in the cell-array for use later in printing. The
cell-array looks as follows:
fdates =
’Jan82’ ’Feb82’ ’Mar82’ ’Apr82’ ’May82’ ’Jun82’
Further illustrations will be provided in Chapter 5 that demonstrate how

these utility functions are used in conjunction with the vector autoregres-
sive modeling functions. To summarize, we provide the following example
program that utilizes the functions to read and print time-series data.
% ----- Example 3.4 Reading and printing time-series
dates = cal(1982,1,12);
load test.dat; % monthly time-series data starting in Jan,1982
[nobs nvar] = size(test);
endd = tsdate(dates,nobs); % date of the last observation
fprintf(1,’The ending date of the sample is %10s \n\n’,endd);
begin_prt = ical(1990,1,dates); % print data with dates for 1990
end_prt = ical(1990,12,dates);
fprintf(1,’The data for 1990 is: \n\n’);
for i=begin_prt:end_prt; fprintf(1,’%10s ’,tsdate(dates,i));
for j=1:nvar fprintf(1,’%8.0f ’,test(i,j)); end;
fprintf(1,’\n’); end;
One comment is that we construct a function tsprint in the next sec-

tion that makes it easier to print time-series data than the approach demon-
strated in the example above.
3.2 Printing and plotting matrices

Printing matrices and time-series is a frequent occurrence in econometric
work, as is the task of producing tabular results for inclusion in word-
processing. This section presents three utility functions to help with this
.::UdecomBooks::.
task. We also provide a function to plot time-series with dates on the time-
axis.
The function lprint transforms a MATLAB matrix by adding formatting
symbols needed to produce a table in LaTeX, a widely-used mathematical
typesetting program.
PURPOSE: print an (nobs x nvar) matrix in LaTeX table format

---------------------------------------------------
USAGE: lprint(x,info)
where:
x = (nobs x nvar) matrix (or vector) to be printed
info = a structure containing printing options
info.begr = beginning row to print, (default = 1)
info.endr = ending row to print, (default = nobs)
info.begc = beginning column to print, (default = 1)
info.endc = ending column to print, (default = nvar)
info.cnames = an (nvar x 1) string of names for columns (optional)
e.g. info.cnames = strvcat(’col1’,’col2’);
(default = no column headings)
info.rnames = an (nobs+1 x 1) string of names for rows (optional)
e.g. info.rnames = strvcat(’Rows’,’row1’,’row2’);
(default = no row labels)
info.fmt = a format string, e.g., ’%12.6f’ or ’%12d’ (default = %10.4f)
or an (nvar x 1) string containing formats
e.g., info.fmt=strvcat(’%12.6f’,’%12.2f’,’%12d’);
info.fid = file-id for printing results to a file
(defaults to the MATLAB command window)
e.g. fid = fopen(’file.out’,’w’);
info.rflag = 1 row #’s printed, 0 no row #’s (default = 0)
---------------------------------------------------
e.g. in.cnames = strvcat(’col1’,’col2’);
in.rnames = strvcat(’rowlabel’,’row1’,’row2’);
lprint(y,in), prints entire matrix, column and row headings
in2.endc = 3; in2.cnames = strvcat(’col1’,’col2’,’col3’);
or: lprint(y,in2), prints 3 columns of the matrix, just column headings
or: lprint(y), prints entire matrix, no column headings or row labels
NOTES: - defaults are used for info-elements not specified
- wrapping occurs at (80/format) columns, which varies with
format used, e.g. %10.2f will wrap after 8 columns
---------------------------------------------------
SEE ALSO: tsprint, mprint, lprint_d
---------------------------------------------------
This is the first example of a function that uses the MATLAB structure
variable as an input argument. This allows us to provide a large number of
input arguments using a single structure variable. It also simplifies parsing
the input arguments in the function and setting default options.
.::UdecomBooks::.
As an example of using the function, consider the following program

that generates a series of matrices containing normally distributed random
numbers and transforms them to LaTeX format using lprint and various
formatting options. Note that you can name the structure variable used to
input the options anything — it is the fieldnames that the function lprint
parses to find the options.
% ----- Example 3.5 Using the lprint() function
table = randn(5,3); fprintf(1,’using no options \n’);
lprint(table); % no printing options used
table2 = round(table)*1000; fprintf(1,’using fmt option \n’);
option.fmt = ’%10.0f’; lprint(table2,option); % format option used
fprintf(1,’using column names and fmt option \n’);
inarg.cnames = strvcat(’Illinois’,’Ohio’,’Indiana’);
inarg.fmt = ’%12.3f’; lprint(table2,inarg); % format, column names
fprintf(1,’row and column labels \n’);
inarg.rnames = strvcat(’Rows’,’row1’,’row2’,’row3’,’row4’,’row5’);
lprint(table2,inarg); % adding row-labels
fprintf(1,’wrapped output for large matrices \n’);
vnames = strvcat(’IL’,’OH’,’IN’,’WV’,’PA’,’AK’,’HI’,’NY’,’MS’,’TN’);
table3 = randn(5,10); option2.fmt = ’%20.4f’; option2.cnames = vnames;
lprint(table3,option2); % wrapping matrices
fprintf(1,’generic row labels \n’);
table4 = randn(4,4)*100; option3.fmt = ’%10.4f’; option3.rflag = 1;
option3.cnames = strvcat(’column1’,’column2’,’column3’,’column4’);
lprint(table4,option3); % generic row-labels
fprintf(1,’variable column formats \n’);
table4(:,1) = round(table4(:,1)); table4(:,3) = round(table4(:,3));
option4.fmt = strvcat(’%10d’,’%8.3f’,’%10d’,’%16.3f’);
option4.cnames = cnames; lprint(table4,option4); % variable column formats
fprintf(1,’demo of printing selected rows and columns \n’);
clear in; table = randn(5,10);
in.begc = 5; in.endc = 9; % specify selected columns to print
in.begr = 2; in.endr = 5; % specify selected rows to print
cnames = strvcat(’col1’,’col2’,’col3’,’col4’);
cnames = strvcat(cnames,’col5’,’col6’,’col7’,’col8’,’col9’,’col10’);
rnames = strvcat(’Rows’,’row1’,’row2’,’row3’,’row4’,’row5’);
% NOTE we need column and row names for all rows and columns
% not just those selected for printing
in.cnames = cnames; in.rnames = rnames;
lprint(table,in); % print selected rows and columns
The output from this example is shown below. Matrices that have a
large number of columns are automatically wrapped when printed, with
wrapping determined by both the numeric format and column names. To
understand how wrapping works, note that the lprint function examines
the width of column names as well as the width of the numeric format
.::UdecomBooks::.
supplied. The numeric format is padded at the left to match column names
that are wider than the format. If column names are not as wide as the
numeric format, these are right-justified. For example, if you supply a 15
character column name and a ‘10.6f’ numeric format, the columns will be
15 characters wide with the numbers printed in the last 10 character places
in the ‘10.6f’ format. On the other hand, if you supply 8 character column
names and ‘10.6f’ numeric format, the columns will be 10 characters wide
with the column names right-justified to fill the right-most 8 characters.
Wrapping occurs after a number of columns determined by 80 divided
by the numeric format width. Depending on the numeric format, you can
achieve a different number of columns for your table. For example, using a
format ‘10f.3’ would allow 8 columns in your table, whereas using ‘5d’ would
allow up to 16 columns. The format of ‘20.4f’ in the example, produces
wrapping after 4 columns. Note that the width of the printout may be
wider than 80 columns because of the column name padding considerations
discussed in the previous paragraph.
using no options
-0.2212 & 0.3829 & -0.6628 \\
-1.5460 & -0.7566 & -1.0419 \\
-0.7883 & -1.2447 & -0.6663 \\
-0.6978 & -0.9010 & 0.2130 \\
0.0539 & -1.1149 & -1.2154 \\
using fmt option
-0 & 0 & -1000 \\
-2000 & -1000 & -1000 \\
-1000 & -1000 & -1000 \\
-1000 & -1000 & 0 \\
0 & -1000 & -1000 \\
using column names and fmt option
Illinois & Ohio & Indiana \\
-0.000 & 0.000 & -1000.000 \\
-2000.000 & -1000.000 & -1000.000 \\
-1000.000 & -1000.000 & -1000.000 \\
-1000.000 & -1000.000 & 0.000 \\
0.000 & -1000.000 & -1000.000 \\
row and column labels
Rows Illinois & Ohio & Indiana \\
row1 -0.000 & 0.000 & -1000.000 \\
row2 -2000.000 & -1000.000 & -1000.000 \\
row3 -1000.000 & -1000.000 & -1000.000 \\
row4 -1000.000 & -1000.000 & 0.000 \\
row5 0.000 & -1000.000 & -1000.000 \\
wrapped output for large matrices
IL & OH & IN & WV &
-0.6453 & 0.6141 & 0.1035 & 1.5466 &
1.0657 & -0.6160 & 0.4756 & 0.6984 &
-0.1516 & 1.0661 & 1.2727 & 0.8227 &
0.8837 & -0.8217 & 1.6452 & 1.1852 &
0.8678 & -1.0294 & 0.8200 & 0.4893 &
.::UdecomBooks::.
PA & AK & HI & NY &

0.7344 & 0.0380 & 0.2682 & -0.0662 &
-0.7534 & 0.3395 & -0.8468 & 0.8473 &
-1.2320 & 0.1657 & -0.4726 & 0.4700 &
1.1433 & -1.7245 & -1.4360 & -1.4543 &
1.0019 & -0.0296 & 0.0132 & 1.1019 &
MS & TN \\
-0.3047 & 1.8723 \\
1.5178 & -0.6233 \\
0.6368 & -0.3576 \\
0.3657 & -0.0250 \\
0.9989 & -0.1637 \\
generic row labels
Obs# column1 & column2 & column3 & column4 \\
1 2.9275 & 71.3611 & 133.5464 & -53.9951 \\
2 9.2330 & -76.0785 & 139.0374 & -55.8497 \\
3 -42.6711 & 94.7476 & -107.4995 & -78.1427 \\
4 -61.9797 & -15.0058 & 40.9104 & 52.3172 \\
variable column formats
column1 & column2 & column3 & column4 \\
3 & 71.361 & 134 & -53.995 \\
9 & -76.079 & 139 & -55.850 \\
-43 & 94.748 & -107 & -78.143 \\
-62 & -15.006 & 41 & 52.317 \\
demo of printing selected rows and columns
Rows col5 & col6 & col7 & col8 & col9 \\
row2 -1.0790 & -0.8061 & 0.0273 & -0.5726 & 0.5509 \\
row3 2.0281 & 1.2631 & 2.3123 & -0.4425 & -1.6558 \\
row4 0.0198 & -0.0496 & -0.2238 & -0.2303 & 1.5463 \\
row5 0.7006 & -0.0253 & -1.1709 & 1.2747 & 0.4729 \\
There is also a function lprintf contributed by Mike Cliff that prints a

matrix of data with a criteria-based symbol next to each element. Data are
formatted globally or column-wise, as specified. This function was designed
to all indicating significance levels of parameter estimates in LaTeX tables.
For printing general matrices to the MATLAB command window or a
file, we provide the function mprint. This function works identically to
lprint, but does not provide the LaTeX formatting symbols. Both lprint
and mprint permit only integer and decimal formats such as ‘%6d’ and
‘%10.4f’. The documentation is:
PURPOSE: print an (nobs x nvar) matrix in formatted form

---------------------------------------------------
USAGE: mprint(x,info)
where:
x = (nobs x nvar) matrix (or vector) to be printed
info = a structure containing printing options
info.begr = beginning row to print, (default = 1)
info.endr = ending row to print, (default = nobs)
info.begc = beginning column to print, (default = 1)
info.endc = ending column to print, (default = nvar)
.::UdecomBooks::.
info.cnames = an (nvar x 1) string of names for columns (optional)

e.g. info.cnames = strvcat(’col1’,’col2’);
(default = no column headings)
info.rnames = an (nobs+1 x 1) string of names for rows (optional)
e.g. info.rnames = strvcat(’Rows’,’row1’,’row2’);
(default = no row labels)
info.fmt = a format string, e.g., ’%12.6f’ or ’%12d’ (default = %10.4f)
or an (nvar x 1) string containing formats
e.g., info.fmt=strvcat(’%12.6f’,’%12.2f’,’%12d’);
info.fid = file-id for printing results to a file
e.g. fid = fopen(’file.out’,’w’);
info.rflag = 1 row #’s printed, 0 no row #’s (default = 0)
info.width = # of columns before wrapping occurs (default = 80)
---------------------------------------------------
e.g. in.cnames = strvcat(’col1’,’col2’);
in.rnames = strvcat(’rowlabel’,’row1’,’row2’);
mprint(y,in), prints entire matrix, column and row headings
in2.endc = 3; in2.cnames = strvcat(’col1’,’col2’,’col3’);
or: mprint(y,in2), prints 3 columns of the matrix, with headings
or: mprint(y), prints entire matrix, no headings
NOTES: - defaults are used for info-elements not specified
- wrapping occurs at 80 columns, which varies depending on
format used, e.g. %10.2f will wrap after 8 columns
---------------------------------------------------
SEE ALSO: tsprint, mprint_d, lprint
---------------------------------------------------
The use of a structure variable to input arguments to functions is a useful

MATLAB programming construct that we employ in many Econometric
Toolbox functions. To see how this is accomplished consider the following
code from mprint that parses the input structure fields for the structure
variable named ‘info’ in the function declaration.
function mprint(y,info)
% note structure variable named info in argument declaration
[nobs nvars] = size(y);
fid = 1; rflag = 0; cflag = 0; rnum = 0; nfmts = 1; % setup defaults
begr = 1; endr = nobs; begc = 1; endc = nvars; fmt = ’%10.4f’; cwidth = 80;
if nargin == 1
% rely on defaults
elseif nargin == 2
if ~isstruct(info)
error(’mprint: you must supply the options as a structure variable’);
end;
fields = fieldnames(info);
nf = length(fields);
for i=1:nf
.::UdecomBooks::.
if strcmp(fields{i},’fmt’)
fmts = info.fmt; [nfmts junk] = size(fmts);
if nfmts <= nvars, fmt = fmts;
else
error(’mprint: wrong # of formats in string -- need nvar’);
end;
elseif strcmp(fields{i},’fid’), fid = info.fid;
elseif strcmp(fields{i},’begc’), begc = info.begc;
elseif strcmp(fields{i},’begr’), begr = info.begr;
elseif strcmp(fields{i},’endc’), endc = info.endc;
elseif strcmp(fields{i},’endr’), endr = info.endr;
elseif strcmp(fields{i},’cnames’), cnames = info.cnames; cflag = 1;
elseif strcmp(fields{i},’rnames’), rnames = info.rnames; rflag = 1;
elseif strcmp(fields{i},’rflag’), rnum = info.rflag;
elseif strcmp(fields{i},’width’); cwidth = info.width;
end;
end;
else
error(’Wrong # of arguments to mprint’);
end; % end of if-elseif input checking
After setting up default values, we use the MATLAB function field-

names to extract the fields from the structure variable into a cell-array
named ‘fields’. These are parsed in a ‘for-loop’ over the length of the ‘fields’
cell-array using the MATLAB function strcmp. Input fields specified by
the user will be detected by the string comparisons and extracted to over-
write the default argument values set prior to parsing the input structure
variable.
As with the case of lprint, Mike Cliff has provided an enchanced version
of mprint in the form of a function named mprint3 which prints a set
of matrices together by stacking the (i,j) elements of each. This allows for
separate formatting of each matrix so one can print parameter estimates,
standard errors and t-statistics.
To facilitate printing time-series data we have a function tsprint that
will print time-series with calendar dates as labels. The function documen-
tation is:
PURPOSE: print time-series data with dates and column labels

---------------------------------------------------
USAGE: tsprint(y,cstruc,begp,endp,vnames,fmt)
or: tsprint(y,cstruct,vnames), prints entire series with names
or: tsprint(y,cstruct), entire series, no variable names
or: tsprint(y,cstruct,fmt) entire series, using fmt
or: tsprint(y,cstruct,vnames,fmt) entire series w/names and fmt
or: tsprint(y,cstruct,begp,endp) partial series no names or fmt
.::UdecomBooks::.
where: y = matrix (or vector) of series to be printed

cstruc = a structure returned by cal()
begp = the beginning observation to print
endp = the ending period to print,
vnames = a string matrix of names for a header (optional)
fmt = a format string, e.g., ’%12.6f’ or ’%12d’
---------------------------------------------------
NOTES: cstr = cal(1980,1,12);
tsprint(y,cstr,13,24), would print data for 1981
or: tsprint(y,cstr,ical(1981,1,cstr),ical(1981,12,cstr)),
which would print the same data for 1981
---------------------------------------------------
This function takes a dates structure, begin and end period and format
as arguments. As indicated in the function documentation, the user can
call this function using a number of different input arguments in almost any
order. Example 3.6 demonstrates some of the alternatives.
% ----- Example 3.6 Using the tsprint() function

names =strvcat(’Illinois’,’Ohio’,’Indiana’,’West Virginia’,’Pennsylvania’);
d1 = randn(120,5); dates = cal(1980,1,12); fmt = ’%14.3f’;
begp = ical(1985,1,dates); endp = ical(1985,12,dates);
fprintf(1,’all options \n’); tsprint(d1,dates,begp,endp,names,fmt);
fprintf(1,’default format \n’); tsprint(d1,dates,begp,endp,names);
fprintf(1,’integer format and printing the whole time-series \n’);
fmt = ’%13d’; d2 = round(data)*100; tsprint(d2,dates,names,fmt);
fprintf(1,’d format with floating point #s \n’); tt=1:90;
d3 = [tt’ d1(:,2:5)]; fmt = ’%13d’; tsprint(d3,dates,begp,endp,fmt);
fprintf(1,’format option, monthly dates and wrapping \n’);
d4 = randn(12,10); fmt = ’%16.8f’; tsprint(d4,dates,fmt);
fprintf(1,’format option, quarterly dates and wrapping \n’);
dates2 = cal(1980,1,4); tsprint(d4,dates2,fmt);
fprintf(1,’default format, annual dates and wrapping \n’);
dates3 = cal(1980,1,1); tsprint(d4,dates3);
This flexibility to allow user input of any number of the options in almost
any order is achieved using the MATLAB varargin variable. This is a
generally useful approach to crafting functions, so we provide the details of
how this is accomplished. The function declaration contains the MATLAB
keyword ‘varargin’ which is a cell-array that we parse inside the function. It
is our responsibility to correctly deduce the number and order of the input
arguments supplied by the user.
function tsprint(y,cstruc,varargin)
% NOTE the use of varargin keyword in the function declaration
[nobs nvar] = size(y); fmt = ’%10.4f’; % set defaults
.::UdecomBooks::.
begp = 1; endp = nobs; nflag = 0;

nargs = length(varargin); % find the # of input arguments
if nargs == 0 % no user-supplied vnames or dates or fmt
% rely on defaults
elseif nargs == 1 % no dates but vnames or fmt
[testf testg] = size(varargin{1});
if testf == 1 % we have a format
fmt = varargin{1}; % replace default fmt with user fmt
else % we have vnames
nflag = 1; % set flag for vnames
vnames = varargin{1}; % pull-out user variable names
end;
elseif nargs == 2; % either begp,endp or names and fmt
if testf == 1 % we have a format or begp
if isnumeric(varargin{1}) % we have begp, endp
begp = varargin{1}; % pull-out begp
endp = varargin{2}; % pull-out endp
end;
else % we have vnames, fmt
vnames = varargin{1}; % pull-out vnames
fmt = varargin{2}; % pull-out format
nflag = 1; % set flag for vnames
end;
elseif nargs == 3 % begp,endp with either vnames or fmt
if testf == 1 % we have a format
begp = varargin{1}; endp = varargin{2}; fmt = varargin{3};
else % we have vnames
nflag = 1; begp = varargin{1}; endp = varargin{2}; vnames = varargin{3};
end;
elseif nargs == 4 % we have everything
nflag = 1; begp = varargin{1}; endp = varargin{2};
vnames = varargin{3}; fmt = varargin{4};
end; % end of input checking
Our deductions are based on logic and the use of MATLAB functions
isnumeric to detect ‘begp’ input arguments and the size command to dis-
tinguish formats from variable names supplied as input arguments. Needless
to say, there are limits to what can be accomplished with this approach de-
pending on the nature of the input arguments. In addition, writing this type
of function is more prone to errors than the use of the structure variable as
an input argument demonstrated by the mprint and lprint functions. The
advantage is that the user need not define an input structure, but can pass
input arguments directly to the function. Passing input arguments directly
may be more convenient when typing commands to the MATLAB COM-
MAND WINDOW, which is likely to be the case when using the tsprint
.::UdecomBooks::.
function. After implementing time-series data transformations on a matrix,

the user may wish to verify the transformations using tsprint typed in the
COMMAND WINDOW.
It might be of interest that tsprint calls the mprint function to do the
printing. This is done after parsing the input and forming a set of row-names
using our tsdate function as follows:
rnames = ’Date’;
for k=begp:endp;
rnames = strvcat(rnames,tsdate(cstruc,k));
end;
in.rnames = rnames;
in.fmt = fmt;
in.cnames = vnames;
mprint(y(begp:endp,:),in);
Because tsprint relies on mprint, the wrapping of time-series matrices

containing a large number of columns takes the same format as in mprint
and lprint. The number of columns printed will vary depending on the
numeric format used. Example 3.7 illustrates using various numeric formats
and the wrapping results.
% ----- Example 3.7 Various tsprint() formats

vnames = strvcat(’illinos’,’indiana’,’kentucky’,’michigan’,’ohio’, ...
’pennsylvania’,’tennessee’,’west virginia’);
dates = cal(1982,1,12); load test.dat; y = test;
fmt = ’%16.0f’; tsprint(y,dates,begp,endp,vnames,fmt);
vnames = strvcat(’IL’,’IN’,’KY’,’MI’,’OH’,’PA’,’TN’,’WV’);
fmt = ’%6.2f’; tsprint(y,dates,begp,endp,vnames,fmt);
fmt = ’%20d’; tsprint(y,dates,begp,endp,vnames,fmt);
The output varies due to the different formats and variable names sup-
plied as shown below.
Date illinos indiana kentucky michigan ohio
Jan90 192 76 348 97 171
Feb90 190 76 350 96 170
Mar90 192 78 356 95 172
Apr90 192 80 357 97 174
May90 194 82 361 104 177
Jun90 199 83 361 106 179
Jul90 200 84 354 105 181
Aug90 200 84 357 84 181
Sep90 199 83 360 83 179
Oct90 199 83 353 82 177
Nov90 197 82 350 81 176
Dec90 196 82 353 96 171
.::UdecomBooks::.
Date pennsylvania tennessee west virginia

Jan90 267 61 352
Feb90 266 57 349
Mar90 271 62 350
Apr90 273 62 351
May90 277 63 355
Jun90 279 63 361
Jul90 279 63 359
Aug90 281 63 358
Sep90 280 62 358
Oct90 277 62 357
Nov90 268 61 361
Dec90 264 59 360
Date IL IN KY MI OH PA TN WV
Jan90 192.00 76.00 348.00 97.00 171.00 267.00 61.00 352.00
Feb90 190.00 76.00 350.00 96.00 170.00 266.00 57.00 349.00
Mar90 192.00 78.00 356.00 95.00 172.00 271.00 62.00 350.00
Apr90 192.00 80.00 357.00 97.00 174.00 273.00 62.00 351.00
May90 194.00 82.00 361.00 104.00 177.00 277.00 63.00 355.00
Jun90 199.00 83.00 361.00 106.00 179.00 279.00 63.00 361.00
Jul90 200.00 84.00 354.00 105.00 181.00 279.00 63.00 359.00
Aug90 200.00 84.00 357.00 84.00 181.00 281.00 63.00 358.00
Sep90 199.00 83.00 360.00 83.00 179.00 280.00 62.00 358.00
Oct90 199.00 83.00 353.00 82.00 177.00 277.00 62.00 357.00
Nov90 197.00 82.00 350.00 81.00 176.00 268.00 61.00 361.00
Dec90 196.00 82.00 353.00 96.00 171.00 264.00 59.00 360.00
Date IL IN KY MI
Jan90 192 76 348 97
Feb90 190 76 350 96
Mar90 192 78 356 95
Apr90 192 80 357 97
May90 194 82 361 104
Jun90 199 83 361 106
Jul90 200 84 354 105
Aug90 200 84 357 84
Sep90 199 83 360 83
Oct90 199 83 353 82
Nov90 197 82 350 81
Dec90 196 82 353 96
Date OH PA TN WV
Jan90 171 267 61 352
Feb90 170 266 57 349
Mar90 172 271 62 350
Apr90 174 273 62 351
May90 177 277 63 355
Jun90 179 279 63 361
Jul90 181 279 63 359
Aug90 181 281 63 358
Sep90 179 280 62 358
Oct90 177 277 62 357
Nov90 176 268 61 361
Dec90 171 264 59 360
.::UdecomBooks::.
A warning about using the cal and tsprint functions. When you trun-
cate time-series to account for transformations, you should re-set the calen-
dar with another call to the cal function based on the new dates associated
with the truncated series. Example 3.8 provides an illustration of this.
% ----- Example 3.8 Truncating time-series and the cal() function
dates = cal(1982,1,12); load test.dat; y = growthr(test,12);
% define beginning and ending print dates
tsprint(y,dates,begp,endp,vnames);
ynew = trimr(y,dates.freq,0); % truncate initial observations
tdates = cal(1983,1,12); % reset the calendar after truncation
% re-define beginning and ending print dates based on the new calendar
begp = ical(1983,1,tdates); endp = ical(1983,12,tdates);
tsprint(ynew,tdates,begp,endp,vnames);
In the code above, growthr (a function discussed in the next section)

transforms data to annual growth rates setting observations for the initial
year to zero. The function trimr (also discussed in the next section) trims
off the rows associated with the initial year, so our data matrix ‘ynew’ starts
in 1983, not 1982 like ‘y’.
If you don’t take responsibility for doing this, the date labels on your
printouts or plots may be incorrect. Some effort has been made to place
error-checking in the function ical so that you cannot make a call with
a ‘year’ argument that is less than the ‘beg yr’ contained in the structure
returned by the cal function. This should protect against the type of mistake
shown in example 3.9.
% ----- Example 3.9 Common errors using the cal() function
dates = cal(1982,1,12); % define calendar to start in 1982,1
load test.dat; y = growthr(test,12); y = trimr(y,12,0);
dates = cal(1983,1,12); % re-define calendar to start in 1983,1
% try to define beginning and ending print dates
% forgetting that the calendar now starts in 1983,1
begp = ical(1982,1,dates);
% ical will produce an error message in the above call
One nice feature of using the cal and ical functions is that dates informa-
tion is documented in the file. It should be fairly clear what the estimation,
forecasting, truncation, etc. dates are — just by examining the file.
In addition to printing time-series data we might wish to plot time-series
variables. A function tsplot was created for this purpose. The usage format
is similar to tsprint and relies on a structure from cal. The documentation
for this function is:
.::UdecomBooks::.
PURPOSE: time-series plot with dates and labels

---------------------------------------------------
USAGE: tsplot(y,cstruc,begp,endp,vnames)
or: tsplot(y,cal_struc,vnames), which plots the entire series
or: tsplot(y,cal_struc), entire series no variable names
where: y = matrix (or vector) of series to be plotted

cstruc = a structure returned by cal()
begp = the beginning observation to plot
endp = the ending observation to plot
vnames = a string matrix of names for a legend (optional)
---------------------------------------------------
NOTES: cstr = cal(1970,1,12);
tsplot(y,cstr); would plot all data
or: tsplot(y,cstr,ical(1981,1,cstr),ical(1981,12,cstr)),
which would plot data for 1981 only
---------------------------------------------------
tsplot produces graphs that look like that shown in Figure 3.1. (Colors
are used to distinguish the lines in MATLAB). The time-axis labeling is not
ideal, but this is due to limitations in MATLAB. The function attempts to
distinguish between graphs that involve fewer observations and those that
involve a large number of observations. A vertical grid is used for the case
where we have a small number of observations.
If you invoke multiple calls to tsplot, you are responsible for placing
MATLAB pause statements as in the case of the plt function from the
regression library. You can add titles as well as use the MATLAB subplot
feature with this function. Example 3.10 produces time-series plots using
the subplot command to generate multiple plots in a single figure.
% ----- Example 3.10 Using the tsplot() function

dates = cal(1982,1,12); load test.dat; y = growthr(test,12);
yt = trimr(y,dates.freq,0);
dates = cal(1983,1,12); % redefine the calendar
subplot(211), % a plot of the whole time-series sample
tsplot(yt,dates,vnames);
title(’Entire data sample’); % add a title to the first plot
subplot(212), % produce a plot showing 1990,1 to 1992,12
% find obs # associated with 1990,1 and 1992,12
tsplot(yt,dates,begp,endp,vnames);
xlabel(’a sub-sample covering 1990-1992’); % add a label in bottom plot
Note, we never really used the ‘dates’ structure returned by the first call
to cal. Nonetheless, this command in the file provides documentation that
.::UdecomBooks::.
Entire data sample

20
IL
10 IN
-10
-20
Jul82 Jan85 Jul87 Jan90 Jul92 Jan95 Jul97
20
10 IL
IN
0
-10
-20
Jan90 Jan91 Jan92 Jan93
a sub-sample covering 1990-1992
20
IL
10 IN
-10
-20
Jan83 Jan84 Jan85 Jan86 Jan87 Jan88
quarterly averages produced with mth2qtr()
Figure 3.1: Output from tsplot() function
the file ‘test.dat’ contains monthly data beginning in January, 1982, so it

does serve a purpose.
In addition to tsplot, there is a function fturns that finds turning points
in time-series and plt that uses the ‘result’ structure variable from fturns
to produce a graphical presentation of the turning points.
Zellner, Hong and Gulati (1988), and Zellner and Hong (1988) formu-
lated the problem of forecasting turning points in economic time series using
a Bayesian decision theoretic framework. The innovative aspect of the Zell-
ner and Hong study was the use of time-series observations along with an
explicit definition of a turning point, either a downturn (DT) or upturn
(UT). This allows for a Bayesian computation of probabilities of a DT or
UT given the past data and a model’s predictive probability density function
(pdf) for future observations.
The function fturns allows one to define a turning point event based
.::UdecomBooks::.
on a sequence of time-series observations, an idea originating with Wecker

(1979). For example, let the conditions for a downturn event be:
yt−k < . . . < yt−2 < yt−1 < yt (3.1)

Which indicates that when we observe a strictly increasing sequence of ob-
servations in our time-series, the conditions are right for a downturn event.
The relationship between yt and future values, yt+j is used to determine if
a downturn event actually took place. As an example, we might define a
downturn event based on the following future sequence:
yt+j < . . . < yt+1 < yt (3.2)

This indicates that we define a downturn event at time t as a sequence of
future observations falling strictly below the current observation at time
t. Our definition is conditional on the initial rising sequence up to time t,
followed by falling values after time t. If future values do not fall below
those at time t, we define a “no downturn event”.
A similar approach can be taken to define an upturn event, for example
the necessary conditions for an upturn event might be a sequence of strictly
falling time-series observations:
yt−k > . . . > yt−2 > yt−1 > yt (3.3)

followed by a sequence of strictly rising observations after time t.
yt+j > . . . > yt+2 > yt+1 > yt (3.4)

As in the case of a downturn event, if the future observations do not rise
after time t, we specify a “no upturn event”.
The function fturns allows the user to define the number of observations
to examine prior to and after an upturn and downturn event as well as
whether a strict sequence of rising or falling values is to be used. A non-
sequential definition of the conditions necessary for a turning point event
would be defined as:
yt−k , . . . , yt−2 , yt−1 > yt (3.5)

The documentation for fturns is:
PURPOSE: finds turning points in a time-series

---------------------------------------------------
USAGE: result = fturns(y,in)
.::UdecomBooks::.
where: y = an (nobs x 1) time-series vector

in = a structure variable with options
in.but, # of down-periods before an upturn (default = 4)
in.aut, # of up-periods after an upturn (default = 1)
in.bdt, # of up-periods before a downturn (default = 4)
in.adt, # of down-periods after a downturn (default = 1)
in.seq, 1 = sequence inequality, default = 1
0 = simple inequality
in.eq 1 = inequality with <=, >=, default = 1
0 = strict inequality <, >
e.g. in.seq=0, in.eq=1 in.but = 3, in.aut = 1 would define:
y(t-3), y(t-2), y(t-1), >= y(t) [but=3, eq=1, seq=0]
and y(t+1) > y(t) as UT [aut=1]
and y(t+1) <= y(t) as NUT
e.g. in.seq=0, in.eq=1 in.bdt = 3, in.adt = 2 would define:
y(t-3), y(t-2), y(t-1), <= y(t) [bdt=3, eq=1, seq=0]
and y(t+2), y(t+1) < y(t) as DT [adt=2]
and y(t+2), y(t+1) >= y(t) as NDT
e.g. in.seq=1, in.eq=1, in.but = 3, in.aut = 1 would define:
y(t-3) >= y(t-2) >= y(t-1) >= y(t) [but=3, eq=1, seq=1]
and y(t+1) > y(t) as UT [aut=1]
and y(t+1) <= y(t) as NUT
e.g. in.seq=1, in.eq=0, in.bdt = 3, in.adt = 2 would define:
y(t-3) < y(t-2) < y(t-1) < y(t) [bdt=3, eq=0, seq=1]
and y(t+2) > y(t+1) < y(t) as DT [adt=2]
and y(t+2) >= y(t+1) >= y(t) as NDT
---------------------------------------------------
RETURNS:
results = a structure variable
results.ut = (nobs x 1) vector with 1 = UT periods
results.dt = (nobs x 1) vector with 1 = DT periods
results.nut = (nobs x 1) vector with 1 = NUT periods
results.ndt = (nobs x 1) vector with 1 = NDT periods
results.y = time-series vector input
(NUT = no upturn, NDT = no downturn)
--------------------------------------------------
SEE ALSO: plt_turns (which will plot turning points)
--------------------------------------------------
An example showing various settings for the input structure variable

best illustrates the use of the alternative turning point definitions that can
be set using the function.
% ----- Example 3.11 Finding time-series turning points with fturns()

dates = cal(1982,1,12); load test.dat; y = growthr(test,dates);
yt = trimr(y,dates.freq,0); % truncate initial zeros
tdates = cal(1983,1,12); % update calendar for truncation
ytime = yt(:,1); % pull out state 1 time-series
.::UdecomBooks::.
% DT definition: y(t-4), y(t-3), y(t-2), y(t-1) <= y(t)

% If y(t+1) < y(t) we have a downturn
in.bdt = 4; in.adt = 1; in.eq = 1; in.seq = 0;
% UT definition: y(t-4), y(t-3), y(t-2), y(t-1) >= y(t)
% If y(t+1) > y(t) we have an upturn
in.but = 4; in.aut = 1;
results = fturns(ytime,in); plt(results,tdates,’employment’);
title(’loose non-sequential definition --- finds lots of turns’); pause;
in.seq = 1; % Now, change to a sequential definition
% DT definition: y(t-4) <= y(t-3) <= y(t-2) <= y(t-1) <= y(t)
% If y(t+1) < y(t) we have a downturn
% UT definition: y(t-4) >= y(t-3) >= y(t-2) >= y(t-1) >= y(t)
% If y(t+1) > y(t) we have an upturn
title(’sequential definition --- produces fewer turns’); pause;
% Now, illustrate requiring many points after the turns
% (should rule out some of the turning points in the above graph)
in.seq = 1; in.aut = 4; % 4 periods of up after an upturn
in.adt = 4; % 4 periods of down after a downturn
% DT definition: y(t-4), y(t-3), y(t-2), y(t-1) <= y(t)
% If y(t+4) < y(t+3) < y(t+2) < y(t+1) < y(t) we have a downturn
% UT definition: y(t-4), y(t-3), y(t-2), y(t-1) >= y(t)
% If y(t+4) > y(t+3) > y(t+2) > y(t+1) > y(t) we have an upturn
title(’sequential aut=4, adt=4 definition --- finds less turns’); pause;
% Now turn off the sequential requirement
% ( should produce more turns that meet this looser definition)
in.seq = 0;
title(’non-sequential definition aut=4, adt=4 --- finds more turns’); pause;
An example of the graphs produced by submitting the ‘results’ structure

from fturns to plt turns or plt is shown in Figure 3.2.
3.3 Data transformation utilities

Typical data transformations for time-series analysis involve constructing
lagged values of variables, differences, growth-rate transformations and sea-
sonal differencing. A problem encountered with these transformations is
that they produce result vectors with missing observations for the initial
periods. For example, when constructing a lagged variable, yt−1 , we face an
initialization problem where the lagged value for the first observation in our
sample data represents a period prior to the beginning of the sample.
A design decision must be made regarding how to handle this situation.
There are two options, one is to return a lagged variable vector or matrix
.::UdecomBooks::.
tight sequential definition of turns --- produces fewer turns

15
employment
upturn
10
downturn
-5
-10
-15
-20
Mar82 Dec84 Sep87 Jun90 Mar93 Dec95 Sep98
Figure 3.2: Graph of turning point events
that is truncated to “feed the lags”. This approach would return different
length vectors or matrices depending on the type of transformation and
frequency of the data. For example, in the case of seasonal differencing, we
loose the first ‘freq’ observations where freq=4 or 12 depending on whether
we are working with quarterly or monthly data. When we consider that
after implementing a seasonal differencing transformation, we might then
construct lagged values of the seasonally differenced series, it becomes clear
that this approach has the potential to raise havoc.
The second design option is to always return the same size vector or
matrix that is used as an input argument to the function. This requires that
the user take responsibility for truncating the vector or matrix returned by
the function. We supply an auxiliary function, trimr than can help with
the truncation task facing the user.
An example to illustrate using a seasonal differencing function sdiff and
.::UdecomBooks::.
trimr to accomplish truncation is:

% ----- Example 3.12 Seasonal differencing with sdiff() function
dates = cal(1982,1,12); load test.dat;
ysdiff = sdiff(test,dates); % seasonally difference the data
ysdiff = trimr(ysdiff,dates.freq,0); % truncate first freq observations
The function trimr is designed after a similar function from the Gauss
programming language, its documentation is:
PURPOSE: return a matrix (or vector) x stripped of the specified rows.
-----------------------------------------------------
USAGE: z = trimr(x,n1,n2)
where: x = input matrix (or vector) (n x k)
n1 = first n1 rows to strip
n2 = last n2 rows to strip
NOTE: modeled after Gauss trimr function
-----------------------------------------------------
RETURNS: z = x(n1+1:n-n2,:)
-----------------------------------------------------
Note that we utilize our cal function structure in the sdiff function, but
an integer ‘freq’ argument works as well. That is: ysdiff = sdiff(test,12);
would return the same matrix of seasonally differenced data as in the exam-
ple above.
In addition to the sdiff function, we create an ordinary differencing func-
tion tdiff. (Note, we cannot use the name diff because there is a MATLAB
function named diff). This function will produce traditional time-series
differences and takes the form:
ydiff = tdiff(y,k);
which would return the matrix or vector y differenced k times. For example:
k = 1 produces: yt − yt−1 , and k = 2 returns: ∆2 yt = yt − 2yt−1 + yt−2 .
Another time-series data transformation function is the lag function that
produces lagged vectors or matrices. The format is:
xlag1 = lag(x); % defaults to 1-lag
xlag12 = lag(x,12); % produces 12-period lag
where here again, we would have to use the trimr function to eliminate the
zeros in the initial positions of the matrices or vectors xlag1, and xlag12
returned by the lag function.
For VAR models, we need to create matrices containing a group of con-
tiguous lags that are used as explanatory variables in these models. A special
.::UdecomBooks::.
lag function mlag is used for this purpose. Given a matrix Yt containing
two time-series variable vectors:
Yt = (y1t , y2t ) (3.6)

the mlag function call: ylag = mlag(y,k); will return a matrix ‘ylag’ equal
to:
ylag = (y1t−1 , y1t−2 , . . . , y1t−k , y2t−1 , y2t−2 , . . . , y2t−k ) (3.7)

which takes the form of a typical set of explanatory variables for these mod-
els.
Another data transformation we might wish to employ in time-series
analysis is one that converts the time-series to the form of annual growth-
rates, i.e., for quarterly time-series, [(yt − yt−4 )/yt−4 ] ∗ 100. The function
growthr carries out this task. Again, we design this function to use either
the ‘freq’ argument or the structure returned by the cal function. Examples
of using this function would be:
% ----- Example 3.13 Annual growth rates using growthr() function

% Using the cal() structure
dates = cal(1982,1,12); % initialize a calendar
load test.dat; % load some data
y = growthr(test,dates); % transform to growth rates
y = trimr(y,dates.freq,0); % truncate the first freq observations
% Using ‘freq=12’ as an input argument

y = growthr(test,12); % transform to growth rates
y = trimr(y,12,0); % truncate the first freq observations
Seasonal dummy variables can be created using the sdummy function.

It returns a matrix of seasonal dummy variables based on the frequency of
the time-series. The function call is:
% ----- Example 3.14 Seasonal dummy variables using sdummy() function

% Using the cal() structure
dates = cal(1982,1,12); % initialize a calendar
nobs = ical(1995,12,dates);
dummies = sdummy(nobs,dates);
% Using an input ‘freq’
freq = 12; dummies = sdummy(nobs,freq); in.fmt = ’%6.0f’;
mprint(dummies,in);
Finally, a function mth2qtr converts monthly time-series to quarterly

averages or sums. The documentation for the function is:
.::UdecomBooks::.
PURPOSE: converts monthly time-series to quarterly averages

---------------------------------------------------
USAGE: yqtr = mth2qtr(ymth,flag)
where: ymth = monthly time series vector or matrix (nobs x k)
flag = 0 for averages (default) and 1 for sums
---------------------------------------------------
RETURNS: yqtr = quarterly time-series vector or matrix
[floor(nobs/3) + 1] in length by k columns
---------------------------------------------------
NOTES: the last observation is the actual month or
the average (sum) of the last 2 months in cases where
nobs/3 has a remainder
---------------------------------------------------
3.4 Gauss functions

There are a great number of Gauss routines in the public domain available
on the Internet. To make it easier to convert these functions to MATLAB
syntax, we provide a set of functions that operate in an identical fashion to
functions by the same name in Gauss.
We have already seen the function trimr that behaves in a fashion iden-
tical to the function provided in the Gauss programming language. Other
functions that work in an identical fashion to Gauss functions of the same
name are:
rows - returns the # of rows in a matrix or vector
cols - returns the # of columns in a matrix or vector
trimr - trims rows of a matrix
trimc - trims columns of a matrix
invpd - inverse of a positive-definite matrix
cumprodc - returns cumulative product of each column of a matrix
cumsumc - returns cumulative sum of each column of a matrix
prodc - returns product of each column of a matrix
sumc - returns sum of each column
delif - select matrix values for which a condition is false
selif - select matrix values for which a condition is true
.::UdecomBooks::.
indexcat - extract indices equal to a scalar or an interval
recserar - computes recursive autoregressive series
recsercp - computes recursive series involving products
seqa - produces a sequence of numbers with a beginning and increment
stdc - standard deviations of the columns of a matrix returned as a

column vector
matdiv - divides matrices that are not of the same dimension but are
row or column compatible. (NOTE: there is no Gauss function like
this, but Gauss allows this type of operation on matrices.)
matmul - multiplies matrices that are not of the same dimension but
are row or column compatible.
matadd - adds matrices that are not of the same dimension but are
row or column compatible.
matsub - divides matrices that are not of the same dimension but are
row or column compatible.
To illustrate the use of these functions we provide an example of convert-

ing a Gauss function for least-absolute deviations regression to a MATLAB
function. The Gauss version of the function is shown below, with line #’s
added to facilitate our discussion.
/* Least Absolute Deviation

Author: Ron Schoenberg rons@u.washington.edu
Date: May 29, 1995
Provided without guarantee for public non-commercial use.
Gauss code for least absolute deviation
estimation: given y_i and x_i, i = 1,...,n,
estimate b such that \sum |y_i - b*x_i| is minimized.
The following code solves this as a re-iterated weighted least squares
problem where the weights are the inverse of the absolute values of the
residuals.*/
1 rndseed 45456;
2 nobs = 100;
3 numx = 3;
4 b = .5 * ones(numx+1,1);
5 x = ones(nobs,1)~rndn(nobs,numx);
6 y = x * b + rndn(nobs,1)^4; /* 4th power for kurtotic error term */
7 print "LS estimates " (y / x)’;
.::UdecomBooks::.
8 print;
9 b_old = 0;
10 b_new = b;
11 w = x;
12 do until abs(b_new - b_old) < 1e-3;
13 b_old = b_new;
14 b_new = invpd(w’x)*(w’y);
15 w = x./abs(y - x * b_new);
16 endo;
17 print "LAD estimates " b_new’;
Some things to note about the Gauss code are:
1. use of matrix divisions that are non-conformable are allowed in Gauss

if the two arguments are conformable in either the column or row
dimensions. This is not the case in MATLAB, so we have a function
matdiv that carries out these operations. This type of operation
occurs on line #15 when scaling the explanatory variables matrix x
by the residual vector abs(y − x ∗ bnew ).
2. The function invpd on line #14 examines a matrix for positive-definite

status and inverts it if positive definite. If it is not positive definite,
augmentation of the smallest eigenvalues occurs to make the matrix
positive definite, then inversion takes place. We have a MATLAB
equivalent function invpd to carry out this task, but it should be noted
that it most likely doesn’t operate the same as the Gauss function.
3. Gauss doesn’t require asterisk symbols for all matrix multiplications,

so we have statements like w0 x and w0 y on line #14 that we need to
change.
4. Gauss carries out regression estimation using: (y/x) on line #7, which
we need to replace with a MATLAB ‘backslash’ operator.
5. We need to replace the do until loop with a while loop.
The MATLAB program corresponding to the Gauss program is shown

in example 3.15.
% ----- Example 3.15 Least-absolute deviations using lad() function

randn(’seed’,45456); nobs = 100; numx = 3; b = .5 * ones(numx+1,1);
x = [ones(nobs,1) randn(nobs,numx)];
y = x * b + randn(nobs,1).^4; % 4th power for kurtotic error term
fprintf(’LS estimates = %12.6f \n’,x\y);
b_old = zeros(numx+1,1); b_new = b; w = x;
.::UdecomBooks::.
while max(abs(b_new - b_old)) > 1e-3;

b_old = b_new;
b_new = invpd(w’*x)*(w’*y);
w = matdiv(x,abs(y - x * b_new));
end;
fprintf(’LAD estimates %12.6f \n’,b_new);
Of course, beyond simply converting the code, we can turn this into a
full-blown regression procedure and add it to the regression library. The
code for this is shown below.
function results = lad(y,x,maxit,crit)

% PURPOSE: least absolute deviations regression
% --------------------------------------------------
% USAGE: results = lad(y,x,itmax,convg)
% x = explanatory variables matrix (nobs x nvar)
% itmax = maximum # of iterations (default=500)
% convg = convergence criterion (default = 1e-15)
% --------------------------------------------------
% results.meth = ’lad’
% results.dw = Durbin-Watson Statistic
% results.iter = # of iterations
% results.conv = convergence max(abs(bnew-bold))
% --------------------------------------------------
% NOTES: minimizes sum(abs(y - x*b)) using re-iterated weighted
% least-squares where the weights are the inverse of
% the absolute values of the residuals
if nargin == 2 % error checking on inputs
crit = 1e-15; maxit = 500;
elseif nargin == 3, crit = 1e-15;
elseif nargin == 4, % do nothing
else, error(’Wrong # of arguments to lad’);
end;
[nobs nvar] = size(x);
b_old = zeros(nvar,1); % starting values
b_new = ones(nvar,1); iter = 0; w = x;
.::UdecomBooks::.
conv = max(abs(b_new-b_old));
while (conv > crit) & (iter <= maxit);
b_old=b_new; b_new = invpd(w’*x)*(w’*y);
resid = (abs(y-x*b_new)); ind = find(resid < 0.00001);
resid(ind) = 0.00001; w = matdiv(x,resid);
iter = iter+1; conv = max(abs(b_new-b_old));
end;
results.meth = ’lad’; results.beta = b_new;
results.y = y; results.nobs = nobs;
results.nvar = nvar; results.yhat = x*results.beta;
results.resid = y - results.yhat;
results.sige = sigu/(nobs-nvar);
tmp = (results.sige)*(diag(inv(w’*x)));
ym = y - ones(nobs,1)*mean(y); rsqr1 = sigu; rsqr2 = ym’*ym;
rsqr1 = rsqr1/(nobs-nvar); rsqr2 = rsqr2/(nobs-1.0);
results.dw = (ediff’*ediff)/sigu’; % durbin-watson
results.iter = iter; results.conv = conv;
3.5 Wrapper functions

When developing a set of econometric estimation routines, it is simplest to
develop a corresponding function to print the results structures returned
by the estimation functions. For example, if we were developing vector-
autoregressive estimation routines, we might devise a function prt var to
print the results structures returned by these estimation functions, and have
another function prt eqs to print results from simultaneous equations esti-
mation functions, and so on. Of course, this places a burden on the user to
remember the function names associated with printing results from each of
these different types of estimation procedures. It would be desirable to have
a single function, say prt that prints results from all econometric estimation
and testing functions.
Developing a single function prt to do this is inconvenient and becomes
unwieldy. Another option is to develop a “wrapper function” named prt
that examines the ‘meth’ field of the structure it is passed and then calls
the appropriate printing function associated with the method.
As an illustration, consider the following plt function that works to
call the appropriate plotting function for the user. Note, we should still
leave error checking on inputs in the individual functions like plt reg, as
.::UdecomBooks::.
some users may directly call these functions in lieu of using the generic plt
function.
function plt(results,vnames)
% PURPOSE: Plots results structures returned by most functions
% by calling the appropriate plotting function
%---------------------------------------------------
% USAGE: plt(results,vnames)
% Where: results = a structure returned by an econometric function
% vnames = an optional vector of variable names
% --------------------------------------------------
% NOTES: this is simply a wrapper function that calls another function
% --------------------------------------------------
% RETURNS: nothing, just plots the results
% --------------------------------------------------

error(’plt: requires a structure input’);
elseif nargin == 3, arg = 3;
else, error(’Wrong # of inputs to plt’);
end;
method = results(1).meth;
% call appropriate plotting routine
switch method
case {’arma’,’hwhite’,’lad’,’logit’,’nwest’,’ols’,’olsc’,’olsar1’,’olst’,...
’probit’,’ridge’,’robust’,’theil’,’tobit’,’tsls’}
% call plt_reg
if arg == 1, plt_reg(results);
elseif arg == 2, plt_reg(results,vnames);
end;
case {’thsls’,’sur’}
% call prt_eqs
if arg == 1, plt_eqs(results);
elseif arg == 2, plt_eqs(results,vnames);
end;
case {’var’,’bvar’,’rvar’,’ecm’,’becm’,’recm’}
% call prt_var
if arg == 1, plt_var(results);
elseif arg == 2, plt_var(results,vnames);
end;
otherwise
error(’results structure not known by plt function’);
end;
Note that the function plt contains no code to carry out plotting, it con-
sists of a large switch statement with multiple case arguments that key on
.::UdecomBooks::.
the ‘meth’ field of the structure variable input. This allows you to construct
your own plotting functions that plot information contained in ‘results struc-
tures’ and add another ‘case’ to the large switch statement in the function
plt. We illustrate how to do this using another wrapper function prt below.
In addition to plotting results, many of the econometric estimation and
testing methods produce results that are printed using an associated prt reg,
prt var, etc., function. A wrapper function named prt seems appropriate
here as well and we have constructed one as part of the Econometrics Tool-
box. All functions that provide ‘results structures’ with information that
can be printed have been added to the function prt, so a simple call to prt
with the ‘results structure’ will provide printed results. As in the case of
plt, the function prt contains no code for actually printing results, it is sim-
ply a large ‘switch’ statement that queries the ‘meth’ field of the structure
variable input and contains many ‘case statements’ where the appropriate
printing functions are called.
All regression results structures and statistical tests results can be printed
using prt(results), or prt(results,vnames,fid). This was accomplished
by providing a specialized printing function to provide formatted output
for each regression or statistical testing function. A call to the specialzed
function is then added as a ‘case’ argument in the wrapper function prt.
You should of course take the same approach when constructing your own
functions.
As an illustration of this, consider the functions johansen, adf, cadf,
phillips that implement cointegration tests. A function prt coint was con-
structed that provides nicely formatted printing of the results structures
returned by each of these functions. We then add to a case to the switch
statement in the prt function that looks as follows:
case {’johansen’,’adf’,’cadf’,’phillips’}
% call prt_coint
if arg == 1
prt_coint(results);
elseif arg == 2
prt_coint(results,vnames);
elseif arg == 3
prt_coint(results,vnames,fid);
else
prt_coint(results,[],fid);
end;
This ensures that a call to the function prt will produce printed re-
sults for the results structures produced by these four cointegration testing
.::UdecomBooks::.
functions by calling the appropriate prt coint function. It saves users the
trouble of knowing or remembering the name of your printing function.
When users submit econometric functions for inclusion in the Econo-
metrics Toolbox, they should provide a printing or plotting function where
appropriate and I will add a ‘case’ statement to the prt or plt function in
the toolbox to call your printing or plotting functions.
3.6 Chapter summary

Basic utilities for printing and plotting matrices will be used often, so con-
siderable effort should be directed towards allowing flexible forms of in-
put arguments in these functions. Use of the MATLAB cell-array variable
‘varargin’ places the burden on the programmer to parse a variable number
of input arguments into variables used by the function. This effort seems
worthwhile for frequently used basic utility functions.
We also demonstrated using structure variables as input arguments to
functions that contain a large number of alternative input options. Default
options can be set in the function and user inputs can be parsed easily with
the user of the MATLAB fieldnames and strcmp functions.
To handle time-series dates associated with observations we devised func-
tions that package this information in a structure variable that can be passed
as an argument to various time-series related functions. An alternative de-
sign would be to use ‘global scope’ variables that are available in MATLAB.
This would eliminate the need to explicitly pass the structure variables as
arguments to the functions.
An overall design decision was made to completely avoid the use of global
variables in the Econometrics toolbox. My experience has been that students
can produce the most creative and difficult to detect bugs when global scope
variables are at their disposal.
Wrapper functions prt and plt were described that allow users to pro-
vide their own printing and plotting functions which can be accessed by
users by simply calling prt or plt with the ‘results’ structure returned by
the function. This enabled easy extention of the toolbox to include new es-
timation and statistical testing functions submitted by others that provide
printed or plotted output using the single prt or plt functions.
.::UdecomBooks::.
Chapter 3 Appendix
The utility functions discussed in this chapter (as well as others not dis-
cussed) are in a subdirectory util.
utility function library
-------- utility functions -------------
accumulate - accumulates column elements of a matrix

blockdiag - creates a block diagonal matrix
cal - associates obs # with time-series calendar
ccorr1 - correlation scaling to normal column length
ccorr2 - correlation scaling to unit column length
crlag - circular lag function
diagrv - replaces main diagonal of square matrix with vector
findnear - finds matrix element nearest a scalar value
fturns - finds turning-points in a time-series
growthr - converts time-series matrix to growth rates
ical - associates time-series dates with obs #
indicator - converts a matrix to indicator variables
invccorr - inverse for ccorr1, ccorr2
invpd - makes a matrix positive-definite, then inverts
kernel_n - normal kernal density estimates
lag - generates a lagged variable vector or matrix
levels - generates factor levels variable
lprint - prints a matrix in LaTeX table-formatted form
lprintf - enhanced lprint function
matadd - adds non-conforming matrices, row or col compatible.
matdiv - divides non-conforming matrices, row or col compatible.
matmul - multiplies non-conforming matrices, row or col compatible.
matsub - divides non-conforming matrices, row or col compatible.
mlag - generates a var-type matrix of lags
mode - calculates the mode of a distribution
77
.::UdecomBooks::.
mprint - prints a matrix

mprint3 - prints coefficient, t-statistics matrices
mth2qtr - converts monthly to quarterly data
nclag - generates a matrix of non-contiguous lags
plt - wrapper function, plots all result structures
prt - wrapper function, prints all result strucutres
recserar - recursive AR series (like Gauss)
recsercp - recursive series product (like Gauss)
roundoff - rounds matrix to fixed number of decimal digits
sacf - sample autocorrelation function estimates
sdiff - seasonal differencing
sdummy - generates seasonal dummy variables
seqa - a sequence of numbers with a beginning and increment
shist - plots spline smoothed histogram
spacf - sample partial autocorrelation estimates
stdc - std deviations of columns returned as a column vector
tally - computes frequencies of distinct levels
tdiff - time-series differencing
trimc - trims columns of a matrix (or vector) like Gauss
trimr - trims rows of a matrix (or vector) like Gauss
tsdates - time-series dates function
tsprint - print time-series matrix
unsort - unsorts a sorted vector or matrix
vec - turns a matrix into a stacked vector
vech - matrix from lower triangular columns of a matrix
xdiagonal - spreads x(nxk) out to X(n*n x n*k) diagonal matrix
yvector - repeats y(nx1) to form Y(n*n x 1)
-------- demonstration programs -------------
cal_d.m - demonstrates cal function

fturns_d - demonstrates fturns and plt
ical_d - demonstrates ical function
lprint_d - demonstrates lprint function
lprintf_d - demonstrates lprintf function
mprint_d - demonstrates mprint function
mprint3_d - demonstrates mprint3 function
sacf_d - demonstrates sacf
spacf_d - demonstrates spacf
tsdate_d - demonstrates tsdate function
tsprint_d - demonstrates tsprint function
util_d - demonstrated some of the utility functions
The graphing functions are in a subdirectory graphs.
.::UdecomBooks::.
graphing function library
-------- graphing programs ---------------------------
pairs - scatter plot (uses histo)

plt - plots all results structures
pltdens - density plots
tsplot - time-series graphs
-------- demonstration programs ---------------------------
pairs_d - demonstrates pairwise scatter

pltdens_d - demonstrates pltdens
tsplot_d - demonstrates tsplot
------- support routines used
histo - used by pairs

plt_turns - used by plt to plot turning points from fturns function
.::UdecomBooks::.
Chapter 4
Regression Diagnostics
This chapter presents functions from a regression diagnostics library that

can be used to diagnose collinearity and outliers in regression. Two com-
mon references to this literature are: Regression Diagnostics, Belsley, Kuh
and Welsch (1980) and Residuals and Influence in Regression, Cook and
Weisberg (1982).
The first section of this chapter introduces functions for diagnosing and
correcting collinearity problems. The last section discusses functions to de-
tect and correct for outliers and influential observations in regression prob-
lems.
4.1 Collinearity diagnostics and procedures

Simply stated, the collinearity problem is that near linear relations among
the explanatory variable vectors tends to degrade the precision of the esti-
mated parameters. Degraded precision refers to a large variance associated
with the parameter estimates. Why should this be a matter of concern?
Quite often the motivation for estimating an economic model is hypothesis
testing to draw inferences about values of the model parameters. A large
variance, or a lack of precision, may inhibit our ability to draw inferences
from the hypothesis tests.
One way to illustrate the increase in dispersion of the least-squares es-
timates is with a Monte Carlo experiment. We generate a set of y vectors
from a model where the explanatory variables are reasonably orthogonal,
involving no near linear dependencies. Alternative sets of y vectors are then
generated from a model where the explanatory variables become increas-
ingly collinear. An examination of the variances of the β̂ estimates from
80
.::UdecomBooks::.
CHAPTER 4. REGRESSION DIAGNOSTICS 81
these experimental models should illustrate the increase in dispersion of the

estimates arising from increasing the severity of the collinear relations be-
tween the explanatory variables.
The specific experiment involved using three explanatory variables in a
model shown in (4.1).
Y = αι + βX1 + γX2 + θX3 + ε (4.1)

Initially, the three explanatory variables X1 , X2 , X3 , were generated as
random numbers from a uniform distribution. This ensures that they will
be reasonably orthogonal or independent, not involved in any near linear
dependencies. We followed a typical Monte Carlo procedure, producing
1000 different y vectors by adding a normally distributed random ε vector
to the same three fixed X’s multiplied times the parameters β, γ, θ, whose
values were set to unity.
After estimating the parameters for the 1000 data sets we find the mean
and variance of the distribution of the 1000 sets of estimates. The mean
and variance of each parameter will characterize the distribution of out-
comes from parameter estimation for a model that obeys the Gauss-Markov
assumptions and contains no collinearity problems. This also provides a
benchmark against which to judge the impact on the distribution of esti-
mates from near linear dependencies that we introduce next.
To create collinear relations we used the scheme shown in (4.2) where
we no longer generate the X2 and X3 vectors independently.
X2 = X3 + u (4.2)
Instead, we generate the X2 vector from the X3 vector with an added random
error vector u. Equation (4.2) represents X2 as a near linear combination of
X3 where the strength of the linear dependency is determined by the size of
the u vector. To generate data sets with an increasing amount of collinearity
between X2 and X3 , we adopted the following strategy:
1. First set the variance of the random normal error vector u at 1.0 and
generate the X2 vector from the X3 vector.
2. Use the three vectors X1 , X2 , X3 to generate a set of 1000 Y vectors

by adding the exact same ε vector that we used in the benchmark
generation to these three fixed X’s. The virtue of using the ε vector
from the benchmark is that, we hold the noise in the data generation
process constant. This should provide a ceteris paribus experiment
.::UdecomBooks::.
where the only change between these 1000 Y vectors and those from
the benchmark generation is the collinear nature of the X2 and X3
vectors.
3. Two additional sets of 1000 Y vectors were generated in the same

manner based on the same X3 and X1 vectors, but with two new
versions of the X2 vector generated from X3 . The new X2 vectors
were produced by decreasing the variance of the random vector u to
0.5 and 0.1, respectively.
In summary, we have four sets of 1000 Y vectors, one benchmark set,

where the three explanatory variables are reasonably independent, and three
sets where the collinear relation between the X2 and X3 vectors becomes
increasingly severe.
The MATLAB code to produce this experiment is:
% ----- Example 4.1 Collinearity experiment

n=100; k=4; u1 = randn(n,1); u2=u1*0.5; u3 = u1*0.1;
x1 = [ones(n,1) rand(n,k-1)]; % orthogonal x’s
x2 = [x1(:,1:2) x1(:,4)+u1 x1(:,4)]; % collinear set 1
ndraws = 1000; beta = ones(k,1);
bsave1 = zeros(ndraws,k); bsave2 = zeros(ndraws,k);
bsave3 = zeros(ndraws,k); bsave3 = zeros(ndraws,k);
for i=1:ndraws; % do 1000 experiments
e = randn(n,1);
y = x1*beta + e; res = ols(y,x1); b1save(i,:) = res.beta’;
end;
% compute means and std deviations for betas
mtable = zeros(4,k); stable = zeros(4,k);
mtable(1,:) = mean(b1save); mtable(2,:) = mean(b2save);
mtable(3,:) = mean(b3save); mtable(4,:) = mean(b4save);
stable(1,:) = std(b1save); stable(2,:) = std(b2save);
stable(3,:) = std(b3save); stable(4,:) = std(b4save);
% print tables
in.cnames = strvcat(’alpha’,’beta’,’gamma’,’theta’);
in.rnames = strvcat(’beta means’,’bench’,’sigu=1.0’,’sigu=0.5’,’sigu=0.1’);
in.fmt = ’%10.4f’;
mprint(mtable,in);
in.rnames = strvcat(’stand dev’,’bench’,’sigu=1.0’,’sigu=0.5’,’sigu=0.1’);
mprint(stable,in);
.::UdecomBooks::.
The results of the experiment showing both the means and standard
deviations from the distribution of estimates are:
beta means alpha beta gamma theta
benchmark 1.0033 1.0027 1.0047 0.9903
sigu=1.0 1.0055 1.0027 1.0003 0.9899
sigu=0.5 1.0055 1.0027 1.0007 0.9896
sigu=0.1 1.0055 1.0027 1.0034 0.9868
standard dev alpha beta gamma theta

benchmark 0.3158 0.3285 0.3512 0.3512
sigu=1.0 0.2697 0.3286 0.1025 0.3753
sigu=0.5 0.2697 0.3286 0.2049 0.4225
sigu=0.1 0.2697 0.3286 1.0247 1.1115
A first point to note about the experimental outcomes is that the means
of the estimates are unaffected by the collinearity problem. Collinearity cre-
ates problems with regard to the variance of the distribution of the estimates,
not the mean. A second point is that the benchmark data set produced pre-
cise estimates, with standard deviations for the distribution of outcomes
around 0.33. These standard deviations would result in t-statistics around
3, allowing us to infer that the true parameters are significantly different
from zero.
Turning attention to the standard deviations from the three collinear
data sets we see a clear illustration that increasing the severity of the near
linear combination between X2 and X3 produces an increase in the standard
deviation of the resulting distribution for the γ and θ estimates associated
with X2 and X3 . The increase is about three-fold for the worse case where
σu2 = 0.1 and the strength of the collinear relation between X2 and X3 is
the greatest.
A diagnostic technique presented in Chapter 3 of Regression Diagnos-
tics by Belsley, Kuh, and Welsch (1980) is implemented in the function
bkw. The diagnostic is capable of determining the number of near linear
dependencies in a given data matrix X, and the diagnostic identifies which
variables are involved in each linear dependency. This diagnostic technique
is based on the Singular Value Decomposition that decomposes a matrix
X = U DV 0 , where U contains the eigenvectors of X and D is a diagonal
matrix containing eigenvalues.
For diagnostic purposes the singular value decomposition is applied to
the variance-covariance matrix of the least-squares estimates and rearranged
to form a table of variance-decomposition proportions. The procedure for a
k variable least-squares model is described in the following. The variance of
the estimate β̂k can be expressed as shown in (4.3).
.::UdecomBooks::.
X
k
var(β̂k ) = σ̂ε2 2
(Vkj /λ2j ) (4.3)
j=1
The diagnostic value of this expression lies in the fact that it decomposes
var(β̂k ) into a sum of components, each associated with one of the k singular
values λ2j that appear in the denominator. Expression (4.4) expands the
summation in (4.3) to show this more clearly.
var(β̂k ) = σ̂ε2 {(V11

2 2
/λ1 ) + (V12 2
/λ2 ) + (V13 2
/λ3 ) + . . . + (V1k /λk )} (4.4)
Since small λj are associated with near linear dependencies, an unusually

large proportion of the variance of the coefficients of variables involved in
the linear dependency will be concentrated in the components associated
with the small λj . The table is formed by defining the terms φ and π shown
in (4.5) and (4.6).
φij = (Vij2 /λ2j ) (4.5)

X
k
φi = φij , i = 1, . . . , k
j=1
πji = (φij /φi ), i, j = 1, . . . , k (4.6)

The term πji is called a variance-decomposition proportion. These magni-
tudes are placed in a table as shown in Table 3.1.
Table 4.1: Variance-decomposition proportions table

Condition index var(β̂1 ) var(β̂2 ) ... var(β̂k )
λmax /λmax π11 π12 ... π1k

λmax /λ2 π21 π22 ... π2k
.. .. .. .. ..
. . . . .
λmax /λmin πk1 πk2 ... πkk
It is shown in Belsley, Kuh and Welsch (1980) that a large value of

the condition index, κ(X) = λmax /λi is associated with each near linear
dependency, and the variates involved in the dependency are those with
.::UdecomBooks::.
large proportions of their variance associated with large κ(X) magnitudes.

Empirical tests performed in Chapter 3 of Belsley, Kuh, and Welsch (1980)
determined that variance-decomposition proportions in excess of 0.5 indicate
the variates involved in specific linear dependencies. The joint condition of
magnitudes for κ(X) > 30, and πij values > 0.5 diagnose the presence of
strong collinear relations as well as determining the variates involved.
Table 3.2 shows an example of how the variance-decomposition propor-
tions table might look for an actual data set. The example in the table
would indicate that there exists one condition index, κ(X), of 87 and an-
other of 98. For these two condition indices we examine the variance pro-
portions looking for those that exceed 0.5. We find that for κ(X) = 87,
two variance-proportions exceed 0.5 pointing towards a near linear relation-
ship between the x1 and x5 variable. The κ(X) = 98 also contains two
variance-proportions that exceed 0.5 indicating the presence of a collinear
relation involving x2 and x6 . From Table 3.2 then we would conclude that
two possible near linear relations existed in the data set.
Table 4.2: BKW collinearity diagnostics example

κ(X) x1 x2 x3 x4 x5 x6
1 .00 .00 .01 .01 .00 .00
28 .02 .03 .07 .03 .04 .03
58 .04 .00 .78 .02 .28 .01
60 .01 .02 .03 .76 .02 .22
87 .58 .28 .04 .10 .55 .07
98 .36 .66 .07 .09 .11 .67
A function bkw implements these tests, with the documentation shown

below:
PURPOSE: computes and prints BKW collinearity diagnostics

variance-decomposition proportions matrix
---------------------------------------------------
USAGE: bkw(x,vnames,fmt)
where: x = independent variable matrix (from a regression model)
vnames = (optional) variable name vector
fmt = (optional) format string, e.g., ’%12.6f’ or ’%12d’
default = %10.2f
---------------------------------------------------
NOTE: you can use either x-variable names or an ols
vnames argument containing y-variable + x-variable names
e.g. vnames = strvcat(’y’,’x1’,’x2’) or
.::UdecomBooks::.
vnames = strvcat(’x1’,’x2’)
---------------------------------------------------
RETURNS:
nothing, just prints the table out
--------------------------------------------------
SEE ALSO: dfbeta, rdiag, diagnose
---------------------------------------------------
REFERENCES: Belsley, Kuh, Welsch, 1980 Regression Diagnostics
----------------------------------------------------
The function allows a variable name vector and format as optional inputs.
As a convenience, either a variable name vector with names for the variables
in the data matrix X or one that includes a variable name for y as well as
the variables in X can be used. This is because the bkw function is often
called in the context of regression modeling, so we need only construct a
single variable name string vector that can be used for printing regression
results as well as labelling variables in the bkw output.
As an example of using the bkw function to carry out tests for collinear-
ity, the program below generates a collinear data set and and uses the bkw
function to test for near linear relationships.
% ----- Example 4.2 Using the bkw() function

n = 100; k = 5;
x = randn(n,k);
% generate collinear data
x(:,1) = x(:,2) + x(:,4) + randn(n,1)*0.1;
bkw(x);
The results of the program are shown below. They detect the near linear
relationship between variables 1, 2 and 4 which we generated in the data
matrix X.
Belsley, Kuh, Welsch Variance-decomposition

K(x) var 1 var 2 var 3 var 4 var 5
1 0.00 0.00 0.00 0.00 0.00
15 0.00 0.00 0.26 0.00 0.40
17 0.00 0.00 0.24 0.00 0.00
20 0.00 0.00 0.47 0.00 0.59
31 1.00 0.99 0.03 0.99 0.01
A common corrective procedure for this problem is ridge regression,

which is implemented by the function ridge. Ridge regression attacks the
problem of small eigenvalues in the X 0 X matrix by augmenting or inflat-
ing the smallest values to create larger magnitudes. The increase in small
eigenvalues is accomplished by adding a diagonal matrix γIk to the X 0 X
.::UdecomBooks::.
matrix before inversion. The scalar term γ is called the ‘ridge’ parameter.
The ridge regression formula is shown in (4.7).
β̂R = (X 0 X + γIk )−1 X 0 y (4.7)

Recall that X 0 X is of dimension (k x k), where k is the number of ex-
planatory variables in the model. Of the k eigenvalues associated with X 0 X,
the smallest are presumably quite small as a result of the collinear relation-
ship between some of the explanatory variables. To see how addition of the
diagonal matrix γIk to the X 0 X matrix increases the smallest eigenvalues,
consider using the singular value decomposition of X 0 X. This allows us to
rewrite (4.7) as:
β̂R = (V 0 DV + γIk )−1 X 0 y (4.8)

Since γIk is a diagonal matrix, containing zeros on the off-diagonal ele-
ments, adding this to the V 0 DV matrices will only affect the elements of the
diagonal matrix D. Noting this, we find that the ridge estimation formula
can be written as in (4.9)
β̂R = (V 0 (D + γIk )V )−1 X 0 y (4.9)

The diagonal matrix D from the Singular Value Decomposition contains
the eigenvalues of the X 0 X matrix, and equation (4.9) shows that the ridge
parameter γ is added to each and every eigenvalue on the diagonal of the
matrix D. An expansion of the matrix (D + γIk ) shown in (4.10) should
make this clear.
 
λ1 + γ 0 0 ... 0
 
 0 λ2 + γ 0 ... 0 
(D + γIk ) = 
 .. .. .. 
 (4.10)
 . . . 
0 0 λk + γ
To illustrate how addition of the γ parameter to the eigenvalues im-
pacts the estimates, consider the following numerical example. The Monte
Carlo experiment for the strongest collinear relationship (where σU2 = 0.1)
produced the eigenvalues shown in Table 3.3
There are two things to note concerning the impact of the addition of
a γ = 0.2 value to the X 0 X matrix. First, the large eigenvalues are not
affected very much, since we are adding a small value to these large magni-
tudes. Second, the smallest eigenvalue will be increased dramatically, from
0.0456 to 0.2456. The impact of adding the small value to the eigenvalues
.::UdecomBooks::.
Table 4.3: Ridge Regression for our Monte Carlo example

Eigenvalues of Condition Index Eigenvalues of Condition Index
X 0X X 0X (X 0 X + 0.2Ik ) (X 0 X + 0.2Ik )
201.2950 1.0 201.4950 1.0

10.0227 20.0 10.2227 19.7
5.6565 35.5 5.8565 34.4
0.0456 4414.3 0.2456 820.4
of the original data matrix X is to dramatically improve the conditioning of

the estimation problem. In the example shown in Table 3.3, the condition
index of the X 0 X matrix falls dramatically to 820.4, producing a condition
index for the X matrix equal to the square root of 820.4, which equals 28.6.
(Note, the eigenvalues of the X 0 X matrix are the square of those from the
X matrix).
As an example, consider the following MATLAB program that generates
a collinear data set and produces both least-squares and ridge regression
estimates. The ridge estimates are produced using a value recommended by
Hoerl and Kennard (1970), but the user has an option of entering a value
for the ridge parameter as well.
% ----- Example 4.3 Using the ridge() function

n = 100; k = 5;
x = randn(n,k); e = randn(n,1); b = ones(k,1);
x(:,1) = x(:,2) + x(:,4) + randn(n,1)*0.1;
y = x*b + e;
% ols regression
res = ols(y,x);
prt(res);
% ridge regression
rres = ridge(y,x);
prt(rres);
The results from ols and ridge estimation are shown below. From these
results we see that the near linear relationship between variables x1 , x2 and
x4 lead to a decrease in precision for these estimates. The ridge estimates
increase the precision as indicated by the larger t−statistics.

R-squared = 0.9055
.::UdecomBooks::.
sigma^2 = 0.9237
***************************************************************
variable 1 -0.293563 -0.243685 0.808000
variable 2 2.258060 1.871842 0.064305
variable 3 1.192133 11.598932 0.000000
variable 4 2.220418 1.796384 0.075612
variable 5 0.922009 8.674158 0.000000
Ridge Regression Estimates

R-squared = 0.9051
sigma^2 = 0.9272
Ridge theta = 0.0022705069
***************************************************************
variable 1 0.430781 0.552496 0.581906
variable 2 1.531597 1.956202 0.053378
variable 3 1.186535 11.541312 0.000000
variable 4 1.476339 1.841555 0.068661
variable 5 0.900636 8.680297 0.000000
A point to note about ridge regression is that it does not produce un-
biased estimates. The amount of bias in the estimates is a function of how
large the value of the ridge parameter γ is. Larger values of γ lead to im-
proved precision in the estimates — at a cost of increased bias.
A function rtrace helps to assess the trade-off between bias and effi-
ciency by plotting the ridge estimates for a range of alternative values of the
ridge parameter. The documentation for rtrace is:
PURPOSE: Plots ntheta ridge regression estimates

---------------------------------------------------
USAGE: rtrace(y,x,thetamax,ntheta,vnames)
x = independent variables matrix
thetamax = maximum ridge parameter to try
ntheta = number of values between 0 and thetamax to try
vnames = optional variable names vector
e.g. vnames = strvcat(’y,’x1’,’x2’);
---------------------------------------------------
RETURNS:
nothing, plots the parameter estimates as a function
.::UdecomBooks::.
of the ntheta alternative ridge parameter values

--------------------------------------------------
As an example of using this function, consider the following program

where we recover the ridge parameter determined using the Hoerl and Ken-
nard formula, double it and produce a trace plot using values between θ = 0
and 2θ. A value of θ = 0 represents least-squares estimates, and 2θ is twice
the value we found using the Hoerl-Kennard formula.
% ----- Example 4.4 Using the rtrace() function
n = 100; k = 5;
x(:,1) = x(:,2) + x(:,4) + randn(n,1)*0.1;
y = x*b + e;
% ridge regression
res = ridge(y,x);
theta = res.theta;
tmax = 2*theta;
ntheta = 50;
vnames = strvcat(’y’,’x1’,’x2’,’x3’,’x4’,’x5’);
rtrace(y,x,tmax,ntheta,vnames);
To the extent that the parameter values vary greatly from those associ-
ated with values of θ = 0, we can infer that a great deal of bias has been
introduced by the ridge regression. A graph produced by rtrace is shown
in Figure 4.1, indicating a fair amount of bias associated with 3 of the 5
parameters in this example.
Another solution for the problem of collinearity is to use a Bayesian
model to introduce prior information. A function theil produces a set of
estimates based on the ‘mixed estimation’ method set forth in Theil and
Goldberger (1961). The documentation for this function is:
PURPOSE: computes Theil-Goldberger mixed estimator
y = X B + E, E = N(0,sige*IN)
c = R B + U, U = N(0,v)
---------------------------------------------------
USAGE: results = theil(y,x,c,R,v)
x = independent variables matrix of rank(k)
c = a vector of prior mean values, (c above)
R = a matrix of rank(r) (R above)
v = prior variance-covariance (var-cov(U) above)
---------------------------------------------------
results.meth = ’theil’
.::UdecomBooks::.
Values of Regression Coefficients as a Function of θ

1.4
x1
x2
1.3 x3
x4
x5
1.2
Regression Coefficients
1.1
0.9
0.8
0.7
0.6
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016
Value of θ, vertical line shows H-K θ value
Figure 4.1: Ridge trace plot
results.beta = bhat estimates

results.tstat = t-statistics
results.pmean = prior means
results.pstd = prior std deviations
results.yhat = predicted values
results.sige = e’e/(n-k)
results.rsqr = r-squared
results.rbar = r-squared adjusted
results.dw = Durbin Watson
results.nobs = # of observations
results.nvar = # of variables
results.y = actual observations
--------------------------------------------------
SEE ALSO: prt, plt, ols_g
---------------------------------------------------
The user can specify subjective prior information in the form of a normal
.::UdecomBooks::.
prior for the parameters β in the model. Theil and Goldberger showed
that this prior information can be expressed as stochastic linear restrictions
taking the form:
c = Rβ + u, (4.11)
These matrices are used as additional dummy or fake data observations in
the estimation process. The original least-squares model in matrix form can
be rewritten as in (4.12) to show the role of the matrices defined above.
     
y X ε
     
 ...  =  ... β +  ...  (4.12)
c R u
The partitioning symbol, (. . .), in the matrices and vectors of (4.12)
designates that we are adding the matrix R and the vectors c and u to the
original matrix X and vectors y and ε. These additional observations make
it clear we are augmenting the weak sample data with our prior information.
At this point we use an OLS estimation algorithm on the modified y vector
and X matrix to arrive at the “mixed estimates”. One minor complication
arises here however, the theoretical disturbance vector no longer consists of
the simple ε vector which obeys the Gauss-Markov assumptions, but has an
additional u vector added to it. We need to consider the variance-covariance
structure of this new disturbance vector which requires a minor modification
of the OLS formula producing the resulting “mixed estimator” shown in
(4.13).
β̂M = (σε−2 X 0 X + R0 Σ−1 −1 −2 0 0 −1

u R) (σε X y + R Σu c) (4.13)
Where Σu in (4.13) represents a diagonal matrix containing the variances
2 , i = 1, . . . , k on the diagonal.
σui
As an illustration, consider using a normal prior centered on the true
parameter values of unity in our previous example. Prior variances of unity
are also assigned, indicating a fairly large amount of uncertainty in our prior
beliefs. The following program sets up the prior and calls theil to estimate
the model.
% ----- Example 4.5 Using the theil() function

n = 100; k = 5;
x(:,1) = x(:,2) + x(:,4) + randn(n,1)*0.1;
y = x*b + e;
.::UdecomBooks::.
% set up prior information

c = ones(k,1); % prior means
sigu = eye(k); % prior variances
R = eye(k);
reso = ols(y,x); % ols
prt(reso);
res = theil(y,x,c,R,sigu); %theil
prt(res);
This program produced the following results indicating that use of prior
information improved the precision of the estimates compared to the least-
squares estimates.

R-squared = 0.9202
sigma^2 = 1.0139
***************************************************************
variable 1 0.445405 0.486733 0.627569
variable 2 1.470548 1.575896 0.118374
variable 3 1.280441 12.802818 0.000000
variable 4 1.451183 1.569624 0.119828
variable 5 0.864678 7.908782 0.000000
Theil-Goldberger Regression Estimates

R-squared = 0.9200
sigma^2 = 1.0155
***************************************************************
Variable Prior Mean Std Deviation
variable 1 1.000000 1.000000
variable 2 1.000000 1.000000
variable 3 1.000000 1.000000
variable 4 1.000000 1.000000
variable 5 1.000000 1.000000
***************************************************************
Posterior Estimates
variable 1 0.795158 1.619936 0.108561
variable 2 1.115213 2.215848 0.029090
variable 3 1.280282 12.789435 0.000000
variable 4 1.098660 2.202581 0.030043
variable 5 0.863902 7.899374 0.000000
.::UdecomBooks::.
4.2 Outlier diagnostics and procedures

Outlier observations are known to adversely impact least-squares estimates
because the aberrant observations generate large errors. The least-squares
criterion is such that observations associated with large errors receive more
weight or exhibit more influence in determining the estimates of β.
A number of procedures have been proposed to diagnose the presence of
outliers and numerous alternative estimation procedures exist that attempt
to “robustify” or downweight aberrant or outlying observations. Function
dfbeta produces a set of diagnostics discussed in Belsley, Kuh and Welsch
(1980). They suggest examining the change in least-squares estimates β̂ that
arise when we omit each observation from the regression sample sequentially.
The basic idea is that eliminating an influential observation will produce a
large change in the β̂ estimates, allowing us to detect these observations
graphically.
The function dfbeta returns a structure that can be used to produce
graphical output with plt dfb, plt dff or plt. The function documentation
is:
PURPOSE: computes BKW (influential observation diagnostics)

dfbetas, dffits, hat-matrix, studentized residuals
---------------------------------------------------
USAGE: result = dfbeta(y,x)
where: y = dependent variable vector (from a regression model)
x = independent variable matrix
---------------------------------------------------
results.meth = ’dfbeta’
results.dfbeta = df betas
results.dffits = df fits
results.hatdi = hat-matrix diagonals
results.stud = studentized residuals
results.nvar = # of variables in x-matrix
--------------------------------------------------
SEE ALSO: plt_dfb, plt_dff, bkw
---------------------------------------------------
An example where we generate a data set and then artificially create two
outliers at observations #50 and #70 is shown below. The graphical output
from plt dfb in Figure 4.2 shows a graph of the change in β̂ associated with
omitting each observation. We see evidence of the outliers at observations
#50 and #70 in the plot.
.::UdecomBooks::.
% ----- Example 4.6 Using the dfbeta() function

n = 100; k = 4;
y = x*b + e;
% now add a few outliers
y(50,1) = 10.0; y(70,1) = -10.0;
vnames = strvcat(’y’,’x1’,’x2’,’x3’,’x4’);
res = dfbeta(y,x);
plt_dfb(res,vnames);
pause;
plt_dff(res);
Figure 4.3 shows another diagnostic ‘dffits’ produced by the function

dfbeta that indicates how the fitted values change when we sequentially
eliminate each observation from the regression data set. A similar function
diagnose computes many of the traditional statistics from the regression
diagnostics literature and prints this information for candidate outlier ob-
servations.
df betas
0.2
x1
-0.2
0 10 20 30 40 50
df betas 60 70 80 90 100
0.1
x2
-0.1
0 10 20 30 40 50
df betas 60 70 80 90 100
0.5
x3
-0.5
0 10 20 30 40 50
df betas 60 70 80 90 100
0.1
x4
-0.1
0 10 20 30 40 50 60 70 80 90 100
Figure 4.2: Dfbeta plots
.::UdecomBooks::.
dffits
4
-2
0 10 20 30 40 50 60 70 80 90 100
studentized residuals
10
-5
-10
0 10 20 30 40 50 60 70 80 90 100
hat-matrix diagonals
0.2
0.15
0.1
0.05
0
0 10 20 30 40 50 60 70 80 90 100
Figure 4.3: Dffits plots
Another routine, rdiag, produces graphical diagnostics for the residuals

taking the form of 4 plots. A normal plot of the residuals, an I-chart of the
residuals, a histogram of the residuals and the residuals versus the fitted
values.
A number of alternative estimation methods exist that attempt to down-
weight outliers. The regression library contains a function robust and olst
as well as lad that we developed in Chapter 3. The documentation for
robust is:
PURPOSE: robust regression using iteratively reweighted

least-squares
---------------------------------------------------
USAGE: results = robust(y,x,wfunc,wparm)
where: y = dependent variable vector (nobs x 1)
x = independent variables matrix (nobs x nvar)
wfunc = 1 for Huber’s t function
2 for Ramsay’s E function
.::UdecomBooks::.
3 for Andrew’s wave function

4 for Tukey’s biweight
wparm = weighting function parameter
---------------------------------------------------
results.meth = ’robust’
results.beta = bhat
results.yhat = yhat
results.iter = # of iterations
results.nobs = nobs
results.wfunc = ’huber’, ’ramsay’, ’andrew’, ’tukey’
results.wparm = wparm
results.weight = nobs - vector of weights
The function incorporates four alternative weighting schemes that have

been proposed in the literature on iteratively re-weighted regression meth-
ods. As an illustration of all four methods, consider the following program
that produces estimates using all methods.
% ----- Example 4.7 Using the robust() function
nobs = 100; nvar = 3; x = randn(nobs,nvar); x(:,1) = ones(nobs,1);
beta = ones(nvar,1); evec = randn(nobs,1); y = x*beta + evec;
y(75,1) = 10.0; y(90,1) = -10.0; % put in 2 outliers
% parameter weighting from OLS (of course you’re free to do differently)
reso = ols(y,x); sige = reso.sige;
bsave = zeros(nvar,5); bsave(:,1) = ones(nvar,1); % storage for results
for i=1:4; % loop over all methods producing estimates
wfunc = i; wparm = 2*sige; % set weight to 2 sigma
res = robust(y,x,wfunc,wparm);
bsave(:,i+1) = res.beta;
end;
in.cnames = strvcat(’Truth’,’Huber t’,’Ramsay’,’Andrews’,’Tukey’);
in.rnames = strvcat(’Coefficients’,’beta1’,’beta2’,’beta3’);
in.fmt = ’%10.4f’;
mprint(bsave,in);
res = robust(y,x,4,2);
prt(res); plt(res); % demonstrate prt and plt functions
Note that we can use our function mprint from Chapter 3 to produce
a formatted printout of the results that looks as follows:
.::UdecomBooks::.
Coefficients Truth Huber t Ramsay Andrews Tukey

beta1 1.0000 1.1036 0.9529 1.0985 1.0843
beta2 1.0000 0.9996 1.1731 0.9987 0.9253
beta3 1.0000 1.1239 0.9924 1.1232 1.0650
The routine olst performs regression based on an assumption that the

errors are t−distributed rather than normal, which allows for “fat-tailed”
error distributions. The documentation is:
PURPOSE: ols with t-distributed errors

--------------------------------------------------
USAGE: results = olst(y,x,itmax,convg)
x = explanatory variables matrix (nobs x nvar)
itmax = maximum # of iterations (default=500)
convg = convergence criterion (default = 1e-8)
--------------------------------------------------
results.meth = ’olst’
results.beta = bhat
results.yhat = yhat
results.nobs = nobs
results.conv = convergence max(abs(bnew-bold))
--------------------------------------------------
NOTES: uses iterated re-weighted least-squares
to find maximum likelihood estimates
--------------------------------------------------
SEE ALSO: prt, plt
---------------------------------------------------
REFERENCES: Section 22.3 Introduction to the Theory and Practice
of Econometrics, Judge, Hill, Griffiths, Lutkepohl, Lee
Another graphical tool for regression diagnostics is the pairs function

that produces pairwise scatterplots for a group of variables as well as his-
tograms of the distribution of observations for each variable. This function
is a modified version of a function by Anders Holtsberg’s public domain
statistics toolbox. The documentation was altered to conform to that for
.::UdecomBooks::.
other functions in the Econometrics Toolbox and a variable names capabil-

ity was added to the function. The following program illustrates use of this
function and Figure 4.4 shows the resulting pairwise scatterplot.
% ----- Example 4.8 Using the pairs() function

n = 100;
y1 = randn(n,1);
y2 = 2*y1 + randn(n,1);
y3 = -2*y2 + randn(n,1);
y4 = randn(n,1); % uncorrelated with y1,y2,y3
y5 = randn(n,1).^4; % leptokurtic variable
y = [y1 y2 y3 y4 y5];
vnames = strvcat(’apples’,’oranges’,’pairs’,’ha ha’,’leptoku’);
pairs(y,vnames);
Figure 4.4: Pairwise scatter plots
.::UdecomBooks::.
4.3 Chapter summary

A large number of regression diagnostic tests and corrective estimation pro-
cedures have been proposed in the econometrics literature. MATLAB and
Gauss code for implementing these methods can be found on many Internet
sites. The Econometrics Toolbox design allows these routines to be imple-
mented and documented in a way that provides a consistent user interface
for printing and plotting the diagnostics.
.::UdecomBooks::.
Chapter 4 Appendix
A library of routines in the subdirectory diagn contain many of the functions

discussed in this chapter.
regression diagnostics library
arch - ARCH(p) test

bkw - BKW collinearity diagnostics
bpagan - Breusch-Pagan heteroscedasticity test
cusums - Brown,Durbin,Evans cusum squares test
dfbeta - BKW influential observation diagnostics
diagnose - compute diagnostic statistics
plt_dfb - plots dfbetas
plt_dff - plots dffits
plt_cus - plots cusums
recresid - compute recursive residuals
rdiag - graphical residuals diagnostics
studentize - standarization transformation
unstudentize - reverses studentize transformation
qstat2 - Box-Ljung Q-statistic
------- demonstration programs -----------
arch_d - demonstrates arch

bkw_d - demonstrates bkw
bpagan_d - demonstrates bpagan
cusums_d - demonstrates cusums
dfbeta_d - demonstrates dfbeta, plt_dfb, plt_dff
diagnose_d - demonstrates diagnose
recresid_d - demonstrates recresid
rdiag_d - demonstrates rdiag
unstudentize_d - demonstrates studentize, unstudentize
qstat2_d - demonstrates qstat2
------- support functions
ols - least-squares regression

plt - plots everything
101
.::UdecomBooks::.
plt_cus - plots cusums test results

Functions from the regression library discussed in this chapter were:

rtrace - ridge trace plots
-------- Demo programs ---------------------------
olst_d - olst demo

ridge_d - ridge regression and rtrace demo
.::UdecomBooks::.
Chapter 5
VAR and Error Correction

Models
This chapter describes the design and use of MATLAB functions to imple-
ment vector autoregressive (VAR) and error correction (EC) models. The
MATLAB functions described here provide a consistent user-interface in
terms of the MATLAB help information, and related routines to print and
plot results from the various models. One of the primary uses of VAR and
EC models is econometric forecasting, for which we provide a set of func-
tions.
Section 5.1 describes the basic VAR model and our function to estimate
and print results for this method. Section 5.2 turns attention to EC models
while Section 5.3 discusses Bayesian variants on these models. Finally, we
take up forecasting in Section 5.4. All of the functions implemented in our
vector autoregressive function library are documented in the Appendix to
this chapter.
5.1 VAR models

A VAR model is shown in (5.1) that contains n variables. The εit denote
independent disturbances, Ci represent constants and yit , i = 1, . . . , n denote
the n variables in the model at time t. Model parameters Aij (`) take the
P
form, m k k
k=1 aijk ` , where ` is the lag operator defined by ` yt = yt−k , and
m is the lag length specified by the modeler.
103
.::UdecomBooks::.
CHAPTER 5. VAR AND ERROR CORRECTION MODELS 104
       
y1t   y1t C1 ε1t
  A11 (`) . . . A1n (`)      
 y2t   .. ..  y2t   C2   ε2t 
 = ..  + +  (5.1)
 ..  . . .  ..   ..   .. 
 .   .   .   . 
An1 (`) . . . Ann (`)
ynt ynt Cn εnt
The VAR model posits a set of relationships between past lagged values
of all variables in the model and the current value of each variable in the
model. For example, if the yit represent employment in state i at time
t, the VAR model structure allows employment variation in each state to
be explained by past employment variation in the state itself, yit−k , k =
1, . . . , m as well as past employment variation in other states, yjt−k , k =
1, . . . , m, j 6= i. This is attractive since regional or state differences in
business cycle activity suggest lead/lag relationships in employment of the
type set forth by the VAR model structure.
The model is estimated using ordinary least-squares, so we can draw
on our ols routine from the regression library. A function var produces
estimates for the coefficients in the VAR model as well as related regression
statistics and Granger-causality test statistics.
The documentation for the var function is:
PURPOSE: performs vector autoregressive estimation

---------------------------------------------------
USAGE: result = var(y,nlag,x)
where: y = an (nobs x neqs) matrix of y-vectors
nlag = the lag length
x = optional matrix of variables (nobs x nx)
NOTE: constant vector automatically included
---------------------------------------------------
RETURNS a structure
results.meth = ’var’
results.nobs = nobs, # of observations
results.neqs = neqs, # of equations
results.nlag = nlag, # of lags
results.nvar = nlag*neqs+nx+1, # of variables per equation
--- the following are referenced by equation # ---
results(eq).beta = bhat for equation eq
results(eq).tstat = t-statistics
results(eq).tprob = t-probabilities
results(eq).resid = residuals
results(eq).yhat = predicted values
results(eq).y = actual values
results(eq).sige = e’e/(n-k)
results(eq).rsqr = r-squared
.::UdecomBooks::.
results(eq).rbar = r-squared adjusted

results(eq).boxq = Box Q-statistics
results(eq).ftest = Granger F-tests
results(eq).fprob = Granger marginal probabilities
---------------------------------------------------
SEE ALSO: varf, prt_var, pgranger, pftests
---------------------------------------------------
This function utilizes a new aspect of MATLAB structure variables, ar-

rays that can store information for each equation in the VAR model. Esti-
mates of the β̂ parameters for the first equation can be accessed from the
results structure using: result(1).beta as can other results that are equation-
specific.
In most applications, the user would simply pass the results structure
on to the prt var (or prt) function that will provide an organized printout
of the regression results for each equation. Here is an example of a typical
program to estimate a VAR model.
% ----- Example 5.1 Using the var() function
dates = cal(1982,1,12);
load test.dat; % monthly mining employment in 8 states
y = growthr(test(:,1:2),12); % convert to growth-rates
yt = trimr(y,dates.freq,0); % truncate
dates = cal(1983,1,1); % redefine calendar for truncation
vnames = strvcat(’illinos’,’indiana’);
nlag = 2;
result = var(yt,nlag); % estimate 2-lag VAR model
prt(result,vnames); % printout results
It would produce the following printout of the estimation results:

***** Vector Autoregressive Model *****
Dependent Variable = illinos
R-squared = 0.9044
sige = 3.3767
Q-statistic = 0.2335
******************************************************************
illinos lag1 1.042540 13.103752 0.000000
illinos lag2 -0.132170 -1.694320 0.092226
indiana lag1 0.228763 3.790802 0.000215
indiana lag2 -0.213669 -3.538905 0.000531
constant -0.333739 -1.750984 0.081940
****** Granger Causality Tests *******
.::UdecomBooks::.
Variable F-value Probability

illinos 363.613553 0.000000
indiana 7.422536 0.000837
Dependent Variable = indiana

R-squared = 0.8236
sige = 6.5582
Q-statistic = 0.0392
******************************************************************
illinos lag1 0.258853 2.334597 0.020856
illinos lag2 -0.195181 -1.795376 0.074555
indiana lag1 0.882544 10.493894 0.000000
indiana lag2 -0.029384 -0.349217 0.727403
constant -0.129405 -0.487170 0.626830

illinos 2.988892 0.053272
indiana 170.063761 0.000000
There are two utility functions that help analyze VAR model Granger-
causality output. The first is pgranger, which prints a matrix of the
marginal probabilities associated with the Granger-causality tests in a con-
venient format for the purpose of inference. The documentation is:
PURPOSE: prints VAR model Granger-causality results
--------------------------------------------------
USAGE: pgranger(results,varargin);
where: results = a structure returned by var(), ecm()
varargin = a variable input list containing
vnames = an optional variable name vector
cutoff = probability cutoff used when printing
usage example 1: pgranger(result,0.05);
example 2: pgranger(result,vnames);
example 3: pgranger(result,vnames,0.01);
example 4: pgranger(result,0.05,vnames);
----------------------------------------------------
e.g. cutoff = 0.05 would only print
marginal probabilities < 0.05
---------------------------------------------------
NOTES: constant term is added automatically to vnames list
you need only enter VAR variable names plus deterministic
---------------------------------------------------
As example of using this function, consider our previous program to esti-

mate the VAR model for monthly mining employment in eight states. Rather
.::UdecomBooks::.
than print out the detailed VAR estimation results, we might be interested
in drawing inferences regarding Granger-causality from the marginal proba-
bilities. The following program would produce a printout of just these prob-
abilities. It utilizes an option to suppress printing of probabilities greater
than 0.1, so that our inferences would be drawn on the basis of a 90% con-
fidence level.
% ----- Example 5.2 Using the pgranger() function
dates = cal(1982,1,12); % monthly data starts in 82,1
load test.dat; % monthly mining employment in 8 states
y = growthr(test,12); % convert to growth-rates
yt = trimr(y,dates.freq,0); % truncate
dates = cal(1983,1,1); % redefine the calendar for truncation
vname = strvcat(’il’,’in’,’ky’,’mi’,’oh’,’pa’,’tn’,’wv’);
nlag = 12;
res = var(yt,nlag); % estimate 12-lag VAR model
cutoff = 0.1; % examine Granger-causality at 90% level
pgranger(res,vname,cutoff); % print Granger probabilities
We use the ‘NaN’ symbol to replace marginal probabilities above the

cutoff point (0.1 in the example) so that patterns of causality are easier to
spot. The results from this program would look as follows:
****** Granger Causality Probabilities *******
Variable il in ky mi oh pa tn wv
il 0.00 0.01 0.01 NaN NaN 0.04 0.09 0.02
in 0.02 0.00 NaN NaN NaN NaN NaN NaN
ky NaN NaN 0.00 NaN 0.10 NaN 0.07 NaN
mi NaN 0.01 NaN 0.00 NaN NaN NaN NaN
oh NaN 0.05 0.08 NaN 0.00 0.01 NaN 0.01
pa 0.05 NaN NaN NaN NaN 0.00 NaN 0.06
tn 0.02 0.05 NaN NaN NaN 0.09 0.00 NaN
wv 0.02 0.05 0.06 0.01 NaN 0.00 NaN 0.03
The format of the output is such that the columns reflect the Granger-
causal impact of the column-variable on the row-variable. That is, Indiana,
Kentucky, Pennsylvania, Tennessee and West Virginia exert a significant
Granger-causal impact on Illinois employment whereas Michigan and Ohio
do not. Indiana exerts the most impact, affecting Illinois, Michigan, Ohio,
Tennessee, and West Virginia.
The second utility is a function pftest that prints just the Granger-
causality joint F-tests from the VAR model. Use of this function is similar
to pgranger, we simply call the function with the results structure returned
by the var function, e.g., pftest(result,vnames), where the ‘vnames’ ar-
gument is an optional string-vector of variable names. This function would
.::UdecomBooks::.
produce the following output for each equation of a VAR model based on
all eight states:

Equation illinois F-value F-probability
illinois 395.4534 0.0000
indiana 3.3255 0.0386
kentucky 0.4467 0.6406
michigan 0.6740 0.5112
ohio 2.9820 0.0536
pennsylvania 6.6383 0.0017
tennessee 0.9823 0.3768
west virginia 3.0467 0.0504
A few points regarding how the var function was implemented. We rely
on the ols function from the regression library to carry out the estimation
of each equation in the model and transfer the ‘results structure’ returned
by ols to a new var results structure array for each equation. Specifically
the code looks as follows:
% pull out each y-vector and run regressions

for j=1:neqs;
yvec = y(nlag+1:nobs,j);
res = ols(yvec,xmat);
results(j).beta = res.beta; % bhats
results(j).tstat = res.tstat; % t-stats
% compute t-probs
tstat = res.tstat;
tout = tdis_prb(tstat,nobse-nvar);
results(j).tprob = tout; % t-probs
results(j).resid = res.resid; % resids
results(j).yhat = res.yhat; % yhats
results(j).y = yvec; % actual y
results(j).rsqr = res.rsqr; % r-squared
results(j).rbar = res.rbar; % r-adjusted
results(j).sige = res.sige; % sig estimate
The explanatory variables matrix, ‘xmat’ is the same for all equations of
the VAR model, so we form this matrix before entering the ‘for-loop’ over
equations. Structure arrays can be formed by simply specifying:
struct(i).fieldname, where i is the array index. This makes them just as
easy to use as the structure variables we presented in the discussion of the
regression function library in Chapter 2.
The most complicated part of the var function is implementation of the
Granger-causality tests which requires that we produce residuals based on
.::UdecomBooks::.
models that sequentially omit each variable in the model from the explana-
tory variables matrix in each equation. These residuals reflect the restricted
model in the joint F-test, whereas the VAR model residuals represent the
unrestricted model. The Granger-causality tests represent a series of tests
for the joint significance of each variable in each equation of the model. This
sequence of calculations is illustrated in the code below:
% form x matrices for joint F-tests

% exclude each variable from the model sequentially
for r=1:neqs;
xtmp = [];
for s=1:neqs;
if s ~= r
xlag = mlag(y(:,s),nlag);
xtmp = [xtmp trimr(xlag,nlag,0)];
end;
end; % end of for s-loop
% we have an xtmp matrix that excludes 1 variable
% add deterministic variables (if any) and constant term
if nx > 0
xtmp = [xtmp x(1:nobse,:) ones(nobse,1)];
else
xtmp = [xtmp ones(nobse,1)];
end;
% get ols residual vector
b = xtmp\yvec; % using Cholesky solution
etmp = yvec-xtmp*b;
sigr = etmp’*etmp;
% joint F-test for variables r
ftest(r,1) = ((sigr - sigu)/nlag)/(sigu/(nobse-k));
end; % end of for r-loop
results(j).ftest = ftest;
results(j).fprob = fdis_prb(ftest,nlag,nobse-k);
The loop over r=1:neqs and s=1:neqs builds up an explanatory vari-

ables matrix, xtmp by sequentially omitting each of the variables in the
model. This programming construct is often useful. We start with a blank
matrix xtmp = [] and then continue to add matrix columns to this blank
matrix during the loop. Each time through the loop a set of columns rep-
resenting lags for another variable in the model are added to the existing
matrix xtmp, until we have ‘built-up’ the desired matrix. If s == r we skip
adding lags for that variable to the matrix xtmp, to create the necessary
exclusion of 1 variable at a time.
Next we need to add deterministic variables (if they exist) and a constant
term to the explanatory variables matrix. Finally, we carry out least-squares
.::UdecomBooks::.
using the matrix xtmp with a Cholesky solution provided by the MATLAB
‘backslash’ operator. We take the Cholesky approach rather than the qr
matrix decomposition because a profile of the var function showed that
over 50% of the time spent in the routine was devoted to this step. In
contrast, only 12% of the time was spent determining the VAR regression
information using the ols command. Finally, note that the code above is
embedded in a loop over all equations in the model, so we store the ‘ftest’
and ‘fprob’ results in the structure for equation j.
Although the illustrations so far have not involved use of deterministic
variables in the VAR model, the var function is capable of handling these
variables. As an example, we could include a set of seasonal dummy variables
in the VAR model using:
% ----- Example 5.3 VAR with deterministic variables

dates = cal(1982,1,12); % monthly data starts in 82,1
load test.dat;
y = test; % use levels data
[nobs neqs] = size(test);
sdum = sdummy(nobs,dates.freq); % create seasonal dummies
sdum = trimc(sdum,1,0); % omit 1 column because we have
% a constant included by var()
dnames = strvcat(’dum1’,’dum2’,’dum3’,’dum4’,’dum5’,’dum6’,’dum7’, ...
’dum8’,’dum9’,’dum10’,’dum11’);
vnames = strvcat(vnames,dnames);
nlag = 12;
result = var(y,nlag,sdum);
prt(result,vnames);
A handy option on the prt var (and prt) function is the ability to print
the VAR model estimation results to an output file. Because these results
are quite large, they can be difficult to examine in the MATLAB command
window. Note that the wrapper function prt described in Chapter 3 also
works to print results from VAR model estimation, as does plt.
PURPOSE: Prints vector autoregression output

from: var,bvar,rvar,ecm,becm,recm models
---------------------------------------------------
USAGE: prt_var(result,vnames,fid)
where: results = a structure returned by:
var,bvar,rvar,ecm,becm,recm
vnames = optional vector of variable names
fid = file-id for printing results to a file
.::UdecomBooks::.
---------------------------------------------------
NOTE: - constant term is added automatically to vnames list
you need only enter VAR variable names plus deterministic
- you may use prt_var(results,[],fid) to print
output to a file with no vnames
---------------------------------------------------
In addition to the prt and prt var functions, there are plt and plt var
functions that produce graphs of the actual versus predicted and residuals
for these models.
One final issue associated with specifying a VAR model is the lag length
to employ. A commonly used approach to determining the lag length is to
perform statistical tests of models with longer lags versus shorter lag lengths.
We view the longer lag models as an unrestricted model versus the restricted
shorter lag version of the model, and construct a likelihood ratio statistic
to test for the significance of imposing the restrictions. If the restrictions
are associated with a statistically significant degradation in model fit, we
conclude that the longer lag length model is more appropriate, rejecting the
shorter lag model.
Specifically, the chi-squared distributed test statistic which has degrees
of freedom equation to the number of restrictions imposed is:
LR = (T − c)(log|Σr | − log|Σu |) (5.2)

where T is the number of observations, c is a degrees of freedom correction
factor proposed by Sims (1980), and |Σr |, |Σu | denote the determinant of
the error covariance matrices from the restricted and unrestricted models
respectively. The correction factor, c, recommended by Sims was the number
of variables in each unrestricted equation of the VAR model.
A function lrratio implements a sequence of such tests beginning at a
maximum lag (specified by the user) down to a minimum lag (also specified
by the user). The function prints results to the MATLAB command window
along with marginal probability levels. As an example, consider the following
program to determine the ‘statistically optimal’ lag length to use for our
VAR model involving the eight-state sample of monthly employment data
for the mining industry.
% ----- Example 5.4 Using the lrratio() function

load test.dat;
y = test; % use all eight states
maxlag = 12;
minlag = 3;
% Turn on flag for Sim’s correction factor
.::UdecomBooks::.
sims = 1;
disp(’LR-ratio results with Sims correction’);
lrratio(y,maxlag,minlag,sims);
The output from this program is:
LR-ratio results with Sims correction

nlag = 12 11, LR statistic = 75.6240, probability = 0.1517
nlag = 8 7, LR statistic = 125.7240, probability = 6.573e-06
There exists an option flag to use the degrees of freedom correction

suggested by Sims, whereas the default behavior of lrratio is to set c = 0.
Example 4.4 turns on the correction factor by setting a flag that we named
‘sims’ equal to 1. The results suggest that the lag length of 11 cannot be
viewed as significantly degrading the fit of the model relative to a lag of
12. For the comparison of lags 11 and 10, we find that at the 0.05 level,
we might reject lag 10 in favor of lag 11 as the optimal lag length. On the
other hand, if we employ a 0.01 level of significance, we would conclude that
the optimal lag length is 9, because the likelihood ratio tests reject lag 8 as
significantly degrading the fit of the model at the 0.01 level of confidence.
Impulse response functions represent a method used to examine the in-
teractions between variables in a VAR model. A shock to the system is im-
posed and the dynamic interactions of the variables based on the estimated
coefficients are simulated for all variables in the system. The function irf
contributed by Mike Cliff performs the calcuations and provides graphical
output of the impulse responses.
% ----- Example 5.5 Using the irf() function

load test.dat; % a test data set containing
% monthly mining employment for
% il,in,ky,mi,oh,pa,tn,wv
% data covers 1982,1 to 1996,5
neqs = 8; y = test(:,1:neqs); % use neqs states
nlag = 6; % number of lags in var-model
% estimate the model
results = var(y,nlag);
vnames = strvcat(’illinois’,’indiana’,’kentucky’,’michigan’,’ohio’, ...
.::UdecomBooks::.
nperiods = 48;
[m1 m2] = irf(results,nperiods,’o1’,vnames);
The results are provided graphically as eight graphs for the example
involving 8 equations, where a typical graph is shown in Figure 5.1. This
function puts up multiple MATLAB graph windows with a varying number
of the impulse response graphs in each window. This allows the user to
examine all of the responses and compare them using the multiple MATLAB
figure windows. The figure presents only a single graph window showing
three of the eight graphs produced by the example program.
Figure 5.1: Impulse response graphs
5.2 Error correction models

We provide a cursory introduction to co-integration and error correction
models and refer the reader to an excellent layman’s introduction by Dickey,
Jansen and Thornton (1991) as well as a more technical work by Johansen
.::UdecomBooks::.
(1995). LeSage (1990) and Shoesmith (1995) cover co-integration and EC

models in the context of forecasting.
Focusing on the practical case of I(1), (integrated of order 1) series,
let yt be a vector of n time-series that are I(1). An I(1) series requires one
difference to transform it to a zero mean, purely non-deterministic stationary
process. The vector yt is said to be co-integrated if there exists an n x r
matrix α such that:
zt = α 0 y t (5.3)
Engle and Granger (1987) provide a Representation Theorem stating
that if two or more series in yt are co-integrated, there exists an error cor-
rection representation taking the following form:
∆yt = A(`)∆yt + γzt−1 + εt (5.4)

where γ is a matrix of coefficients of dimension n x r of rank r, zt−1 is of
dimension r x 1 based on r ≤ n − 1 equilibrium error relationships, zt =
α0 yt from (5.3), and εt is a stationary multivariate disturbance. The error
correction (EC) model in (5.4) is simply a VAR model in first-differences
with r lagged error correction terms (zt−1 ) included in each equation of the
model. If we have deterministic components in yt , we add these terms as
well as the error correction variables to each equation of the model.
With the case of only two series yt and xt in the model, a two-step
procedure proposed by Engle and Granger (1987) can be used to determine
the co-integrating variable that we add to our VAR model in first-differences
to make it an EC model. The first-step involves a regression: yt = θ+αxt +zt
to determine estimates of α and zt . The second step carries out tests on zt to
determine if it is stationary, I(0). If we find this to be the case, the condition
yt = θ + αxt is interpreted as the equilibrium relationship between the two
series and the error correction model is estimated as:
∆yt = −γ1 zt−1 + lagged(∆xt , ∆yt ) + c1 + ε1t

∆xt = −γ2 zt−1 + lagged(∆xt , ∆yt ) + c2 + ε2t
where: zt−1 = yt−1 − θ − αxt−1 , ci are constant terms and εit denote distur-
bances in the model.
We provide a function adf, (augmented Dickey-Fuller) to test time-series
for the I(1), I(0) property, and another routine cadf (co-integrating aug-
mented Dickey-Fuller) to carry out the tests from step two above on zt to
.::UdecomBooks::.
determine if it is stationary, I(0). These routines as well as the function

johansen that implements a multivariate extension of the two-step Engle
and Granger procedure were designed to mimic a set of Gauss functions by
Sam Quilaris named coint.
The adf function documentation is:
PURPOSE: carry out DF tests on a time series vector
---------------------------------------------------
USAGE: results = adf(x,p,nlag)
where: x = a time-series vector
p = order of time polynomial in the null-hypothesis
p = -1, no deterministic part
p = 0, for constant term
p = 1, for constant plus time-trend
p > 1, for higher order polynomial
nlags = # of lagged changes of x included
---------------------------------------------------
RETURNS: a results structure
results.meth = ’adf’
results.alpha = estimate of the autoregressive parameter
results.adf = ADF t-statistic
results.crit = (6 x 1) vector of critical values
[1% 5% 10% 90% 95% 99%] quintiles
results.nlag = nlag
This would be used to test a time-series vector for I(1) or I(0) status.
Allowance is made for polynomial time trends as well as constant terms in
the function and a set of critical values are returned in a structure by the
function. A function prt coint (as well as prt) can be used to print output
from adf, cadf and johansen, saving users the work of formatting and
printing the result structure output.
The function cadf is used for the case of two variables, yt , xt , where we
wish to test whether the condition yt = αxt can be interpreted as an equi-
librium relationship between the two series. The function documentation
is:
PURPOSE: compute augmented Dickey-Fuller statistic for residuals
from a cointegrating regression, allowing for deterministic
polynomial trends
------------------------------------------------------------
USAGE: results = cadf(y,x,p,nlag)
where: y = dependent variable time-series vector
x = explanatory variables matrix
.::UdecomBooks::.

nlag = # of lagged changes of the residuals to include in regression
------------------------------------------------------------
RETURNS: results structure
results.meth = ’cadf’
results.alpha = autoregressive parameter estimate
results.adf = ADF t-statistic
results.crit = (6 x 1) vector of critical values
[1% 5% 10% 90% 95% 99%] quintiles
results.nvar = cols(x)
results.nlag = nlag
As an illustration of using these two functions, consider testing our two

monthly time-series on mining employment in Illinois and Indiana for I(1)
status and then carrying out a test of whether they exhibit an equilibrating
relationship. The program would look as follows:
% ----- Example 5.6 Using the adf() and cadf() functions
dates = cal(1982,1,12);
load test.dat;
y = test(:,1:2); % use only two series
vnames = strvcat(’illinois’,’indiana’);
% test Illinois for I(1) status
nlags = 6;
for i=1:nlags;
res = adf(y(:,1),0,i);
prt(res,vnames(1,:));
end;
% test Indiana for I(1) status
nlags = 6;
for i=1:nlags;
res = adf(y(:,2),0,i);
prt(res,vnames(2,:));
end;
% test if Illinois and Indiana are co-integrated
for i=1:nlags;
res = cadf(y(:,1),y(:,2),0,i);
prt(res,vnames);
end;
The program sets a lag length of 6, and loops over lags of 1 to 6 to

provide some feel for how the augmented Dickey-Fuller tests are affected
by the number of lags used. We specify p = 0 because the employment
time-series do not have zero mean, so we wish to include a constant term.
The result structures returned by the adf and cadf functions are passed on
to prt for printing. We present the output for only lag 6 to conserve on
.::UdecomBooks::.
space, but all lags produced the same inferences. One point to note is that
the adf and cadf functions return a set of 6 critical values for significance
levels 1%,5%,10%,90%,95%,99% as indicated in the documentation for these
functions. Only three are printed for purposes of clarity, but all are available
in the results structure returned by the functions.
Augmented DF test for unit root variable: illinois
ADF t-statistic # of lags AR(1) estimate
-0.164599 6 0.998867
1% Crit Value 5% Crit Value 10% Crit Value
-3.464 -2.912 -2.588
Augmented DF test for unit root variable: indiana

ADF t-statistic # of lags AR(1) estimate
-0.978913 6 0.987766
-3.464 -2.912 -2.588
Augmented DF test for co-integration variables: illinois,indiana

CADF t-statistic # of lags AR(1) estimate
-1.67691570 6 -0.062974
-4.025 -3.404 -3.089
We see from the adf function results that both Illinois and Indiana are
I(1) variables. We reject the augmented Dickey-Fuller hypothesis of I(0)
because our t-statistics for both Illinois and Indiana are less than (in absolute
value terms) the critical value of -2.588 at the 90% level.
From the results of cadf we find that Illinois and Indiana mining employ-
ment are not co-integrated, again because the t-statistic of -1.67 does not
exceed the 90% critical value of -3.08 (in absolute value terms). We would
conclude that an EC model is not appropriate for these two time-series.
For most EC models, more than two variables are involved so the Engle
and Granger two-step procedure needs to be generalized. Johansen (1988)
provides this generalization which takes the form of a likelihood-ratio test.
We implement this test in the function johansen. The Johansen procedure
provides a test statistic for determining r, the number of co-integrating
relationships between the n variables in yt as well as a set of r co-integrating
vectors that can be used to construct error correction variables for the EC
model.
As a brief motivation for the work carried out by the johansen function,
we start with a reparameterization of the EC model:
∆yt = Γ1 ∆yt−1 + . . . + Γk−1 ∆yt−k+1 − Ψyt−k + εt (5.5)
.::UdecomBooks::.
where Ψ = (In − A1 − A2 − . . . − Ak ). If the matrix Ψ contains all zeros, (has

rank=0), there are no co-integrating relationships between the variables in
yt . If Ψ is of full-rank, then we have n long-run equilibrium relationships,
so all variables in the model are co-integrated. For cases where the matrix
Ψ has rank r < n, we have r co-integrating relationships. The Johansen
procedure provides two tests for the number of linearly independent co-
integrating relationships among the series in yt , which we have labeled r in
our discussion. Both tests are based on an eigenvalue-eigenvector decom-
position of the matrix Ψ, constructed from canonical correlations between
∆yt and yt−k with adjustments for intervening lags, and taking into account
that the test is based on estimates that are stochastic. The test statistics
are labeled the ‘trace statistic’ and the ‘maximal eigenvalue statistic’.
Given the value of r, (the number of co-integrating relationships), we
can use the eigenvectors provided by the johansen function along with the
levels of yt lagged one period to form a set of error correction variables for
our EC model. In practice, the function ecm does this for you, so you need
not worry about the details.
The documentation for johansen is shown below. A few points to note.
The published literature contains critical values for the trace statistic for
VAR models with up to 12 variables in Johansen (1995), and for the maximal
eigenvalue statistic, Johansen and Juselius (1988) present critical values for
VAR models containing up to 5 variables. To extend the number of variables
for which critical values are available, a procedure by MacKinnon (1996) was
used to generate critical values for both the trace and maximal eigenvalue
statistics for models with up to 12 variables. MacKinnon’s method is an
approximation, but it produces values close to those in Johansen (1995). The
critical values for the trace statistic have been entered in a function c sjt
and those for the maximal eigenvalue statistic are in c sja. The function
johansen calls these two functions to obtain the necessary critical values. In
cases where the VAR model has more than 12 variables, zeros are returned
as critical values in the structure field ‘result.cvt’ for the trace statistic and
the ‘result.cvm’ field for the maximal eigenvalue.
Another less serious limitation is that the critical values for these statis-
tics are only available for trend transformations where −1 ≤ p ≤ 1. This
should not present a problem in most applications where p will take on
values of -1, 0 or 1.
PURPOSE: perform Johansen cointegration tests
-------------------------------------------------------
USAGE: result = johansen(x,p,k)
where: x = input matrix of time-series in levels, (nobs x m)
.::UdecomBooks::.

k = number of lagged difference terms used when
computing the estimator
-------------------------------------------------------
RETURNS: a results structure:
result.eig = eigenvalues (m x 1)
result.evec = eigenvectors (m x m), where first
r columns are normalized coint vectors
result.lr1 = likelihood ratio trace statistic for r=0 to m-1
(m x 1) vector
result.lr2 = maximum eigenvalue statistic for r=0 to m-1
(m x 1) vector
result.cvt = critical values for trace statistic
(m x 3) vector [90% 95% 99%]
result.cvm = critical values for max eigen value statistic
(m x 3) vector [90% 95% 99%]
result.ind = index of co-integrating variables ordered by
size of the eigenvalues from large to small
-------------------------------------------------------
NOTE: c_sja(), c_sjt() provide critical values generated using
a method of MacKinnon (1994, 1996).
critical values are available for n<=12 and -1 <= p <= 1,
zeros are returned for other cases.
As an illustration of the johansen function, consider the eight-state

sample of monthly mining employment. We would test for the number of
co-integrating relationships using the following code:
% ----- Example 5.7 Using the johansen() function

y = load(’test.dat’); % use all eight states
nlag = 9;
pterm = 0;
result = johansen(y,pterm,nlag);
prt(result,vnames);
The johansen function is called with the y matrix of time-series vari-

ables for the eight states, a value of p = 0 indicating we have a constant
term in the model, and 9 lags. (We want p = 0 because the constant term
is necessary where the levels of employment in the states differ.) The lag of
9 was determined to be optimal using the lrratio function in the previous
section.
.::UdecomBooks::.
The johansen function will return results for a sequence of tests against
alternative numbers of co-integrating relationships ranging from r ≤ 0 up
to r ≤ m − 1, where m is the number of variables in the matrix y.
The function prt provides a printout of the trace and maximal eigenvalue
statistics as well as the critical values returned in the johansen results
structure.
Johansen MLE estimates

NULL: Trace Statistic Crit 90% Crit 95% Crit 99%
r <= 0 illinos 307.689 153.634 159.529 171.090
r <= 1 indiana 205.384 120.367 125.618 135.982
r <= 2 kentucky 129.133 91.109 95.754 104.964
r <= 3 ohio 83.310 65.820 69.819 77.820
r <= 4 pennsylvania 52.520 44.493 47.855 54.681
r <= 5 tennessee 30.200 27.067 29.796 35.463
r <= 6 west virginia 13.842 13.429 15.494 19.935
r <= 7 michigan 0.412 2.705 3.841 6.635
NULL: Eigen Statistic Crit 90% Crit 95% Crit 99%

r <= 0 illinos 102.305 49.285 52.362 58.663
r <= 1 indiana 76.251 43.295 46.230 52.307
r <= 2 kentucky 45.823 37.279 40.076 45.866
r <= 3 ohio 30.791 31.238 33.878 39.369
r <= 4 pennsylvania 22.319 25.124 27.586 32.717
r <= 5 tennessee 16.359 18.893 21.131 25.865
r <= 6 west virginia 13.430 12.297 14.264 18.520
r <= 7 michigan 0.412 2.705 3.841 6.635
The printout does not present the eigenvalues and eigenvectors, but they
are available in the results structure returned by johansen as they are
needed to form the co-integrating variables for the EC model. The focus of
co-integration testing would be the trace and maximal eigenvalue statistics
along with the critical values. For this example, we find: (using the 95%
level of significance) the trace statistic rejects r ≤ 0 because the statistic of
307.689 is greater than the critical value of 159.529; it also rejects r ≤ 1,
r ≤ 2, r ≤ 3, r ≤ 4, and r ≤ 5 because these trace statistics exceed the
associated critical values; for r ≤ 6 we cannot reject H0, so we conclude that
r = 6. Note that using the 99% level, we would conclude r = 4 as the trace
statistic of 52.520 associated with r ≤ 4 does not exceed the 99% critical
value of 54.681.
We find a different inference using the maximal eigenvalue statistic. This
statistic allows us to reject r ≤ 0 as well as r ≤ 1 and r ≤ 2 at the 95% level.
We cannot reject r ≤ 3, because the maximal eigenvalue statistic of 30.791
does not exceed the critical value of 33.878 associated with the 95% level.
.::UdecomBooks::.
This would lead to the inference that r = 3, in contrast to r = 6 indicated

by the trace statistic. Using similar reasoning at the 99% level, we would
infer r = 2 from the maximal eigenvalue statistics.
After the johansen test determines the number of co-integrating rela-
tionships, we can use these results along with the eigenvectors returned by
the johansen function, to form a set of error correction variables. These are
constructed using yt−1 , (the levels of y lagged one period) multiplied by the
r eigenvectors associated with the co-integrating relationships to form r co-
integrating variables. This is carried out by the ecm function, documented
below.
PURPOSE: performs error correction model estimation

---------------------------------------------------
USAGE: result = ecm(y,nlag,r)
where: y = an (nobs x neqs) matrix of y-vectors in levels
r = # of cointegrating relations to use
(optional: this will be determined using
Johansen’s trace test at 95%-level if left blank)
NOTES: constant vector automatically included
x-matrix of exogenous variables not allowed
error correction variables are automatically
constructed using output from Johansen’s ML-estimator
---------------------------------------------------
RETURNS a structure
results.meth = ’ecm’
results.nvar = nlag*neqs+nx+1, # of variables per equation
results.coint= # of co-integrating relations (or r if input)
results.index= index of co-integrating variables ranked by
size of eigenvalues large to small
results(eq).beta = bhat for equation eq (includes ec-bhats)
results(eq).yhat = predicted values (levels) (nlag+2:nobs,1)
results(eq).dyhat = predicted values (differenced) (nlag+2:nobs,1)
results(eq).y = actual y-level values (nobs x 1)
results(eq).dy = actual y-differenced values (nlag+2:nobs,1)
results(eq).ftest = Granger F-tests
results(eq).fprob = Granger marginal probabilities
.::UdecomBooks::.
---------------------------------------------------
The ecm function allows two options for implementing an EC model.

One option is to specify the number of co-integrating relations to use, and the
other is to let the ecm function determine this number using the johansen
function and the trace statistics along with the critical values at the 95%
level of significance. The motivation for using the trace statistic is that it
seems better suited to the task of sequential hypotheses for our particular
decision problem. If you find this decision problematical, you can specify r
as an input to the ecm function, or change the code presented below to rely
on the maximal eigenvalue statistic.
To illustrate how this is carried out by the ecm function consider the
following code from the function.
if nargin == 3 % user is specifying the # of error correction terms to
% include -- get them using johansen()
jres = johansen(y,0,nlag);
% recover error correction vectors
ecvectors = jres.evec;
index = jres.ind; % recover index of ec variables
% construct r-error correction variables
x = mlag(y(:,index),1)*ecvectors(:,1:r);
[nobs2 nx] = size(x);
elseif nargin == 2 % we need to find r

jres = johansen(y,0,nlag);
% find r = # significant co-integrating relations using
% the trace statistic output
trstat = jres.lr1; % max eigenvalue stats are in jres.lr2
tsignf = jres.cvt; % max eigenvalue crit values are in jres.cvm
r = 0;
for i=1:neqs;
if trstat(i,1) > tsignf(i,2)
r = i;
end;
end;
% recover error correction vectors
ecvectors = jres.evec;
index = jres.ind; % recover index of ec variables
% construct r error correction variables
x = mlag(y(:,index),1)*ecvectors(:,1:r);
[junk nx] = size(x);
else
error(’Wrong # of arguments to ecm’);
end;
Based on the number of input arguments, we know if the user is specifying
.::UdecomBooks::.
the number of co-integrating relations (the case where nargin == 3) or

wishes us to determine this, (the case where nargin == 2). The function
calls johansen which computes eigenvalues and eigenvectors along with the
trace and maximal eigenvalue statistics. In the case where the user specifies
the number of co-integrating relations as an input argument r, we simply
use this number of eigenvectors multiplied times the levels lagged 1 period
to form the r error correction variables. One implementation detail is that
the variables involved in the co-integrating relationships are those associated
with the largest eigenvalues. The johansen function sorts the eigenvalues
and eigenvectors by size and returns an index vector that points to the
variables involved in the co-integrating relationships. We use this index to
arrange the levels variables in y in a similar order as the sorted eigenvectors
before carrying out the multiplication to produce error correction variables.
For the case where the user leaves determination of r to the ecm function,
we loop over the trace statistics and find the value of r using the 95%
significance level. This value is then used to form r error correction variables.
From this point, we need simply transform the y matrix to first differ-
ences and call our var estimation routine using y transformed to first dif-
ferences along with the r error correction variables plus a constant entered
as deterministic variables in the call to var.
An identical approach can be taken to implement a Bayesian variant
of the EC model based on the Minnesota prior as well as a more recent
Bayesian variant based on a “random-walk averaging prior”. Both of these
are discussed in the next section. To summarize, we can implement EC
models using Bayesian priors as well as the usual case with no priors by
simply calling johansen to determine the eigenvectors and number of co-
integrating relationships if the user request this. We then transform the
data to first-difference form and call the associated vector autoregressive
estimation procedure using the differenced data and the error correction
variables along with a constant term.
The prt var function will produce a printout of the results structure
returned by ecm showing the autoregressive plus error correction variable
coefficients along with Granger-causality test results as well as the trace,
maximal eigenvalue statistics, and critical values from the johansen proce-
dure. As an example, we show a program to estimate an EC model based
on our eight-state sample of monthly mining employment, where we have
set the lag-length to 2 to conserve on the amount of printed output.
% ----- Example 5.8 Estimating error correction models

y = load(‘test.dat’); % monthly mining employment for
.::UdecomBooks::.
% il,in,ky,mi,oh,pa,tn,wv 1982,1 to 1996,5

vnames = strvcat(’il’,’in’,’ky’,’mi’,’oh’,’pa’,’tn’,’wv’);
% estimate the model, letting ecm determine # of co-integrating vectors
result = ecm(y,nlag);
prt(result,vnames); % print results to the command window
The printed output is shown below for a single state indicating the pres-
ence of two co-integrating relationships involving the states of Illinois and
Indiana. The estimates for the error correction variables are labeled as such
in the printout. Granger causality tests are printed, and these would form
the basis for valid causality inferences in the case where co-integrating rela-
tionships existed among the variables in the VAR model.
Dependent Variable = wv
R-squared = 0.1975
sige = 341.6896
******************************************************************
il lag1 0.141055 0.261353 0.794176
il lag2 0.234429 0.445400 0.656669
in lag1 1.630666 1.517740 0.131171
in lag2 -1.647557 -1.455714 0.147548
ky lag1 0.378668 1.350430 0.178899
ky lag2 0.176312 0.631297 0.528801
mi lag1 0.053280 0.142198 0.887113
mi lag2 0.273078 0.725186 0.469460
oh lag1 -0.810631 -1.449055 0.149396
oh lag2 0.464429 0.882730 0.378785
pa lag1 -0.597630 -2.158357 0.032480
pa lag2 -0.011435 -0.038014 0.969727
tn lag1 -0.049296 -0.045237 0.963978
tn lag2 0.666889 0.618039 0.537480
wv lag1 -0.004150 -0.033183 0.973572
wv lag2 -0.112727 -0.921061 0.358488
ec term il -2.158992 -1.522859 0.129886
ec term in -2.311267 -1.630267 0.105129
constant 8.312788 0.450423 0.653052
il 0.115699 0.890822
in 2.700028 0.070449
ky 0.725708 0.485662
mi 0.242540 0.784938
oh 1.436085 0.241087
pa 2.042959 0.133213
.::UdecomBooks::.
tn 0.584267 0.558769
wv 1.465858 0.234146
Johansen MLE estimates
NULL: Trace Statistic Crit 90% Crit 95% Crit 99%
r <= 0 il 214.390 153.634 159.529 171.090
r <= 1 in 141.482 120.367 125.618 135.982
r <= 2 ky 90.363 91.109 95.754 104.964
r <= 3 oh 61.555 65.820 69.819 77.820
r <= 4 tn 37.103 44.493 47.855 54.681
r <= 5 wv 21.070 27.067 29.796 35.463
r <= 6 pa 10.605 13.429 15.494 19.935
r <= 7 mi 3.192 2.705 3.841 6.635
NULL: Eigen Statistic Crit 90% Crit 95% Crit 99%
r <= 0 il 72.908 49.285 52.362 58.663
r <= 1 in 51.118 43.295 46.230 52.307
r <= 2 ky 28.808 37.279 40.076 45.866
r <= 3 oh 24.452 31.238 33.878 39.369
r <= 4 tn 16.034 25.124 27.586 32.717
r <= 5 wv 10.465 18.893 21.131 25.865
r <= 6 pa 7.413 12.297 14.264 18.520
r <= 7 mi 3.192 2.705 3.841 6.635
The results indicate that given the two lag model, two co-integrating
relationships were found leading to the inclusion of two error correction
variables in the model. The co-integrating relationships are based on the
trace statistics compared to the critical values at the 95% level. From the
trace statistics in the printed output we see that, H0: r ≤ 2 was rejected
at the 95% level because the trace statistic of 90.363 is less than the asso-
ciated critical value of 95.754. Keep in mind that the user has the option
of specifying the number of co-integrating relations to be used in the ecm
function as an optional argument. If you wish to work at the 90% level of
significance, we would conclude from the johansen results that r = 4 co-
integrating relationships exist. To estimate an ecm model based on r = 4
we need simply call the ecm function with:
% estimate the model, using 4 co-integrating vectors

result = ecm(y,nlag,4);
5.3 Bayesian variants

Despite the attractiveness of drawing on cross-sectional information from
related economic variables, the VAR model has empirical limitations. For
example, a model with eight variables and six lags produces 49 independent
variables in each of the eight equations of the model for a total of 392
.::UdecomBooks::.
coefficients to estimate. Large samples of observations involving time series

variables that cover many years are needed to estimate the VAR model,
and these are not always available. In addition, the independent variables
represent lagged values, e.g., y1t−1 , y1t−2 , . . . , y1t−6 , which tend to produce
high correlations that lead to degraded precision in the parameter estimates.
To overcome these problems, Doan, Litterman and Sims (1984) proposed the
use of Bayesian prior information. The Minnesota prior means and variances
suggested take the following form:
βi ∼ N (1, σβ2i )
βj ∼ N (0, σβ2j ) (5.6)
where βi denotes the coefficients associated with the lagged dependent vari-
able in each equation of the VAR and βj represents any other coefficient.
The prior means for lagged dependent variables are set to unity in belief
that these are important explanatory variables. On the other hand, a prior
mean of zero is assigned to all other coefficients in the equation, βj in (5.6),
indicating that these variables are viewed as less important in the model.
The prior variances, σβ2i , specify uncertainty about the prior means β̄i =
1, and σβ2j indicates uncertainty regarding the means β̄j = 0. Because
the VAR model contains a large number of parameters, Doan, Litterman
and Sims (1984) suggested a formula to generate the standard deviations
as a function of a small number of hyperparameters: θ, φ and a weighting
matrix w(i, j). This approach allows a practitioner to specify individual
prior variances for a large number of coefficients in the model using only a
few parameters that are labeled hyperparameters. The specification of the
standard deviation of the prior imposed on variable j in equation i at lag k
is:
µ ¶
σ̂uj
σijk = θw(i, j)k −φ (5.7)
σ̂ui
where σ̂ui is the estimated standard error from a univariate autoregression
involving variable i, so that (σ̂uj /σ̂ui ) is a scaling factor that adjusts for vary-
ing magnitudes of the variables across equations i and j. Doan, Litterman
and Sims (1984) labeled the parameter θ as ‘overall tightness’, reflecting the
standard deviation of the prior on the first lag of the dependent variable.
The term k −φ is a lag decay function with 0 ≤ φ ≤ 1 reflecting the decay
rate, a shrinkage of the standard deviation with increasing lag length. This
has the effect of imposing the prior means of zero more tightly as the lag
.::UdecomBooks::.
length increases, based on the belief that more distant lags represent less
important variables in the model. The function w(i, j) specifies the tight-
ness of the prior for variable j in equation i relative to the tightness of the
own-lags of variable i in equation i.
The overall tightness and lag decay hyperparameters used in the stan-
dard Minnesota prior have values θ = 0.1, φ = 1.0. The weighting matrix
used is:
 
1 0.5 . . . 0.5

 0.5 1 0.5 

W=
 ..  (5.8)
. . 
.. ..
 . 
0.5 0.5 . . . 1
This weighting matrix imposes β̄i = 1 loosely, because the lagged de-
pendent variable in each equation is felt to be an important variable. The
weighting matrix also imposes the prior mean of zero for coefficients on
other variables in each equation more tightly since the βj coefficients are
associated with variables considered less important in the model.
A function bvar will provide estimates for this model. The function
documentation is:
PURPOSE: Performs a Bayesian vector autoregression of order n

---------------------------------------------------
USAGE: result = bvar(y,nlag,tight,weight,decay,x)
tight = Litterman’s tightness hyperparameter
weight = Litterman’s weight (matrix or scalar)
decay = Litterman’s lag decay = lag^(-decay)
x = an optional (nobs x nx) matrix of variables
---------------------------------------------------
results.meth = ’bvar’
results.nvar = nlag*neqs+1+nx, # of variables per equation
results.tight = overall tightness hyperparameter
results.weight = weight scalar or matrix hyperparameter
results.decay = lag decay hyperparameter
results(eq).beta = bhat for equation eq
.::UdecomBooks::.
results(eq).yhat = predicted values
results(eq).y = actual values
---------------------------------------------------
SEE ALSO: bvarf, var, ecm, rvar, plt_var, prt_var
---------------------------------------------------
The function bvar allows us to input a scalar weight value or a more

general matrix. Scalar inputs will be used to form a symmetric prior, where
the scalar is used on the off-diagonal elements of the matrix. A matrix will
be used in the form submitted to the function.
As an example of using the bvar function, consider our case of monthly
mining employment for eight states. A program to estimate a BVAR model
based on the Minnesota prior is shown below:
% ----- Example 5.9 Estimating BVAR models

nlag = 2;
tight = 0.1; % hyperparameter values
weight = 0.5;
decay = 1.0;
result = bvar(y,nlag,tight,weight,decay);
prt(result,vnames);
The printout shows the hyperparameter values associated with the prior.
It does not provide Granger-causality test results as these are invalid given
the Bayesian prior applied to the model. Results for a single equation of the
mining employment example are shown below.
***** Bayesian Vector Autoregressive Model *****

***** Minnesota type Prior *****
PRIOR hyperparameters
tightness = 0.10
decay = 1.00
Symmetric weights based on 0.50
Dependent Variable = il
R-squared = 0.9942
sige = 12.8634
.::UdecomBooks::.
******************************************************************
il lag1 1.134855 11.535932 0.000000
il lag2 -0.161258 -1.677089 0.095363
in lag1 0.390429 1.880834 0.061705
in lag2 -0.503872 -2.596937 0.010230
ky lag1 0.049429 0.898347 0.370271
ky lag2 -0.026436 -0.515639 0.606776
mi lag1 -0.037327 -0.497504 0.619476
mi lag2 -0.026391 -0.377058 0.706601
oh lag1 -0.159669 -1.673863 0.095996
oh lag2 0.191425 2.063498 0.040585
pa lag1 0.179610 3.524719 0.000545
pa lag2 -0.122678 -2.520538 0.012639
tn lag1 0.156344 0.773333 0.440399
tn lag2 -0.288358 -1.437796 0.152330
wv lag1 -0.046808 -2.072769 0.039703
wv lag2 0.014753 0.681126 0.496719
constant 9.454700 2.275103 0.024149
There exists a number of attempts to alter the fact that the Minnesota
prior treats all variables in the VAR model except the lagged dependent
variable in an identical fashion. Some of the modifications suggested have
focused entirely on alternative specifications for the prior variance. Usually,
this involves a different (non-symmetric) weight matrix W and a larger value
of 0.2 for the overall tightness hyperparameter θ in place of the value θ = 0.1
used in the Minnesota prior. The larger overall tightness hyperparameter
setting allows for more influence from other variables in the model. For
example, LeSage and Pan (1995) constructed a weight matrix based on
first-order spatial contiguity to emphasize variables from neighboring states
in a multi-state agricultural output forecasting model. LeSage and Magura
(1991) employed interindustry input-output weights to place more emphasis
on related industries in a multi-industry employment forecasting model.
These approaches can be implemented using the bvar function by con-
structing an appropriate weight matrix. For example, the first order conti-
guity structure for the eight states in our mining employment example can
be converted to a set of prior weights by placing values of unity on the main
diagonal of the weight matrix, and in positions that represent contiguous
entities. An example is shown in (5.9), where row 1 of the weight matrix
is associated with the time-series for the state of Illinois. We place a value
of unity on the main diagonal to indicate that autoregressive values from
Illinois are considered important variables. We also place values of one in
columns 2 and 3, reflecting the fact that Indiana (variable 2) and Kentucky
(variable 3) are states that have borders touching Illinois. For other states
.::UdecomBooks::.
that are not neighbors to Illinois, we use a weight of 0.1 to downweight

their influence in the BVAR model equation for Illinois. A similar scheme
is used to specify weights for the other seven states based on neighbors and
non-neighbors.
 
1.0 1.0 1.0 0.1 0.1 0.1 0.1 0.1
 
 1.0 1.0 1.0 1.0 1.0 0.1 0.1 0.1 
 
 1.0 1.0 1.0 0.1 1.0 0.1 1.0 1.0 
 
 0.1 1.0 0.1 1.0 1.0 0.1 0.1 0.1 
W =


 (5.9)
 0.1 1.0 1.0 1.0 1.0 1.0 0.1 1.0 
 
 0.1 0.1 0.1 0.1 1.0 1.0 0.1 1.0 
 
 0.1 0.1 1.0 0.1 0.1 0.1 1.0 0.1 
0.1 0.1 1.0 0.1 1.0 1.0 0.1 1.0
The intuition behind this set of weights is that we really don’t believe
the prior means of zero placed on the coefficients for mining employment
in neighboring states. Rather, we believe these variables should exert an
important influence. To express our lack of faith in these prior means, we
assign a large prior variance to the zero prior means for these states by
increasing the weight values. This allows the coefficients for these time-
series variables to be determined by placing more emphasis on the sample
data and less emphasis on the prior.
This could of course be implemented using bvar with a weight matrix
specified, e.g.,
% ----- Example 5.10 Using bvar() with general weights

dates = cal(1982,1,12);
nlag = 2;
tight = 0.1;
decay = 1.0;
w = [1.0 1.0 1.0 0.1 0.1 0.1 0.1 0.1

1.0 1.0 1.0 1.0 1.0 0.1 0.1 0.1
1.0 1.0 1.0 0.1 1.0 0.1 1.0 1.0
0.1 1.0 0.1 1.0 1.0 0.1 0.1 0.1
0.1 1.0 1.0 1.0 1.0 1.0 0.1 1.0
0.1 0.1 0.1 0.1 1.0 1.0 0.1 1.0
0.1 0.1 1.0 0.1 0.1 0.1 1.0 0.1
0.1 0.1 1.0 0.1 1.0 1.0 0.1 1.0];
result = bvar(y,nlag,tight,w,decay);
prt(result,vnames);
.::UdecomBooks::.
Another more recent approach to altering the equal treatment character

of the Minnesota prior is a “random-walk averaging prior” suggested by
LeSage and Krivelyova (1999).
As noted above, previous attempts to alter the fact that the Minnesota
prior treats all variables in the VAR model except the first lag of the de-
pendent variable in an identical fashion have focused entirely on alternative
specifications for the prior variance. The prior proposed by LeSage and
Krivelyova (1999) involves both prior means and variances motivated by the
distinction between important and unimportant variables in each equation
of the VAR model. To motivate the prior means, consider the weighting
matrix for a five variable VAR model shown in (5.10). The weight matrix
contains values of unity in positions associated with important variables in
each equation of the VAR model and values of zero for unimportant vari-
ables. For example, the important variables in the first equation of the
VAR model are variables 2 and 3 whereas the important variables in the
fifth equation are variables 4 and 5.
Note that if we do not believe that autoregressive influences reflected by
lagged values of the dependent variable are important, we have a zero on
the main diagonal of the weight matrix. In fact, the weighting matrix shown
in (5.10) classifies autoregressive influences as important in only two of the
five equations in the VAR system, equations three and five. As an example
of a case where autoregressive influences are totally ignored LeSage and
Krivelyova (1999) constructed a VAR system based on spatial contiguity
that relies entirely on the influence of neighboring states and ignores the
autoregressive influence associated with past values of the variables from
the states themselves.
 
0 1 1 0 0
 
 1 0 1 0 0 
 
W = 1 1 1 0 0  (5.10)
 
 0 0 1 0 1 
0 0 0 1 1
The weight matrix shown in (5.10) is standardized to produce row-sums
of unity resulting in the matrix labeled C shown in (5.11).
 
0 0.5 0.5 0 0
 
 0.5 0 0.5 0 0 
 
C= 0.33 0.33 0.33 0 0  (5.11)
 
 0 0 0.5 0 0.5 
0 0 0 0.5 0.5
.::UdecomBooks::.
Using the row-standardized matrix C, we consider the random-walk with

drift that averages over the important variables in each equation i of the
VAR model as shown in (5.12).
X
n
yit = αi + Cij yjt−1 + uit (5.12)
j=1
Expanding expression (5.12) we see that multiplying yjt−1 , j = 1, . . . , 5

containing 5 variables at time t−1 by the row-standardized weight matrix C
shown in (5.11) produces a set of explanatory variables for each equation of
the VAR system equal to the mean of observations from important variables
in each equation at time t − 1 as shown in (5.13).
       
y1t α1 0.5y2t−1 + 0.5y3t−1 u1t
 y2t   α2   0.5y1t−1 + 0.5y3t−1   u2t 
       
 y3t  =  α3 + 0.33y1t−1 + 0.33y2t−1 + 0.33y3t−1 + u3t (5.13)
 y4t   α4   0.5y3t−1 + 0.5y5t−1   u4t 
y5t α5 0.5y4t−1 + 0.5y5t−1 u5t
This suggests a prior mean for the VAR model coefficients on variables
associated with the first own-lag of important variables equal to 1/ci , where
ci is the number of important variables in each equation i of the model. In
the example shown in (5.13), the prior means for the first own-lag of the
important variables y2t−1 and y3t−1 in the y1t equation of the VAR would
equal 0.5. The prior means for unimportant variables, y1t−1 , y4t−1 and y5t−1
in this equation would be zero.
This prior is quite different from the Minnesota prior in that it may
downweight the lagged dependent variable using a zero prior mean to dis-
count the autoregressive influence of past values of this variable. In contrast,
the Minnesota prior emphasizes a random-walk with drift model that relies
on prior means centered on a model: yit = αi + yit−1 + uit , where the in-
tercept term reflects drift in the random-walk model and is estimated using
a diffuse prior. The random-walk averaging prior is centered on a random-
walk model that averages over important variables in each equation of the
model and allows for drift as well. As in the case of the Minnesota prior,
the drift parameters αi are estimated using a diffuse prior.
Consistent with the Minnesota prior, LeSage and Krivelyova use zero
as a prior mean for coefficients on all lags other than first lags. Litterman
(1986) motivates reliance on zero prior means for many of the parameters
of the VAR model by appealing to ridge regression. Recall, ridge regression
can be interpreted as a Bayesian model that specifies prior means of zero
.::UdecomBooks::.
for all coefficients, and as we saw in Chapter 4 can be used to overcome

collinearity problems in regression models.
One point to note about the random walk averaging approach to speci-
fying prior means is that the time series for the variables in the model need
to be scaled or transformed to have similar magnitudes. If this is not the
case, it would make little sense to indicate that the value of a time series
observation at time t was equal to the average of values from important
related variables at time t − 1. This should present no problem as time se-
ries data can always be expressed in percentage change form or annualized
growth rates which meets our requirement that the time series have similar
magnitudes.
The prior variances LeSage and Krivelyova specify for the parameters in
the model differ according to whether the coefficients are associated with
variables that are classified as important or unimportant as well as the lag
length. Like the Minnesota prior, they impose lag decay to reflect a prior
belief that time series observations from the more distant past exert a smaller
influence on the current value of the time series we are modeling. Viewing
variables in the model as important versus unimportant suggests that the
prior variance (uncertainty) specification should reflect the following ideas:
1. Parameters associated with unimportant variables should be assigned
a smaller prior variance, so the zero prior means are imposed more
‘tightly’ or with more certainty.
2. First own-lags of important variables are given a smaller prior variance,
so the prior means force averaging over the first own-lags of important
variables.
3. Parameters associated with unimportant variables at lags greater than
one will be given a prior variance that becomes smaller as the lag length
increases to reflect our belief that influence decays with time.
4. Parameters associated with lags other than first own-lag of important
variables will have a larger prior variance, so the prior means of zero
are imposed ‘loosely’. This is motivated by the fact that we don’t
really have a great deal of confidence in the zero prior mean specifica-
tion for longer lags of important variables. We think they should exert
some influence, making the prior mean of zero somewhat inappropri-
ate. We still impose lag decay on longer lags of important variables
by decreasing our prior variance with increasing lag length. This re-
flects the idea that influence decays over time for important as well as
unimportant variables.
.::UdecomBooks::.
It should be noted that the prior relies on inappropriate zero prior means
for the important variables at lags greater than one for two reasons. First, it
is difficult to specify a reasonable alternative prior mean for these variables
that would have universal applicability in a large number of VAR model
applications. The difficulty of assigning meaningful prior means that have
universal appeal is most likely the reason that past studies relied on the
Minnesota prior means while simply altering the prior variances. A prior
mean that averages over previous period values of the important variables
has universal appeal and widespread applicability in VAR modeling. The
second motivation for relying on inappropriate zero prior means for longer
lags of the important variables is that overparameterization and collinearity
problems that plague the VAR model are best overcome by relying on a
parsimonious representation. Zero prior means for the majority of the large
number of coefficients in the VAR model are consistent with this goal of par-
simony and have been demonstrated to produce improved forecast accuracy
in a wide variety of applications of the Minnesota prior.
A flexible form with which to state prior standard deviations for variable
j in equation i at lag length k is shown in (5.14).
π(aijk ) = N (1/ci , σc ), j ∈ C, k = 1, i, j = 1, . . . , n
π(aijk ) = N (0, τ σc /k), j ∈ C, k = 2, . . . , m, i, j = 1, . . . , n
π(aijk ) = N (0, θσc /k), j¬ ∈ C, k = 1, . . . , m, i, j = 1, . . . , n
(5.14)
where:
0 < σc < 1 (5.15)

τ > 1 (5.16)
0<θ < 1 (5.17)
For variables j = 1, . . . , m in equation i that are important in explaining

variation in variable i, (j ∈ C), the prior mean for lag length k = 1 is
set to the average of the number of important variables in equation i and
to zero for unimportant variables (j¬ ∈ C). The prior standard deviation
is set to σc for the first lag, and obeys the restriction set forth in (5.15),
reflecting a tight imposition of the prior mean that forces averaging over
important variables. To see this, consider that the prior means 1/ci range
between zero and unity so typical σc values might be in the range of 0.1
to 0.25. We use τ σc /k for lags greater than one which imposes a decrease
in this variance as the lag length k increases. Equation (5.16) states the
.::UdecomBooks::.
restriction necessary to ensure that the prior mean of zero is imposed on the
parameters associated with lags greater than one for important variables
loosely, relative to a tight imposition of the prior mean of 1/ci on first own-
lags of important variables. We use θσc /k for lags on unimportant variables
whose prior means are zero, imposing a decrease in the variance as the lag
length increases. The restriction in (5.17) would impose the zero means for
unimportant variables with more confidence than the zero prior means for
important variables.
This mathematical formulation adequately captures all aspects of the
intuitive motivation for the prior variance specification enumerated above.
A quick way to see this is to examine a graphical depiction of the prior mean
and standard deviation for an important versus unimportant variable. An
artificial example was constructed for an important variable in Figure 5.1
and an unimportant variable in Figure 5.2. Figure 5.1 shows the prior
mean along with five upper and lower limits derived from the prior standard
deviations in (5.14). The five standard deviation limits shown in the figure
reflect ± 2 standard deviation limits resulting from alternative settings for
the prior hyperparameter τ ranging from 5 to 9 and a value of σc = 0.25.
Larger values of τ generated the wider upper and lower limits.
The solid line in Figure 5.1 reflects a prior mean of 0.2 for lag 1 indicating
five important variables, and a prior mean of zero for all other lags. The prior
standard deviation at lag 1 is relatively tight producing a small band around
the averaging prior mean for this lag. This imposes the ‘averaging’ prior
belief with a fair amount of certainty. Beginning at lag 2, the prior standard
deviation is increased to reflect relative uncertainty about the new prior
mean of zero for lags greater than unity. Recall, we believe that important
variables at lags greater than unity will exert some influence, making the
prior mean of zero not quite appropriate. Hence, we implement this prior
mean with greater uncertainty.
Figure 5.2 shows an example of the prior means and standard deviations
for an unimportant variable based on σc = 0.25 and five values of θ ranging
from .35 to .75. Again, the larger θ values produce wider upper and lower
limits. The prior for unimportant variables is motivated by the Minnesota
prior that also uses zero prior means and rapid decay with increasing lag
length.
A function rvar implements the random-walk averaging prior and a re-
lated function recm carries out estimation for an EC model based on this
prior. The documentation for the rvar function is shown below, where we
have eliminated information regarding the results structure variable returned
by the function to save space.
.::UdecomBooks::.
0.5
mean
0.4 upper
lower
0.3
0.2
prior standard deviation
0.1
-0.1
-0.2
-0.3
-0.4
-0.5
0 2 4 6 8 10 12
lag length
Figure 5.2: Prior means and precision for important variables
PURPOSE: Estimates a Bayesian vector autoregressive model

using the random-walk averaging prior
---------------------------------------------------
USAGE: result = rvar(y,nlag,w,freq,sig,tau,theta,x)
where: y = an (nobs x neqs) matrix of y-vectors (in levels)
w = an (neqs x neqs) matrix containing prior means
(rows should sum to unity, see below)
freq = 1 for annual, 4 for quarterly, 12 for monthly
sig = prior variance hyperparameter (see below)
tau = prior variance hyperparameter (see below)
theta = prior variance hyperparameter (see below)
x = an (nobs x nx) matrix of deterministic variables
(in any form, they are not altered during estimation)
(constant term automatically included)
priors important variables: N(w(i,j),sig) for 1st own lag
N( 0 ,tau*sig/k) for lag k=2,...,nlag
priors unimportant variables: N(w(i,j),theta*sig/k) for lag 1
N( 0 ,theta*sig/k) for lag k=2,...,nlag
e.g., if y1, y3, y4 are important variables in eq#1, y2 unimportant
w(1,1) = 1/3, w(1,3) = 1/3, w(1,4) = 1/3, w(1,2) = 0
.::UdecomBooks::.
0.2
mean
upper
0.15 lower
0.1
prior standard deviation
0.05
-0.05
-0.1
-0.15
-0.2
0 2 4 6 8 10 12
lag length
Figure 5.3: Prior means and precision for unimportant variables
typical values would be: sig = .1-.3, tau = 4-8, theta = .5-1
---------------------------------------------------
NOTES: - estimation is carried out in annualized growth terms because
the prior means rely on common (growth-rate) scaling of variables
hence the need for a freq argument input.
- constant term included automatically
---------------------------------------------------
Because this model is estimated in growth-rates form, an input argument

for the data frequency is required. As an illustration of using both the rvar
and recm functions, consider the following example based on the eight-state
mining industry data. We specify a weight matrix for the prior means using
first-order contiguity of the states.
% ----- Example 5.11 Estimating RECM models
y = load(’test.dat’); % a test data set
sig = 0.1;
tau = 6;
theta = 0.5;
.::UdecomBooks::.
freq = 12; % monthly data

% this is an example of using 1st-order contiguity
% of the states as weights to produce prior means
W=[0 0.5 0.5 0 0 0 0 0
0.25 0 0.25 0.25 0.25 0 0 0
0.20 0.20 0 0 0.20 0 0.20 0.20
0 0.50 0 0 0.50 0 0 0
0 0.20 0.20 0.20 0 0.20 0.20 0.20
0 0 0 0 0.50 0 0 0.50
0 0 1 0 0 0 0 0
0 0 0.33 0 0.33 0.33 0 0];
% estimate the rvar model
results = rvar(y,nlag,W,freq,sig,tau,theta);
% print results to a file
fid = fopen(’rvar.out’,’wr’);
prt(results,vnames,fid);
% estimate the recm model letting the function
% determine the # of co-integrating relationships
results = recm(y,nlag,W,freq,sig,tau,theta);
% print results to a file
fid = fopen(’recm.out’,’wr’);
prt(results,vnames,fid);
5.3.1 Theil-Goldberger estimation of these models

VAR models based on both the Minnesota prior and the random-walk aver-
aging prior can be estimated using the ‘mixed estimation’ method set forth
in Theil and Goldberger (1961). Letting the matrix X represent the lagged
values of yit , i = 1, . . . , n and the vector A denote the coefficients aijk (`)
from (5.1), we can express a single equation of the model as:
y1 = XA + ε1 (5.18)
where it is assumed, var(ε1 ) = σ 2 I. The stochastic prior restrictions for this
single equation can be written as:
      
m111 σ/σ111 0 ... 0 a111 u111
      
 m112   0 σ/σ112 0 0  a112   u112 
 =  + 
 ..   ..  ..   .. 
 .   0 . 0  .   . 
mnnk 0 0 0 σ/σnnk annk
unnk
(5.19)
2
where we assume, var(u) = σ I and the σijk take the form shown in (5.7)
for the Minnesota prior, and that set forth in (5.14) for the random-walk
.::UdecomBooks::.
averaging prior model. Similarly, the prior means mijk take the form de-
scribed for the Minnesota and averaging priors. Noting that (5.19) can be
written in the form suggested by Theil-Goldberger:
r = RA + u, (5.20)
the estimates for a typical equation are derived using (5.21).
Â = (X 0 X + R0 R)−1 (X 0 y1 + R0 r) (5.21)
The difference in prior means specified by the Minnesota prior and the
random-walk averaging prior resides in the mijk terms found on the left-
hand-side of the equality in (5.20). The Minnesota prior indicates values:
(σ/σ111 , 0, . . . , 0)0 , where the non-zero value occurs in the position represent-
ing the lagged dependent variable. The averaging prior would have non-zero
values in locations associated with important variables and zeros elsewhere.
5.4 Forecasting the models

A set of forecasting functions are available that follow the format of the
var, bvar, rvar, ecm, becm, recm functions named varf, bvarf, rvarf,
ecmf, becmf, recmf. These functions all produce forecasts of the time-
series levels to simplify accuracy analysis and forecast comparison from the
alternative models. They all take time-series levels arguments as inputs and
carry out the necessary transformations. As an example, the varf documen-
tation is:
PURPOSE: estimates a vector autoregression of order n

and produces f-step-ahead forecasts
-------------------------------------------------------------
USAGE:yfor = varf(y,nlag,nfor,begf,x,transf)
where: y = an (nobs * neqs) matrix of y-vectors in levels
nfor = the forecast horizon
begf = the beginning date of the forecast
(defaults to length(x) + 1)
x = an optional vector or matrix of deterministic
variables (not affected by data transformation)
transf = 0, no data transformation
= 1, 1st differences used to estimate the model
= freq, seasonal differences used to estimate
= cal-structure, growth rates used to estimate
e.g., cal(1982,1,12) [see cal() function]
-------------------------------------------------------------
.::UdecomBooks::.
NOTE: constant term included automatically

-------------------------------------------------------------
RETURNS:
yfor = an nfor x neqs matrix of level forecasts for each equation
-------------------------------------------------------------
Note that you input the variables y in levels form, indicate any of four
data transformations that will be used when estimating the VAR model,
and the function varf will carry out this transformation, produce estimates
and forecasts that are converted back to levels form. This greatly simplifies
the task of producing and comparing forecasts based on alternative data
transformations.
Of course, if you desire a transformation other than the four provided,
such as logs, you can transform the variables y prior to calling the function
and specify ‘transf=0’. In this case, the function does not provide levels
forecasts, but rather forecasts of the logged-levels will be returned. Setting
‘transf=0’, produces estimates based on the data input and returns forecasts
based on this data.
As an example of comparing alternative VAR model forecasts based on
two of the four alternative transformations, consider the program in example
5.12.
% ----- Example 5.12 Forecasting VAR models
y = load(’test.dat’); % a test data set containing
% monthly mining employment for
% il,in,ky,mi,oh,pa,tn,wv
dates = cal(1982,1,12); % data covers 1982,1 to 1996,5
nfor = 12; % number of forecast periods
begf = ical(1995,1,dates); % beginning forecast period
endf = ical(1995,12,dates); % ending forecast period
% no data transformation example
fcast1 = varf(y,nlag,nfor,begf);
% seasonal differences data transformation example
freq = 12; % set frequency of the data to monthly
fcast2 = varf(y,nlag,nfor,begf,[],freq);
% compute percentage forecast errors
actual = y(begf:endf,:);
error1 = (actual-fcast1)./actual;
error2 = (actual-fcast2)./actual;
fdates = cal(1995,1,12);
fprintf(1,’VAR model in levels percentage errors \n’);
tsprint(error1*100,fdates,vnames,’%7.2f’);
fprintf(1,’VAR - seasonally differenced data percentage errors \n’);
tsprint(error2*100,fdates,vnames,’%7.2f’);
.::UdecomBooks::.
The results from the program are:
VAR model in levels percentage errors

Date il in ky mi oh pa tn wv
Jan95 -3.95 -2.86 -1.15 -6.37 -5.33 -7.83 -0.19 -0.65
Feb95 -5.63 -2.63 -3.57 -7.77 -7.56 -8.28 -0.99 0.38
Mar95 -3.62 -1.75 -4.66 -5.49 -5.67 -6.69 2.26 2.30
Apr95 -3.81 -4.23 -7.11 -4.27 -5.18 -5.41 2.14 0.17
May95 -4.05 -5.60 -8.14 -0.92 -5.88 -3.93 2.77 -1.11
Jun95 -4.10 -3.64 -8.87 0.10 -4.65 -4.15 2.90 -2.44
Jul95 -4.76 -3.76 -10.06 1.99 -1.23 -5.06 3.44 -3.67
Aug95 -8.69 -3.89 -9.86 4.85 -2.49 -5.41 3.63 -3.59
Sep95 -8.73 -3.63 -12.24 0.70 -4.33 -6.28 3.38 -4.04
Oct95 -11.11 -3.23 -12.10 -7.38 -4.74 -8.34 3.21 -5.57
Nov95 -11.79 -4.30 -11.53 -8.93 -4.90 -7.27 3.60 -5.69
Dec95 -12.10 -5.56 -11.12 -13.11 -5.57 -8.78 2.13 -9.38
VAR - seasonally differenced data percentage errors

Date il in ky mi oh pa tn wv
Jan95 -6.53 -0.52 -3.75 3.41 -1.49 -0.06 3.86 0.05
Feb95 -4.35 1.75 -6.29 0.35 -3.53 -2.76 4.46 2.56
Mar95 -1.12 2.61 -6.83 1.53 -2.72 2.24 2.96 3.97
Apr95 -0.38 -2.36 -7.03 -4.30 -1.28 0.70 5.55 2.73
May95 0.98 -5.05 -3.90 -4.65 -1.18 2.02 6.49 -0.43
Jun95 -0.73 -2.55 -2.04 -0.30 2.30 0.81 3.96 -1.44
Jul95 -1.41 -0.36 -1.69 0.79 4.83 -0.06 7.68 -4.24
Aug95 -3.36 2.36 -1.78 7.99 4.86 -1.07 8.75 -3.38
Sep95 -3.19 3.47 -3.26 6.91 2.31 -1.44 8.30 -3.02
Oct95 -2.74 3.27 -2.88 -2.14 2.92 -0.73 9.00 0.08
Nov95 -2.47 1.54 -2.63 -5.23 4.33 0.36 9.02 0.64
Dec95 -1.35 0.48 -3.53 -7.89 4.38 1.33 7.03 -3.92
It is also possible to build models that produce forecasts that “feed-in” to

another model as deterministic variables. For example, suppose we wished
to use national employment in the primary metal industry (SIC 33) as a
deterministic variable in our model for primary metal employment in the
eight states. The following program shows how to accomplish this.
% ----- Example 5.13 Forecasting multiple related models

dates = cal(1982,1,12); % data starts in 1982,1
y=load(’sic33.states’); % industry sic33 employment for 8 states
[nobs neqs] = size(y);
load sic33.national; % industry sic33 national employment
ndates = cal(1947,1,12);% national data starts in 1947,1
begs = ical(1982,1,ndates); % find 1982,1 for national data

ends = ical(1996,5,ndates); % find 1996,5 for national data
.::UdecomBooks::.
x = sic33(begs:ends,1); % pull out national employment in sic33

% for the time-period corresponding to
% our 8-state sample
begf = ical(1990,1,dates); % begin forecasting date
endf = ical(1994,12,dates); % end forecasting date
nfor = 12; % forecast 12-months-ahead
nlag = 6;
xerror = zeros(nfor,1);
yerror = zeros(nfor,neqs);
cnt = 0; % counter for the # of forecasts we produce
for i=begf:endf % loop over dates producing forecasts
xactual = x(i:i+nfor-1,1); % actual national employment
yactual = y(i:i+nfor-1,:); % actual state employment
% first forecast national employment in sic33
xfor = varf(x,nlag,nfor,i); % an ar(6) model
xdet = [x(1:i-1,1) % actual national data up to forecast period
xfor ]; % forecasted national data
% do state forecast using national data and forecast as input
yfor = varf(y,nlag,nfor,i,xdet);
% compute forecast percentage errors
xerror = xerror + abs((xactual-xfor)./xactual);
yerror = yerror + abs((yactual-yfor)./yactual);
cnt = cnt+1;
end; % end loop over forecasting experiment dates
% compute mean absolute percentage errors
xmape = xerror*100/cnt; ymape = yerror*100/cnt;
% printout results
in.cnames = strvcat(’national’,’il’,’in’,’ky’,’mi’,’oh’,’pa’,’tn’,’wv’);
rnames = ’Horizon’;
for i=1:12; rnames = strvcat(rnames,[num2str(i),’-step’]); end;
in.rnames = rnames;
in.fmt = ’%6.2f’;
fprintf(1,’national and state MAPE percentage forecast errors \n’);
fprintf(1,’based on %d 12-step-ahead forecasts \n’,cnt);
mprint([xmape ymape],in);
Our model for national employment in SIC33 is simply an autoregressive

model with 6 lags, but the same approach would work for a matrix X of
deterministic variables used in place of the vector in the example. We can
also provide for a number of deterministic variables coming from a variety
of models that are input into other models, not unlike traditional structural
econometric models. The program produced the following output.
national and state MAPE percentage forecast errors
based on 60 12-step-ahead forecasts
Horizon national il in ky mi oh pa tn wv
1-step 0.27 0.70 0.78 1.00 1.73 0.78 0.56 0.88 1.08
2-step 0.46 1.02 1.10 1.15 1.95 1.01 0.78 1.06 1.58
.::UdecomBooks::.
3-step 0.68 1.22 1.26 1.39 2.34 1.17 1.00 1.16 1.91
4-step 0.93 1.53 1.45 1.46 2.81 1.39 1.25 1.35 2.02
5-step 1.24 1.84 1.63 1.74 3.27 1.55 1.57 1.53 2.10
6-step 1.55 2.22 1.70 2.05 3.41 1.53 1.81 1.64 2.15
7-step 1.84 2.62 1.59 2.24 3.93 1.68 1.99 1.76 2.49
8-step 2.21 3.00 1.56 2.34 4.45 1.82 2.10 1.89 2.87
9-step 2.55 3.30 1.59 2.58 4.69 1.93 2.33 1.99 3.15
10-step 2.89 3.64 1.74 2.65 5.15 2.08 2.51 2.12 3.39
11-step 3.25 3.98 1.86 2.75 5.75 2.29 2.70 2.27 3.70
12-step 3.60 4.36 1.94 2.86 6.01 2.40 2.94 2.23 3.96
Consider that it would be quite easy to assess the contribution of using

national employment as a deterministic variable in the model by running
another model that excludes this deterministic variable.
As a final example, consider an experiment where we wish to examine
the impact of using different numbers of error correction variables on the
forecast accuracy of the EC model. Shoesmith (1995) suggests that one
should employ the number of error correction variables associated with the
Johansen likelihood ratio statistics, but he provides only limited evidence
regarding this contention.
The experiment uses time-series on national monthly employment from
12 manufacturing industries covering the period 1947,1 to 1996,12. Forecasts
are carried out over the period from 1970,1 to 1995,12 using the number
of error correction terms suggested by the Johansen likelihood ratio trace
statistics, as well as models based on +/-1 and +/-2 error correction terms
relative to the value suggested by the trace statistic.
We then compare the relative forecast accuracy of these models by exam-
ining the ratio of the MAPE forecast error from the models with +/-1 and
+/-2 terms to the errors from the model based on r relationships suggested
by the trace statistic.
Here is the program code:
% ----- Example 5.14 comparison of forecast accuracy as a function of

% the # of co-integrating vectors used
load level.mat; % 20 industries national employment
y = level(:,1:12); % use only 12 industries
[nobs neqs] = size(y); dates = cal(1947,1,12);
begf = ical(1970,1,dates); % beginning forecast date
endf = ical(1995,12,dates); % ending forecast date
nfor = 12; % forecast horizon
nlag = 10; cnt = 1; % nlag based on lrratio() results
for i=begf:endf;
jres = johansen(y,0,nlag); trstat = jres.lr1; tsignf = jres.cvt;
r = 0;
.::UdecomBooks::.
for j=1:neqs; % find r indicated by trace statistic

if trstat(j,1) > tsignf(j,2), r = j; end;
end;
% set up r-1,r-2 and r+1,r+2 forecasts in addition to forecasts based on r
if (r >= 3 & r <=10)
frm2 = ecmf(y,nlag,nfor,i,r-2); frm1 = ecmf(y,nlag,nfor,i,r-1);
fr = ecmf(y,nlag,nfor,i,r); frp1 = ecmf(y,nlag,nfor,i,r+1);
frp2 = ecmf(y,nlag,nfor,i,r+2); act = y(i:i+nfor-1,1:12);
% compute forecast MAPE
err(cnt).rm2 = abs((act-frm2)./act); err(cnt).rm1 = abs((act-frm1)./act);
err(cnt).r = abs((act-fr)./act); err(cnt).rp1 = abs((act-frp1)./act);
err(cnt).rp2 = abs((act-frp2)./act); cnt = cnt+1;
else
fprintf(1,’time %d had %d co-integrating relations \n’,i,r);
end; % end if-else; end; % end of loop over time
rm2 = zeros(12,12); rm1 = rm2; rm0 = rm2; rp1 = rm2; rp2 = rm2;
for i=1:cnt-1;
rm2 = rm2 + err(i).rm2; rm1 = rm1 + err(i).rm1;
rm0 = rm0 + err(i).r; rp1 = rp1 + err(i).rp1;
rp2 = rp2 + err(i).rp2;
end;
rm2 = rm2/(cnt-1); rm1 = rm1/(cnt-1);
rm0 = rm0/(cnt-1); rp1 = rp1/(cnt-1);
rp2 = rp2/(cnt-1);
rnames = ’Horizon’; cnames = [];
for i=1:12;
rnames = strvcat(rnames,[num2str(i),’-step’]);
cnames = strvcat(cnames,[’IND’,num2str(i)]);
end;
in.rnames = rnames; in.cnames = cnames; in.fmt = ’%6.2f’;
fprintf(1,’forecast errors relative to error by ecm(r) model \n’);
fprintf(1,’r-2 relative to r \n’);
mprint(rm2./rm0,in);
fprintf(1,’r-1 relative to r \n’);
mprint(rm2./rm0,in);
fprintf(1,’r+1 relative to r \n’);
mprint(rp1./rm0,in);
fprintf(1,’r+2 relative to r \n’);
mprint(rp2./rm0,in);
The program code stores the individual MAPE forecast errors in a struc-
ture variable using: err(cnt).rm2 = abs((actual-frm2)./actual);, which
will have fields for the errors from all five models. These fields are matrices
of dimension 12 x 12, containing MAPE errors for each of the 12-step-ahead
forecasts for time cnt and for each of the 12 industries. We are not re-
ally interested in these individual results, but present this as an illustration.
As part of the illustration, we show how to access the individual results to
.::UdecomBooks::.
compute the average MAPE errors for each horizon and industry. If you
wished to access industry number 2’s forecast errors based on the model us-
ing r co-integrating relations, for the first experimental forecast period you
would use: err(1).rm(:,2). The results from our experiment are shown
below. These results represent an average over a total of 312 twelve-step-
ahead forecasts. Our simple MATLAB program produced a total of 224,640
forecasts, based on 312 twelve-step-ahead forecasts, for 12 industries, times
5 models!
Our experiment indicates that using more than the r co-integrating re-
lationships determined by the Johansen likelihood trace statistic degrades
the forecast accuracy. This is clear from the large number of forecast error
ratios greater than unity for the two models based on r + 1 and r + 2 ver-
sus those from the model based on r. On the other hand, using a smaller
number of co-integrating relationships than indicated by the Johansen trace
statistic seems to improve forecast accuracy. In a large number of industries
at many of the twelve forecast horizons, we see comparison ratios less than
unity. Further, the forecast errors associated with r − 2 are superior to those
from r − 1, producing smaller comparison ratios in 9 of the 12 industries.
5.5 Chapter summary

We found that a library of functions can be constructed to produce esti-
mates and forecasts for a host of alternative vector autoregressive and error
correction models. An advantage of MATLAB over a specialized program
like RATS is that we have more control and flexibility to implement special-
ized priors. The prior for the rvar model cannot be implemented in RATS
software as the vector autoregressive function in that program does not al-
low you to specify prior means for variables other than the first own-lagged
variables in the model.
Another advantage is the ability to write auxiliary functions that process
the structures returned by our estimation functions and present output in a
format that we find helpful. As an example of this, the function pgranger
produced a formatted table of Granger-causality probabilities making it easy
to draw inferences.
.::UdecomBooks::.
forecast errors relative to error by ecm(r) model

r-2 relative to r
Horizon I1 I2 I3 I4 I5 I6 I7 I8 I9 I10 I11 I12
1-step 1.01 0.99 1.00 1.01 1.00 1.00 1.01 0.99 1.00 0.98 0.97 0.99
2-step 0.92 1.01 0.99 0.96 1.03 1.00 1.02 0.99 1.01 1.03 0.99 0.94
3-step 0.89 1.04 1.00 0.94 1.03 1.02 1.01 0.98 0.99 1.03 1.00 0.93
4-step 0.85 1.03 0.99 0.94 1.05 1.03 1.02 1.00 0.97 1.01 1.00 0.91
5-step 0.82 1.03 0.98 0.94 1.03 1.03 1.04 1.00 0.97 0.98 1.02 0.92
6-step 0.81 1.05 0.97 0.94 1.01 1.04 1.04 0.99 0.97 0.96 1.03 0.92
7-step 0.79 1.07 0.96 0.93 0.99 1.03 1.05 0.98 0.97 0.94 1.03 0.92
8-step 0.78 1.04 0.95 0.93 0.98 1.02 1.04 0.96 0.96 0.93 1.03 0.93
9-step 0.76 1.03 0.93 0.92 0.97 1.01 1.02 0.95 0.95 0.91 1.01 0.94
10-step 0.76 1.01 0.92 0.91 0.96 0.99 1.01 0.94 0.94 0.90 0.99 0.94
11-step 0.75 1.00 0.91 0.91 0.95 0.98 1.01 0.95 0.94 0.90 0.99 0.95
12-step 0.74 0.99 0.90 0.91 0.94 0.98 0.99 0.94 0.93 0.89 0.98 0.95
r-1 relative to r
1-step 1.01 0.99 1.00 1.01 1.00 1.00 1.01 0.99 1.00 0.98 0.97 0.99
2-step 0.92 1.01 0.99 0.96 1.03 1.00 1.02 0.99 1.01 1.03 0.99 0.94
3-step 0.89 1.04 1.00 0.94 1.03 1.02 1.01 0.98 0.99 1.03 1.00 0.93
4-step 0.85 1.03 0.99 0.94 1.05 1.03 1.02 1.00 0.97 1.01 1.00 0.91
5-step 0.82 1.03 0.98 0.94 1.03 1.03 1.04 1.00 0.97 0.98 1.02 0.92
6-step 0.81 1.05 0.97 0.94 1.01 1.04 1.04 0.99 0.97 0.96 1.03 0.92
7-step 0.79 1.07 0.96 0.93 0.99 1.03 1.05 0.98 0.97 0.94 1.03 0.92
8-step 0.78 1.04 0.95 0.93 0.98 1.02 1.04 0.96 0.96 0.93 1.03 0.93
9-step 0.76 1.03 0.93 0.92 0.97 1.01 1.02 0.95 0.95 0.91 1.01 0.94
10-step 0.76 1.01 0.92 0.91 0.96 0.99 1.01 0.94 0.94 0.90 0.99 0.94
11-step 0.75 1.00 0.91 0.91 0.95 0.98 1.01 0.95 0.94 0.90 0.99 0.95
12-step 0.74 0.99 0.90 0.91 0.94 0.98 0.99 0.94 0.93 0.89 0.98 0.95
r+1 relative to r
1-step 1.01 1.00 1.02 1.00 1.00 1.01 1.01 0.99 1.00 1.01 1.02 1.01
2-step 0.99 1.02 1.01 0.99 0.99 1.03 1.00 0.99 0.99 1.05 1.03 1.04
3-step 0.99 1.01 1.01 0.99 1.00 1.04 1.00 0.99 0.98 1.07 1.03 1.04
4-step 0.99 0.99 1.01 0.98 1.01 1.05 1.01 1.01 0.97 1.08 1.04 1.03
5-step 0.98 0.98 1.03 0.99 1.01 1.05 1.01 1.03 0.97 1.08 1.04 1.04
6-step 0.98 0.98 1.03 0.99 1.01 1.06 1.00 1.03 0.97 1.07 1.04 1.04
7-step 0.98 0.98 1.04 1.00 1.01 1.06 1.00 1.04 0.97 1.08 1.04 1.04
8-step 0.98 0.96 1.05 1.00 1.02 1.06 0.99 1.05 0.97 1.06 1.04 1.04
9-step 0.97 0.95 1.05 1.01 1.02 1.07 0.99 1.05 0.96 1.05 1.04 1.04
10-step 0.97 0.96 1.05 1.01 1.02 1.07 0.98 1.05 0.96 1.04 1.04 1.03
11-step 0.97 0.97 1.05 1.01 1.02 1.07 0.98 1.06 0.95 1.05 1.04 1.03
12-step 0.97 0.97 1.05 1.01 1.02 1.07 0.98 1.07 0.95 1.05 1.04 1.03
r+2 relative to r
1-step 1.00 1.01 1.02 1.01 0.99 1.01 1.01 0.99 0.99 1.05 1.03 1.01
2-step 1.00 1.05 1.00 0.97 1.00 1.03 1.01 1.00 0.99 1.11 1.03 1.06
3-step 1.00 1.02 1.01 0.96 1.01 1.06 1.02 1.02 0.98 1.13 1.04 1.06
4-step 1.00 0.99 1.01 0.97 1.02 1.07 1.02 1.04 0.97 1.14 1.05 1.05
5-step 1.01 0.97 1.03 0.98 1.04 1.08 1.02 1.07 0.97 1.15 1.05 1.04
6-step 1.01 0.95 1.04 0.99 1.04 1.10 1.02 1.08 0.97 1.15 1.06 1.04
7-step 1.01 0.96 1.06 1.00 1.05 1.10 1.01 1.09 0.96 1.15 1.06 1.03
8-step 1.00 0.93 1.08 0.99 1.05 1.10 1.00 1.10 0.95 1.15 1.07 1.02
9-step 1.01 0.92 1.09 0.99 1.06 1.11 0.99 1.11 0.95 1.14 1.08 1.02
10-step 1.01 0.92 1.09 0.99 1.05 1.11 0.98 1.11 0.94 1.13 1.08 1.01
11-step 1.01 0.93 1.09 0.99 1.05 1.12 0.98 1.13 0.94 1.14 1.09 1.00
12-step 1.00 0.93 1.09 0.99 1.05 1.12 0.97 1.13 0.94 1.15 1.09 0.99
.::UdecomBooks::.
Finally, many of the problems encountered in carrying out forecast ex-

periments involve transformation of the data for estimation purposes and
reverse transformations needed to compute forecast errors on the basis of the
levels of the time-series. Our functions can perform these transformations
for the user, making the code necessary to carry out forecast experiments
quite simple. In fact, one could write auxiliary functions that compute al-
ternative forecast accuracy measures given matrices of forecast and actual
values.
We also demonstrated how the use of structure array variables can facil-
itate storage of individual forecasts or forecast errors for a large number of
time periods, horizons and variables. This would allow a detailed examina-
tion of the accuracy and characteristics associated with individual forecast
errors for particular variables and time periods. As noted above, auxiliary
functions could be constructed to carry out this type of analysis.
.::UdecomBooks::.
Chapter 5 Appendix
The vector autoregressive library functions are in a subdirectory var bvar.
vector autoregressive function library
------- VAR/BVAR program functions -----------
becm_g - Gibbs sampling BECM estimates

becmf - Bayesian ECM model forecasts
becmf_g - Gibbs sampling BECM forecasts
bvar - BVAR model
bvar_g - Gibbs sampling BVAR estimates
bvarf - BVAR model forecasts
bvarf_g - Gibbs sampling BVAR forecasts
ecm - ECM (error correction) model estimates
ecmf - ECM model forecasts
irf - impulse response functions
lrratio - likelihood ratio lag length tests
pftest - prints Granger F-tests
pgranger - prints Granger causality probabilities
recm - ecm version of rvar
recm_g - Gibbs sampling random-walk averaging estimates
recmf - random-walk averaging ECM forecasts
recmf_g - Gibbs sampling random-walk averaging forecasts
rvar - Bayesian random-walk averaging prior model
rvar_g - Gibbs sampling RVAR estimates
rvarf - Bayesian RVAR model forecasts
rvarf_g - Gibbs sampling RVAR forecasts
var - VAR model
varf - VAR model forecasts
------ demonstration programs -----------
becm_d - BECM model demonstration

becm_gd - Gibbs sampling BECM estimates demo
becmf_d - becmf demonstration
becmf_gd - Gibbs sampling BECM forecast demo
bvar_d - BVAR model demonstration
148
.::UdecomBooks::.
bvar_gd - Gibbs sampling BVAR demonstration

bvarf_d - bvarf demonstration
bvarf_gd - Gibbs sampling BVAR forecasts demo
ecm_d - ECM model demonstration
ecmf_d - ecmf demonstration
irf_d - impulse response function demo
irf_d2 - another irf demo
lrratio_d - demonstrates lrratio
pftest_d - demo of pftest function
recm_d - RECM model demonstration
recm_gd - Gibbs sampling RECM model demo
recmf_d - recmf demonstration
recmf_gd - Gibbs sampling RECM forecast demo
rvar_d - RVAR model demonstration
rvar_g - Gibbs sampling rvar model demo
rvarf_d - rvarf demonstration
rvarf_gd - Gibbs sampling rvar forecast demo
var_d - VAR model demonstration
varf_d - varf demonstration
------- support functions -----------
johansen - used by ecm,ecmf,becm,becmf,recm,recmf

lag - does ordinary lags
mlag - does var-type lags
nclag - does contiguous lags (used by rvar,rvarf,recm,recmf)
ols - used for VAR estimation
prt - prints results from all functions
prt_coint - used by prt_var for ecm,becm,recm
prt_var - prints results of all var/bvar models
prt_varg - prints results of all Gibbs var/bvar models
rvarb - used for RVARF forecasts
scstd - does univariate AR for BVAR
theil_g - used for Gibbs sampling estimates and forecasts
theilbf - used for BVAR forecasts
theilbv - used for BVAR estimation
trimr - used by VARF,BVARF, johansen
vare - used by lrratio
The co-integration library functions are in a subdirectory coint.
co-integration library
------ co-integration testing routines --------
adf - carries out Augmented Dickey-Fuller unit root tests

cadf - carries out ADF tests for co-integration
johansen - carries out Johansen’s co-integration tests
phillips - Phillips-Peron unit root test
.::UdecomBooks::.
------ demonstration programs ----------
adf_d - demonstrates adf

cadf_d - demonstrates cadf
johansen_d - demonstrates johansen
phillips_d - demonstrates phillips
------ support functions -------------
c_sja - returns critical values for SJ maximal eigenvalue test

c_sjt - returns critical values for SJ trace test
cols - (like Gauss cols)
detrend - used by johansen to detrend data series
prt_coint - prints results from adf,cadf,johansen
ptrend - used by adf to create time polynomials
rows - (like Gauss rows)
rztcrit - returns critical values for cadf test
tdiff - time-series differences
trimr - (like Gauss trimr)
ztcrit - returns critical values for adf test
.::UdecomBooks::.
Chapter 6
Markov Chain Monte Carlo

Models
A gibbs sampling library of utility functions along with various estimation

functions are described in this chapter. Additional examples that demon-
strate gibbs library functions for limited dependent variable logit, probit,
and tobit models are presented in the next chapter that discusses limited
dependent variable estimation.
A recent methodology known as Markov Chain Monte Carlo is based
on the idea that rather than compute a probability density, say p(θ|y), we
would be just as happy to have a large random sample from p(θ|y) as to
know the precise form of the density. Intuitively, if the sample were large
enough, we could approximate the form of the probability density using
kernel density estimators or histograms. In addition, we could compute
accurate measures of central tendency and dispersion for the density, using
the mean and standard deviation of the large sample. This insight leads to
the question of how to efficiently simulate a large number of random samples
from p(θ|y).
Metropolis, et al. (1953) showed that one could construct a Markov
chain stochastic process for (θt , t ≥ 0) that unfolds over time such that: 1)
it has the same state space (set of possible values) as θ, 2) it is easy to
simulate, and 3) the equilibrium or stationary distribution which we use to
draw samples is p(θ|y) after the Markov chain has been run for a long enough
time. Given this result, we can construct and run a Markov chain for a very
large number of iterations to produce a sample of (θt , t = 1, . . .) from the
posterior distribution and use simple descriptive statistics to examine any
features of the posterior in which we are interested.
151
.::UdecomBooks::.
CHAPTER 6. MARKOV CHAIN MONTE CARLO MODELS 152
This approach, known as Markov Chain Monte Carlo, (MCMC) to deter-

mining posterior densities has greatly reduced the computational problems
that previously plagued application of the Bayesian methodology. Gelfand
and Smith (1990), as well as a host of others, have popularized this method-
ology by demonstrating its use in a wide variety of statistical applications
where intractable posterior distributions previously hindered Bayesian anal-
ysis. A simple introduction to the method can be found in Casella and
George (1990) and an expository article dealing specifically with the normal
linear model is Gelfand, Hills, Racine-Poon and Smith (1990). Two recent
books that deal in detail with all facets of these methods are: Gelman,
Carlin, Stern and Rubin (1995) and Gilks, Richardson and Spiegelhalter
(1996).
The most widely used approach to MCMC is due to Hastings (1970)
which generalizes a method of Metropolis et al. (1953). A second approach
(that we focus on) is known as Gibbs sampling due to Geman and Geman
(1984). Hastings approach suggests that given an initial value θ0 we can
construct a chain by recognizing that any Markov chain that has found
its way to a state θt can be completely characterized by the probability
distribution for time t + 1. His algorithm relies on a proposal or candidate
distribution, f (θ|θt ) for time t + 1, given that we have θt . A candidate point
θ? is sampled from the proposal distribution and:
1. This point is accepted as θt+1 = θ? with probability:

· ¸
? p(θ? |y)f (θt |θ? )
αH (θt , θ ) = min 1, (6.1)
p(θt |y)f (θ? |θt )
2. otherwise, θt+1 = θt , that is we stay with the current value of θ.
In other words, we can view the Hastings algorithm as indicating that

we should toss a Bernoulli coin with probability αH of heads and make
a move to θt+1 = θ? if we see a heads, otherwise set θt+1 = θt . Hastings
demonstrates that this approach to sampling represents a Markov chain with
the correct equilibrium distribution capable of producing samples from the
posterior p(θ|y) we are interested in.
Gibbs sampling dates to the work of Geman and Geman (1984) in im-
age analysis and is related to the EM algorithm (Dempster, Laird and
Rubin, 1977) which dealt with maximum likelihood and Bayesian estima-
tion with missing information. Assume a parameter vector θ = (θ1 , θ2 ),
a prior p(θ), and likelihood l(θ|y), that produces a posterior distribution
p(θ|y) = cp(θ)l(θ|y), with c a normalizing constant. It is often the case that
.::UdecomBooks::.
the posterior distribution over all parameters is difficult to work with. On

the other had, if we partition our parameters into two sets θ1 , θ2 and had ini-
tial estimates for θ1 (treated like missing information in the EM algorithm),
we could estimate θ2 conditional on θ1 using p(θ2 |y, θ̂1 ). (Presumably, this
estimate is much easier to derive. We will provide details illustrating this
case in section 6.2.) Denote the estimate, θ̂2 derived by using the posterior
mean or mode of p(θ2 |y, θ̂1 ), and consider than we are now able to construct
a new estimate of θ1 based on the conditional distribution p(θ1 |y, θ̂2 ), which
can be used to construct another value for θ2 , and so on.
In general, for the case of k parameters, the algorithm can be summarized
as:
Initialize θ0
Repeat {
Sample θ1t+1 ∼ p[θ1 |y, (θ2t , . . . , θkt )]

Sample θ2t+1 ∼ p[θ2 |y, (θ1t+1 , θ3t , . . . , θkt )]
..
.
Sample θkt+1 ∼ p[θk |y, (θ1t+1 , θ2t+1 , . . . , θk−1
t+1
)]
t=t+1
Geman and Geman (1984) demonstrated that the stochastic process θt

from this approach to sampling the complete sequence of conditional distri-
butions represents a Markov chain with the correct equilibrium distribution.
Gibbs sampling is in fact closely related to Hastings and Metropolis MCMC
methods. This chapter only deals with Gibbs sampling, providing a simple
introduction that draws heavily on work by Geweke (1993), who introduced
the approach presented here in Section 6.3 for dealing with heteroscedastic
regression models.
Section 6.1 introduces a simple Bayesian regression example, and the fol-
lowing section takes the reader through an application and computational
code necessary to carry out Gibbs sampling estimation. MATLAB func-
tions for diagnosing convergence of the sampling draws are the subject of
Section 6.2. The Bayesian regression model from Section 6.1 is extended
in Section 6.3 to the case of t−distributed errors that can accommodate
outliers and non-constant variance. A Markov Chain Monte Carlo model
for estimating autoregressive models with stability restrictions imposed on
.::UdecomBooks::.
the parameters is used in Section 6.4 to illustrate construction of MATLAB

functions that carry out Gibbs sampling estimation of econometric models.
In Section 6.5 we illustrate a ‘Metropolis step within Gibbs sampling’ al-
gorithm that is useful in many Gibbs sampling models. Metropolis sampling
represents a special case of the Hastings method where the proposal distri-
bution is restricted to be symmetric. The Metropolis (or Hastings) within
Gibbs sampling approach is frequently used when one encounters a condi-
tional distribution that is not readily identified. In this case, we cannot draw
samples using a conventional distribution and must rely on the Metropolis
(or Hastings) algorithm to carry out these draws within the sequence of
Gibbs sampling.
A concluding section briefly describes alternative econometric estimation
methods that have been implemented in the gibbs sampling library. Gibbs
sampling for probit and tobit model estimation is deferred until Chapter 7
where limited dependent variable models are discussed.
6.1 The Bayesian Regression Model

As an introduction to Gibbs sampling for Bayesian regression models, we
consider the case of a linear regression model with an informative prior. The
multiple normal linear regression model can be written as in (6.2).
y = Xβ + ε (6.2)
ε ∼ N (0, σ In ) 2
Where y is an nx1 vector of dependent variables and X represents the nxk

matrix of explanatory variables. We assume that ε is an nx1 vector of
independent identically distributed normal random variates.
The parameters to be estimated in (6.2) are (β, σ), for which we assign
a prior density of the form π(β, σ) = π1 (β)π2 (σ). That is, we assume our
prior for the parameters β is independent from that for the parameter σ.
The normal prior density we assign to β requires that we specify a mean
and variance for these parameters which can be expressed in terms of linear
combinations of β as shown in (6.3).
Rβ ∼ N (r, T ) ⇔ π1 (β) ∝ exp{−(1/2)(Rβ − r)0 T −1 (Rβ − r)} (6.3)
Where the number of linear combinations of β can be equal to m ≤ k, so

that R is an mxk matrix that establishes the m linear relations, r is an mx1
.::UdecomBooks::.
vector containing the prior means and T is an mxm matrix containing the
prior variances and covariances.
As is well-known, when m < k, the prior in (6.3) is improper, but can
be justified as the limiting case of a set of proper priors. For our purposes
it is convenient to express (6.3) in an alternative (equivalent) form based on
a factorization of T −1 into Q0 Q = T −1 , and q = Qr leading to (6.4).
Qβ ∼ N (q, Im ) ⇔ π1 (β) ∝ exp{−(1/2)(Qβ − q)0 (Qβ − q)} (6.4)
For simplicity, we assume the diffuse prior for σ, π2 (σ) ∝ (1/σ), but all
of our results would follow for the case of an informative conjugate gamma
prior for this parameter. Following the usual Bayesian methodology, we
combine the likelihood function for our simple model:
L(β, σ) ∝ (1/σ n )exp[(y − Xβ)0 (y − Xβ)/2σ 2 ] (6.5)

with the priors π1 (β) and π2 (σ) to produce the posterior density for (β, σ)
shown in (6.6).
p(β, σ) ∝ (1/σ n+1 )exp[β − β̂(σ)]0 [V (σ)]−1 [β − β̂(σ)] (6.6)

0 2 0 −1 0 2 0
β̂(σ) = (X X + σ Q Q) (X y + σ Q q)
V (σ) = σ (X X + σ Q Q)−1
2 0 2 0
In (6.6), we have used the notation β̂(σ) to convey that the mean of the
posterior, β̂, is conditional on the parameter σ, as is the variance, denoted
by V (σ). This single parameter prevents analytical solution of the Bayesian
regression problem. In order to overcome this problem, Theil and Gold-
berger (1961) observed that conditional on σ, the posterior density for β is
multivariate normal. They proposed that σ 2 be replaced by an estimated
value, σ̂ 2 = (y − X β̂)0 (y − X β̂)/(n − k), based on least-squares estimates β̂.
The advantage of this solution is that the estimation problem can be solved
using existing least-squares regression software. Their solution produces a
point estimate which we label β̂T G and an associated variance-covariance
estimate, both of which are shown in (6.7). This estimation procedure is
implemented by the function theil discussed in Chapter 4.
β̂T G = (X 0 X + σ̂ 2 Q0 Q)−1 (X 0 y + σ̂ 2 Q0 q) (6.7)

2 0 2 0 −1
var(β̂T G ) = σ̂ (X X + σ̂ Q Q)
.::UdecomBooks::.
6.2 The Gibbs Sampler

The Gibbs sampler provides a way to sample from a multivariate probability
density based only on the densities of subsets of vectors conditional on all
others. The attractiveness of this approach is that it provides a solution
to the Bayesian multiple integration problem when the conditional densities
are simple and easy to obtain.
The Gibbs sampling approach set forth here specifies the complete con-
ditional distributions for all parameters in the model and proceeds to carry
out random draws from these distributions to collect a large sample of pa-
rameter draws. Gelfand and Smith (1990) demonstrate that Gibbs sampling
from the sequence of complete conditional distributions for all parameters
in the model, we achieve a set of draws that converge in the limit to the
true (joint) posterior distribution of the parameters. That is, despite the
use of conditional distributions in our sampling scheme, a large sample of
the draws can be used to produce valid posterior inferences about the mean
and moments of the parameters (or any function of the parameters) one is
interested in.
The method is most easily described by developing and implementing a
two-step Gibbs sampler for the posterior distribution (6.6) of our Bayesian
regression model based on the distribution of β conditional on σ and the
distribution of σ conditional on β. For our regression problem, the posterior
density for β conditional on σ, p(β|σ), is multivariate normal with mean
equal to (6.8) and variance as indicated in (6.9).
β̂(σ) = (X 0 X + σ 2 Q0 Q)−1 (X 0 y + σ 2 Q0 q) (6.8)

2 0 2 0 −1
V (σ) = σ (X X + σ Q Q) (6.9)
The posterior density for σ conditional on β, p(σ|β) is:
p(σ|β) ∝ (1/σ (n+1) )exp[−(y − Xβ)0 (y − Xβ)] (6.10)

Which can be manipulated to obtain:
[(y − Xβ)0 (y − Xβ)/σ 2 ] | β ∼ χ2 (n) (6.11)

The Gibbs sampler suggested by these two conditional posterior distri-
butions involves the following computations.
1. Begin with arbitrary values for the parameters β 0 and σ 0 , which we

designate with the superscript 0.
.::UdecomBooks::.
2. Compute the mean and variance of β using (6.8) and (6.9) conditional
on the initial value σ 0 .
3. Use the computed mean and variance of β to draw a multivariate

normal random vector, which we label β 1 .
4. Use the value β 1 along with a random χ2 (n) draw to determine σ 1

using (6.11).
The above four steps are known as a ‘single pass’ through our (two-step)
Gibbs sampler, where we have replaced the initial arbitrary values of β 0 and
σ 0 with new values labeled β 1 and σ 1 . We now return to step 1 using the
new values β 1 and σ 1 in place of the initial values β 0 and σ 0 , and make
another ‘pass’ through the sampler. This produces a new set of values, β 2
and σ 2 .
Gelfand and Smith (1990) outline fairly weak conditions under which
continued passes through our Gibbs sampler will produce a distribution of
(β i , σ i ) values that converges to the joint posterior density in which we are
interested, p(β, σ). Given independent realizations of β i , σ i , the strong law
of large numbers suggests we can approximate the expected value of the β, σ
parameters using averages of these sampled values.
To illustrate the Gibbs sampler for our Bayesian regression model, we
generate a regression model data set containing 100 observations and 3 ex-
planatory variables; an intercept term, and two uncorrelated explanatory
variables generated from a standard normal distribution. The true values
of β0 for the intercept term, and the two slope parameters β1 and β2 , were
set to unity. A standard normal error term (mean zero, variance equal to
unity) was used in generating the data.
The prior means for the β parameters were set to unity and the prior
variance used was also unity, indicating a fair amount of uncertainty. The
following MATLAB program implements the Gibbs sampler for this model.
% ----- Example 6.1 A simple Gibbs sampler

n=100; k=3; % set nobs and nvars
x = randn(n,k); b = ones(k,1); % generate data set
y = x*b + randn(n,1);
r = [1.0 1.0 1.0]’; % prior means
R = eye(k); T = eye(k); % prior variance
nu = 0; d0 = 0; % diffuse prior for sige
Q = chol(inv(T)); Qpr = Q*r;
b0 = (x’*x)\(x’*y); % use ols as initial values
sige = (y-x*b0)’*(y-x*b0)/(n-k);
ndraw = 1100; nomit = 100; % set the number of draws
.::UdecomBooks::.
bsave = zeros(ndraw,k); % allocate storage for results

ssave = zeros(ndraw,1);
tic;
for i=1:ndraw; % Start the sampling
xpxi = inv(x’*x + sige*Q); xpy = (x’*y + sige*Qpr);
% update b
b = xpxi*xpy;
a = chol(xpxi);
b = sqrt(sige)*a’*randn(k,1) + b;
bsave(i,:) = b’;
% update sige
nu1 = n + nu;
e = y - x*b;
d1 = d0 + e’*e;
chi = chis_rnd(1,nu1);
sige = d1/chi;
ssave(i,1) = sige;
end; % End the sampling
toc;
bhat = mean(bsave(nomit+1:ndraw,:)); % calculate means and std deviations
bstd = std(bsave(nomit+1:ndraw,:)); tstat = bhat./bstd;
sighat = mean(ssave(nomit+1:ndraw,1));
tout = tdis_prb(tstat’,n); % compute t-stat significance levels
% set up for printing results
in.cnames = strvcat(’Coefficient’,’t-statistic’,’t-probability’);
in.rnames = strvcat(’Variable’,’variable 1’,’variable 2’,’variable 3’);
in.fmt = ’%16.6f’;
tmp = [bhat’ tstat’ tout];
fprintf(1,’Gibbs estimates \n’); % print results
mprint(tmp,in);
result = theil(y,x,r,R,T); % compare to Theil-Goldberger estimates
prt(result);
We rely on MATLAB functions norm rnd and chis rnd to provide the
multivariate normal and chi-squared random draws which are part of the
distributions function library discussed in Chapter 9. Note also, we omit
the first 100 draws at start-up to allow the Gibbs sampler to achieve a
steady state before we begin sampling for the parameter distributions.
The results are shown below, where we find that it took only 1.75 seconds
to carry out the 1100 draws and produce a sample of 1000 draws on which
we can base our posterior inferences regarding the parameters β and σ. For
comparison purposes, we produced estimates using the theil function from
the regression function library.
elapsed_time =
1.7516
Gibbs estimates
.::UdecomBooks::.

variable 1 0.725839 5.670182 0.000000
variable 2 0.896861 7.476760 0.000000
variable 3 1.183147 9.962655 0.000000

R-squared = 0.6598
sigma^2 = 1.4518
***************************************************************
variable 1 1.000000 1.000000
variable 2 1.000000 1.000000
variable 3 1.000000 1.000000
***************************************************************
Posterior Estimates
variable 1 0.724510 4.653860 0.000010
variable 2 0.898450 6.231785 0.000000
variable 3 1.188448 8.840263 0.000000
6.2.1 Monitoring convergence of the sampler

An important issue in using Gibbs sampling is convergence of the sampler to
the posterior distribution. We know from theory that the sampler converges
in the limit as n → ∞, but in any applied problem one must determine how
many draws to make with the sampler. Ad-hoc convergence tests have been
used in applied work and found to work well in simple regression models of
the type considered here. For example, Smith and Roberts (1992) proposed
a test they label the ‘felt-tip pen test’, that compares smoothed histograms
or distributions from earlier draws in the sequence of passes through the
sampler to those from later draws. If the two distributions are similar,
within the tolerance of the felt-tip pen, convergence is assumed to have
taken place.
There is some evidence that linear regression models exhibit rapid con-
vergence, confirmed in our example by the Raftery and Lewis (1992a) proce-
dure which we demonstrate below. It should be noted that once convergence
occurs, we need to continue making passes to build up a sample from the
posterior distribution which we use to make inferences about the parame-
ters. As one might expect, convergence is a function of how complicated
the set of conditional distributions are. For example, Geweke (1992) found
that Gibbs sampling the Tobit censured regression model produced poor
.::UdecomBooks::.
results with 400 passes and much better results with 10,000 passes. We will
illustrate Tobit censured regression in Chapter 7.
Best et al., 1995 provide a set of Splus functions that implement six
different MCMC convergence diagnostics, some of which have been imple-
mented in a MATLAB function coda. This function provides: autocorre-
lation estimates, Rafterty-Lewis (1995) MCMC diagnostics, Geweke (1992)
NSE, (numerical standard errors) RNE (relative numerical efficiency) esti-
mates, Geweke Chi-squared test on the means from the first 20% of the
sample versus the last 50%. We describe the role of each of these diagnostic
measures using an applied example.
First, we have implemented the Gibbs sampler for the Bayesian regres-
sion model as a MATLAB function ols g that returns a structure variable
containing the draws from the sampler along with other information. Details
regarding this function are presented in Section 6.4.
Some options pertain to using the function to estimate heteroscedastic
linear models, a subject covered in the next section. For our purposes we can
use the function to produce Gibbs samples for the Bayesian homoscedastic
linear model by using a large value of the hyperparameter r. Note that this
function utilizes a structure variable named ‘prior’ to input information to
the function. Here is an example that uses the function to produce Gibbs
draws that should be similar to those illustrated in the previous section
(because we use a large hyperparameter value of r = 100).
% ----- Example 6.2 Using the coda() function

n=100; k=3; % set number of observations and variables
randn(’seed’,10101);
randn(’seed’,20201); y = x*b + randn(n,1);
r = [1.0 1.0 1.0]’; % prior b means
T = eye(k); % prior b variance
rval = 100; % homoscedastic prior for r-value
prior.beta = r;
prior.bcov = T;
prior.rval = rval;
result = ols_g(y,x,prior,ndraw,nomit);
vnames = strvcat(’beta1’,’beta2’,’beta3’);
coda(result.bdraw,vnames);
The sample Gibbs draws for the parameters β are in the results structure
variable, result.bdraw which we send down to the coda function to produce
convergence diagnostics. This function uses a MATLAB variable ‘nargout’
to determine if the user has called the function with an output argument. If
.::UdecomBooks::.
so, the function returns a result structure variable that can be printed later
using the prt (or prt gibbs) functions. In the case where the user sup-
plies no output argument, (as in the example code above) the convergence
diagnostics will be printed to the MATLAB command window.
Note that if we wished to analyze convergence for the estimates of the σ
parameters in the model, we could call the function with these as arguments
in addition to the draws for the β parameters, using:
coda([result.bdraw result.sdraw]);
A partial listing of the documentation for the function coda is shown

below, where the form of the structure variable returned by coda when
called with an output argument has been eliminated to save space.
PURPOSE: MCMC convergence diagnostics, modeled after Splus coda

------------------------------------------------------
USAGE: coda(draws,vnames,info)
or: result = coda(draws)
where: draws = a matrix of MCMC draws (ndraws x nvars)
vnames = (optional) string vector of variable names (nvar x 1)
info = (optional) structure setting input values
info.q = Raftery quantile (default = 0.025)
info.r = Raftery level of precision (default = 0.005)
info.s = Rafterty probability for r (default = 0.950)
info.p1 = 1st % of sample for Geweke chi-sqr test (default = 0.2)
info.p2 = 2nd % of sample for Geweke chi-sqr test (default = 0.5)
------------------------------------------------------
NOTES: you may supply only some of the info-structure arguments
the remaining ones will take on default values
------------------------------------------------------
RETURNS: output to command window if nargout = 0
autocorrelation estimates
Rafterty-Lewis MCMC diagnostics
Geweke NSE, RNE estimates
Geweke chi-sqr prob on means from info.p1 vs info.p2
a results structure if nargout = 1
The results from executing the example program as are follows:
MCMC CONVERGENCE diagnostics

Based on sample size = 1000
Autocorrelations within each parameter chain
Variable Lag 1 Lag 5 Lag 10 Lag 50
beta1 0.031 0.009 -0.071 -0.058
beta2 0.024 -0.041 -0.004 0.072
beta3 -0.023 0.019 0.029 0.006
.::UdecomBooks::.
Raftery-Lewis Diagnostics for each parameter chain

(q=0.0250, r=0.005000, s=0.950000)
Variable Thin Burn Total(N) (Nmin) I-stat
beta1 1 2 3869 3746 1.033
beta2 1 2 3869 3746 1.033
beta3 1 2 3869 3746 1.033
Geweke Diagnostics for each parameter chain

Variable Mean std dev NSE iid RNE iid
beta1 1.041488 0.111201 0.003516 1.000000
beta2 0.980489 0.104348 0.003300 1.000000
beta3 0.938445 0.113434 0.003587 1.000000
Variable NSE 4% RNE 4% NSE 8% RNE 8% NSE 15% RNE 15%

beta1 0.003529 0.992811 0.003092 1.293611 0.002862 1.509356
beta2 0.003653 0.816099 0.003851 0.734157 0.003594 0.842994
beta3 0.003135 1.309284 0.002790 1.652990 0.003150 1.296704
Geweke Chi-squared test for each parameter chain

First 20% versus Last 50% of the sample
Variable beta1
NSE estimate Mean N.S.E. Chi-sq Prob
i.i.d. 1.040547 0.004257 0.842837
4% taper 1.040811 0.004374 0.829737
8% taper 1.040860 0.003910 0.808189
15% taper 1.040747 0.003354 0.782689
Variable beta2
i.i.d. 0.980194 0.004015 0.078412
4% taper 0.980483 0.004075 0.090039
8% taper 0.980873 0.003915 0.088567
15% taper 0.982157 0.003365 0.095039
Variable beta3
i.i.d. 0.940191 0.004233 0.961599
4% taper 0.940155 0.003972 0.957111
8% taper 0.940150 0.003735 0.954198
15% taper 0.940160 0.003178 0.946683
The role of each convergence diagnostic measure is described in the fol-

lowing sub-sections.
.::UdecomBooks::.
6.2.2 Autocorrelation estimates

The role of the time-series autocorrelation estimates is to provide an indica-
tion of how much independence exists in the sequence of each β parameter
draws. From time-series analysis we know that if β, i = 1, . . . , n is a sta-
P
tionary correlated process, then β̄ = (1/n) ni=1 βi is a consistent estimate
of E(β) as n → ∞, so it is permissible to simulate correlated draws from
the posterior distribution in order to summarize the features of the poste-
rior. This is provided that we produce a large enough sample of draws, and
the amount of correlation plays a role in determining the number of draws
necessary. For example, if β follows an AR(1) process with serial correla-
tion parameter ρ, mean β̄ and standard deviation σ, we know that β̄ has a
standard deviation given by:
q
σβ = σ/(n)1/2 (1 + ρ)/(1 − ρ) (6.12)
If ρ = 0, so that the draws represent an independent identically dis-
tributed (iid) process, the standard deviation has the usual form: σ/(n)1/2 ,
but if ρ = 0.9, the standard errorp becomes quite large. In this case, we
would have to run the sampler 1.9/0.1 = 4.35 times longer than if the
Gibbs draws represented iid draws. A high degree of autocorrelation indi-
cates that we may need to carry out more draws to achieve a sample of
sufficient size to draw accurate posterior estimates.
The coda results indicate that our draws for the parameters β exhibit
small autocorrelations at lags 1, 5, 10 and 50, so we need not be concerned
about this problem in our particular application.
6.2.3 Raftery-Lewis diagnostics

Raftery and Lewis (1992a, 1992b, 1995) proposed a set of diagnostics that
they implemented in a FORTRAN program named Gibbsit, which were
converted to a MATLAB function raftery. This function is called by coda,
but can also be used independently of coda. Given some output from a
Gibbs (or MCMC) sampler, Raftery and Lewis provide an answer regarding
how long to monitor the chain of draws that is based on the accuracy of the
posterior summaries desired by the user. They require that the user specify
three pieces of information that are set to default values by coda, or can be
user-defined using the ‘info’ structure variable as an input argument to the
function.
1. Which quantiles of the marginal posteriors are you interested in? Usu-
ally the answer is the 2.5% and 97.5% points, because these are the
.::UdecomBooks::.
basis for a 95% interval estimate. This information is set using ‘info.q’,
which has a default value of 0.025.
2. What is the minimum probability needed to archive the accuracy goals.

A default value of 95% is used, or can be set by the user with ‘info.s’,
which has a default value of 0.95.
3. How much accuracy is desired in the estimated quantiles? Raftery and

Lewis specify this using the area to the left (in the case of info.q =
0.025) or right (in the case of info.q=0.975) of the reported quantile in
the CDF. By default info.r=0.005, so that nominal reporting based on
a 95% interval using the 0.025 and 0.975 quantile points should result
in actual posterior values that lie between 0.95 and 0.96.
Given our draws for β, raftery dichotomizes the draws using a binary
time-series that is unity if βi ≤ ‘info.q’ and zero otherwise. This binary
chain should be approximately Markovian so standard results for two-state
Markov chains can be used to estimate how long the chain should be run to
achieve the desired accuracy for the chosen quantile ‘info.q’.
The function coda prints out three different estimates from the raftery
function. A thinning ratio which is a function of the amount of autocorrela-
tion in the draws, the number of draws to use for ‘burn-in’ before beginning
to sample the draws for purposes of posterior inference, and the total number
of draws needed to achieve the accuracy goals.
Some terminology that will help to understand the raftery output. It is
always a good idea to discard a number of initial draws referred to as “burn-
in” draws for the sampler. Starting from arbitrary parameter values makes
it unlikely that initial draws come from the stationary distribution needed
to construct posterior estimates. Another practice followed by researchers
involves saving only every third, fifth, tenth, etc. draw since the draws from
a Markov chain are not independent. This practice is labeled “thinning”
the chain. Neither thinning or burn-in are mandatory in Gibbs sampling
and they tend to reduce the effective number of draws on which posterior
estimates are based.
From the coda output, we see that the thinning estimate provided by
raftery in the second column is 1, which is consistent with the lack of
autocorrelation in the sequence of draws. The third column reports that
only 2 draws are required for burn-in, which is quite small. Of course, we
started our sampler using the default least-squares estimates provided by
ols g which should be close to the true values of unity used to generate the
regression data set. In the fourth column, we find the total number of draws
.::UdecomBooks::.
needed to achieve the desired accuracy for each parameter. This is given as
3869 for β1 , β2 and β3 , which exceeds the 1,000 draw we used, so it would
be advisable to run the sampler again using this larger number of draws.
On the other hand, a call to the function raftery with the desired accu-
racy (‘info.r’) set to 0.01, so that nominal reporting based on a 95% interval
using the 0.025 and 0.975 quantile points should result in actual posterior
values that lie between 0.95 and 0.97, produces the results shown below.
These indicate that our 1,000 draws would be adequate to produce this
desired level of accuracy for the posterior.
% ----- Example 6.3 Using the raftery() function

q = 0.025;
r = 0.01;
s = 0.95;
res = raftery(result.bdraw,q,r,s);
prt(res,vnames);
Raftery-Lewis Diagnostics for each parameter chain
(q=0.0250, r=0.010000, s=0.950000)
Variable Thin Burn Total(N) (Nmin) I-stat
beta1 1 2 893 937 0.953
beta2 1 2 893 937 0.953
beta3 1 2 893 937 0.953
The Nmin reported in the fifth column represents the number of draws
that would be needed if the draws represented an iid chain, which is virtually
true in our case. Finally, the i−statistic is the ratio of the fourth to the fifth
column. Raftery and Lewis indicate that values exceeding 5 for this statistic
are indicative of convergence problems with the sampler.
6.2.4 Geweke diagnostics

The function coda also produces estimates of the numerical standard errors
(NSE) and relative numerical efficiency (RNE) proposed by Geweke (1992).
Using spectral analysis of time-series methods, we can produce an estimate
of the variance of the β parameters we are interested in based on the sampled
values using:
var(β̂i ) = S(0)/k (6.13)

where S(0) is the spectral density of βi evaluated at ω = 0. Issues arise in
how one approximates S(ω), so alternative tapering of the spectral window
is used. The coda function reports estimates of the NSE and RNE based
on 4%, 8% and 15% tapering or truncation of the periodgram window. The
.::UdecomBooks::.
MATLAB functions that implement these calculations are adaptions of rou-

tines provided by Geweke and Chib, that can be found on the internet at
http://www.econ.umn.edu/bacc.
The first set of NSE and RNE estimates reported are based on the as-
sumption that the draws come from an iid process. These are reported along
with the means and standard deviations of the chain of draws. A second set
of NSE and RNE estimates are reported based on alternative tapering of
the spectral window, where the non-iid nature of the draws is taken into ac-
count by the NSE and RNE estimates. Dramatic differences between these
estimates would lead one to rely on the latter set of estimates, as these
differences would reflect autocorrelation in the draws.
The RNE estimates provide an indication of the number of draws that
would be required to produce the same numerical accuracy if the draws rep-
resented had been made from an iid sample drawn directly from the posterior
distribution. In our example, the RNE’s are close to unity, indicative of the
iid nature of our sample. RNE estimates greater than unity, say around 3,
would indicate that only 33% of the number of draws would be required to
achieve the same accuracy from an iid set of draws.
These results are produced by a call to the MATLAB function mo-
mentg, which is called by coda. As with the function raftery, this function
can be called by itself without invoking coda. As an example:
% ----- Example 6.4 Geweke’s convergence diagnostics

result = ols_g(y,x,prior,ndraw,nomit);
vnames = strvcat(’beta1’,’beta2’,’beta3’);
geweke = momentg(result.bdraw);
prt(geweke,vnames);
Geweke Diagnostics for each parameter chain
Variable Mean std dev NSE iid RNE iid
beta1 1.040659 0.111505 0.003526 1.000000
beta2 0.980891 0.104660 0.003310 1.000000
beta3 0.938394 0.113805 0.003599 1.000000
Variable NSE 4% RNE 4% NSE 8% RNE 8% NSE 15% RNE 15%

beta1 0.003509 1.009612 0.003309 1.135733 0.003229 1.192740
beta2 0.003423 0.934811 0.003453 0.918728 0.003194 1.073586
beta3 0.003235 1.237219 0.002876 1.565849 0.003250 1.225889
A second set of diagnostics suggested by Geweke are printed by coda

after the NSE and RNE estimates. These diagnostics represent a test of
whether the sample of draws has attained an equilibrium state based on the
means of the first 20% of the sample of draws versus the last 50% of the
sample. If the Markov chain of draws from the Gibbs sampler has reached
.::UdecomBooks::.
an equilibrium state, we would expect the means from these two splits of
the sample to be roughly equal. A Z−test of the hypothesis of equality of
these two means is carried out and the chi-squared marginal significance is
reported. For our illustrative example, the second β parameter does not
fair well on these tests. We cannot reject the hypothesis of equal means at
the 95% level of significance, but we can at the 90% level. Increasing the
number of draws to 4,100 (suggested by the Rafterty and Lewis diagnostics)
and discarding the first 100 for burn-in produced the following results for
the chi-squared test of equality of the means from the first 20% versus the
last 50% of the 4,000 draws.

First 20% versus Last 50% of the sample
Variable beta1
i.i.d. 1.042974 0.002132 0.453998
4% taper 1.042788 0.002274 0.461874
8% taper 1.042987 0.002008 0.428475
15% taper 1.043294 0.001643 0.394748
Variable beta2
i.i.d. 0.982636 0.001962 0.596623
4% taper 0.982807 0.001856 0.612251
8% taper 0.982956 0.001668 0.623582
15% taper 0.982959 0.001695 0.630431
Variable beta3
i.i.d. 0.942324 0.002137 0.907969
4% taper 0.942323 0.001813 0.891450
8% taper 0.942284 0.001823 0.885642
15% taper 0.942288 0.001936 0.892751
Here we see that the means are equal, indicating no problems with con-
vergence. The coda function allows the user to specify the proportions of
the sample used to carry out this test as ‘info.p1’ and ‘info.p2’ in the struc-
ture variable used to input user-options to coda. The default values based
on the first 20% of the sample versus the last 50% are values used by the
Splus version of CODA.
The chi-squared tests are implemented by a call inside coda to a MAT-
LAB function apm. This function allows one to produce posterior moment
estimates that represent an average over two sets of draws. This function
can be used without invoking coda and would be useful in cases where one
.::UdecomBooks::.
wished to combine smaller samples from multiple MCMC or Gibbs sam-

pling runs and examine the impact of the additional draws on NSE or test
for equality of the means from the two samples. The documentation for
apm is:
PURPOSE: computes Geweke’s chi-squared test for

two sets of MCMC sample draws
------------------------------------------------
USAGE: result = apm(results1,results2)
where: results1 = a structure returned by gmoment
results2 = a structure returned by gmoment
------------------------------------------------
results(i).pmean(k) = posterior mean for variable i
for k = nse, nse1,nse2,nse3
results(i).nse(k) = nse for variable i
results(i).prob(k) = chi-sq test prob for variable i
------------------------------------------------
As an illustration, suppose we started our ols g sampler at different

starting values. A structure variable can be used as input to ols g to set
starting values for the sampler as illustrated in Example 6.5 below. Each call
to the function will also use different seeds for the random number generators
that produce the normal and chi-squared random deviate draws. To test
convergence, we call the apm function with results structures returned by
momentg based on the two sets of draws. We then use prt to print the
means, NSE and chi-squared test results.
% ----- Example 6.5 Using the momentg() function

y = x*b + randn(n,1);
ndraw1 = 600; ndraw2 = 1100; nomit = 100;
r = [1.0 1.0 1.0]’; % prior b means
R = eye(k); T = eye(k); % prior b variance
rval = 100; % homoscedastic prior for r-value
prior.beta = r; prior.bcov = T;
prior.rmat = R; prior.rval = rval;
start1.b = zeros(k,1); start1.sig = 1.0; start1.V = ones(n,1);
result1 = ols_g(y,x,prior,ndraw1,nomit,start1);
gres1 = momentg(result1.bdraw);
start2.b = ones(k,1); start2.sig = 10.0; start2.V = start1.V;
.::UdecomBooks::.
result2 = ols_g(y,x,prior,ndraw2,nomit,start2);
gres2 = momentg(result2.bdraw);
result = apm(gres1,gres2);
prt(result)
First 33 % versus Last 67 % of the sample
Variable variable 1
NSE estimate Mean N.S.E. Equality chi sq
i.i.d. 1.04442600 0.00285852 0.6886650
4% taper 1.04414401 0.00303774 0.6864577
8% taper 1.04428137 0.00303411 0.6934830
15% taper 1.04455890 0.00267794 0.6867602
Variable variable 2
i.i.d. 0.97690447 0.00271544 0.9589017
4% taper 0.97686937 0.00233080 0.9581417
8% taper 0.97684701 0.00199957 0.9586626
15% taper 0.97683040 0.00172392 0.9614142
Variable variable 3
i.i.d. 0.93683408 0.00298336 0.7947394
4% taper 0.93733319 0.00260842 0.7662486
8% taper 0.93728792 0.00243342 0.7458718
15% taper 0.93728776 0.00227986 0.7293916
Another useful function for examining MCMC output is pltdens, a func-

tion from the graphing library that produces density plots. This function
is part of a public domain statistics toolbox written by Anders Holtsberg
with the documentation altered to conform to that used by all Economet-
rics Toolbox functions. Samples of MCMC draws can be used to produce
posterior density plots with a simple call such as:
pltdens(result.bdraw(:,1));
% demo of pltdens options
bandwidth = 0.2; % a kernel density smoothing parameter option
positive = 1; % a flag for densities with zero mass at negative values
kerneltype = 1; % a Gaussian kernel type
pltdens(result.sdraw,bandwidth,positive,kerneltype);
6.3 A heteroscedastic linear model

There is a history of Bayesian literature that deals with heteroscedastic
and leptokurtic disturbances. A non-Bayesian regression methodology was
introduced by Lange, Little and Taylor (1989), which assumes an indepen-
dent Student-t distribution for the regression disturbances. Geweke (1993)
.::UdecomBooks::.
provides an equivalent Bayesian approach which he labels a heteroscedastic

normal linear regression model. We adopt the approach of Geweke (1993)
to extend our basic model from Section 6.2.
The regression model we wish to implement is shown in (6.14).
y = Xβ + ε (6.14)
ε ∼ N(0, σ V ),
2
V = diag(v1 , v2 , . . . , vn )
β ∼ N (c, T )
σ ∼ (1/σ)
r/vi ∼ ID χ2 (r)/r
r ∼ Γ(m, k)
Where y is an nx1 vector of dependent variable observations and X is an

nxk matrix of explanatory variables. As before, we place a multivariate
normal prior on β and a diffuse prior on σ. The parameters to be estimated
are β, σ and the relative variance terms (v1 , v2 , . . . , vn ), which are assumed
fixed but unknown. The thought of estimating n parameters, v1 , v2 , . . . , vn ,
in addition to the k+1 parameters, β, σ using n data observations may seem
quite problematical. However, a Bayesian approach is taken which assigns
an independent χ2 (r)/r prior distribution to the vi terms that depends on
a single hyperparameter r. This allows us to estimate these additional n
model parameters by adding the single parameter r to our model estimation
procedure.
This type of prior has been used by Lindley (1971) for cell variances
in an analysis of variance problem, and Geweke (1993) in modeling het-
eroscedasticity and outliers. The specifics regarding the prior assigned to
the Vi terms can be motivated by considering that the mean of prior, which
we designate Ep equals unity, that is: Ep (1/vij ) = 1, and the prior vari-
ance also designated with the subscript p is: varp (1/vij ) = 2/r. This im-
plies that as r becomes very large, the prior reflects the special case where
εi ∼ N (0, σ 2 In ). We will see that the role of Vi 6= In is to robustify against
outliers and observations containing large variances by downweighting these
observations. Large r values are associated with a prior belief that outliers
and non-constant variances do not exist.
We choose to control the values assigned to the hyperparameter r by
assigning a Γ(m, k) prior distribution to this parameter. This distribution
has a mean of m/k and variance m/k 2 , so using m = 8, k = 2 would assign
a prior to r centered on a small r = 4 with variance of r equal to 2. For
.::UdecomBooks::.
small values of r, we can see the impact of the prior distribution assumed
for vij by considering that, the mean of the prior is r/(r − 2) and the mode
of the prior equals r/(r + 2). Small values of the hyperparameter r allow the
vij to take on a skewed form where the mean and mode are quite different.
This is illustrated in Figure 6.1 where distributions for vi associated with
various values of the parameter r are presented.
1.4
r=2
r=5
1.2
r=20
1
prior probability density
0.8
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3
Vi values
Figure 6.1: Prior Vi distributions for various values of r
As an intuitive motivation for the conditional distributions in this model,

note that, given values for V = diag(v1 , . . . , vn ), we could proceed to con-
struct estimates for β and σ using a version of the Theil-Goldberger mixed
estimator based on generalized least-squares. It is often the case with the
Gibbs sampling methodology that complex estimation problems are simpli-
fied considerably by conditioning on unknown parameters, that is, assuming
these values are known.
The conditional posterior density for β is:
.::UdecomBooks::.
β|(σ, V ) ∼ N [H(X 0 V −1 y + σ 2 R0 T −1 c) , σ 2 H]. (6.15)

0 −1 0 −1 −1
H = (X V X +RT R)
If we let ei = yi − x0i β, the conditional posterior density for σ is
X
n
[ (e2i /vi )/σ 2 ]|(β, V ) ∼ χ2 (n) (6.16)
i=1
This result parallels our simple case from section 6.2 where we adjust the ei
using the relative variance terms vi .
Geweke (1993) shows that the posterior distribution of V conditional on
(β, σ) is proportional to:
[(σ −2 e2i + r)/vi ]|(β, σ) ∼ χ2 (r + 1) (6.17)

Given the three conditional posterior densities in (6.15), through (6.17),
we can formulate a Gibbs sampler for this model using the following steps:
1. Begin with arbitrary values for the parameters β 0 , σ 0 , vi0 and r0 , which
we designate with the superscript 0.
2. Compute the mean and variance of β using (6.15) conditional on the

initial values σ 0 , vi0 and r0 .
3. Use the computed mean and variance of β to draw a multivariate

normal random vector, which we label β 1 .
4. Calculate expression (6.16) using β 1 determined in step 3 and use this

value along with a random χ2 (n) draw to determine σ 1 .
5. Using β 1 and σ 1 , calculate expression (6.17) and use the value along
with an n−vector of random χ2 (r + 1) draws to determine vi , i =
1, . . . , n.
6. Draw a Γ(m, k) value to update r0 .
These steps constitute a single pass of the Gibbs sampler. As before, we

wish to make a large number of passes to build up a sample (β j , σ j , vij , rj )
of j values from which we can approximate the posterior distributions for
our parameters.
The implementation of the heteroscedastic linear model is similar to our
previous example from Section 6.2 with the addition of an updating step
.::UdecomBooks::.
to handle draws for the vi parameters and the Γ(m, k) draw to update the
hyperparameter r. To demonstrate MATLAB code for this model, we gener-
ate a heteroscedastic data set by multiplying normally distributed constant
variance disturbances for the last 50 observations in a sample of 100 by the
square root of a time trend variable. Prior means for β were set equal to
the true values of unity and prior variances equal to unity as well. A diffuse
prior for σ was employed and the Γ(m, k) prior for r was set to m = 8, k = 2
indicating a prior belief in heteroscedasticity.
% ----- Example 6.6 Heteroscedastic Gibbs sampler

randn(’seed’,202112); % set seed for random number generator
tt = ones(n,1); tt(51:100,1) = [1:50]’;
y = x*b + randn(n,1).*sqrt(tt); % heteroscedastic disturbances
bsave = zeros(ndraw,k); % allocate storage for results
ssave = zeros(ndraw,1); rsave = zeros(ndraw,1);
vsave = zeros(ndraw,n);
r = [1.0 1.0 1.0]’; % prior b means
R = eye(k); T = eye(k); % prior b variance
Q = inv(T); Qpr = Q*r;
b0 = (x’*x)\(x’*y); % use ols starting values
sige = (y-x*b0)’*(y-x*b0)/(n-k);
V = ones(n,1); in = ones(n,1); % initial value for V
rval = 4; % initial value for rval
mm=8; kk=2; % prior for r-value
% mean(rvalue) = 4, var(rvalue) = 2
tic; % start timing
ys = y.*sqrt(V); xs = matmul(x,sqrt(V));
xpxi = inv(xs’*xs + sige*Q);
b = xpxi*(xs’*ys + sige*Qpr); % update b
a = chol(xpxi);
b = sqrt(sige)*a’*randn(k,1) + b; % draw MV normal mean(b), var(b)
bsave(i,:) = b’; % save b draws
e = y - x*b; ssr = e’*e; % update sige
chi = chis_rnd(1,n); % do chisquared(n) draw
sige = ssr/chi; ssave(i,1) = sige; % save sige draws
chiv = chis_rnd(n,rval+1); % update vi
vi = ((e.*e./sige) + in*rval)./chiv;
V = in./vi; vsave(i,:) = vi’; % save the draw
rval = gamm_rnd(1,1,mm,kk); % update rval
rsave(i,1) = rval; % save the draw
toc; % stop timing
bhat = mean(bsave(nomit+1:ndraw,:)); % calculate means and std deviations
.::UdecomBooks::.
bstd = std(bsave(nomit+1:ndraw,:)); tstat = bhat./bstd;

smean = mean(ssave(nomit+1:ndraw,1)); vmean = mean(vsave(nomit+1:ndraw,:));
rmean = mean(rsave(nomit+1:ndraw,1));
tout = tdis_prb(tstat’,n); % compute t-stat significance levels
% set up for printing results
in.cnames = strvcat(’Coefficient’,’t-statistic’,’t-probability’);
in.rnames = strvcat(’Variable’,’variable 1’,’variable 2’,’variable 3’);
in.fmt = ’%16.6f’; tmp = [bhat’ tstat’ tout];
fprintf(1,’Gibbs estimates \n’); % print results
mprint(tmp,in);
fprintf(1,’Sigma estimate = %16.8f \n’,smean);
fprintf(1,’rvalue estimate = %16.8f \n’,rmean);
result = theil(y,x,r,R,T); % compare to Theil-Goldberger estimates
prt(result); plot(vmean); % plot vi-estimates
title(’mean of vi-estimates’);
The program makes use of the fact that Vp−1 is a diagonal matrix, so we
transform the vector y by multiplying it by (V −1 ) and we use the Gauss
function matmul to carry out the same transformation on the matrix X.
Using this transformation saves space in RAM memory and speeds up the
Gibbs sampler. The function gamm rnd from the distributions function
library (see Chapter 9) is used to produce the random Γ(m, k) draws and
chis rnd is capable of generating a vector of random χ2 draws.
The results based on 1,100, 2,100 and 10,100 draws with the first 100
omitted for burn-in are shown below along with the Theil-Goldberger es-
timates. We can see that convergence is not a problem as the means and
standard deviations from the sample of 2,100 and 10,100 draws are quite
close. The time required to carry out the draws on a MacIntosh G3, 266
Mhz. computer are also reported for each set of draws. The heteroscedastic
linear model estimates are slightly more accurate than the Theil-Goldberger
estimates, presumably because the latter do not take the heteroscedastic na-
ture of the disturbances into account. Note that both estimation procedures
are based on the same prior information.
elapsed_time = 13.2920 seconds (1,100 draws)

Gibbs estimates
variable 1 1.510766 4.382439 0.000029
variable 2 1.056020 3.026450 0.003146
variable 3 0.975192 2.967255 0.003759
Sigma estimate = 11.70012398
rvalue estimate = 4.05923847

Gibbs estimates
.::UdecomBooks::.

variable 1 1.500910 4.437038 0.000023
variable 2 1.064960 2.977550 0.003645
variable 3 0.995109 2.956360 0.003883

Gibbs estimates
variable 1 1.506098 4.505296 0.000018
variable 2 1.060568 3.027138 0.003140
variable 3 0.992387 2.952396 0.003930

R-squared = 0.2935
sigma^2 = 11.4540
***************************************************************
variable 1 1.000000 0.707107
variable 2 1.000000 0.707107
variable 3 1.000000 0.707107
***************************************************************
Posterior Estimates
variable 1 1.543398 1.429473 0.156081
variable 2 1.104141 0.996164 0.321649
variable 3 1.025292 0.953436 0.342739
A plot of the mean over the 2000 samples of vi estimates is shown in Fig-
ure 6.2. These estimates appear to capture the nature of the heteroscedastic
disturbances introduced in the model for observations 51 to 100.
6.4 Gibbs sampling functions

To make implementation of these models more convenient, we can construct
a function that implements the Markov Chain Monte Carlo sampling for the
Bayesian heteroscedastic linear model. The documentation for this function
ols g is shown below.
PURPOSE: Gibbs estimates for the Bayesian heteroscedastic linear model
.::UdecomBooks::.
6
Vi estimates
1
0 10 20 30 40 50 60 70 80 90 100
Observations
Figure 6.2: Mean of Vi draws
y = X B + E, E = N(0,sige*V),
V = diag(v1,v2,...vn), r/vi = ID chi(r)/r, r = Gamma(m,k)
B = N(c,T), sige = gamma(nu,d0)
---------------------------------------------------
USAGE: results = ols_g(y,x,,ndraw,nomit,prior,start)
x = independent variables matrix of rank(k)
ndraw = # of draws
nomit = # of initial draws omitted for burn-in
prior = a structure for prior information input:
prior.beta, prior means for beta, c above
prior.bcov, prior beta covariance , T above
prior.rval, r prior hyperparameter, default=4
prior.m, informative Gamma(m,k) prior on r
prior.k, informative Gamma(m,k) prior on r
prior.nu, informative Gamma(nu,d0) prior on sige
prior.d0 informative Gamma(nu,d0) prior on sige
default for above: nu=0,d0=0 (diffuse prior)
.::UdecomBooks::.
start = (optional) structure containing starting values:

defaults: OLS beta,sige, V= ones(n,1)
start.b = beta starting values (nvar x 1)
start.sig = sige starting value (scalar)
start.V = V starting values (n x 1)
---------------------------------------------------
results.meth = ’ols_g’
results.bdraw = bhat draws (ndraw-nomit x nvar)
results.vmean = mean of vi draws (nobs x 1)
results.sdraw = sige draws (ndraw-nomit x 1)
results.yhat = mean of draws from posterior for y-predicted
results.rdraw = r draws (ndraw-nomit x 1), if Gamma(m,k) prior
results.pmean = b prior means (prior.beta from input)
results.pstd = b prior std deviation, sqrt(prior.bcov)
results.m = prior m-value for r hyperparameter (if input)
results.k = prior k-value for r hyperparameter (if input)
results.r = value of hyperparameter r (if input)
results.nu = prior nu-value for sige prior
results.d0 = prior d0-value for sige prior
results.nvar = # of variables
results.ndraw = # of draws
results.nomit = # of initial draws omitted
results.x = x-matrix
results.time = time taken for sampling
--------------------------------------------------
NOTE: use either improper prior.rval
or informative Gamma prior.m, prior.k, not both of them
---------------------------------------------------
SEE ALSO: coda, gmoment, prt_gibbs(results)
---------------------------------------------------
REFERENCES: Geweke (1993) ’Bayesian Treatment of the
Independent Student-t Linear Model’, Journal of Applied
Econometrics, 8, s19-s40.
----------------------------------------------------
The function allows an improper prior for the hyperparameter r to be

entered by simply setting a value, or parameter values m, k for a proper
Γ(m, k) prior can be input. We use a structure variable ‘prior’ to input the
prior parameters and then rely on a series of ‘strcmp’ comparisons to parse
the prior information input by the user. An informative prior for σ 2 can be
set using the ‘prior.nu’ and ‘prior.d0’ structure variable fields, or a diffuse
prior (the default) can be input by setting these fields to zero values.
A design decision was made to return the draws (excluding the first
‘nomit’ values) so the user can construct posterior densities and graphically
.::UdecomBooks::.
examine the sampling process. Note that the format of the draws returned in
the structure is such that mean(results.bdraw) or std(results.bdraw)
will produce posterior means and standard deviations.
A corresponding function prt gibbs is used to process the results struc-
ture returned by ols g and print output in the form of a regression as shown
below:
Bayesian Heteroscedastic Linear Model Gibbs Estimates

R-squared = 0.1556
sigma^2 = 4.6415
ndraws,nomit = 1100, 100
time in secs = 12.1118
r-value = 4
***************************************************************
variable 1 0.000000 31.622777
variable 2 0.000000 31.622777
variable 3 0.000000 31.622777
***************************************************************
Posterior Estimates
variable 1 1.124002 3.917272 0.000164
variable 2 1.045441 3.664259 0.000399
variable 3 0.812168 2.882300 0.004833
As an example of constructing such a function, consider the case of an

AR(m) autoregressive model:
(1 − φ1 L − φ2 L2 − . . . − φm Lm )yt = c + εt (6.18)
where we wish to impose the restriction that the mth order difference equa-
tion is stable. This requires that the roots of
(1 − φ1 z − φ2 z 2 − . . . − φm z m ) = 0 (6.19)
lie outside the unit circle. Restrictions such as this, as well as non-linear
restrictions, can be imposed on the parameters during Gibbs sampling by
simply rejecting values that do not meet the restrictions (see Gelfand, Hills,
Racine-Poon and Smith, 1990). Below is a function ar g that implements
a Gibbs sampler for this model and imposes the stability restrictions using
rejection sampling. Information regarding the results structure is not shown
to save space, but the function returns a structure variable containing draws
and other information in a format similar to the ols g function.
.::UdecomBooks::.
PURPOSE: estimates Bayesian heteroscedastic AR(m) model imposing

stability using Gibbs sampling
y = A(L)y B + E, E = N(0,sige*V),
---------------------------------------------------
USAGE: results = ar_g(y,nlag,ndraw,nomit,prior,start)
nlag = # of lagged values
ndraw = # of draws
prior = a structure for prior information input:
prior.const, a switch for constant term,
default = 1 (a constant included)
prior.nu, a prior parameter for sige
prior.d0, a prior parameter for sige
(default = diffuse prior for sige)
defaults: OLS beta,sige, V= ones(n,1)
start.sig = sige starting value (scalar)
--------------------------------------------------
NOTES: a constant term is automatically included in the model
unless you set prior.const = 0;
--------------------------------------------------
SEE ALSO: prt, prt_gibbs(results), coda
----------------------------------------------------
REFERENCES: Chib (1993) ‘Bayes regression with autoregressive
errors: A Gibbs sampling approach,’ Journal of Econometrics, pp. 275-294.
----------------------------------------------------
[n junk] = size(y); results.y = y;
if ~isstruct(prior) % error checking on input
error(’ar_g: must supply the prior as a structure variable’);
elseif nargin == 6 % user-supplied starting values
if ~isstruct(start)
error(’ar_g: must supply starting values in a structure’);
end;
b0 = start.b; sige = start.sig; V = start.V; sflag = 1;
elseif nargin == 5, sflag = 0; % we supply ols starting values
else
error(’Wrong # of arguments to ar_g’);
end;
fields = fieldnames(prior); nf = length(fields); % parse prior info
.::UdecomBooks::.
mm = 0; rval = 4; const = 1; nu = 0; d0 = 0; % set defaults

for i=1:nf
if strcmp(fields{i},’rval’), rval = prior.rval;
elseif strcmp(fields{i},’m’), mm = prior.m; kk = prior.k;
rval = gamm_rnd(1,1,mm,kk); % initial value for rval
elseif strcmp(fields{i},’const’), const = prior.const;
elseif strcmp(fields{i},’nu’), nu = prior.nu; d0 = prior.d0;
end;
end;
if sflag == 0 % we supply ols starting values
if const == 1, x = [ones(n,1) mlag(y,nlag)];
else, x = mlag(y,nlag);
end;
x = trimr(x,nlag,0); y = trimr(y,nlag,0); % feed the lags
nadj = length(y);
b0 = x\y; % Find ols values as initial starting values
k = nlag+const; sige = (y-x*b0)’*(y-x*b0)/(nadj-k);
V = ones(nadj,1); in = ones(nadj,1); % initial value for V
else
if const == 1, x = [ones(n,1) mlag(y,nlag)];
else, x = mlag(y,nlag);
end;
x = trimr(x,nlag,0); y = trimr(y,nlag,0); % feed the lags
nadj = length(y); in = ones(nadj,1); % initial value for V
end;
c = prior.beta; [checkk,junk] = size(c);
if checkk ~= k, error(’ar_g: prior means are wrong’);
elseif junk ~= 1, error(’ar_g: prior means are wrong’);
end;
T = prior.bcov; [checkk junk] = size(T);
if checkk ~= k, error(’ar_g: prior bcov is wrong’);
elseif junk ~= k, error(’ar_g: prior bcov is wrong’);
end;
Q = inv(T); Qpc = Q*c;
% storage for draws
ssave = zeros(ndraw-nomit,1); rsave = zeros(ndraw-nomit,1);
bsave = zeros(ndraw-nomit,k); vmean = zeros(nadj,1); yhat = zeros(nadj,1);
hwait = waitbar(0,’Gibbs sampling ...’);
t0 = clock; iter = 1; counter = 0;
while iter <= ndraw; % Start sampling
% generate beta conditional on sige
ys = y.*sqrt(V); xs = matmul(x,sqrt(V));
xpx = inv(xs’*xs + sige*Q); beta1 = xpx*(xs’*ys + sige*Qpc);
c = chol(sige*xpx); accept = 0; % rejection sampling
while accept == 0;
beta = beta1 + c’*randn(k,1); betap = beta’;
coef = [-fliplr(betap(2:k)) 1]; root = roots(coef);
rootmod = abs(root);
if min(rootmod) >= 1.0001; accept = 1;
.::UdecomBooks::.
else, counter = counter+1; % counts acceptance rate; accept = 0;

end;
end; % end of while loop
% generate sige conditional on beta
nu1 = nadj + nu; e = ys - xs*beta; d1 = d0 + e’*e;
chi = chis_rnd(1,nu1); t2 = chi/d1; sige = 1/t2;
chiv = chis_rnd(nadj,rval+1); % update vi
vi = ((e.*e./sige) + in*rval)./chiv; V = in./vi;
if mm ~= 0, rval = gamm_rnd(1,1,mm,kk); % update rval; end;
if iter > nomit; % save draws
vmean = vmean + vi; ssave(iter-nomit,1) = sige;
yhat = yhat + randn(nadj,1).*sqrt(sige*vi) + x*beta;
bsave(iter-nomit,:) = beta’;
if mm~= 0, rsave(i-nomit,1) = rval; end;
end; % end of if
iter = iter+1;
waitbar(iter/ndraw);
end; % end of while iter < ndraw
gtime = etime(clock,t0);
results.accept = 1 - counter/(iter+counter); % find acceptance rate
results.meth = ’ar_g’; results.bdraw = bsave;
results.sdraw = ssave; results.vdraw = vsave;
results.pmean = prior.beta; results.pstd = sqrt(diag(T));
if mm~= 0, results.rdraw = rsave; results.m = mm; results.k = kk;
else, results.r = rval; results.rdraw = rsave;
end;
results.nobs = n; results.nadj = nadj;
results.nvar = nlag+const; results.ndraw = ndraw;
results.nomit = nomit; results.time = gtime;
results.x = x; results.nu = nu; results.d0 = d0;
The function allows for informative or diffuse priors on the noise vari-
ance σε2 and allows for a homoscedastic or heteroscedastic implementation
using the chi-squared prior described for the Bayesian heteroscedastic lin-
ear model. Processing this input information from the user is accomplished
using the MATLAB ‘fieldnames’ command and ‘strcmp’ statements, with
defaults provided by the function for cases where the user inputs no infor-
mation for a particular field.
One point to note is the rejection sampling code, where we use the MAT-
LAB function roots to examine the roots of the polynomial for stability. We
also rely on a MATLAB function fliplr that ‘flips’ a vector or matrix from
‘left to right’. This is needed because of the format assumed for the poly-
nomial coefficients by the roots function. If the stability condition is not
met, we simply carry out another multivariate random draw to obtain a new
vector of coefficients and check for stability again. This process continues
until we obtain a coefficient draw that meets the stability conditions. One
.::UdecomBooks::.
could view this as discarding (or rejecting) draws where the coefficients are
inconsistent with stability.
Use of rejection sampling can produce problems if the acceptance rate
becomes excessively low. Consider that if we need a sample of 1000 draws
to obtain posterior inferences and the acceptance rate is 1 of 100 draws,
we would need to make 100,000 multivariate random draws for the autore-
gressive parameters to obtain a usable sample of 1000 draws on which to
based our posterior inferences. To help the user monitor problems that
might arise due to low acceptance rates, we calculate the acceptance rate
and return this information in the results structure. Of course, the function
prt gibbs was modified to produce a printout of the results structure from
the ar g estimation procedure.
In addition to returning the time required in seconds, the function places
a ‘waitbar’ on the screen that shows graphically the progress of the Gibbs
sampling. This is accomplished with the MATLAB function waitbar().
The following program generates data based on an AR(2) model which
is on the edge of the stability conditions. The AR(1)+AR(2) coefficients
equal unity, where the restriction for stability of this model requires that
these two coefficients sum to less than unity. The demonstration program
compares estimates from ols, ar g and theil.
% ----- Example 6.7 Using the ar_g() function
n = 200; k = 3; e = randn(n,1)*2; y = zeros(n,1);
for i=3:n
y(i,1) = 1 + y(i-1,1)*0.25 + y(i-2,1)*0.75 + e(i,1);
end;
x = [ones(n,1) mlag(y,2)];
yt = trimr(y,100,0); xt = trimr(x,100,0); % omit first 100 for startup
vnames = strvcat(’y-variable’,’constant’,’ylag1’,’ylag2’);
res1 = ols(yt,xt);
prt(res1,vnames);
ndraw = 1100; nomit = 100;
bmean = zeros(k,1); bcov = eye(k)*100;
prior.bcov = bcov; % diffuse prior variance
prior.beta = bmean; % prior means of zero
res2 = ar_g(yt,2,ndraw,nomit,prior);
prt(res2,’y-variable’);
res3 = theil(yt,xt,bmean,eye(k),bcov);
prt(res3,vnames);
The results from running the program are shown below. Running the
program 100 times produced results where the stability conditions were vi-
olated 10 percent of the time by the least-squares and Theil-Goldberger
estimates. Figure 6.3 shows a histogram of the distribution of φ1 + φ2 for
.::UdecomBooks::.
the 100 sets of estimates, where the 10 violations of stability for least-squares
and Theil-Goldberger appear as the two bars farthest to the right. The mean
acceptance rate for the Gibbs sampler over the 100 runs was 0.78, and the
median was 0.86.

Dependent Variable = y-variable
R-squared = 0.9790
sigma^2 = 3.5352
***************************************************************
constant -0.682783 -0.561070 0.576044
ylag1 0.350730 4.429152 0.000025
ylag2 0.666715 8.256773 0.000000
Bayesian Autoregressive Model Gibbs Estimates

R-squared = 0.9786
sigma^2 = 3.6092
nu,d0 = 0, 0
accept rate = 0.2942
rvalue = 4.0000
***************************************************************
constant 0.000000 10.000000
y-variable lag 1 0.000000 10.000000
y-variable lag 2 0.000000 10.000000
***************************************************************
Posterior Estimates
constant 1.565553 2.085717 0.039550
y-variable lag 1 0.325273 3.903220 0.000172
y-variable lag 2 0.663394 7.935940 0.000000

R-squared = 0.9790
sigma^2 = 3.5353
.::UdecomBooks::.
***************************************************************
constant 0.000000 10.000000
ylag1 0.000000 10.000000
ylag2 0.000000 10.000000
***************************************************************
Posterior Estimates
constant -0.672706 -0.296170 0.767733
ylag1 0.350736 2.355845 0.020493
ylag2 0.666585 4.391659 0.000029
45
40 ols
gibbs
theil
35
30
Frequency of outcomes
25
20
15
10
0
0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 1.05
φ1+φ2
Figure 6.3: Distribution of φ1 + φ2
6.5 Metropolis sampling

Econometric estimation problems amenable to Gibbs sampling can take one
of two forms. The simplest case is where all of the conditional distributions
.::UdecomBooks::.
are from well-known distributions allowing us to sample random deviates

using standard computational algorithms. This circumstance is sometimes
referred to as a ‘standard conjugate prior-to-posterior updating’, evoking
the sense of conjugate priors from Bayesian analysis. This was the situation
with the Gibbs samplers described in the previous sections.
A second more complicated case that one sometimes encounters in Gibbs
sampling is where one or more of the conditional distributions can be ex-
pressed mathematically, but they take an unknown form. It is still possible
to implement a Gibbs sampler for these models using a host of alternative
methods that are available to produce draws from distributions taking a
non-standard form.
One of the more commonly used ways to deal with this situation is
known as the ‘Metropolis algorithm’. To illustrate this situation, we draw
on a first-order spatial autoregressive (FAR) model which takes the following
form:
y = ρW y + ε (6.20)
ε ∼ N (0, σ In ) 2
where y contains an nx1 vector of dependent variables collected from points

in space, e.g., counties, states, or neighborhoods. W is an nxn known spatial
weight matrix, usually containing first-order contiguity relations or functions
of distance. A first-order contiguity matrix has zeros on the main diagonal,
rows that contain zeros in positions associated with non-contiguous obser-
vational units and ones in positions reflecting neighboring units that are
(first-order) contiguous. An example is shown in (6.21) for a sample of five
areas.
 
0 1 1 0 0
 
 1 0 1 0 0 
 
C= 1 1 0 1 0  (6.21)
 
 0 0 1 0 1 
0 0 0 1 0
Information regarding first-order contiguity is recorded for each obser-
vation as ones for areas that are neighbors (e.g., observations 2 and 3 are
neighbors to 1) and zeros for those that are not (e.g., observations 4 and 5
are not neighbors to 1). By convention, zeros are placed on the main diag-
onal of the spatial weight matrix. Standardization to produce row-sums of
unity results in the matrix W shown in (6.22) that is used in the model.
.::UdecomBooks::.
 
0 0.5 0.5 0 0
 
 0.5 0 0.5 0 0 
 
W = 0.33 0.33 0 0.33 0  (6.22)
 
 0 0 0.5 0 0.5 
0 0 0 1 0
The parameter ρ is a coefficient on the spatially lagged dependent vari-
able W y that reflects the influence of this explanatory variable on variation
in the dependent variable y. The model is called a first order spatial autore-
gression because its represents a spatial analogy to the first order autoregres-
sive model from time series analysis, yt = ρyt−1 + εt . Multiplying a vector
y containing 5 areas cross-sectional data observations by the standardized
spatial contiguity matrix W produces an explanatory variable equal to the
mean of observations from contiguous states.
Using diffuse priors, π(ρ) and π(σ) for the parameters (ρ, σ) shown in
(6.23),
π(ρ) ∝ constant (6.23)

π(σ) ∝ (1/σ), 0 < σ < +∞
which can be combined with the likelihood for this model, we arrive at a
joint posterior distribution for the parameters, p(ρ, σ|y).
1
p(ρ, σ|y) ∝ |In − ρW |σ −(n+1) exp{− (y − ρW y)0 (y − ρW y)} (6.24)
2σ 2
If we treat ρ as known, the kernel for the conditional posterior for σ
given ρ takes the form:
1 0
p(σ|ρ, y) ∝ σ −(n+1) exp{− ε ε} (6.25)
2σ 2
where ε = y − ρW y. It is important to note that by conditioning on ρ
(treating it as known) we can subsume the determinant, |In − ρW |, as part
of the constant of proportionality, leaving us with one of the standard dis-
tributional forms. From (6.25) we conclude that σ 2 ∼ χ2 (n).
Unfortunately, the conditional distribution of ρ given σ takes the follow-
ing non-standard form:
p(ρ|σ, y) ∝ |In − ρW |{(y − ρW y)0 (y − ρW y)}−n/2 (6.26)
.::UdecomBooks::.
To sample from (6.26) we can rely on a method called ‘Metropolis sampling’,

within the Gibbs sampling sequence, hence it is often labeled ‘Metropolis-
within-Gibbs’.
Metroplis sampling is described here for the case of a symmetric normal
candidate generating density. This should work well for the conditional
distribution of ρ because, as Figure 6.4 shows, the function |In − ρW | is
everywhere positive and exhibits a form similar to the normal distribution.
0.9
0.8
0.7
0.6
|In - ρ W|
0.5
0.4
0.3
0.2
0.1
0
-2 -1.5 -1 -0.5 0 0.5 1
ρ values
Figure 6.4: |In − ρW | as a function of ρ
To describe Metropolis sampling in general, suppose we are interested in

sampling from a density f () and x0 denotes the current draw from f . Let
the candidate value be generated by y = x0 + cZ, where Z is a draw from
a standard normal distribution and c is a known constant. An acceptance
probability is computed using: p = min{1, f (y)/f (x0 }. We then draw a
uniform random deviate we label U , and if U < p, the next draw from f is
given by x1 = y. If on the other hand, U ≥ p, the draw is taken to be the
current value, x1 = x0 .
A MATLAB program to implement this approach for the case of the first-
order spatial autoregressive model is shown in Example 6.8. The function
c rho evaluates the conditional distribution for ρ at any value of ρ. A
.::UdecomBooks::.
further complication that arises in producing estimates for this model is

that Anselin (1988) shows that:
1/λmin < ρ < 1/λmax

where λmin and λmax are the minimum and maximum eigenvalues of the
standardized spatial weight matrix W . We impose this restriction using
rejection sampling as demonstrated in the previous section.
% ----- Example 6.8 Metropolis within Gibbs sampling
n=49; ndraw = 1100; nomit = 100; nadj = ndraw-nomit;
% generate data based on a given W-matrix
load wmat.dat; W = wmat; IN = eye(n); in = ones(n,1); weig = eig(W);
lmin = 1/min(weig); lmax = 1/max(weig); % bounds on rho
rho = 0.7; % true value of rho
y = inv(IN-rho*W)*randn(n,1); Wy = W*y;
% set starting values
rho = 0.5; % starting value for the sampler
sige = 10.0; % starting value for the sampler
c = 0.5; % for the Metropolis step (adjusted during sampling)
rsave = zeros(nadj,1); % storage for results
ssave = zeros(nadj,1); rtmp = zeros(nomit,1);
iter = 1; cnt = 0;
while (iter <= ndraw); % start sampling;
e = y - rho*Wy; ssr = (e’*e); % update sige;
chi = chis_rnd(1,n); sige = (ssr/chi);
% metropolis step to get rho update
rhox = c_rho(rho,n,y,W); % c_rho evaluates conditional
rho2 = rho + c*randn(1); accept = 0;
while accept == 0; % rejection bounds on rho
if ((rho2 > lmin) & (rho2 < lmax)); accept = 1; end;
rho2 = rho + c*randn(1); cnt = cnt+1;
end; % end of rejection for rho
rhoy = c_rho(rho2,n,y,W); % c_rho evaluates conditional
ru = unif_rnd(1,0,1); ratio = rhoy/rhox; p = min(1,ratio);
if (ru < p)
rho = rho2; rtmp(iter,1) = rho; iter = iter+1;
end;
if (iter >= nomit);
if iter == nomit % update c based on initial draws
c = 2*std(rtmp(1:nomit,1));
end;
ssave(iter-nomit+1,1) = sige; rsave(iter-nomit+1,1) = rho;
end; % end of if iter > nomit
end; % end of sampling loop
% printout results
fprintf(1,’hit rate = %6.4f \n’,ndraw/cnt);
fprintf(1,’mean and std of rho %6.3f %6.3f \n’,mean(rsave),std(rsave));
fprintf(1,’mean and std of sig %6.3f %6.3f \n’,mean(ssave),std(ssave));
.::UdecomBooks::.
Some points to note about the code are:
1. We generate a model using a first-order spatial contiguity matrix W

from 49 neighborhoods in Columbus, Ohio presented in Anselin (1988).
2. We adjust the parameter c used in the Metropolis step after the initial
‘nomit’ passes through the sampler based on a two standard deviation
measure of the ρ values sampled up to this point.
3. Since ‘rho2’ is the candidate value that might become are updated
value of ρ depending on the outcome of the Metropolis step, we carry
out rejection sampling on this draw to ensure that sampled values will
meet the restriction. In the event that ‘rho2’ does not meet the re-
striction, we discard it and draw another value. This process continues
until we obtain a value for ‘rho2’ that meets the restriction.
4. The function c rho evaluates (6.26) for a given value of ρ.
function yout = c_rho(rho,n,y,W)

% evaluates conditional distribution of rho
% for the spatial autoregressive model
IN = eye(n); B = IN - rho*W;
term0 = log(det(B)); yt = y - rho*W*y;
term3 = (n/2)*log(yt’*yt);
tmp = term0-term3;
yout = exp(tmp);
This estimator has been incorporated in a function far g that allows the
user to input an informative prior for the spatial autocorrelation parameter
ρ.
We carried out an experiment to illustrate Metropolis sampling on the
conditional distribution of ρ in the first-order spatial autoregressive model.
A series of ten models were generated using values of ρ ranging from -0.9
to 0.9 in increments of 0.2. (The bounds on ρ for this standardized spatial
weight matrix were -1.54 and 1.0, so these values of ρ used to generate the
data were within the bounds.) Estimates for all ten models were derived
using a sample of 2,100 draws with the first 100 omitted for “burn-in”. In
addition to producing Gibbs estimates, based on a diffuse prior centered on
the true value of ρ having a variance of 10. We also produced maximum
likelihood estimates for the ten models using the maximum likelihood meth-
ods presented in Anselin (1988). Timing results as well as acceptance rates
for the Gibbs sampler are reported in Table 5.1.
.::UdecomBooks::.
Table 6.1: A Comparison of FAR Estimators

True ρ Metropolis ρ̂ Time Accept Max Likel ρ̂
(Seconds) Rate
-0.900 -0.899 12.954 0.915 -1.094
-0.700 -0.701 13.069 0.975 -1.028
-0.500 -0.500 13.373 0.999 -0.307
-0.300 -0.314 13.970 0.999 -0.797
-0.100 -0.100 13.729 0.999 -0.020
0.100 0.100 13.163 1.000 0.195
0.300 0.300 13.118 0.994 0.426
0.500 0.500 13.029 0.992 0.523
0.700 0.700 13.008 0.700 0.857
0.900 0.900 13.283 0.586 0.852
From the table we see that the Metropolis algorithm produced estimates
close to the true value of ρ used to generate the data vector y, as did the
maximum likelihood method. The acceptance rates are lower for the value
of ρ = 0.9 because this magnitude is near the upper bound of unity on ρ for
this particular weight matrix.
6.6 Functions in the Gibbs sampling library

This section describes some of the Gibbs sampling functions available in
addition to ar g, ols g and far g already described in the chapter.
The use of Gibbs sampling to estimate vector autoregressive models
should be intuitively clear given the description of the Bayesian heteroscedas-
tic linear model and its comparison to the Theil-Goldberger estimates in
Section 6.3. A series of functions bvar g, rvar g, becm g, recm g imple-
ment Gibbs sampling estimation of the bvar, rvar, becm, recm models
described in Chapter 5. These functions should be useful in cases where
outliers or non-constant variance is present. In cases where the disturbances
obey the Gauss-Markov assumptions, the estimates produced by Gibbs sam-
pling should be very similar to those from Theil-Goldberger estimation of
these models.
It makes little sense to implement these models with a homoscedastic
prior represented by a large value for the hyperparameter r, so a default
value of r = 4 is set indicating heteroscedasticity or outliers.
.::UdecomBooks::.
The input format for these functions rely on a structure variable ‘prior’
for inputting prior hyperparameters. For example, the documentation for
the function bvar g is shown below, where information regarding the results
structure returned by the function was eliminated to save space.
PURPOSE: Gibbs sampling estimates for Bayesian vector
autoregressive model using Minnesota-type prior
y = A(L) Y + X B + E, E = N(0,sige*V),
c = R A(L) + U, U = N(0,Z), Minnesota prior
a diffuse prior is used for B associated with deterministic
variables
---------------------------------------------------
USAGE: result = bvar_g(y,nlag,prior,ndraw,nomit,x)
ndraw = # of draws
prior = a structure variable
prior.tight, Litterman’s tightness hyperparameter
prior.weight, Litterman’s weight (matrix or scalar)
prior.decay, Litterman’s lag decay = lag^(-decay)
x = an optional (nobs x nx) matrix of variables
---------------------------------------------------
The output from these functions can be printed using the wrapper func-
tion prt which calls a function prt varg that does the actual working of
printing output results.
In addition, Gibbs forecasting functions bvarf g, rvarf g, becmf g,
recmf g analogous to the forecasting functions with names corresponding
to those described in Chapter 5 are in the vector autoregressive function
library.
Probit and tobit models can be estimated using Gibbs sampling, a sub-
ject covered in the next chapter. Functions probit g and tobit g implement
this method for estimating these models and allow for heteroscedastic errors
using the chi-squared prior described in Section 6.3.
A Bayesian model averaging procedure set forth in Raftery, Madigan
and Hoeting (1997) is implemented in the function bma g. Leamer (1983)
suggested that a Bayesian solution to model specification uncertainty is to
average over all possible models using the posterior probabilities of the mod-
els as weights. Raftery, Madigan and Hoeting (1997) devise a way to imple-
.::UdecomBooks::.
ment this approach using a Metropolis sampling scheme to systematically

move through model space, producing an MCMC sample from the space of
all possible models.
A key insight they exploit is that the Bayes factor or posterior odds
ratios from two alternative model specifications can be used to construct
a Markov chain that traverses model neighborhoods. They argue that this
systematic movement of the Markov Chain via a Hastings step generates a
stochastic process that moves through model space. Under certain regularity
conditions the average of the parameters from the sampled models converges
to the model uncertainty solution proposed by Leamer (1983).
As an illustration, consider the following example program that generates
a dependent variable y based on a constant term and variables x1 , x2 and
two dummy variables x5 , x6 . Six variables are input to the bma g function
and 5,000 draws are carried out. The example program is:
% ----- Example 6.9 Bayesian model averaging with the bma_g() function
nobs = 200; % true model based on variables iota,x1,x2,x5,x6
vin = [1 2]; % variables in the model
vout = [3 4]; % variables not in the model
nv1 = length(vin); nv2 = length(vout);
nvar = nv1+nv2; % total # of variables
x1 = randn(nobs,nvar);
xdum1 = zeros(nobs,1); xdum1(150:nobs,1) = 1.0;
xdum2 = zeros(nobs,1); xdum2(125:150,1) = 1.0;
x2 = [xdum1 xdum2];
x1t = x1(:,vin); x2t = x2(:,1:2);
xtrue = [x1t x2t]; % true model based on variables iota,x1,x2,x5,x6
[junk nvart] = size(xtrue);
btrue = ones(nvart,1); btrue(3,1) = 5.0;
y = 10*ones(nobs,1) + xtrue*btrue + 0.5*randn(nobs,1);
ndraw = 5000;
result = bma_g(ndraw,y,x1,x2);
prt(result);
The printed output presents models with posterior probabilities greater

than 1 percent and the explanatory variables associated with these models
in the format shown below. In addition, estimates and t−statistics based
on a posterior probability weighted average over all unique models found
during the MCMC sampling are produced and printed.
Bayesian Model Averaging Estimates
R-squared = 0.961
sigma^2 = 0.248
Nobs = 200
Nvars = 6
.::UdecomBooks::.
# of draws = 5000
nu,lam,phi = 4.000, 0.250, 3.000
# of models = 20
time(seconds) = 30.8
***************************************************************
Model averaging information
Model v1 v2 v3 v4 v5 v6 Prob Visit
model 1 1 1 0 1 1 1 1.624 479
model 2 1 1 1 0 1 1 33.485 425
model 3 1 1 0 0 1 1 64.823 212
***************************************************************
const 10.069325 222.741085 0.000000
v1 0.947768 25.995923 0.000000
v2 0.996021 26.755817 0.000000
v3 -0.026374 -0.734773 0.463338
v4 0.000018 0.000462 0.999632
v5 4.785116 58.007254 0.000000
v6 1.131978 10.349354 0.000000
From the example, we see that the correct model is found and assigned
a posterior probability of 64.8%. The printout also shows how many times
the MCMC sampler ‘visited’ each particular model. The remaining 4,000
MCMC draws spent time visiting other models with posterior probabilities
less than one percent. All unique models visited are returned in the results
structure variable by the bma g function so you can examine them. It is
often the case that even when the procedure does not find the true model,
the estimates averaged over all models are still quite close to truth. The
true model is frequently among the high posterior probability models.
This function was created from Splus code provided by Raftery, Madigan
and Hoeting available on the internet. A few points to note about using the
function bma g. The function assigns a diffuse prior based on data vectors
y and explanatory matrices X that are standardized inside the function so
the default prior should work well in most applied settings. Example 6.8
above relies on these default prior settings, but they can be changed as
an input option. Priors for dichotomous or polychotomous variables are
treated differently from continuous variables, requiring that these two types
of variables be entered separately as in Example 6.9. The documentation
for the function is:
PURPOSE: Bayes model averaging estimates of Raftery, Madigan and Hoeting
-----------------------------------------------------------------
USAGE: result = bma_g(ndraw,y,x1,x2,prior)
or: result = bma_g(ndraw,y,x1)
.::UdecomBooks::.
where: ndraw = # of draws to carry out

y = dependent variable vector (nobs x 1)
x1 = continuous explanatory variables (nobs x k1)
x2 = (optional) dummy variables (nobs x k2)
prior = (optional) structure variable with prior information
prior.nu = nu hyperparameter (see NOTES)
prior.lam = lam hyperparameter (see NOTES)
prior.phi = phi hyperparameter (see NOTES)
-----------------------------------------------------------------
RETURNS: a structure results:
results.meth = ’bma’
results.nmod = # of models visited during sampling
results.beta = bhat averaged over all models
results.tstat = t-statistics averaged over all models
results.prob = posterior prob of each model (nmod x 1)
results.model = indicator variables for each model (nmod x k1+k2)
results.yhat = yhat averaged over all models
results.resid = residuals based on yhat averaged over models
results.sige = averaged over all models
results.rsqr = rsquared based on yhat averaged over models
results.nobs = nobs
results.nvar = nvars = k1+k2
results.k1 = k1, # of continuous explanatory variables
results.k2 = k2, # of dichotomous variables
results.visit = visits to each model during sampling (nmod x 1)
results.time = time taken for MCMC sampling
results.ndraw = # of MCMC sampling draws
results.nu = prior hyperparameter
results.phi = prior hyperparameter
results.lam = prior hyperparameter
--------------------------------------------------
NOTES: prior is: B = N(m,sig*V), nu*lam/sig = chi(nu)
m = (b0,0,...,0), b0 = ols intercept estimate
V = diag[var(y), phi^2/var(x1), phi^2/var(x2) ...]
defaults: nu=4, lam = 0.25, phi=3
--------------------------------------------------
SEE ALSO: prt(results), plt(results)
---------------------------------------------------
REFERENCES: Raftery, Madigan and Hoeting (1997) ’Bayesian model averaging
for linear regression models’, JASA 92, pp. 179-191
-----------------------------------------------------------------
As an illustration of how this function might be useful, consider a data

set on ninth grade proficiency scores for 610 Ohio school districts. A host of
15 potential explanatory variables were explored in an attempt to explain
variation in proficiency scores across the school districts. The bma g func-
tion explored 503 distinct models in 1,000 MCMC draws. From the results
.::UdecomBooks::.
printed below, we see that 30 of the 503 distinct models exhibited posterior
probabilities exceeding 1 percent.
Certain variables such as student attendance rates (attend), median in-
come (medinc), teachers average salary (tpay), a large city dummy variable
(bcity) and a dummy variable for the northwest region of the state (northw)
appeared in all of the 30 high posterior probability models. Other vari-
ables such as small city dummy variable (scity) as well as regional dummy
variables (southe, northe, northc) never appeared in the high posterior prob-
ability models.
Perhaps most interesting are the increased posterior probabilities asso-
ciated with the class size variable (csize) entering models 21 to 30 and the
welfare (welf) variable entering models 23 to 30. Also of interest is that
expenditures per pupil (expnd) enters and exits the 30 highest posterior
probability models. The may explain why economists disagree about the
impact of resources on student performance. Different model specifications
will find this variable to be significant or insignificant depending on the other
variables entered in the model.
Bayesian Model Averaging Estimates

Dependent Variable = proficiency scores
R-squared = 0.554
sigma^2 = 107.795
Nobs = 610
Nvars = 15
# of draws = 1000
nu,lam,phi = 4.000, 0.250, 3.000
# of models = 503
time(seconds) = 10004.5
***************************************************************
Model averaging information
Model unemp nwhite medinc welf expnd csize tpay attend
model 1 1 1 1 0 1 0 1 1
model 2 1 1 1 1 1 0 1 1
model 3 1 1 1 1 1 0 1 1
model 4 0 1 1 1 1 0 1 1
model 5 0 1 1 0 1 0 1 1
model 6 1 1 1 0 1 0 1 1
model 7 1 1 1 1 0 0 1 1
model 8 0 1 1 1 1 0 1 1
model 9 1 1 1 1 0 0 1 1
model 10 0 1 1 0 0 0 1 1
model 11 1 1 1 0 0 0 1 1
model 12 0 1 1 1 0 0 1 1
model 13 0 1 1 0 1 0 1 1
model 14 1 1 1 0 1 1 1 1
.::UdecomBooks::.
model 15 0 1 1 0 1 1 1 1
model 16 0 1 1 0 0 1 1 1
model 17 0 1 1 1 0 0 1 1
model 18 1 1 1 0 1 1 1 1
model 19 1 1 1 0 0 1 1 1
model 20 0 1 1 0 0 0 1 1
model 21 0 1 1 0 1 1 1 1
model 22 0 1 1 0 0 1 1 1
model 23 1 1 1 1 1 1 1 1
model 24 1 1 1 1 0 1 1 1
model 25 0 1 1 1 1 1 1 1
model 26 1 1 1 1 1 1 1 1
model 27 1 1 1 1 0 1 1 1
model 28 0 1 1 1 0 1 1 1
model 29 0 1 1 1 1 1 1 1
model 30 0 1 1 1 0 1 1 1
Model bcity scity subur southe northc northe northw Prob Visit
model 1 1 0 1 0 0 0 1 1.415 2
model 2 1 0 1 0 0 0 1 1.466 3
model 3 1 0 1 0 0 0 1 1.570 2
model 4 1 0 1 0 0 0 1 1.620 3
model 5 1 0 1 0 0 0 1 1.622 1
model 6 1 0 1 0 0 0 1 1.643 4
model 7 1 0 1 0 0 0 1 1.794 2
model 8 1 0 1 0 0 0 1 1.900 2
model 9 1 0 1 0 0 0 1 1.945 1
model 10 1 0 1 0 0 0 1 1.971 1
model 11 1 0 1 0 0 0 1 1.997 2
model 12 1 0 1 0 0 0 1 2.061 3
model 13 1 0 1 0 0 0 1 2.201 2
model 14 1 0 1 0 0 0 1 2.338 1
model 15 1 0 1 0 0 0 1 2.358 1
model 16 1 0 1 0 0 0 1 2.519 2
model 17 1 0 1 0 0 0 1 2.606 2
model 18 1 0 1 0 0 0 1 2.757 5
model 19 1 0 1 0 0 0 1 2.938 5
model 20 1 0 1 0 0 0 1 3.119 4
model 21 1 0 1 0 0 0 1 3.179 3
model 22 1 0 1 0 0 0 1 3.329 4
model 23 1 0 1 0 0 0 1 3.781 2
model 24 1 0 1 0 0 0 1 4.247 6
model 25 1 0 1 0 0 0 1 4.571 4
model 26 1 0 1 0 0 0 1 4.612 7
model 27 1 0 1 0 0 0 1 5.083 4
model 28 1 0 1 0 0 0 1 5.147 5
model 29 1 0 1 0 0 0 1 7.420 7
model 30 1 0 1 0 0 0 1 8.172 5
.::UdecomBooks::.
***************************************************************
const -374.606723 -7.995893 0.000000
unemp 0.029118 0.088849 0.929231
nwhite -0.287751 -6.413557 0.000000
medinc 0.000748 5.498948 0.000000
welf -0.113650 -1.286391 0.198795
expnd -0.000138 -0.352608 0.724504
csize -0.287077 -1.475582 0.140572
tpay 0.000478 3.661464 0.000273
attend 4.358154 9.260027 0.000000
bcity 17.116312 3.692185 0.000242
scity -0.000026 -0.000009 0.999993
subur 1.935253 1.543629 0.123197
southe 0.002544 0.001778 0.998582
northc 0.000012 0.000008 0.999994
northe 0.000123 0.000109 0.999913
northw 5.651431 4.328001 0.000018
A set of functions that implement Gibbs sampling estimation of Bayesian

spatial autoregressive models are also included in the spatial econometrics
function library. The function sar g estimates a heteroscedastic spatial
autoregressive model, sart g estimates a tobit version of the spatial au-
toregressive model and sarp g implements the probit version of the spatial
autoregressive model. LeSage (1997, 1998) describes these models and their
Gibbs sampling estimation.
6.7 Chapter summary

An introduction to Bayesian methods for exploring posterior distributions
via MCMC and Gibbs sampling was provided along with illustrations of
how these estimation methods can be encapsulated in MATLAB functions.
Although the functions presented here make these methods easy to imple-
ment, diagnostics for convergence are important when using this approach
to estimation. A set of MATLAB functions that address convergence by
providing a battery of diagnostic tests for convergence was also described.
These functions can be used to process the output from the MCMC estima-
tion functions contained in the results structure variables returned by the
estimation functions.
One drawback to the design approach taken here is that a large amount
of RAM memory is needed to handle problems where say 10,000 draws are
carried out for an estimation problem involving 10 or 20 parameters. Keep
in mind that you always have the option of carrying out a smaller number
.::UdecomBooks::.
of draws and writing the output to a file for storage using the mprint
function. Additional sequences of draws can then be generated and placed
in files for storage until a sufficient sample of draws has been accumulated.
Most of the MCMC estimation functions allow the user to input starting
values which can also be useful in this circumstance. A series of draws
based on alternative starting values is another approach that is often used
to diagnose convergence of the sampler.
.::UdecomBooks::.
Chapter 6 Appendix
The Gibbs convergence diagnostic functions are in a subdirectory gibbs.
Gibbs sampling convergence diagnostics library functions
--------------- convergence testing functions -----------------
apm - Geweke’s chi-squared test

coda - convergence diagnostics
momentg - Geweke’s NSE, RNE
raftery - Raftery and Lewis program Gibbsit for convergence
--------------- demonstration programs -----------------
apm_d - demonstrates apm

coda_d - demonstrates coda
momentg_d - demonstrates momentg
raftery_d - demonstrates raftery
--------------- support functions -----------------
prt_coda - prints coda, raftery, momentg, apm output (use prt)

empquant - These were converted from:
indtest - Rafferty and Lewis FORTRAN program.
mcest - These function names follow the FORTRAN subroutines
mctest -
ppnd -
thin -
The Gibbs regression estimation functions discussed are in the regress

subdirectory.
Gibbs sampling regression functions
------- Gibbs sampling regression functions -----------
ar_g.m - Bayesian autoregressive model (homo/heteroscedastic)
199
.::UdecomBooks::.
bma_g.m - Bayesian model averaging

ols_g.m - Bayesian linear model (homo/heteroscedastic)
probit_g.m - Bayesian probit model
tobit_g.m - Bayesian tobit model
ar_gd.m - demonstration of ar_g

bma_gd.m - demonstrates Bayesian model averaging
ols_gd.m - demonstration of ols_g
probit_gd.m - demonstrates Bayesian probit model
tobit_gd.m - demonstrates Bayesian tobit model
bmapost.m - used by bma_g

find_new.m - used by bma_g
prt_gibbs - prints results for bma_g,ar_g,ols_g,tobit_g,probit_g
sample.m - used by bma_g
The Gibbs sampling spatial econometric estimation functions discussed

are in a subdirectory spatial.
spatial econometrics function library
------- spatial regression program functions -----------
casetti - Casetti’s spatial expansion model

darp - Casetti’s darp model
far - 1st order spatial AR model - y = pWy + e
far_g - Gibbs sampling Bayesian far model
gwr - geographically weighted regression
gwr_logit - logit version of gwr
gwr_probit - probit version of gwr
bgwr - Bayesian geographically weighted regression
bgwrv - robust geographically weighted regression
lmerror - LM error statistic for regression model
lmsar - LM error statistic for sar model
lratios - Likelihood ratio statistic for regression models
moran - Moran’s I-statistic
normxy - isotropic normalization of x-y coordinates
normw - normalizes a spatial weight matrix
sac - spatial model - y = p*W1*y + X*b + u, u = c*W2*u + e
sac_g - Gibbs sampling Bayesian sac model
sacp_g - Gibbs sampling Bayesian sac probit model
sact_g - Gibbs sampling Bayesian sac tobit model
sar - spatial autoregressive model - y = p*W*y + X*b + e
sar_g - Gibbs sampling Bayesian sar model
sarp_g - Gibbs sampling Bayesian sar probit model
.::UdecomBooks::.
sart_g - Gibbs sampling Bayesian sar tobit model

sem - spatial error model - y = X*b +u, u=c*W + e
sem_g - Gibbs sampling Bayesian spatial error model
semp_g - Gibbs sampling Bayesian spatial error probit model
semt_g - Gibbs sampling Bayesian spatial error tobit model
semo - spatial error model (optimization solution)
sdm - spatial Durbin model y = a + X*b1 + W*X*b2 + e
sdm_g - Gibbs sampling Bayesian sdm model
sdmp_g - Gibbs sampling Bayesian sdm probit model
sdmt_g - Gibbs sampling Bayesian sdm tobit model
slag - creates spatial lags
walds - Wald test for regression models
xy2cont - constructs a contiguity matrix from x-y coordinates
casetti_d - Casetti model demo

darp_d - Casetti darp demo
darp_d2 - darp for all data observations
far_d - demonstrates far using a small data set
far_d2 - demonstrates far using a large data set
far_gd - far Gibbs sampling with small data set
far_gd2 - far Gibbs sampling with large data set
gwr_d - geographically weighted regression demo
gwr_d2 - GWR demo with Harrison-Rubinfeld Boston data
bgwr_d - demo of Bayesian GWR
bgwr_d2 - BGWR demo with Harrison-Rubinfeld Boston data
lmerror_d - lmerror demonstration
lmsar_d - lmsar demonstration
lratios_d - likelihood ratio demonstration
moran_d - moran demonstration
sac_d - sac model demo
sac_d2 - sac model demonstration large data set
sac_gd - sac Gibbs sampling demo
sac_gd2 - sac Gibbs demo with large data set
sacp_gd - sac Gibbs probit demo
sact_gd - sac Gibbs tobit demo
sact_gd2 - sac tobit right-censoring demo
sar_d - sar model demonstration
sar_d2 - sar model demonstration large data set
sar_gd - sar Gibbs sampling demo
sar_gd2 - sar Gibbs demo with large data set
sarp_gd - sar probit Gibbs sampling demo
sart_gd - sar tobit model Gibbs sampling demo
sart_gd2 - sar tobit right-censoring demo
sdm_d - sdm model demonstration
sdm_d2 - sdm model demonstration large data set
sdm_gd - sdm Gibbs sampling demo
sdm_gd2 - sdm Gibbs demo with large data set
.::UdecomBooks::.
sdmp_g - sdm Gibbs probit demo

sdmt_g - sdm Gibbs tobit demo
sem_d - sem model demonstration
sem_d2 - sem model demonstration large data set
sem_gd - sem Gibbs sampling demo
sem_gd2 - sem Gibbs demo with large data set
semo_d - semo function demonstration
semo_d2 - semo demo with large data set
semp_gd - sem Gibbs probit demo
semt_gd - sem Gibbs tobit demo
semt_gd2 - sem tobit right-censoring demo
slag_d - demo of slag function
walds_d - Wald test demonstration
xy2cont_d - xy2cont demo
anselin.dat- Anselin (1988) Columbus crime data

boston.dat - Harrison-Rubinfeld Boston data set
latit.dat - latittude for HR data
longi.dat - longitude for HR data
c_far - used by far_g
c_sem - used by sem_g
c_sar - used by sar_g
c_sdm - used by sdm_g
c_sac - used by sac_g
darp_lik1 - used by darp
elect.dat - Pace and Barry 3,107 obs data set
ford.dat - Pace and Barry 1st order contiguity matrix
f_far - far model likelihood (concentrated)
f_sac - sac model likelihood (concentrated)
f_sar - sar model likelihood (concentrated)
f_sem - sem model likelihood (concentrated)
f_sdm - sdm model likelihood (concentrated)
f2_far - far model likelihood
f2_sac - sac model likelihood
f2_sar - sar model likelihood
f2_sem - sem model likelihood
f3_sem - semo model likelihood
f2_sdm - sdm model likelihood
prt_gwr - prints gwr_reg results structure
prt_spat - prints results from spatial models
scoref - used by gwr
wmat.dat - Anselin (1988) 1st order contiguity matrix
The Gibbs sampling VAR/BVAR estimation functions discussed are in

the var bvar subdirectory.
.::UdecomBooks::.
------- VAR/BVAR Gibbs sampling program functions -----------

------- Gibbs sampling VAR/BVAR demonstrations -----------
becm_g - Gibbs sampling BECM estimates demo

rvar_gd - Gibbs sampling rvar model demo

.::UdecomBooks::.
Chapter 7
Limited Dependent Variable

Models
The regression function library contains routines for estimating limited de-
pendent variable logit, probit, and tobit models. In addition, there are
Gibbs sampling functions to implement recently proposed MCMC methods
for estimating these models proposed by Chib (1992) and Albert and Chib
(1993).
These models arise when the dependent variable y in our regression
model takes values 0, 1, 2, . . . representing counts of some event or a coding
system for qualitative outcomes. For example, y = 0 might represent a cod-
ing scheme indicating a lack of labor force participation by an individual in
our sample, and y = 1 denotes participation in the labor force. As another
example where the values taken on by y represent counts, we might have
y = 0, 1, 2, . . . denoting the number of foreign direct investment projects in
a given year for a sample of states in the U.S.
Regression analysis of these data usually are interpreted in the framework
of a probability model that attempts to describe the P rob(event i occurs) =
F (X: parameters). If the outcomes represent two possibilities, y = 0, 1, the
model is said to be binary, whereas models with more than two outcomes
are referred to as multinomial or polychotomous.
Ordinary least-squares can be used to carry out a regression with a binary
response variable y = 0, 1, but two problems arise. First, the errors are by
construction heteroscedastic. This is because the actual y = 0, 1 minus the
value Xβ = −Xβ or ι − Xβ. Note also that the heteroscedastic errors are a
function of the parameter vector β. The second problem with least-squares
is that the predicted values can take on values outside the (0,1) interval,
204
.::UdecomBooks::.
CHAPTER 7. LIMITED DEPENDENT VARIABLE MODELS 205
which is problematic given a probability model framework. In this setting

we would like to see:
limXβ→+∞ P rob(y = 1) = 1 (7.1)
limXβ→−∞ P rob(y = 1) = 0 (7.2)

Two distributions that have been traditionally used to produce this type
of outcome (that ensures predicted values between zero and one) are the
logisitic and normal distributions resulting in the logit model shown in (7.3)
and probit model shown in (7.4), where Φ denotes the cumulative normal
probability function.
P rob(y = 1) = eXβ /(1 + eXβ ) (7.3)
P rob(y = 1) = Φ(Xβ) (7.4)

The logistic distribution is similar to the normal except in the tails where
it is fatter resembling a Student t−distribution. Green (1997) and others
indicate that the logistic distribution resembles a t−distribution with seven
degrees of freedom. Graphs of the cumulative logistic, normal and two
t−distributions are shown in Figure 7.1. From the figure, it appears that
the logistic distribution is somewhat different from the two t−distributions,
one based on 7 degrees of freedom and the other based on 2 degrees of
freedom.
Albert and Chib (1993) propose a t−distribution in a Bayesian setting
where the degrees of freedom parameter becomes part of the estimation
problem. Data sets that produce a posterior estimate indicating a small
degrees of freedom parameter suggest the logit distribution is more appro-
priate, whereas estimates leading to a large degrees of freedom parameter
indicate a probit distribution. (Keep in mind that the t−distribution with a
large degrees of freedom parameter approaches the normal.) Given the lack
of total similarity between the t−distributions and the logistic illustrated in
Figure 7.1, we shouldn’t take this reasoning too formally.
Geweke and Keane (1997) propose a mixture of normal distributions
and demonstrate how to draw a posterior inference regarding the number of
components in the mixture using posterior odds ratios. They demonstrate
that in an applied multinomial setting, their mixture of normal distributions
is often favored over the traditional logit, probit and Student t distributions.
.::UdecomBooks::.
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
logit
0.1 t dof=7
normal
t dof=2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 7.1: Cumulative distribution functions compared
Section 7.1 presents the logit and probit regression functions from the
regression function library and Section 7.2 takes up Gibbs sampling esti-
mation of these models. In Section 7.3 tobit regression is discussed and
Section 7.4 turns attention to Gibbs sampling this model.
7.1 Logit and probit regressions

To illustrate the logit and probit functions we use a data set from Spec-
tor and Mazzeo (1980) that contains a binary dependent variable indicating
improvement in students grades after exposure to a new teaching method
for economics. The explanatory variables in the regression are: an intercept
term, grade point average, a pre-test for understanding of college-level eco-
nomics (TUCE) and a binary variable indicator whether the student was
exposed to the new teaching method.
The example program simply loads the data and calls the functions logit
and probit in addition to carrying out a least-squares regression for contrast.
% ----- Example 7.1 Logit and probit regression functions
.::UdecomBooks::.
load spector.dat; % data from Spector and Mazzeo, 1980

y = spector(:,1);
x = spector(:,2:5);
vnames = strvcat(’grade’,’const’,’psi’,’tuce’,’gpa’);
reso = ols(y,x);
prt(reso,vnames);
resl = logit(y,x);
prt(resl,vnames);
resp = probit(y,x);
prt(resp,vnames);
The resulting printouts are shown below. The logit function displays
the usual coefficient estimates, t−statistics and marginal probabilities as
well as measures of fit proposed by McFadden (1984) and Estrella (1998).
The maximized value of the log-likelihood function is reported along with
a log-likelihood ratio test statistic and marginal probability that all slopes
in the model are zero. This is based on a comparison of the maximized
log-likelihood (Lu) versus the log-likelihood for a restricted model (Lr) with
only a constant term.

Dependent Variable = grade
R-squared = 0.4159
sigma^2 = 0.1506
Nobs, Nvars = 32, 4
***************************************************************
const -1.498017 -2.859419 0.007929
psi 0.378555 2.720035 0.011088
tuce 0.010495 0.538685 0.594361
gpa 0.463852 2.864054 0.007841
Logit Maximum Likelihood Estimates

McFadden R-squared = 0.3740
Estrella R-squared = 0.4528
LR-ratio, 2*(Lu-Lr) = 15.4042
LR p-value = 0.0015
Log-Likelihood = -12.8896
# of iterations = 7
Convergence criterion = 6.976422e-08
Nobs, Nvars = 32, 4
# of 0’s, # of 1’s = 21, 11
***************************************************************
.::UdecomBooks::.
const -13.021347 -2.640538 0.013382

psi 2.378688 2.234424 0.033617
tuce 0.095158 0.672235 0.506944
gpa 2.826113 2.237723 0.033376
Probit Maximum Likelihood Estimates

LR-ratio, 2*(Lu-Lr) = 15.5459
LR p-value = 0.0014
# of iterations = 7
Nobs, Nvars = 32, 4
# of 0’s, # of 1’s = 21, 11
***************************************************************
const -7.452320 -2.931131 0.006656
psi 1.426332 2.397045 0.023445
tuce 0.051729 0.616626 0.542463
gpa 1.625810 2.343063 0.026459
Solving the logit model for parameter estimates involves non-linear op-
timization of the likelihood function. Fortunately, the likelihood function
takes a form that is globally concave allowing us to use Newton’s method.
This approach would usually converge in just a few iterations unless we
encounter an ill-conditioned data set. (We don’t actually use Newton’s
method, an issue discussed below.)
Since this is our first encounter with optimization, an examination of the
logit function is instructive. The iterative ‘while-loop’ that optimizes the
likelihood function is shown below:
iter = 1;
while (iter < maxit) & (crit > tol)
tmp = (i+exp(-x*b)); pdf = exp(-x*b)./(tmp.*tmp); cdf = i./(i+exp(-x*b));
tmp = find(cdf <=0); [n1 n2] = size(tmp);
if n1 ~= 0; cdf(tmp) = 0.00001; end;
tmp = find(cdf >= 1); [n1 n2] = size(tmp);
if n1 ~= 0; cdf(tmp) = 0.99999; end;
% gradient vector for logit, see page 883 Green, 1997
term1 = y.*(pdf./cdf); term2 = (i-y).*(pdf./(i-cdf));
for kk=1:k;
tmp1(:,kk) = term1.*x(:,kk); tmp2(:,kk) = term2.*x(:,kk);
end;
g = tmp1-tmp2; gs = (sum(g))’; delta = exp(x*b)./(i+exp(x*b));
% Hessian for logit, see page 883 Green, 1997
.::UdecomBooks::.
H = zeros(k,k);
for ii=1:t;
xp = x(ii,:)’; H = H - delta(ii,1)*(1-delta(ii,1))*(xp*x(ii,:));
end;
db = -inv(H)*gs;
% stepsize determination
s = 2; term1 = 0; term2 = 1;
while term2 > term1
s = s/2;
term1 = lo_like(b+s*db,y,x);
term2 = lo_like(b+s*db/2,y,x);
end;
bn = b + s*db; crit = abs(max(max(db)));
b = bn; iter = iter + 1;
end; % end of while
The MATLAB variables ‘maxit’ and ‘tol’ can be set by the user as an
input option, but there are default values supplied by the function. The
documentation for the function shown below makes this clear.
PURPOSE: computes logistic regression estimates
---------------------------------------------------
USAGE: results = logit(y,x,maxit,tol)
x = independent variables matrix (nobs x nvar)
maxit = optional (default=100)
tol = optional convergence (default=1e-6)
---------------------------------------------------
result.meth = ’logit’
result.beta = bhat
result.tstat = t-stats
result.yhat = yhat
result.resid = residuals
result.sige = e’*e/(n-k)
result.r2mf = McFadden pseudo-R^2
result.rsqr = Estrella R^2
result.lratio = LR-ratio test against intercept model
result.lik = unrestricted Likelihood
result.cnvg = convergence criterion, max(max(-inv(H)*g))
result.iter = # of iterations
result.nobs = nobs
result.nvar = nvars
result.zip = # of 0’s
result.one = # of 1’s
result.y = y data vector
--------------------------------------------------
SEE ALSO: prt(results), probit(), tobit()
---------------------------------------------------
.::UdecomBooks::.
The ‘while-loop’ first evaluates the logistic probability density and cumu-
lative density function at the initial least-squares values set for the param-
eter vector β, returned by the ols function. Values for the cdf outside the
(0,1) bounds are set to 0.00001 and 0.99999 to avoid overflow and underflow
computational problems by the following code.
tmp = find(cdf <=0); [n1 n2] = size(tmp);
if n1 ~= 0; cdf(tmp) = 0.00001; end;
tmp = find(cdf >= 1); [n1 n2] = size(tmp);
if n1 ~= 0; cdf(tmp) = 0.99999; end;
Next, the analytical gradient and hessian functions presented in Green

(1997) are computed and used to determine a new vector of β values. Non-
linear optimization problems can be solved by iterative methods that begin
from an initial value β0 . If β0 is not the optimal value for β, a direction
vector ∆0 , and step size λ0 are computed to find a new value β1 for the next
iteration using:
β1 = β0 + λ0 ∆0 (7.5)
Gradient methods of optimization rely on a direction vector ∆ = W g,
where W is a positive definite matrix and g is the gradient of the function
evaluated at the current value β0 , ∂F (β0 )/∂β0 . Newton’s method is based on
a linear Taylor series expansion of the first order conditions: ∂F (β0 )/∂β0 =
0, which leads to W = −H −1 and ∆ = −H −1 g. Note that Newton’s method
implicitly sets λ0 to unity.
Our algorithm determines a stepsize variable for λ, which need not be
unity, using the following code.
db = -inv(H)*gs;
% stepsize determination
s = 2;
term1 = 0; term2 = 1;
while term2 > term1
s = s/2;
term1 = lo_like(b+s*db,y,x);
term2 = lo_like(b+s*db/2,y,x);
end;
bn = b + s*db; crit = abs(max(max(db)));
b = bn; iter = iter + 1;
end; % end of while
After convergence, the logit function evaluates the analytical hessian

at the maximum likelihood parameter values for the purpose of comput-
ing asymptotic variances, t−statistics, measures of fit, as well as the log-
likelihood ratio tests reported in the printout.
.::UdecomBooks::.
In addition to these functions, a function mlogit implements maximum

likelihood estimation for the multinomial logit model.
7.2 Gibbs sampling logit/probit models

Albert and Chib (1993) propose augmenting the data set in logit/probit
models with a set of variables z that are drawn from an underlying contin-
uous distribution such that yi = 1 when zi > 0 and yi = 0 otherwise. They
then point out that the conditional distribution of β given z takes the form of
a simple normal distribution with a mean that can be easily computed. Fur-
thermore, the conditional distribution of the augmenting variables z given
β take the form of truncated normal distributions that are also easy to com-
pute. This leads to a Gibbs sampling approach to estimation of logit/probit
models.
Consider the case of logit/probit models where the observed y can be
viewed as associated with a latent variable zi < 0 which produces an ob-
served variable yi = 0 and zi ≥ 0 associated with yi = 1. Albert and Chib
(1993) show that the posterior distribution of zi conditional on all other
parameters in the model β, σ takes the form of a truncated normal distri-
bution. This truncated normal distribution is constructed by truncating a
N [ỹi , σi2 ] distribution from the right by zero. If we let ỹ denote the predicted
value for the ith row of zi , and let σi2 denote the variance of the prediction,
the pdf of the latent variables zi is:
(
N (ỹi , σi2 ), truncated at the left by 0 if yi = 1
f (zi |ρ, β, σ) ∼ (7.6)
N (ỹi , σi2 ), truncated at the right by 0 if yi = 0
Because the probit model is unable to identify both β and σε2 , we scale our
problem to make σε2 equal to unity.
These expressions simply indicate that we can replace values of yi = 1
with the sampled normals truncated at the left by 0 and values of yi = 0
with sampled normals truncated at the right by 0.
As an intuitive view of why this works, consider the following MATLAB
program that generates a vector of latent values z and then converts these
values to zero or one, depending on whether they take on values greater or
less than zero. The program carries out Gibbs sampling using the function
probit g. During the sampling, generated values for the latent variable z
from each pass through the sampler are saved and returned in a structure
.::UdecomBooks::.
variable results.ydraw. The program then plots the mean of these predic-
tions versus the actual vector of latent z values, which is shown in Figure 7.2.
% ----- Example 7.2 Demonstrate how Albert-Chib latent variable works

nobs=100; nvar=4; % generate a probit data set
x = rand(nobs,nvar);
beta = ones(nvar,1);
evec = randn(nobs,1);
y = x*beta + evec;
ysave = y; % save the generate y-values
for i=1:nobs % convert to 0,1 values
if y(i,1) >= 0, y(i,1) = 1;
else, y(i,1) = 0;
end;
end;
prior.beta = zeros(4,1); % diffuse prior for beta
prior.bcov = eye(4)*10000;
ndraw = 1100; nomit = 100;
prior.rval = 100; % probit prior for r-value
results = probit_g(y,x,ndraw,nomit,prior);
ymean = results.ymean; % recover the mean of draws
tt=1:nobs;
plot(tt,ysave,tt,ymean,’--’);
xlabel(’observations’);
ylabel(’actual y-values vs mean of y values drawn’);
legend(’acutal’,’draws’);
Of course, in an applied problem we would not have access to the latent

variables zi generated in example 7.2, but this provides an illustration of the
insight behind Albert and Chib’s Gibbs sampling approach to estimating
these models.
The documentation for the function probit g which implements the
Gibbs sampler for the logit/probit model is shown below with information
regarding the results structure returned omitted to save space.
PURPOSE: Gibbs sampler for the Bayesian Probit model

y = X B + E, E = N(0,V),
B = N(c,T)
--------------------------------------------------------------
USAGE: results = probit_g(y,x,ndraw,nomit,prior,start)
where: y = nobs x 1 independent variable vector
x = nobs x nvar explanatory variables matrix
ndraw = # of draws
prior = a structure for prior information input
.::UdecomBooks::.
5
acutal
draws
4
actual y-values vs mean of y values drawn
-1
-2
0 10 20 30 40 50 60 70 80 90 100
observations
Figure 7.2: Actual y vs. mean of latent y-draws

defaults: max likelihood beta, V= ones(n,1)
----------------------------------------------------------------
---------------------------------------------------------------
The first point to note is that the function allows a prior for the param-
eters β and implements the heteroscedastic model based on t−distributed
errors discussed in Chapter 6. Albert and Chib (1993) point out that tak-
ing this approach allows one to produce a family of models that encompass
.::UdecomBooks::.
both traditional logit and probit models. Recall that the logistic distribu-
tion is somewhat near the t−distribution with seven degrees of freedom and
the normal distribution assumed by probit models can be represented as a
t−distribution with a very large degrees of freedom parameter.
An implication of this is that the user can rely on a large prior hyperpa-
rameter value for r = 100 say, and a diffuse prior for β to produce estimates
close to a traditional probit model. On the other hand, setting r = 7 or even
r = 3 and relying on a diffuse prior for β should produce estimates close to
those from a traditional logit regression. Green (1997) states that the issue
of which distributional form should be used on applied econometric prob-
lems is unresolved. He further indicates that inferences from either logit or
probit models are often the same.
Example 7.3 illustrates these ideas by comparing maximum likelihood
logit estimates to those from Gibbs sampling with a hyperparameter value
for r = 7.
% ----- Example 7.3 Gibbs vs. maximum likelihood logit

nobs=100; nvar=4; % generate data
x = randn(nobs,nvar)*2; beta = ones(nvar,1);
beta(1,1) = -1; beta(2,1) = -1; evec = randn(nobs,1);
xb = x*beta + evec;
y = exp(xb)./(ones(nobs,1) + exp(xb)); % logit model generation
ysave = y;
for i=1:nobs; % convert to 0,1 values
if ysave(i,1) >= 0.5, y(i,1) = 1;
else, y(i,1) = 0;
end;
end;
prt(logit(y,x)); % print out maximum likelihood estimates
ndraw = 2100; nomit = 100;
prior.beta = zeros(4,1); % diffuse prior means for beta
prior.bcov = eye(4)*10000;% diffuse prior variance for beta
prior.rval = 7; % logit prior r-value
resp = probit_g(y,x,ndraw,nomit,prior); % Gibbs sampling
prt(resp); % print Gibbs probit results
The results shown below indicate that the estimates, t−statistics, and
fit from maximum likelihood logit and Gibbs sampling are quite similar as
they should be.

LR-ratio, 2*(Lu-Lr) = 102.1035
LR p-value = 0.0000
.::UdecomBooks::.
# of iterations = 10
# of 0’s, # of 1’s = 51, 49
***************************************************************
variable 1 -1.852287 -3.582471 0.000537
variable 2 -1.759636 -3.304998 0.001336
variable 3 2.024351 3.784205 0.000268
variable 4 1.849547 3.825844 0.000232
Bayesian Heteroscedastic Probit Model Gibbs Estimates

McFadden R^2 = 0.7256
Estrella R^2 = 0.8334
# 0, 1 y-values = 51, 49
r-value = 7
***************************************************************
variable 1 0.000000 100.000000
variable 2 0.000000 100.000000
variable 3 0.000000 100.000000
variable 4 0.000000 100.000000
***************************************************************
Posterior Estimates
variable 1 -1.495815 -3.657663 0.000408
variable 2 -1.411028 -3.430553 0.000877
variable 3 1.640488 3.766352 0.000280
variable 4 1.482100 3.893146 0.000179
Example 7.4 illustrates that we can rely on the same function probit g
to produce probit estimates using a setting for the hyperparameter r of 100,
which results in the normal probability model.
% ----- Example 7.4 Gibbs vs. maximum likelihood probit

x = randn(nobs,nvar); beta = ones(nvar,1);
beta(1,1) = -1; beta(2,1) = -1; evec = randn(nobs,1);
xb = x*beta + evec; y = norm_cdf(xb); % generate probit model
ysave = y;
for i=1:nobs
if ysave(i,1) >= 0.5, y(i,1) = 1;
else, y(i,1) = 0;
end;
.::UdecomBooks::.
end;
prt(probit(y,x)); % print out maximum likelihood estimates
ndraw = 2100; nomit = 100;
prior.rval = 100; % probit prior r-value
Again, the results shown below indicate that Gibbs sampling can be used
to produce estimates similar to those from maximum likelihood estimation.
Probit Maximum Likelihood Estimates
LR-ratio, 2*(Lu-Lr) = 80.1866
LR p-value = 0.0000
# of iterations = 8
# of 0’s, # of 1’s = 51, 49
***************************************************************
variable 1 -1.148952 -3.862861 0.000203
variable 2 -1.393497 -4.186982 0.000063
variable 3 1.062447 3.167247 0.002064
variable 4 1.573209 4.353392 0.000034

# 0, 1 y-values = 51, 49
r-value = 100
***************************************************************
variable 1 0.000000 100.000000
variable 2 0.000000 100.000000
variable 3 0.000000 100.000000
variable 4 0.000000 100.000000
***************************************************************
Posterior Estimates
variable 1 -1.265604 -4.503577 0.000018
variable 2 -1.532119 -4.838973 0.000005
variable 3 1.191029 4.040334 0.000105
.::UdecomBooks::.
variable 4 1.763315 4.991809 0.000003
In addition to the ability to replicate maximum likelihood estimation

results, the Gibbs sampling approach can produce robust estimates in the
face of outliers and provide a set of variance estimates for every sample
observation. The function probit g returns the draws made for the vi terms
in the results structure variable and these can be averaged to produce a mean
of the posterior distribution of these terms. Inferences about the presence
of outliers at certain observations can be based on a graphical presentation
of the posterior vi estimates.
Example 7.5 illustrates these ideas by generating a data set that contains
outliers at observations 50 and 75. Note that we need to use a prior r−value
that is small, indicating our prior belief in outliers, consistent with a logit
model.
% ----- Example 7.5 Heteroscedastic probit model

x = randn(nobs,nvar);
x(50,:) = 20*abs(x(50,:)); % insert outliers
x(75,:) = 20*abs(x(75,:)); % at observations 50,75
beta = ones(nvar,1); beta(1,1) = -1; beta(2,1) = -1;
evec = randn(nobs,1); xb = x*beta + evec;
y = exp(xb)./(ones(nobs,1) + exp(xb)); % logit model generation
ysave = y;
for i=1:nobs; % convert to 0,1 values
if ysave(i,1) >= 0.5,y(i,1) = 1;
else, y(i,1) = 0;
end;
end;
prt(logit(y,x)); % print out maximum likelihood estimates
ndraw = 2100; nomit = 100;
prior.rval = 5; % prior r-value for outliers
vmean = mean(resp.vdraw);
tt=1:nobs;
plot(tt,vmean);
ylabel(’v_i estimates’);
xlabel(’observations’);
The maximum likelihood as well as Gibbs estimates and a graph of the

vi estimates are shown below the code for example 7.5. The estimates for
the variance scalars vi clearly point to the outliers at observations 50 and
.::UdecomBooks::.
75. The Gibbs estimates that robustify for the outliers are closer to the true
values of the parameters used to generate the data, as we would expect.
LR-ratio, 2*(Lu-Lr) = 86.4123
LR p-value = 0.0000
# of iterations = 9
# of 0’s, # of 1’s = 44, 56
***************************************************************
variable 1 -1.579212 -3.587537 0.000528
variable 2 -2.488294 -3.743234 0.000310
variable 3 2.633014 4.038189 0.000108
variable 4 1.958486 3.632929 0.000452

# 0, 1 y-values = 44, 56
r-value = 5
***************************************************************
variable 1 0.000000 100.000000
variable 2 0.000000 100.000000
variable 3 0.000000 100.000000
variable 4 0.000000 100.000000
***************************************************************
Posterior Estimates
variable 1 -0.953621 -3.256567 0.001540
variable 2 -1.437799 -3.896968 0.000176
variable 3 1.715255 3.906306 0.000170
variable 4 1.304000 4.119908 0.000078
7.2.1 The probit g function

The function that implements Gibbs sampling estimation of the probit model
relies on two functions, nmrt rnd and nrlt rnd from the distributions
function library discussed in Chapter 9. These carry out the draws from the
.::UdecomBooks::.
4.5
3.5
3
vi estimates
2.5
1.5
1
0 10 20 30 40 50 60 70 80 90 100
observations
Figure 7.3: Posterior mean of vi draws with outliers
right- and left-truncated normal distributions and are based on an approach

set forth in Geweke (1991). The code for the sampling loop is shown below:
hwait = waitbar(0,’Gibbs sampling ...’);
t0 = clock;
% update z
lp=xstar*bhat;
bb=.5*(1+erf(-lp/sqrt(2)));
tt=(bb.*(1-yin)+(1-bb).*yin).*rand(n,1)+bb.*yin;
y=sqrt(2)*erfinv(2*tt-1) + lp;
% update beta
xpxi = inv(xstar’*xstar + Q);
xpy = xstar’*ystar + Qpc;
bhat = xpxi*xpy;
bhat = norm_rnd(xpxi) + bhat;
% update V
e = y - x*bhat;
chiv = chis_rnd(n,rval+1);
.::UdecomBooks::.
vi = ((e.*e) + in*rval)./chiv;
V = in./vi;
if mm ~= 0
rval = gamm_rnd(1,1,mm,kk); % update rval
end;
if i > nomit % if we are past burn-in, save the draws
bsave(i-nomit,:) = bhat’;
ymean = ymean + lp;
vmean = vmean + vi;
yhat = yhat + 1-bb;
if mm~= 0
rsave(i-nomit,1) = rval;
end;
end; % end of if i > nomit
waitbar(i/ndraw);
gtime = etime(clock,t0);
close(hwait);
A few points to note about the sampling routine. First, we allow the
user to rely on an improper prior for the hyperparameter r, or an infor-
mative Gamma(m,k) prior. In the event of an informative Gamma prior,
the MATLAB variable mm 6= 0 and we update the hyperparameter r us-
ing a random gamma draw produced by the function gamm rnd based on
the informative prior parameters stored in mm, kk. Note also that we save
the mean of latent variable draws for y and the mean of the non-constant
variance parameters vi and return these in the result structure variable.
The code that carries out the left-truncated normal random draw is:
% update z
lp=xstar*bhat;
bb=.5*(1+erf(-lp/sqrt(2)));
tt=(bb.*(1-yin)+(1-bb).*yin).*rand(n,1)+bb.*yin;
y=sqrt(2)*erfinv(2*tt-1) + lp;
An inefficient alternative procedure for producing the truncated normal

draws is that suggested by Albert and Chib (1993). They propose making
random normal draws with rejection of draws that do not meet the trunca-
tion restriction.
7.3 Tobit models

The censored regression or tobit model involves situations where truncation
occurs because the sample data are drawn from a subset of a larger popula-
tion of interest. For example, a study of income that draws a sample based
.::UdecomBooks::.
on incomes above or below a poverty level of income would provide invalid

inferences regarding the population as a whole.
The general formulation of a tobit regression model relies on an index
function yi? so that:
yi? = xi β + εi (7.7)
yi = 0 if yi? ≤ 0 (7.8)
yi = yi? if yi? > 0 (7.9)
where the index function produces a set of limit and non-limit observations.
This results in a likelihood function that consists of two parts, one part
corresponding to a classical regression for the non-limit observations and
the other part involving a discrete distribution of the relevant probabilities
for the limiting cases.
The log-likelihood function shown in (7.10) was placed in the MATLAB
function to like which we maximize using the function maxlik from the
optimization function library discussed in Chapter 10.
X X
lnL = −1/2[ln(2π)+lnσ 2 +(yi −xi β)2 /σ 2 ]+ ln[1−Φ(xi β)/σ] (7.10)
yi >0 yi =0
The documentation for the tobit function that produces maximum like-
lihood estimates for this model is shown below.
% PURPOSE: computes Tobit Regression

%---------------------------------------------------
% USAGE: results = tobit(y,x,info)
% where: y = censored dependent variable vector (nobs x 1)
% info = a structure with options:
% info.trunc = ’left’ or ’right’ for censoring (default=left)
% info.limit = value for censoring (default=0)
% info.meth = ’bfgs’,’dfp’ (optional) default = bfgs
% info.btol = tolerance for b convergence (default = 1e-8)
% info.ftol = tolerance for FUN convergence (default = 1e-8)
% info.maxit = maximum # of iterations (default = 500)
% info.b = starting values for parameters (default = ols)
% results.meth = ’tobit’
.::UdecomBooks::.
% results.lik = Log-Likelihood value
% results.iter = # of iterations taken
% results.grad = gradient at solution
% results.opt = method of optimization used
% results.nobsc = # of censored observations
Because the likelihood function represents a nonstandard mixture of dis-

crete and continuous distributions, the tobit function provides user input
options for: different types of hessian updating during optimization, differ-
ent starting values, and alternative convergence criterion. Note also that
the function provides input options to allow for ‘left’ or ‘right’ censoring of
values in the model, (with the default being the case of ‘left’ censoring),
and a user-specified ‘limit value’. Example 7.6 illustrates use of the default
convergence criterion, but implements all three alternative hessian updating
schemes.
% ----- example 7.6 Tobit regression function
% generate uncensored data
n=100; k=2;
x = randn(n,k); x(:,1) = ones(n,1);
beta = ones(k,1)*2.0;
beta(1,1) = -2.0; beta(2,1) = -2.0;
y = x*beta + randn(n,1);
% now censor the data
for i=1:n
if y(i,1) < 0, y(i,1) = 0.0; end;
end;
vnames = strvcat(’y’,’x1’,’x2’);
info.meth = ’bfgs’;
tic; resp = tobit(y,x,info); toc;
prt(resp,vnames);
info.meth = ’bhhh’;
prt(resp,vnames);
info.meth = ’dfp’;
prt(resp,vnames);
The resulting estimates and the CPU time taken are shown below. We
see that the estimates produced by all three optimization methods are iden-
.::UdecomBooks::.
tical to three decimal places. All optimization algorithms took about the
same time. In addition to reporting estimates, t-statistics and marginal
probabilities, the printing routine for tobit estimation reports the gradient
at the solution, which might provide useful information if the user wishes to
change convergence tolerances.
elapsed_time = 1.1433
Tobit Regression Estimates
Dependent Variable = y
R-squared = 0.8028
sigma^2 = 0.8144
# iterations = 13
optimization = bfgs
# of censored = 81
***************************************************************
gradient at solution
Variable Gradient
x1 0.00214371
x2 -0.00233626
sigma 0.00148077
x1 -2.240217 -4.913982 0.000004
x2 -2.194446 -6.957942 0.000000
R-squared = 0.8028
sigma^2 = 0.8144
# iterations = 13
optimization = bhhh
# of censored = 81
***************************************************************
Variable Gradient
x1 0.00214371
x2 -0.00233626
sigma 0.00148077
x1 -2.240217 -4.913982 0.000004
x2 -2.194446 -6.957942 0.000000
.::UdecomBooks::.
R-squared = 0.8028
sigma^2 = 0.8144
# iterations = 13
optimization = dfp
# of censored = 81
***************************************************************
Variable Gradient
x1 0.00214371
x2 -0.00233626
sigma 0.00148077
x1 -2.240217 -4.913982 0.000004
x2 -2.194446 -6.957942 0.000000
7.4 Gibbs sampling Tobit models

Chib (1992) examines the tobit model from a Bayesian perspective and illus-
trates that Gibbs sampling can be used to estimate this model. Intuitively,
we simply sample the latent variable from a right-truncated normal distri-
bution for limit observations. Given the sample of yi observations that are
non-limit plus the sampled limit observations, we can proceed to sample
conditionally for β and σ as in the usual Bayesian regression model.
The function tobit g implements a Gibbs sampling solution for the tobit
regression model and allows for heteroscedastic disturbances as in the case of
the probit g function. The documentation for the function is shown below,
but the results structure returned by the function is omitted.
y = X B + E, E = N(0,sige*V),
----------------------------------------------------------------
USAGE: result = tobit_g(y,x,ndraw,nomit,prior,start)
x = nobs x nvar explanatory variables matrix
ndraw = # of draws
prior = a structure variable for prior information input
prior.beta, prior means for beta, c above (default=0)
.::UdecomBooks::.
priov.bcov, prior beta covariance, T above (default=1e+12)

default for above: not used, rval=4 is used
prior.trunc = ’left’ or ’right’ (default = ’left’)
prior.limit = value for censoring (default = 0)
defaults: max likelihood beta, sige, V= ones(n,1)
start.sige = sige starting value (1x1)
---------------------------------------------------------------
----------------------------------------------------------------
Implementing this model with a diffuse prior for β, σ and assuming a

homoscedastic prior for the disturbances should produce estimates similar
to those from maximum likelihood estimation. Example 7.7 illustrates this
type of application.
% ----- Example 7.7 Gibbs sampling tobit estimation

n=100; k = 4; sige = 10; evec = randn(n,1)*sqrt(sige);
x = randn(n,k); b = ones(k,1); b(1,1) = -1; b(2,1) = -1;
y = x*b + evec; yc = zeros(n,1);
for i=1:n
if y(i,1) >= 0, yc(i,1) = y(i,1); end;
end;
Vnames = strvcat(’y’,’x1’,’x2’,’x3’,’x4’);
prt(tobit(yc,x),Vnames);
prior.bcov = eye(k)*1000; % diffuse prior var-cov for b
prior.beta = zeros(k,1); % diffuse prior means for b
prior.rval = 100; % homoscedastic prior
ndraw = 1100; nomit = 100;
result = tobit_g(yc,x,ndraw,nomit,prior);
prt(result,Vnames);
The maximum likelihood and Gibbs sampling estimates are very similar,
as they should be, and would lead to similar inferences. As in the case of
the Gibbs sampling probit model, the value of the Bayesian tobit model lies
in its ability to incorporate subjective prior information and to deal with
cases where outliers or non-constant variance exist in the model.
.::UdecomBooks::.

R-squared = 0.9874
sigma^2 = 6.7321
# iterations = 7
optimization = bfgs
# of censored = 45
***************************************************************
Variable Gradient
x1 -0.00073896
x2 -0.00084981
x3 0.00075886
x4 0.00066791
sigma 0.00029843
x1 -1.442589 -4.994263 0.000003
x2 -0.849097 -3.461847 0.000803
x3 0.989221 3.159387 0.002115
x4 0.842224 3.111376 0.002453
Bayesian Heteroscedastic Tobit Model Gibbs Estimates

R-squared = 0.9874
sigma^2 = 7.2537
nu,d0 = 0, 0
# of censored = 45
r-value = 100
***************************************************************
x1 0.000000 31.622777
x2 0.000000 31.622777
x3 0.000000 31.622777
x4 0.000000 31.622777
***************************************************************
Posterior Estimates
x1 -1.454368 -4.965893 0.000003
x2 -0.834533 -3.352712 0.001131
x3 1.027901 3.242982 0.001608
x4 0.817880 2.993917 0.003471
Green (1997) discusses the issue of heteroscedasticity in tobit models
.::UdecomBooks::.
and points out that maximum likelihood estimates are problematical in this
circumstance. Studies relying on a single dummy variable or one with group-
wise heteroscedasticity indicate that heteroscedasticity presents a serious
problem for maximum likelihood estimation. An approach to solving this
problem is to replace the constant variance term σ with σi , where specifica-
tion of a particular model for σi needs to be made by the investigator. This
of course complicates the task of maximizing the likelihood function.
The Bayesian approach introduced by Geweke (1993) implemented in
the function tobit g eliminates the need to specify the form of the non-
constant variance and accommodates the case of outliers as well as non-
constant variance. In addition, the estimated parameters vi based on the
mean of Gibbs draws (representing the mean of posterior distribution for
these parameters) can be used to draw inferences regarding the nature of
the non-constant variance.
7.5 Chapter summary

This chapter demonstrated the regression function library programs for max-
imum likelihood estimation of logit, probit, and tobit models. In addition
to these maximum likelihood estimation functions, two powerful functions,
probit g and tobit g that implement a Gibbs sampling approach to esti-
mating these models were discussed and illustrated. These functions have
the advantage of dealing with non-constant variance and outliers that pose
problems for maximum likelihood estimation of these models.
In addition to the functions described in this chapter, there are functions
in the spatial econometrics function library that implement Gibbs sampling
estimation of heteroscedastic spatial autoregressive probit and tobit mod-
els. LeSage (1997) describes these models and methods and compares them
to an alternative approach set forth in McMillen (1992) based on the EM
algorithm.
Although we did not utilize the coda convergence diagnostic functions
(discussed in detail in Chapter 6) in conjunction with probit g and tobit g,
these convergence diagnostics should of course be used in any application of
these functions to diagnose convergence of the Gibbs sampler.
.::UdecomBooks::.
Chapter 7 Appendix
The maximum likelihood logit, mlogit, probit and tobit estimation functions
as well as the Gibbs sampling functions probit g and tobit g discussed in this
chapter are part of the regression function library in subdirectory regress.
Regression functions discussed in this chapter
------- logit, probit, tobit and Gibbs regression functions -----------


mlogit_d - demonstrates mlogit regression
probit_gd - demonstrates Bayesian probit model
tobit_d2 - right-censored tobit regression demo
tobit_gd - demonstrates Bayesian tobit model
tobit_gd2 - demonstrates right-censored Bayesian tobit model

maxlik - used by tobit
228
.::UdecomBooks::.

norm_pdf - used by probit
prt_gibbs - prints results for tobit_g,probit_g
to_rlike - used by right-censored tobit (likelihood)
to_llike - used by left-censored tobit (likelihood)
The spatial autoregressive versions of the probit and tobit models men-
tioned in the chapter summary are in the spatial econometrics function
library in subdirectory spatial.
------- spatial econometrics program functions -----------
sarp_g - Gibbs sampling Bayesian sar Probit model

sart_g - Gibbs sampling Bayesian sar Tobit model
sacp_g - Gibbs sampling Bayesian sac Probit model
semp_g - Gibbs sampling Bayesian sem Probit model
semt_g - Gibbs sampling Bayesian sem tobit model
------- demonstrations -----------
sarp_gd - sar Probit Gibbs sampling demo

sart_gd - sar Tobit model Gibbs sampling demo
sacp_gd - sac Probit Gibbs sampling demo
sact_gd - sac Tobit model Gibbs sampling demo
semp_gd - sem Probit Gibbs sampling demo
semt_gd - sem Tobit model Gibbs sampling demo
anselin - Anselin (1988) Columbus crime data

g_rho - used by sar_g,sart_g,sarp_g
wmat.dat - Anselin (1988) 1st order contiguity
c_sacl - conditional distribution for sacp_g, sact_g
c_sacp - conditional distribution for sacp_g, sact_g
c_sar - conditional distribution for sarp_g, sart_g
c_sem - conditional distribution for semp_g, semt_g
.::UdecomBooks::.
Chapter 8
Simultaneous Equation
Models
The regression function library contains routines for two-stage least-squares,

three-stage least-squares and seemingly unrelated regression models. This
chapter discusses these functions and provides examples that use these esti-
mation routines. Section 8.1 covers two-stage least-squares with three-stage
least-squares demonstrated in Section 8.2. Seemingly unrelated regression
equations are discussed and illustrated in Section 8.3.
8.1 Two-stage least-squares models

The problem of simultaneity in regression models arises when the explana-
tory variables matrix cannot be viewed as fixed during the dependent vari-
able generation process. An example would be a simple Keynesian consump-
tion function model:
Ct = α + βYt + ε (8.1)
Where Ct represents aggregate consumption at time t and Yt is disposable
income at time t.
Because of the national income accounting identity in this simple model:
Yt ≡ Ct + It , we cannot plausibly argue that the variable Yt is fixed in re-
peated samples. The accounting identity implies that every time Ct changes,
so will the variable Yt .
A solution to the ‘simultaneity problem’ is to rely on two-stage least-
squares estimation rather than ordinary least-squares. The function tsls will
230
.::UdecomBooks::.
CHAPTER 8. SIMULTANEOUS EQUATION MODELS 231
implement this estimation procedure. The documentation for the function

is:
PURPOSE: computes Two-Stage Least-squares Regression

---------------------------------------------------
USAGE: results = tsls(y,yendog,xexog,xall)
yendog = endogenous variables matrix (nobs x g)
xexog = exogenous variables matrix for this equation
xall = all exogenous and lagged endogenous variables
in the system
---------------------------------------------------
results.meth = ’tsls’
results.bhat = bhat estimates
results.tstat = t-statistics
results.yhat = yhat predicted values
results.nobs = nobs,
results.nendog = # of endogenous
results.nexog = # of exogenous
results.nvar = results.nendog + results.nexog
--------------------------------------------------
NOTE: you need to put a constant term in the x1 and xall matrices
--------------------------------------------------
SEE ALSO: prt(results), prt_reg(), plt(), thsls()
---------------------------------------------------
As an example of using this function, consider example 8.1, where a

dependent variable y2 is generated that depends on two variables y1 and
x2. The variable y1 is generated using the same error vector evec as in the
generation process for y2 so this variable is correlated with the error term.
This violates the Gauss-Markov assumption that the explanatory variables
in the regression model are fixed in repeated sampling, leading to biased and
inconsistent estimates from least-squares.
% ----- Example 8.1 Two-stage least-squares

nobs = 200;
x1 = randn(nobs,1); x2 = randn(nobs,1);
b1 = 1.0; b2 = 1.0; iota = ones(nobs,1);
y1 = zeros(nobs,1); y2 = zeros(nobs,1);
.::UdecomBooks::.
% create simultaneously determined variables y1,y2

for i=1:nobs;
y1(i,1) = iota(i,1) + x1(i,1)*b1 + evec(i,1);
y2(i,1) = iota(i,1) + y1(i,1) + x2(i,1)*b2 + evec(i,1);
end;
vname1 = strvcat(’y1-eqn’,’y2 variable’,’constant’,’x1 variable’);
vname2 = strvcat(’y2-eqn’,’y1 variable’,’constant’,’x2 variable’);
% use all exogenous in the system as instruments
xall = [iota x1 x2];
% do ols regression for comparison with two-stage estimates
result1 = ols(y2,[y1 iota x2]);
prt(result1,vname2);
% do tsls regression
result2 = tsls(y2,y1,[iota x2],xall);
prt(result2,vname2);
Note that the MATLAB function tsls requires that the user supply a
matrix of variables that will be used as instruments during the first-stage
regression. In example 8.1 all exogenous variables in the system of equations
are used as instrumental variables, a necessary condition for two-stage least-
squares to produce consistent estimates.
Least-squares and two-stage least-squares estimates from example 8.1 are
presented below. The least-squares estimates exhibit the simultaneity bias
by producing estimates that are significantly different from the true values
of β = 1, used to generate the dependent variable y2. In contrast, the two-
stage least-squares estimates are much closer to the true values of unity.
The results from the example also illustrate that the simultaneity bias tends
to center on the parameter estimates associated with the constant term and
the right-hand-side endogenous variable y1. The coefficient estimates for
x2 from least-squares and two-stage least-squares are remarkably similar.
Dependent Variable = y2-eqn
R-squared = 0.9144
sigma^2 = 0.5186
***************************************************************
y1 variable 1.554894 42.144421 0.000000
constant 0.462866 7.494077 0.000000
x2 variable 0.937300 18.154273 0.000000
Two Stage Least-squares Regression Estimates

Dependent Variable = y2-eqn
.::UdecomBooks::.
R-squared = 0.7986
sigma^2 = 1.2205
***************************************************************
y1 variable 0.952439 10.968835 0.000000
constant 1.031016 9.100701 0.000000
x2 variable 0.937037 11.830380 0.000000
Implementation of the MATLAB function tsls is relatively straightfor-

ward as illustrated by the following code:
if (nargin ~= 4); error(’Wrong # of arguments to tsls’); end;

results.meth = ’tsls’;
[nobs1 g] = size(y1); [nobs2 k] = size(x1); [nobs3 l] = size(xall);
results.nendog = g; results.nexog = k; results.nvar = k+g;
if nobs1 == nobs2;
if nobs2 == nobs3, nobs = nobs1; end;
else
error(’tsls: # of observations in yendog, xexog, xall not the same’);
end;
results.y = y; results.nobs = nobs;
% xall contains all explanatory variables
% x1 contains exogenous; % y1 contains endogenous
xapxa = inv(xall’*xall);
% form xpx and xpy
xpx = [y1’*xall*xapxa*xall’*y1 y1’*x1
x1’*y1 x1’*x1];
xpy = [y1’*xall*xapxa*xall’*y
x1’*y ];
xpxi = inv(xpx);
results.beta = xpxi*xpy; % bhat
results.yhat = [y1 x1]*results.beta; % yhat
results.resid = y - results.yhat; % residuals
results.sige = sigu/(nobs-k-g); % sige
tmp = results.sige*(diag(xpxi));
ym = y - ones(nobs,1)*mean(y);
rsqr1 = sigu; rsqr2 = ym’*ym;
rsqr1 = rsqr1/(nobs-k-g);
rsqr2 = rsqr2/(nobs-1.0);
results.dw = (ediff’*ediff)/sigu; % durbin-watson
.::UdecomBooks::.
After error checking on the input arguments, the function simply forms
the matrices needed to carry out two-stage least-squares estimation of the
parameters (e.g., Green, 1997). Given parameter estimates, the usual sum-
mary statistics measuring fit and dispersion of the parameters are calculated
and added to the results structure variable that is returned by the function.
Of course, a corresponding set of code to carry out printing the results was
added to the prt reg function, which is called by the wrapper function prt.
As a final example that clearly demonstrates the nature of inconsistent
estimates, consider example 8.2 where a Monte Carlo experiment is carried
out to compare least-squares and two-stage least-squares over 100 runs. In
the program code for example 8.2, we rely on the utility function mprint
described in Chapter 3, to produce a tabular print-out of our results with
row and column labels. Another point to note regarding the code is that we
dimension our matrices to store the coefficient estimates as (100,3), where we
have 3 parameter estimates and 100 Monte Carlo runs. This facilitates using
the MATLAB functions mean and std to compute means and standard
deviations of the resulting estimates. Recall these functions work down the
columns of matrices to compute averages and standard deviations of each
column in the input matrix.
% ----- Example 8.2 Monte Carlo study of ols() vs. tsls()
nobs = 200;
x1 = randn(nobs,1); x2 = randn(nobs,1);
b1 = 1.0; b2 = 1.0; iota = ones(nobs,1);
y1 = zeros(nobs,1); y2 = zeros(nobs,1);
for i=1:nobs;
y1(i,1) = iota(i,1) + x1(i,1)*b1 + evec(i,1);
y2(i,1) = iota(i,1) + y1(i,1) + x2(i,1)*b2 + evec(i,1);
end;
% use all exogenous in the system as instruments
xall = [iota x1 x2];
niter = 100; % number of Monte Carlo loops
bols = zeros(niter,3); % storage for ols results
b2sls = zeros(niter,3); % storage for 2sls results
disp(’patience -- doing 100 2sls regressions’);
for iter=1:niter; % do Monte Carlo looping
y1 = zeros(nobs,1); y2 = zeros(nobs,1); evec = randn(nobs,1);
for i=1:nobs;
y1(i,1) = iota(i,1)*1.0 + x1(i,1)*b1 + evec(i,1);
y2(i,1) = iota(i,1)*1.0 + y1(i,1)*1.0 + x2(i,1)*b2 + evec(i,1);
end;
result1 = ols(y2,[y1 iota x2]); % do ols regression
.::UdecomBooks::.
result2 = tsls(y2,y1,[iota x2],xall); % do tsls regression

bols(iter,:) = result1.beta’;
b2sls(iter,:) = result2.beta’;
end; % end Monte Carlo looping
% find means and std deviations over the niter runs
bolsm = mean(bols); b2slsm = mean(b2sls);
bolss = std(bols); b2slss = std(b2sls);
% print results
fprintf([’OLS results over ’,num2str(niter),’ runs\n’]);
in.rnames = strvcat(’Coefficients’,’b1’,’b2’,’b3’);
in.cnames = strvcat(’Mean’,’std dev’);
mprint([bolsm’ bolss’],in);
fprintf([’TSLS results over ’,num2str(niter),’ runs\n’]);
mprint([b2slsm’ b2slss’],in);
The biased and inconsistent estimates produced by least-squares in the

face of simultaneity should exhibit estimates that deviate from the true
β = 1 values used in the experiment. The fact that this bias remains even
when averaged over 100 Monte Carlo runs, illustrates an important point
about consistent versus inconsistent estimates.
OLS results over 100 runs
Coefficients Mean std dev
b1 1.4769 0.0233
b2 0.4820 0.0460
b3 1.0581 0.0418
TSLS results over 100 runs

Coefficients Mean std dev
b1 0.9986 0.0694
b2 1.0162 0.1105
b3 0.9972 0.0791
8.2 Three-stage least-squares models

When the disturbances from one equation in a system of simultaneous equa-
tions are contemporaneously correlated with those from other equations, we
can rely on three-stage least-squares as a method of incorporating this cor-
relation structure in the resulting parameter estimates. This will tend to
increase the precision of the estimates. A potential problem with this esti-
mation procedure is that model misspecification in one equation can work
through the correlation structure in the disturbances to contaminate the
specification of other equations in the system.
We face a challenge in constructing a three-stage least-squares routine
because the user must provide input to our MATLAB function thsls that
.::UdecomBooks::.
indicates the equation structure of the entire system as well as all of the
associated variable vectors. This problem was solved using MATLAB struc-
ture variables. An additional challenge is that the results structure must
return results for all equations in the system, which we also solve using a
MATLAB structure variable, as in the case of vector autoregressive models
discussed in Chapter 5.
The documentation for the thsls function is:
PURPOSE: computes Three-Stage Least-squares Regression
for a model with neqs-equations
---------------------------------------------------
USAGE: results = thsls(neqs,y,Y,X)
where:
neqs = # of equations
y = an ’eq’ structure containing dependent variables
e.g. y(1).eq = y1; y(2).eq = y2; y(3).eq = y3;
Y = an ’eq’ structure containing RHS endogenous
e.g. Y(1).eq = []; Y(2).eq = [y1 y3]; Y(3).eq = y2;
X = an ’eq’ structure containing exogenous/lagged endogenous
e.g. X(1).eq = [iota x1 x2];
X(2).eq = [iota x1];
X(3).eq = [iota x1 x2 x3];
---------------------------------------------------
NOTE: X(i), i=1,...,G should include a constant vector
if you want one in the equation
---------------------------------------------------
RETURNS a structure:
result.meth = ’thsls’
result(eq).beta = bhat for each equation
result(eq).tstat = tstat for each equation
result(eq).tprob = tprobs for each equation
result(eq).resid = residuals for each equation
result(eq).yhat = yhats for each equation
result(eq).y = y for each equation
result(eq).rsqr = r-squared for each equation
result(eq).rbar = r-squared adj for each equation
result(eq).nvar = nvar in each equation
result(eq).sige = e’e/nobs for each equation
result(eq).dw = Durbin-Watson
result.nobs = nobs
result.neqs = neqs
result.sigma = sig(i,j) across equations
result.ccor = correlation of residuals across equations
--------------------------------------------------
SEE ALSO: prt, prt_eqs, plt
---------------------------------------------------
As an example of using the function, consider example 8.3 that gener-
.::UdecomBooks::.
ates a three equation system with cross-equation correlation between the

disturbances in the second and third equations. Variable y1 appears as
a right-hand-side endogenous variable in the y2 equation and variable y2
appears in a similar way in the y3 equation.
The function thsls requires that we set up structure variables that take
the form of var name(eq#).eq, where eq# is the number of the equation.
We need to input: left-hand-side or dependent variables for each equation in
the first structure variable input argument to the function, right-hand-side
endogenous variables in the second structure variable argument to the func-
tion, and exogenous variables for each equation as the third input argument
to the function. In the example, the first structure variable was named y, the
second is Y and the third is X, following the usual econometrics textbook
naming conventions.
Some things to note about the input arguments are that:
1. We enter an empty matrix for equations where no right-hand-side vari-

ables exist, for example: Y(1).eq = [];
2. We enter constant term vectors in the structure variable containing

the exogenous variables in the model for all equations where we desire
their presence.
3. The function does not check for identification, that is the user’s re-
sponsibility.
4. The structure variable must have a name ending with .eq.
5. The number of observations in all equations for all variable vectors

input to the function must be equal.
We construct three separate string vectors using the MATLAB strvcat

function that contain the variable names for each equation. This function
is used again to concatenate these three string vectors into a single long
string vector for the printing function. One could of course simply construct
a single long string vector of names, but the approach taken here allows
us to carry out two-stage least-squares estimates or ordinary least-squares
estimates on each equation and print results with the corresponding variable
names.
The variable name strings must of course reflect the order of the variables
in each equation and they need to follow the equation order from first to
last.
.::UdecomBooks::.
% ----- Example 8.3 Three-stage least-squares

nobs = 100; neqs = 3;
x1 = randn(nobs,1); x2 = randn(nobs,1); x3 = randn(nobs,1);
b1 = 1.0; b2 = 1.0; b3 = 1.0; iota = ones(nobs,1);
y1 = zeros(nobs,1); y2 = zeros(nobs,1); y3 = zeros(nobs,1);
e = randn(nobs,3);
e(:,2) = e(:,3) + randn(nobs,1); % create cross-eqs corr
for i=1:nobs;
y1(i,1) = iota(i,1)*10.0 + x1(i,1)*b1 + e(i,1);
y2(i,1) = iota(i,1)*10.0 + y1(i,1)*1.0 + x2(i,1)*b2 + e(i,2);
y3(i,1) = iota(i,1)*10.0 + y2(i,1)*1.0 + x2(i,1)*b2 + x3(i,1)*b3 + e(i,3);
end;
vname1 = strvcat(’y1-LHS’,’constant’,’x1 var’);
vname2 = strvcat(’y2-LHS’,’y1-RHS’,’constant’,’x2 var’);
vname3 = strvcat(’y3-LHS’,’y2-RHS’,’constant’,’x2 var’,’x3 var’);
% set up a structure for y containing y’s for each eqn
y(1).eq = y1; y(2).eq = y2; y(3).eq = y3;
% set up a structure for Y (RHS endogenous) for each eqn
Y(1).eq = []; Y(2).eq = [y1]; Y(3).eq = [y2];
% set up a structure for X (exogenous) in each eqn
X(1).eq = [iota x1]; X(2).eq = [iota x2]; X(3).eq = [iota x2 x3];
result = thsls(neqs,y,Y,X); % do thsls regression
vname = strvcat(vname1,vname2,vname3);
prt(result,vname);
A printout produced by running example 8.3 is shown below. A separate

function prt eqs was developed to carry out the printing, but as with print-
ing results from other functions in the Econometrics Toolbox, the wrapper
function prt will work to produce printed output by calling the appropriate
function.
The printed results are presented in order for each equation in the system
and then the cross-equation error covariance and correlation estimates are
presented. From the results shown for example 8.3 we see that all of the
constant term estimates are near the true value of 10, and all other estimates
are near the true values of unity. We see no evidence of cross-equation
correlation between the disturbances of equations 1 and 2, but in equations
2 and 3 where we generated a correlation, the estimated coefficient is 0.58.
Three Stage Least-squares Estimates -- Equation 1

Dependent Variable = y1-LHS
R-squared = 0.4808
sigma^2 = 3.3261
.::UdecomBooks::.
***************************************************************
constant 9.985283 100.821086 0.000000
x1 var 1.101060 9.621838 0.000000

R-squared = 0.6650
sigma^2 = 3.3261
***************************************************************
y1-RHS 0.959248 7.615010 0.000000
constant 10.389051 8.177007 0.000000
x2 var 0.925196 7.326562 0.000000

R-squared = 0.9156
sigma^2 = 3.3261
***************************************************************
y2-RHS 1.005874 9.667727 0.000000
constant 9.819170 4.705299 0.000008
x2 var 0.969571 6.898767 0.000000
x3 var 1.097547 14.974708 0.000000
Cross-equation sig(i,j) estimates

equation y1-LHS y2-LHS y3-LHS
y1-LHS 0.9779 0.1707 -0.1108
y2-LHS 0.1707 1.4383 0.7660
y3-LHS -0.1108 0.7660 0.9005
Cross-equation correlations
equation y1-LHS y2-LHS y3-LHS
y1-LHS 1.0000 0.1439 -0.1181
y2-LHS 0.1439 1.0000 0.6731
y3-LHS -0.1181 0.6731 1.0000
There is also a function plt eqs that is called by the wrapper function
plt to produce graphs of the actual versus predicted and residuals for each
equation of the system. The graphs are produced in a ‘for loop’ with a
pause between the graphs for each equation in the system. A virtue of the
.::UdecomBooks::.
multiple equation result structure is that printing and plotting functions

written previously for single equation regression methods can be adapted.
For the case of systems of equations we can embed the previous code in ‘for
loops’ that range over the equations in the results structure variable.
As an example of this, consider the following code fragment from the
function plt eqs that loops over the structure variable results accessing
the actual and predicted values for each equation as well as the residuals.
tt=1:nobs;
clf;
cnt = 1;
for j=1:neqs;
nvar = results(j).nvar;
subplot(2,1,1), plot(tt,results(j).y,’-’,tt,results(j).yhat,’--’);
if nflag == 1
title([upper(results(1).meth), ’ Act vs. Predicted ’,vnames(cnt,:)]);
else
title([upper(results(1).meth), ’ Act vs. Predicted eq ’,num2str(j)]);
end;
subplot(2,1,2), plot(tt,results(j).resid)
cnt = cnt+nvar+1;
pause;
end;
8.3 Seemingly unrelated regression models

A function sur exists for estimating seemingly unrelated regression models,
where a series of regression relations exhibit covariation in the disturbances
across equations. An initial least-squares regression is carried out for all
equations in the model and the residuals from this series of regressions are
used to form the cross-equation error covariance matrix estimates. One
design consideration of interest here is that we rely on a special function
olse to carry out the initial least-squares regressions rather than ols. This
function avoids many of the computations performed in ols because our only
concern is a residual vector from least-squares squares estimation.
The function olse is shown below, where we rely on the Cholesky de-
composition approach to solving the least-squares problem, again for speed.
function resid=olse(y,x)
% PURPOSE: OLS regression returning only residual vector
%---------------------------------------------------
% USAGE: residual = olse(y,x)
.::UdecomBooks::.
%---------------------------------------------------
% RETURNS: the residual vector
%---------------------------------------------------
if (nargin ~= 2); error(’Wrong # of arguments to olse’); end;
beta = x\y;
resid = y - x*beta;
The documentation for the sur function is shown below with the infor-
mation regarding the results structure returned by the function omitted to
save space.
PURPOSE: computes seemingly unrelated regression estimates

for a model with neqs-equations
---------------------------------------------------
USAGE: results = sur(neqs,y,x,iflag,info)
or, results = sur(neqs,y,x) (for no iteration)
where:
neqs = # of equations
y = an ’eq’ structure with dependent variables
e.g. y(1).eq = y1; y(2).eq = y2; y(3).eq = y3;
x = an ’eq’ structure with explanatory variables
e.g. x(1).eq = [iota x1 x4];
x(2).eq = [iota x1];
x(3).eq = [iota x1 x2 x5];
iflag = 1 for iteration on error covariance matrix,
0 for no iteration (0 = default)
info = a structure for iteration options:
info.itmax = maximum # of iterations (default = 100)
info.crit = criterion for absolute bhat change
(default = 0.001)
---------------------------------------------------
NOTE: x(i), i=1,...,G should include a constant vector
if you want one in the equation
---------------------------------------------------
SEE ALSO: prt(results), prt_reg(), plt(results)
---------------------------------------------------
As noted in the documentation for sur, there is an option to iterate

on the cross-equation error covariance matrix estimates. Iteration involves
using the residuals from the initial seemingly unrelated regression system
(based on the least-squares residuals) to form another estimate of the error
covariance matrix. This process of using the sur residuals to form an updated
error covariance matrix estimate continues until there is a small change in
the β̂ estimates from iteration to iteration. A structure variable with fields
‘itmax’ and ‘crit’ allows the user to input these iteration options. The ‘itmax’
field allows the user to specify the maximum number of iterations and the
.::UdecomBooks::.
field ‘crit’ allows input of the convergence criterion. The latter scalar input
represents the change in the sum of the absolute value of the β̂ estimates
from iteration to iteration. When the estimates change by less than the
‘crit’ value from one iteration to the next, convergence has been achieved
and iteration stops.
As an example, consider the Grunfeld investment model, where we have
annual investment time-series covering the period 1935-54 for five firms.
In addition, market value of the firm and desired capital stock serve as
explanatory variables in the model. The explanatory variables represent
market value lagged one year and a constructed desired capital stock variable
that is also lagged one year. See Theil (1971) for an explanation of this
model.
Example 8.4 estimates the Grunfeld model using the sur function with
and without iteration.
% ----- Example 8.4 Using the sur() function

load grun.dat; % grunfeld investment data (page 650, Green 1997)
y1 = grun(:,1); x11 = grun(:,2); x12 = grun(:,3); % general electric
y2 = grun(:,4); x21 = grun(:,5); x22 = grun(:,6); % westinghouse
y3 = grun(:,7); x31 = grun(:,8); x32 = grun(:,9); % general motors
y4 = grun(:,10);x41 = grun(:,11);x42 = grun(:,12);% chrysler
y5 = grun(:,13);x51 = grun(:,14);x52 = grun(:,15);% us steel
nobs = length(y1); iota = ones(nobs,1);
vname1 = strvcat(’I gen motors’,’cons’,’fgm’,’cgm’);
vname2 = strvcat(’I chrysler’,’const’,’fcry’,’ccry’);
vname3 = strvcat(’I gen electric’,’const’,’fge’,’cge’);
vname4 = strvcat(’I westinghouse’,’const’,’fwest’,’cwest’);
vname5 = strvcat(’I us steel’,’const’,’fuss’,’cuss’);
% set up a structure for y in each eqn
% (order follows that in Green, 1997)
y(1).eq = y3; % gm
y(2).eq = y4; % chrysler
y(3).eq = y1; % general electric
y(4).eq = y2; % westinghouse
y(5).eq = y5; % us steel
% set up a structure for X in each eqn
X(1).eq = [iota x31 x32];
X(2).eq = [iota x41 x42];
X(3).eq = [iota x11 x12];
X(4).eq = [iota x21 x22];
X(5).eq = [iota x51 x52];
% do sur regression with iteration
neqs = 5; iflag = 1; % rely on default itmax, crit values
result = sur(neqs,y,X,iflag);
% do sur regression with no iteration
result2 = sur(neqs,y,X);
.::UdecomBooks::.
vname = strvcat(vname1,vname2,vname3,vname4,vname5);
prt(result,vname); % print results for iteration
prt(result2,vname);% print results for no iteration
The results indicate that iteration does not make a great deal of different
in the resulting estimates. To conserve on space, we present comparative
results for a two equation model with General Electric and Westinghouse
for which results are reported in Theil (1971).
% RESULTS with iteration

Seemingly Unrelated Regression -- Equation 1
Dependent Variable = I gen electr
System R-sqr = 0.6124
R-squared = 0.6868
sigma^2 = 11961.0900
Nobs, Nvars = 20, 3
***************************************************************
constant -30.748395 -1.124423 0.276460
fge 0.040511 3.021327 0.007696
cge 0.135931 5.772698 0.000023

Dependent Variable = I westinghouse
R-squared = 0.7378
sigma^2 = 2081.1840
Nobs, Nvars = 20, 3
***************************************************************
constant -1.701600 -0.245598 0.808934
fwest 0.059352 4.464549 0.000341
cwest 0.055736 1.143147 0.268821

equation I gen electr I westinghouse
I gen electr 702.2337 195.3519
I westinghouse 195.3519 90.9531
I gen electr 1.0000 0.7730
.::UdecomBooks::.
% RESULTS without iteration

Dependent Variable = I gen electr
R-squared = 0.6926
sigma^2 = 11979.6253
Nobs, Nvars = 20, 3
***************************************************************
constant -27.719317 -1.019084 0.322448
fge 0.038310 2.870938 0.010595
cge 0.139036 5.948821 0.000016

Dependent Variable = I westinghouse
R-squared = 0.7404
sigma^2 = 2082.7474
Nobs, Nvars = 20, 3
***************************************************************
constant -1.251988 -0.180970 0.858531
fwest 0.057630 4.337133 0.000448
cwest 0.063978 1.314408 0.206168

I gen electr 689.4188 190.6363
I gen electr 1.0000 0.7650
8.4 Chapter summary

Estimation functions for systems of equations can be constructed by using
MATLAB structure variables to input data vectors and matrices for each
equation in the model. Structure variables can also be used to return esti-
mation results for each equation. This allows us to use many of the same
approaches developed for single equation regression models for printing and
.::UdecomBooks::.
plotting results by embedding the previous code in a for-loop over the equa-
tions in the model.
.::UdecomBooks::.
Chapter 8 Appendix
The tsls, thsls and sur estimation functions discussed in this chapter are
part of the regression function library in the subdirectory regress.
Regression functions discussed in this chapter
------- tsls, thsls, sur regression functions -----------



246
.::UdecomBooks::.
Chapter 9
Distribution functions library
The distributions function library contains routines to carry out calculations

based on a host of alternative statistical distributions. For each statistical
distribution, functions exist to: calculate the probability density function
(pdf), cumulative density function (cdf), inverse or quantile function (inv),
as well as generate random numbers. A naming convention is used for these
functions that takes the form of: a four-letter abbreviation for the statistical
distribution name followed by an underscore character and the three letter
codes, cdf, pdf, inv or rnd. As an example, the beta statistical distribution
has four functions named: beta cdf, beta pdf, beta inv, beta rnd.
The library contains these four functions for the following 12 distribu-
tions with the indicated four-letter abbreviations.
1. beta distribution (beta)
2. binomial distribution (bino)
3. chi-square distribution (chis)
4. F-distribution (fdis)
5. gamma distribution (gamm)
6. hypergeometric distribution (hypg)
7. log-normal distribution (logn)
8. logistic distribution (logt)
9. multivariate normal distribution (norm)
247
.::UdecomBooks::.
CHAPTER 9. DISTRIBUTION FUNCTIONS LIBRARY 248
10. poisson distribution (pois)
11. standard normal distribution (stdn)
12. Student t-distribution (tdis)
There are also some related more special purpose functions. These are
shown in the list below:
chis prb - computes marginal probabilities for chi-squared distributed

statistics
fdis prb - computes marginal probabilities for F-distributed statistics
tdis prb - computes marginal probabilities for t-distributed statistics
norm crnd - generates normally distributed random variates from a

contaminated normal distribution, (1 − γ)N (0, 1) + γN (0, σ).
nmlt rnd - generates left-truncated normal draws.
nmrt rnd - generates right-truncated normal draws.
unif rnd - generates uniform draws between upper and lower limits.
wish rnd - generates random draws from a Wishart distribution.
Section 9.1 demonstrates use of the pdf, cdf, inv, rnd statistical distri-
bution functions and the specialized functions are taken up in Section 9.2.
9.1 The pdf, cdf, inv and rnd functions

Each of the 12 distributions in the library has a demonstration program
that illustrates use of the four associated functions, pdf,cdf,inv,rnd. For
example, the demonstration function beta d carries out random draws from
the beta distribution, based on parameters (a, b), where the mean of the
distribution is: a/(a + b) and the variance is: ab/((a + b)2 (a + b + 1)).
Given a sample of 1000 random beta(a,b) draws, the mean and variance
are computed and compared to the theoretical values. Next, a grid of 1000
uniform random values is generated and sorted from low to high. The pdf,
cdf and quantiles are generated for this grid of values. Plotting the pdf,
cdf and quantile values against the input grid should produce smooth den-
sities that conform to the beta distribution shape. Example 9.1 shows the
demonstration program for the beta distribution.
.::UdecomBooks::.
% ----- Example 9.1 Beta distribution function example

n = 1000; a = 10; b = 5;
tst = beta_rnd(n,a,b); % generate random draws
% mean should equal a/(a+b)
fprintf(’mean should = %16.8f \n’,a/(a+b));
fprintf(’mean of draws = %16.8f \n’,mean(tst));
% variance should equal a*b/((a+b)*(a+b)*(a+b+1))
fprintf(’variance should = %16.8f \n’,(a*b)/((a+b)*(a+b)*(a+b+1)));
fprintf(’variance of draws = %16.8f \n’,std(tst)*std(tst));
tst = rand(n,1); tsort = sort(tst); % generate a grid of values
pdf = beta_pdf(tsort,a,b); % calculate pdf over the grid
cdf = beta_cdf(tsort,a,b); % calculate cdf over the grid
x = beta_inv(tsort,a,b); % calculate quantiles for the grid
subplot(3,1,1),
plot(tsort,pdf); xlabel([’pdf of beta(’,num2str(a),’,’,num2str(b),’)’]);
subplot(3,1,2),
plot(tsort,cdf); xlabel([’cdf of beta(’,num2str(a),’,’,num2str(b),’)’]);
subplot(3,1,3),
plot(tsort,x); xlabel([’inv of beta(’,num2str(a),’,’,num2str(b),’)’]);
The results from running the example 9.1 program are displayed below
and the plots produced by the program are shown in Figure 9.1.
mean should = 0.66666667
mean of draws = 0.66721498
variance should = 0.01388889
variance of draws = 0.01426892
These demonstration programs also provide a way to check the validity

of the 12 distribution functions. Many of the functions in the distributions
library were found on the internet and have been crafted by others. In all
cases, the function documentation was modified to be consistent with other
functions in the Econometrics Toolbox and in some cases, the functions were
converted to MATLAB from other languages. In cases where the MATLAB
functions were written by others, the names of the authors have been left in
the function source code.
9.2 The specialized functions

Some of the specialized functions need no explanation. For example, the abil-
ity to calculate probabilities associated with t−distributed, F −distributed
and χ2 -distributed statistics provided by tdis prb, fdis prb, chis prb
seems obviously useful.
As we have seen in Chapters 6 and 7, the ability to carry out ran-
dom draws from multivariate normal, chi-squared, gamma distributions and
.::UdecomBooks::.
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
pdf of beta(10,5)
1
0.5
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
cdf of beta(10,5)
1
0.8
0.6
0.4
0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
quantiles of beta(10,5)
Figure 9.1: Beta distribution demonstration program plots
left or right-truncated normal distributions is needed to implement Gibbs

sampling estimation.
Another random number generation function that is useful in this type
of estimation is the Wishart random draw function wish rnd. Consider
the case of a general multivariate normal model where m responses yij , j =
1, . . . , m exist from measurements taken at m different time intervals or
during multiple experimental runs. This situation, where we have a set of
n m-variate observations, leads to errors in the model εij , i = 1, . . . , n, j =
1, . . . , m, that are often assumed to take the form of an m−variate Normal
Nm (0, Σ), that depend on a set of parameters θ.
Consider a Bayesian analysis where we assume: an independence in the
priors for θ and Σ, a locally uniform prior for θ, and the non-informative
reference prior p(Σ) ∝ |Σ|−1/2(m+1) . This produces a posterior distribution
for Σ conditional on θ that takes the form of a Wishart distribution. Specif-
.::UdecomBooks::.
P
ically, Σ−1 ∼ Wm (S −1 (θ), n − m + 1), where S(θ) = ni=1 εij εij , (Box and
Tiao, 1992).
An example from Tanner (1991) provides an applied illustration of how
this distribution arises in the context of Gibbs sampling. Data on 30 young
rats growth measured over a span of five weeks were analyzed in Gelfand,
Hills, Racine-Poon and Smith (1990). Assuming a linear model for individ-
ual rats growth curves:
yij ∼ N (αi , βi xij , σ 2 ) (9.1)

Where, i = 1, . . . , 30 representing individual rats and j = 1, . . . , 5 denoting
the five weeks in the measurement period. The explanatory variable in the
model, xij measures the age of the ith rat at measurement period j. The
prior for the intercept and slope parameters in the model takes the form of
a multivariate normal, where all rats αi , βi are centered on α0 , β0 . The prior
variance-covariance structure (Σ) for these parameters takes the form of a
Wishart (W ) distribution.
This prior information can be written:
Ã ! Ã !
αi α0
=N ,W (9.2)
βi β0
The priors for:
α0 , β0 = N (γ, C)
Σ−1 = W (ρR)−1 , ρ)
σ 2 = IG(ν0 /2, ν0 τ02 /2)
Where IG denotes the inverted gamma distribution. Gelfand et al. (1990)

use C −1 = 0, ν0 = 0, ρ = 2 and R = diag(100, 0.1) to reflect a vague
prior. Letting V = (30Σ−1 + C −1 )−1 , the following sequence of conditional
distributions for the parameters in the model form the Gibbs sampler.
p(αi , βi | . . .) = N {(G(σ −2 Xi0 Yi + Σ−1 µ), G},
where: G = (σ −2 Xi0 Xi + Σ−1 )−1
p(µ| . . .) = N {V (30Σ−1 θ̄ + C −1 ν), V }

P30 0
where: θ̄ = (1/30) i=1 (αi , βi )
P
p(Σ−1 | . . .) = W {[ i (θi − µ)(θi − µ)0 + ρR]−1 , 30 + ρ}
.::UdecomBooks::.
P
p(σ 2 | . . .) = IG{(n + ν0 )/2, (1/2)[ i (Yi − Xi θi )0 (Yi − Xi θi ) + ν0 τ0 ]}
As a demonstration of random draws from the Wishart distribution con-

sider example 9.2, where we generate a symmetric matrix for input and
assign v = 10 degrees of freedom. The mean of the Wishart should equal
v times the input matrix. In the example, we carry out 1000 draws and
average over the random matrix draws to find the mean which is compared
to the theoretical mean.
% ----- Example 9.2 Random Wishart draws

ndraws = 1000;
n = 100; k=5;
x = randn(n,5); xpx = x’*x; xpxi = inv(xpx); v = 10;
w = zeros(k,k);
for i=1:ndraws;
w = w + wish_rnd(xpx,v);
end;
fprintf(’mean of wishart should = \n’);
mprint(v*xpx);
fprintf(’mean of wishart draws = \n’);
mprint(w/ndraws);
The output from example 9.2 is:
mean of wishart should =

749.0517 -20.5071 6.8651 75.1039 -57.3764
-20.5071 899.6492 13.0913 30.3349 -121.7186
6.8651 13.0913 912.6154 -3.2621 -186.6620
75.1039 30.3349 -3.2621 985.6231 27.4726
-57.3764 -121.7186 -186.6620 27.4726 945.3465
mean of wishart draws =

765.3774 -13.8467 26.8851 92.3882 -61.7533
-13.8467 946.6986 43.2879 26.6926 -129.1496
26.8851 43.2879 941.1585 0.2385 -185.0185
92.3882 26.6926 0.2385 997.2087 36.5898
-61.7533 -129.1496 -185.0185 36.5898 886.5470
Another set of specialized functions, nmlt rnd and nmrt rnd were used
to produce left- and right-truncated normal draws when Gibbs sampling
estimates for the probit and tobit models. Example 9.3 shows the use of
these functions and produces a series of three histograms based on draws
from the truncated normal distributions. It should be noted that one can
implement these function by simply drawing from a normal distribution and
rejecting draws that don’t meet the truncation restrictions. This however is
.::UdecomBooks::.
not very efficient and tends to produce a very slow routine. The functions
nmlt rnd and nmrt rnd are based on FORTRAN code that implements
an efficient method described in Geweke (1991).
% ----- Example 9.3 Left- and right-truncated normal draws
n = 1000; x = zeros(n,1);
% generate from -infinity < 0
for i=1:n; x(i,1) = nmrt_rnd(0); end;
subplot(3,1,1), hist(x,20); xlabel(’right-truncated at zero normal’);
% generate from 1 < infinity
for i=1:n; x(i,1) = nmlt_rnd(1); end;
subplot(3,1,2), hist(x,20); xlabel(’left-truncated at +1 normal’);
% generate from -1 < +infinity
for i=1:n; x(i,1) = nmlt_rnd(-1); end;
subplot(3,1,3), hist(x,20); xlabel(’left-truncated at -1 normal’);
1500
1000
500
0
-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0
right-truncated at zero normal
3000
2000
1000
0
1 1.5 2 2.5 3 3.5 4 4.5 5
left-truncated at +1 normal
1500
1000
500
0
-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 4
left-truncated at -1 normal
Figure 9.2: Histograms of truncated normal distributions
Example 9.4 constructs a program to explore the improvement in speed

of nmrt rnd over the inefficient approach based on rejection sampling of
.::UdecomBooks::.
draws from a normal. The program produces a set of 10 draws using 7

different truncation limits ranging from -3 to +3. In addition to using the
nmrt rnd function, we include a ‘while loop’ that simply draws from the
random normal centered on zero until it collects a sample of 10 draws that
meet the truncation restriction.
Consider that simple rejection will encounter very few normal draws that
take on values less than -3, requiring a large amount of time. Further, for
cases where the truncation limit is -3.5 or -4, we would have an even longer
wait.
% ----- Example 9.4 Rejection sampling of truncated normal draws
tt=-3:3; m = length(tt);n = 10;
x = zeros(n,m); x2 = zeros(n,m); time = zeros(m,2);
for j=1:m;
% generate from -infinity to tt(j)
% using nmrt_rnd function and keep track of time
t0 = clock;
for i=1:n; x(i,j) = nmrt_rnd(tt(j)); end;
time(j,1) = etime(clock,t0);
% generate from -infinity to tt(j) using rejection
% of draws that don’t meet the truncation constraint
cnt=1;
t0 = clock;
while cnt <= n
tst = randn(1,1);
if tst < tt(j)
x2(cnt,j) = tst;
cnt = cnt+1;
end;
end;
time(j,2) = etime(clock,t0);
end;
time
The timing results from running the program are shown below, indicating
that for the case of truncation at -3, the simple rejection approach took over
80 times as long as the nmrt rnd function to produce 10 draws. Similarly,
for the truncation limit at -2, rejection sampling took 7 times as long as
nmrt rnd. For truncation limits ranging from -1 to 3 rejection sampling
was equal or up to 3 times faster than the nmrt rnd function.
Rejection sampling versus nmrt_rnd function
(time in seconds)
truncation nmrt_rnd rejection
at
-3 0.0763 6.2548
.::UdecomBooks::.
-2 0.0326 0.2305
-1 0.0340 0.0301
0 0.0321 0.0125
1 0.0244 0.0091
2 0.0248 0.0081
3 0.0294 0.0092
Another special purpose distribution function is norm crnd that pro-

vides random draws from a contaminated normal distribution. This distri-
bution takes the form of a mixture of two normals: (1−γ)N (0, 1)+γN (0, σ).
Typical values for the parameters in the mixture would be: γ = 0.05 and a
large variance σ 2 = 10. This produces a distribution with outliers exhibiting
fatter tails than those of the standard normal distribution.
A graphical illustration is provided in Figure 9.3, where draws from
the contaminated normal distribution are plotted alongside those from a
standard normal. The fatter tails associated with the contaminated normal
are clearly visible in the figure.
0.4
contaminated
normal
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 9.3: Contaminated normal versus a standard normal distribution
.::UdecomBooks::.
9.3 Chapter summary

A large number of algorithms for generating random deviates, evaluating
the probability density, cumulative density and quantile functions exits for
a host of statistical distributions. These have been collected in a library of
functions that employ a common naming convention. The Gibbs sampling
estimation methods discussed in Chapters 6 and 7 rely heavily on this
library of distribution functions.
In addition to the library of 12 statistical distributions, some special-
ized functions exist that are useful in computing probabilities for hypothesis
testing and generating various results related to statistical distributions.
.::UdecomBooks::.
Chapter 9 Appendix
Distribution functions discussed in this chapter are in the subdirectory dis-

trib.
------- pdf, cdf, inverse functions -----------
beta_cdf - beta(a,b) cdf

beta_inv - beta inverse (quantile)
beta_pdf - beta(a,b) pdf
bino_cdf - binomial(n,p) cdf
bino_inv - binomial inverse (quantile)
bino_pdf - binomial pdf
chis_cdf - chisquared(a,b) cdf
chis_inv - chi-inverse (quantile)
chis_pdf - chisquared(a,b) pdf
chis_prb - probability for chi-squared statistics
fdis_cdf - F(a,b) cdf
fdis_inv - F inverse (quantile)
fdis_pdf - F(a,b) pdf
fdis_prb - probability for F-statistics
gamm_cdf - gamma(a,b) cdf
gamm_inv - gamma inverse (quantile)
gamm_pdf - gamma(a,b) pdf
hypg_cdf - hypergeometric cdf
hypg_inv - hypergeometric inverse
hypg_pdf - hypergeometric pdf
logn_cdf - lognormal(m,v) cdf
logn_inv - lognormal inverse (quantile)
logn_pdf - lognormal(m,v) pdf
logt_cdf - logistic cdf
logt_inv - logistic inverse (quantile)
logt_pdf - logistic pdf
norm_cdf - normal(mean,var) cdf
norm_inv - normal inverse (quantile)
norm_pdf - normal(mean,var) pdf
norm_prb - normal probability
257
.::UdecomBooks::.
normt_inv - truncated normal inverse

normrt_inv - right-truncated normal inverse
normlt_inv - left-truncated normal inverse
pois_cdf - poisson cdf
pois_inv - poisson inverse
pois_pdf - poisson pdf
stdn_cdf - std normal cdf
stdn_inv - std normal inverse
stdn_pdf - std normal pdf
tdis_cdf - student t-distribution cdf
tdis_inv - student t inverse (quantile)
tdis_pdf - student t-distribution pdf
tdis_prb - probability for t-statistics
------- random samples -----------
beta_rnd - random beta(a,b) draws

bino_rnd - random binomial draws
chis_rnd - random chi-squared(n) draws
fdis_rnd - random F(a,b) draws
gamm_rnd - random gamma(a,b) draws
hypg_rnd - random hypergeometric draws
logn_rnd - random log-normal draws
logt_rnd - random logistic draws
nmlt_rnd - left-truncated normal draw
nmrt_rnd - right-truncated normal draw
normt_rnd - truncated normal random draws
norm_crnd - contaminated normal random draws
norm_rnd - multivariate normal draws
pois_rnd - poisson random draws
tdis_rnd - random student t-distribution draws
unif_rnd - random uniform draws (lr,rt) interval
wish_rnd - random Wishart draws
-------- demonstration and test programs --------
beta_d - demo of beta distribution functions

bino_d - demo of binomial distribution functions
chis_d - demo of chi-squared distribution functions
fdis_d - demo of F-distribution functions
gamm_d - demo of gamma distribution functions
hypg_d - demo of hypergeometric distribution functions
logn_d - demo of lognormal distribution functions
logt_d - demo of logistic distribution functions
normt_d - demo of normt_rnd
normlt_d - demo of normlt_rnd
normrt_d - demo of normrt_rnd
pois_d - demo of poisson distribution functions
stdn_d - demo of std normal distribution functions
.::UdecomBooks::.
tdis_d - demo of student-t distribution functions

trunc_d - demo of truncated normal distribution function
unif_d - demo of uniform random distribution function
-------- support functions --------
betacfj - used by fdis_prb

betai - used by fdis_prb
bincoef - binomial coefficients
com_size - test and converts to common size
gammalnj - used by fdis_prb
is_scalar - test for scalar argument
.::UdecomBooks::.
Chapter 10
Optimization functions
library
The optimization function library contains routines for maximum like-

lihood estimation of econometric models. Given a likelihood function which
depends on the data as well as the parameters of the estimation problem,
these routines will find a set of parameter values that maximize the likelihood
function. In addition to discussing and illustrating use of the functions in the
optimization function library, we also show how to use the MATLAB
simplex optimization algorithms fmin and fmins as well as Em algorithms
to solve for maximum likelihood estimates in econometric models. An ex-
ample that demonstrates how to incorporate a new optimization algorithm
obtained from the Internet into the library is also provided.
Unlike Gauss software, MATLAB does not contain a single ‘maxlik’ func-
tion that serves as an ‘all purpose’ routine for solving econometric maximum
likelihood optimization problems. In an attempt to remedy this, a function
maxlik was devised for MATLAB. Unfortunately, this function is nowhere
near as robust or professionally crafted as the Gauss function by the same
name. This function as well as univariate and multivariate simplex routines,
algorithms from Numerical Recipes that have been re-coded for MATLAB,
some Internet optimization functions written for MATLAB and a general
purpose function minz contributed by Mike Cliff form the basis of the op-
timization library,.
The MathWorks sells an optimization toolbox that contains a host of al-
gorithms for solving constrained and unconstrained optimization problems.
These functions are not oriented toward econometric models, but rather
solution of very general optimization problems. To use these routines in
260
.::UdecomBooks::.
CHAPTER 10. OPTIMIZATION FUNCTIONS LIBRARY 261
solving econometric optimization problems, an interface similar to that de-

scribed in Section 10.3 would need to be developed to adapt the algorithms
for use in an econometric likelihood function setting.
Section 10.1 illustrates the use of a univariate simplex algorithm fmin
from the MATLAB toolbox. This type of algorithm is useful for cases where
the optimization problem involves a single parameter, a case often encoun-
tered with concentrated likelihood functions. It also demonstrates the use
of a multivariate simplex optimization function fmins from the MATLAB
toolbox.
Section 10.2 demonstrates use of the EM-algorithm approach to solving
maximum likelihood estimation problems. Finally, Section 10.3 deals with
the most general case where the likelihood function is optimized with respect
to a multivariate vector of parameters using numerical gradient methods.
10.1 Simplex optimization

Simplex optimization has an advantage in that it does not require com-
putation of the derivatives of the function being optimized. It also has a
disadvantage in that the usual estimates of dispersion for the parameters
in the model obtained from the numerical hessian matrix in gradient meth-
ods are not available. The MATLAB toolbox contains both a univariate
simplex algorithm fmin and a multivariate function fmins. The use of the
univariate function is taken up in Section 10.1.1 and fmins is illustrated in
Section 10.1.2.
10.1.1 Univariate simplex optimization

As an example of using the MATLAB fmin univariate simplex optimiza-
tion function to solve an econometric maximum likelihood estimation prob-
lem, consider the Box-Cox model. This model relies on a transformation
of the sample data vector y (λ) = (y λ − 1)/λ for the case where λ 6= 0
and y (λ) = ln(y) when λ = 0. The model may also involve an identical
transformation of the explanatory variables matrix, which we denote X (λ) .
A more general model involves different values, λ1 , λ2 associated with the
y (λ1 ) -transformation and X (λ2 ) -transformations, and a very general model
allows for individual λxi , i = 1, . . . , k for each explanatory variable vector in
the sample data matrix X.
We focus initially on the simple model involving a single parameter λ.
This model produces a log-linear model when λ1 = λ2 = 0, a semi-log
model when λ1 = 1, λ2 = 0, as well as more general flexible functional forms
.::UdecomBooks::.
√
associated with values like, λ = 1/2, or λ = 2. Some of the functional
forms associated with values for λ1 , λ2 are:
ln yi = β0 + ln Xi β + εi , λ1 = 0, λ2 = 0 (10.1)
yi = β0 + ln Xi β + εi , λ1 = 1, λ2 = 0 (10.2)
yi = β0 + Xi β + εi , λ1 = 1, λ2 = 1 (10.3)
The log likelihood function is concentrated with respect to β(λ), σ(λ),

producing a function of only a single parameter λ:
X
n
L(λ|y, X) = const + (λ − 1) ln yi − (n/2) ln σ̂ 2 (λ) (10.4)
i=1
where:
σ̂ 2 (λ) = [y (λ) − X (λ) β̂(λ)0 (y (λ) − X (λ) β̂(λ)]/n (10.5)

(λ)0 (λ) −1 (λ)0 (λ)
β̂(λ) = (X X ) X y
To minimize the log of this likelihood function with respect to the param-
eter λ we can use the MATLAB function fmin that implements a simplex
optimization procedure. This function also allows us to set lower and upper
limits on the parameter λ to which the simplex search will be constrained.
It is often the case that, values of −2 ≤ λ ≤ 2 are thought to represent
a reasonable range of feasible values for the parameter λ in the Box-Cox
model.
Our first task in using fmin is to write a log-likelihood function that
evaluates the concentrated log-likelihood for any value of the parameter λ
and returns a scalar value equal to the negative of the log-likelihood function.
(Minimizing the negative log-likelihood is equivalent to maximizing the log-
likelihood.) This function is shown below:
function like = box_lik(lam,y,x,model);

% PURPOSE: evaluate Box-Cox model concentrated likelihood function
%----------------------------------------------------
% USAGE: like = box_lik(lam,y,x,model)
% where: lam = box-cox parameter (scalar)
% y = dependent variable vector (un-transformed)
% x = explanatory variables matrix (un-transformed)
% model = 0 for y-transform only, 1 for y,x both transformed
.::UdecomBooks::.
% NOTE: x should contain intercept vector in 1st column (if desired)

%----------------------------------------------------
% RETURNS: like = (a scalar) = -log likelihood function
%----------------------------------------------------
[n k] = size(x); ys = boxc_trans(y,lam);
if model == 1 % user wants to transform both y,x
% see if an intercept term exists in the model
iota = x(:,1); ifind = find(iota == 1);
if isempty(ifind), xs = boxc_trans(x,lam); % no intercept
else, if length(ifind) == n, % an intercept
xtrans = boxc_trans(x(:,2:k),lam);
xs = [ones(n,1) xtrans];
else, xs = boxc_trans(x,lam); % no intercept
end;
end;
elseif model == 0, xs = x; % transform only y-vector
end;
bhat = inv(xs’*xs)*xs’*ys; e = ys - xs*bhat; sige = (e’*e)/n;
like = (lam - 1)*sum(log(y)) -(n/2)*log(sige); like = -like;
The function relies on another function boxc trans to carry out the
Box-Cox data transformation. It also contains an argument ‘model’ that
allows for a case (‘model=0’) where the y variable alone is transformed and
another case (‘model=1’) where both y and the X variables are transformed.
The function boxc trans is:
function z = boxc_trans(x,lam)
% PURPOSE: compute box-cox transformation
%----------------------------------------------------
% USAGE: bdata = boxc_trans(data,lam)
% where: lam = scalar transformation parameter
% data = matrix nobs x k
%----------------------------------------------------
% RETURNS: bdata = data matrix box-cox transformed
[n k] = size(x); z = zeros(n,k); iota = ones(n,1);
for i=1:k;
if lam ~= 0, z(:,i) = (x(:,i).^lam - iota)/lam;
else, z(:,i) = log(x(:,i)); end;
end;
Now we can turn attention to use of the MATLAB simplex optimization

function fmin, which can be called with: a string containing the function
name ‘box lik’, an upper and lower limit for the simplex optimization search
over values of λ (‘lamlo’,‘lamup’), a 4x1 vector of optimization options (‘fop-
tions’), and arguments that will be passed along to our likelihood function
‘box lik’. In this example, we wish to pass ‘y,x,model’ as arguments to the
‘box lik’ function. An example of the call is:
.::UdecomBooks::.
[lam options] = fmin(’box_lik’,lamlo,lamup,foptions,y,x,model);
This call is made by a Box-Cox regression function box cox which al-
lows the user to input a sample data vector y and matrix X as well as lower
and upper limits on the parameter λ, a flag to indicate whether the transfor-
mation should be applied to just y or both the y and X data, and optional
optimization options. The function documentation is:
PURPOSE: box-cox regression using a single scalar transformation
parameter for both y and (optionally) x
-----------------------------------------
USAGE: results = box_cox(y,x,lam,lam_lo,lam_up,model,foptions)
x = explanatory variables matrix
(intercept vector in 1st column --- if desired)
lam_lo = scalar, lower limit for simplex search
lam_up = scalar, upper limit for simplex search
model = 0 for y-transform only
= 1 for both y, and x-transform
foptions = (optional) a 4x1 vector of optimization information
foptions(1) = flag to display intermediate results while working
foptions(2) = convergence for simplex (default = 1e-4)
foptions(3) = convergence for function value (default = 1e-4)
foptions(4) = maximum number of iterations (default = 500)
-----------------------------------------
results.meth = ’boxcox’
results.beta = bhat estimates
results.lam = lamda estimate
results.tstat = t-stats for bhat
results.yhat = yhat (box-cox transformed)
results.nobs = nobs
results.y = y data vector (box-cox transformed)
results.like = -log likelihood function value
--------------------------------------------------
NOTE: uses MATLAB simplex function fmin
--------------------------------------------------
SEE ALSO: prt(results), plt(results)
---------------------------------------------------
The function box cox is responsible for checking that the input data are
positive (as required by the Box-Cox regression procedure) and providing
.::UdecomBooks::.
default optimization options for the user if they are not input. The function
also constructs estimates of β̂, σ̂ 2 , t−statistics, R−squared, predicted values,
etc., based on the maximum likelihood estimate of the parameter λ. These
values are returned in a structure variable suitable for use with the prt reg
function for printing results and plt reg function for plotting actual vs.
predicted values and residuals.
Example 10.1 demonstrates use of the function box cox. In the example,
two models are generated. One model involves a y-transformation only such
that the estimated value of λ should equal zero. This is accomplished by
transforming a typical regression generated y-variable using the exponential
function, and carrying out the Box-Cox regression on the transformed y =
exp(y). The second model performs no transformation on y or X, so the
estimated value of λ should equal unity.
% --- Example 10.1 Simplex max likelihood estimation for Box-Cox model
% generate box-cox model data
n = 100; k = 2; kp1 = k+1;
x = abs(randn(n,k)) + ones(n,k)*10;
btrue = ones(k,1); epsil = 0.2*randn(n,1); x = [ones(n,1) x];
y = 10*x(:,1) + x(:,2:k+1)*btrue + epsil;
ycheck = find(y > 0); % ensure y-positive
if length(ycheck) ~= n,error(’all y-values must be positive’); end;
yt = exp(y); % should produce lambda = 0 estimate
model = 0; % transform only y-variable
result = box_cox(yt,x,-2,2,model); prt(result);
model = 1; % transform both y,x variables
xcheck = find(x > 0);
if length(xcheck) ~= n*kp1, error(’all x-values must be positive’); end;
yt = y; xt = x; % should produce lambda=1 estimate
result = box_cox(yt,xt,-2,2,model); prt(result); plt(result);
The printed output from the box cox function indicates λ estimates
near zero for the first regression and unity for the second. The parameter
estimates are also reasonably close to the values used to generate the data.
Box-Cox 1-parameter model Estimates

R-squared = 0.9419
sigma^2 = 0.0909
Lambda = 0.0155
# iterations = 13
-log like = 2977.9633
***************************************************************
.::UdecomBooks::.
variable 1 4.506392 5.029223 0.000002

variable 2 1.765483 29.788761 0.000000
variable 3 1.592008 28.310745 0.000000
Box-Cox 1-parameter model Estimates

R-squared = 0.9415
sigma^2 = 0.0185
Lambda = 0.9103
# iterations = 8
-log like = -168.62035
***************************************************************
variable 1 8.500698 21.621546 0.000000
variable 2 0.983223 29.701359 0.000000
variable 3 0.886072 28.204172 0.000000
One problem that arises with univariate simplex optimization algorithm

is that the numerical hessian matrix is not available to provide estimates of
dispersion for the parameters. This presents a problem with respect to the
parameter λ in the Box-Cox model, but asymptotic likelihood ratio statistics
based on -2 ln L are traditionally used to test hypotheses regarding λ
One way to overcome this limitation is to rely on the theoretical infor-
mation matrix to produce estimates of dispersion. An example of this is
the function sar that implements maximum likelihood estimation for the
spatial autoregressive model (already discussed in Chapter 6 in the context
of Gibbs sampling):
y = ρW y + Xβ + ε (10.6)
This model involves a cross-section of observations made at various points
in space. The parameter ρ is a coefficient on the spatially lagged dependent
variable W y, that reflects the influence of neighboring observations on varia-
tion in individual elements of the dependent variable y. The model is called
a first order spatial autoregression because its represents a spatial analogy to
the first order autoregressive model from time series analysis, yt = ρyt−1 +εt .
The matrix W represents a standardized spatial contiguity matrix such that
the row-sums are unity. The product W y produces an explanatory variable
equal to the mean of observations from contiguous areas.
Anselin (1988) provides the theoretical information matrix for this model
along with a concentrated likelihood function that depends on the single
parameter ρ. Using the theoretical information matrix, we can construct a
dispersion estimate and associated t−statistic for the important spatial lag
.::UdecomBooks::.
parameter ρ. An advantage of using the univariate simplex optimization

algorithm in this application is that the value of ρ must lie between:
where λmin and λmax are the minimum and maximum eigenvalues of the
standardized spatial weight matrix W . We can impose this restriction using
fmin as already demonstrated.
Given a maximum likelihood estimate for ρ, we can compute estimates
for β, σ 2 , and construct the theoretical information matrix which we use to
produce a variance-covariance matrix for the estimates. The following code
fragment from the function sar provides an example of this approach.
% construct information matrix, (page 80-81 Anselin, 1982)

B = eye(n) - p*W;
BI = inv(B); WB = W*BI;
pterm = trace(WB*WB + WB*WB’);
xpx = zeros(nvar+2,nvar+2); % bhat,bhat
xpx(1:nvar,1:nvar) = (1/sige)*(x’*x); % bhat,rho
xpx(1:nvar,nvar+1) = (1/sige)*x’*W*BI*x*bhat;
xpx(nvar+1,1:nvar) = xpx(1:nvar,nvar+1)’; % rho,rho
xpx(nvar+1,nvar+1) = (1/sige)*bhat’*x’*BI’*W’*W*BI*x*bhat + pterm;
xpx(nvar+2,nvar+2) = n/(2*sige*sige); %sige,sige
xpx(nvar+1,nvar+2) = (1/sige)*trace(WB); % rho,sige
xpx(nvar+2,nvar+1) = xpx(nvar+1,nvar+2);
xpxi = inv(xpx);
tmp = diag(xpxi(1:nvar+1,1:nvar+1));
bvec = [results.beta
results.rho];
results.tstat = bvec./(sqrt(tmp));
It should be noted that a similar approach could be taken with the Box-
Cox model to produce an estimate for the dispersion of the transformation
parameter λ. Fomby, Hill and Johnson (1984) provide the theoretical infor-
mation matrix for the Box-Cox model.
Summarizing our discussion of the MATLAB function fmin, we should
not have a great deal of difficulty since single parameter optimization prob-
lems should be easy to solve. The ability to impose restrictions on the
domain of the parameter over which we are optimizing is often useful (or re-
quired) given theoretical knowledge of the likelihood function, or the nature
of the problem being solved. Despite the seeming simplicity of a univariate
likelihood function that depends on a single parameter, a number of prob-
lems in econometrics take this form once other parameters are concentrated
out of the likelihood function. The greatest problem with this approach to
.::UdecomBooks::.
optimization is that no measure of dispersion for the estimated parameter

is available without knowledge of the theoretical information matrix.
10.1.2 Multivariate simplex optimization

As an example of multivariate simplex optimization, we illustrate the case
of maximum likelihood estimation of the AR(1) autoregressive error model:
yt = Xβ + ut (10.7)
ut = ρut−1 + ε (10.8)
We maximize the likelihood function over the parameters β, ρ in the model

(with σ 2 concentrated out). Another example of using fmins is the function
sac that estimates a two-parameter version of the spatial autoregressive
model. We also implement a Box-Cox model with separate parameters for
the y− and X−transformations using this approach in a function box cox2.
For the case of the autoregressive error model, we set up the likelihood
function in a file ar1 like where we have concentrated out the parameter
σ2:
function llike = ar1_like(param,y,x)

% PURPOSE: log-likelihood for a regression model with AR(1) errors
%-----------------------------------------------------
% USAGE: like = ar1_like(b,y,x)
% where: b = parameter vector (m x 1)
% x = explanatory variables matrix (n x m)
%-----------------------------------------------------
% NOTE: this function returns a scalar equal to -log(likelihood)
% b(1,1) contains rho parameter
% sige is concentrated out
%-----------------------------------------------------
% REFERENCES: Green, 1997 page 600
%-----------------------------------------------------
[n k] = size(x);
rho = param(1,1); beta = param(2:2+k-1,1);
ys = y - rho*lag(y); xs = x - rho*lag(x);
ys(1,1) = sqrt(1-rho*rho)*y(1,1);
xs(1,:) = sqrt(1-rho*rho)*x(1,:);
term1 = -(n/2)*log((ys - xs*beta)’*(ys - xs*beta));
term2 = 0.5*log(1-rho*rho);
like = term1+term2;
llike = -like;
.::UdecomBooks::.
A regression function ols ar1 performs error checking on the user input,
establishes optimization options for the user, and supplies starting values
based on the Cochrane-Orcutt estimation function olsc. The code to carry
out Cochrane-Orcutt estimation, set up optimization options, and call fmins
looks as follows:
% use cochrane-orcutt estimates as initial values

reso = olsc(y,x);
parm = zeros(nvar+2,1);
parm(1,1) = reso.rho; % initial rho
parm(2:2+nvar-1,1) = reso.beta; % initial bhat’s
parm(2+nvar,1) = reso.sige; % initial sigma
options(1,1) = 0; % no print of intermediate results
options(2,1) = 0.0001; % simplex convergence criteria
options(3,1) = 0.0001; % convergence criteria for function value
options(4,1) = 1000; % default number of function evaluations
[beta foptions] = fmins(’ar1_like’,parm,options,[],y,x);
niter = foptions(1,10);
llike = foptions(1,8);
rho = beta(1,1);
As in the example for the function sar, we obtain an estimate of the

precision of the important parameter ρ using the theoretical information
matrix (in code not shown here).
Some points to note about using the simplex optimization algorithm
fmins. Good starting values are often necessary to produce a reasonable
solution and the time necessary depends to a large extent on the quality of
these initial values. The example provided is somewhat unrealistic in that
the superior quality of Cochrane-Orcutt estimates as starting values creates
a situation where the function fmins has to do a minimum of work.
In more realistic applications such as the Box-Cox model with one param-
eter,for the y−transformation and another for the X−transformation, the
function fmins performs more slowly and less accurately. As an illustration
of this, you can experiment with the function box cox2 that implements
the two-parameter Box-Cox model using fmins.
10.2 EM algorithms for optimization

Another approach to solving maximum likelihood estimation problems is
to rely on an EM, or expectation-maximization algorithm of Dempster et
al. (1977). This algorithm proceeds by successively maximizing the current
conditional expectation of the log likelihood function, where missing data (or
.::UdecomBooks::.
parameters) are replaced by their expectation prior to maximization. Given

estimates from the maximization (the M-step), a new set of expected values
are computed and substituted in the likelihood function (the E-step). This
process continues until convergence to a set of values that do not increase
the likelihood. An advantage of the EM algorithm is that each step is
guaranteed to increase the likelihood function value. A disadvantage is that
the algorithm tends to move slowly during the final steps.
As an example of this type of problem, consider a simple switching re-
gression model:
y1t = X1t β1 + ε1t (10.9)

y2t = X2t β2 + ε2t
y3t = X3t β3 + ε3t
yt = y1t if y3t ≤ 0
yt = y2t if y3t > 0
ε1t ∼ N (0, σ12 )
ε2t ∼ N (0, σ22 )
ε3t ∼ N (0, 1) (10.10)
This model suggests that the observed value of yt is generated by two

different latent unobservable variables y1t , y2t . The variable y3t is also a
latent variable that serves to classify yt into either regime.
An EM algorithm for this model involves iteratively estimating a sys-
tem that relies on simple least-squares (OLS) and weighted least-squares
(WLS) regressions. Given initial values for the parameters in the model:
β1 , β2 , β3 , σ1 , σ2 , we can begin by calculating weights for the WLS regres-
sion using:
w1 (yt ) = λt [f1 (yt )/h(yt )] (10.11)

w2 (yt ) = (1 − λt )[f2 (yt )/h(yt )]
Where w1 denotes a scalar weight associated with observation t and fi , i =

1, 2 is the normal probability density function for each regime evaluated at
observation t:
fi (yt ) = (2π)−1/2 σi−1 exp{−(yt − x0t βi )2 /2σi2 } (10.12)
.::UdecomBooks::.
λt represents the cumulative normal density function of ε3t , which we eval-

uate for each observation t using: Φ(−x03t β3 ) to determine the probability
that yt = y1t . Finally, h(yt ) represents the probability density function for
the observed variable yt calculated from:
h(yt ) = λt f1 (yt ) + (1 − λt )f2 (yt ) (10.13)

Given the above procedure for calculating weights, the EM-algorithm
proceeds as follows:
1. Using the initial values βi , σi , determine the weights, w1i as discussed

above for regimes i = 1, 2 and form: Wi = diag(w1i , w2i , . . . , wni ).
2. Perform WLS regressions:
βi = [Xi0 Wi Xi ]−1 [Xi0 Wi y] (10.14)
3. Compute a value for β3 using and OLS regression:
β3 = (X30 X3 )−1 X30 ε3 (10.15)

ε3t = x03t β3 − w1 (yt )(f3 (0)/λt ) + w2 (yt )f3 (0)/(1 − λt )
4. Compute values for σi2 , i = 1, 2 using the WLS regressions:

X
σi2 = [1/ wi (yt )](y − Xi βi )0 Wi (y − Xi βi ) (10.16)
5. Go to step 1, and replace the initial values with the new values of
βi , σi .
The routine first generates a set of weights based on arbitrary starting

values for all parameters in the model. Given these weights, we can deter-
mine the unobserved y1 , y2 , y3 values and produce estimates β1 , β2 , β3 , σ1 , σ2
that maximize the likelihood function.
Note that by replacing the unobservable y3t with the conditional expec-
tation based on yt values (represented by ε3t ), the problem of classifying
observed yt values into the two regimes is simple. Further, the task of max-
imizing the likelihood function with respect to βi , i = 1, 2, 3, σi , i = 1, 2
involves relatively simple OLS and WLS regressions. Given the estimated
βi , i = 1, 2, 3, σi , i = 1, 2, a new set of weights can be produced and the
process repeated until convergence in the estimated parameters occurred.
.::UdecomBooks::.
This procedure is implemented in the MATLAB function switch em

which has the following documentation:
PURPOSE: Switching Regime regression (EM-estimation)
y1 = x1*b1 + e1
y2 = x2*b2 + e2
y3 = x3*b3 + e3; e3 =N(0,1)
y = y1 if y3 <= 0, y = y2 if y3 > 0
---------------------------------------------------
USAGE: results = switch_em(y,x1,x2,x3,crit)
x1 = independent variables matrix (nobs x k1)
b1 = (optional) initial values for b1
crit = (optional) convergence criteria (default = 0.001)
---------------------------------------------------
results.meth = ’switch_em’
results.beta1 = bhat1
results.t1 = t-stats for bhat1
results.yhat1 = predicted values regime 1
results.yhat2 = predicted values regime 2
results.r1 = residuals regime 1
results.r2 = residuals regime 2
results.sig1 = e1’*e1/(n1-k1)
results.sig2 = e2’*e2/(n2-k2)
results.rsqr1 = rsquared, regime 1
results.rsqr2 = rsquared, regime 2
results.nobs = nobs
results.k1 = k1, variables in x1
results.nvar = k1+k2+k3
results.prob1 = probability of regime 1
results.prob2 = probability of regime 2
results.iter = # of iterations in EM-algorithm
results.crit = convergence criterion
results.like = likelihood function value
--------------------------------------------------
Some points about the code are:
.::UdecomBooks::.
1. We need starting values for the initial parameters β1 , β2 , σ1 , σ2 . These

are based on separate regressions involving the positive and negative
y values on the associated X values. The initial values for β3 are
based on a regression of y on X3 . This approach works well when y
has a mean close to zero, but another method based on Goldfeld and
Quandt’s (1973) method of moments estimation would be preferable.
2. During EM-iteration we need to check that the weights associated

with one of the two regimes do not become zero for all observations,
effectively producing a single regime result. Once the weights for one of
the two regimes become zero at all observations, there is no possibility
of returning to positive weight values for the zero-weight regime, so
we terminate with an error message.
3. We set a default maximum # of iterations and convergence criterion,

but the user can change these values as input arguments. In the event
that the iterations exceed the maximum, we display a warning (using
the MATLAB warn function) to the user and break out of the EM-
loop. The use of a warning rather than an error (based on a call to
the MATLAB error function) allows the user to examine a printout of
the estimates, and perhaps set better starting values for another run.
4. We return a log likelihood function value by evaluating the likelihood

at the converged EM-estimates.
The code for switch em is:

% code for EM-estimation of the switching regime regression
if nargin == 4; bflag = 0; crit = 0.001; maxit = 1000;
elseif nargin == 7; bflag = 1; crit = 0.001; maxit = 1000;
elseif nargin == 8; bflag = 1; maxit = 1000;
elseif nargin == 9; bflag = 1;
else, error(’Wrong # of arguments to switch_em’);
end;
[n1 k1] = size(x1); [n2 k2] = size(x2); [n3 k3] = size(x3);
if n1 ~= n2, error(’switch_em: x1, x2 have different nobs’);
elseif n2 ~= n3, error(’switch_em: x2, x3 have different nobs’);
end;
nobs = n1; converge = 1.0; iter = 0;
if bflag == 0 % user supplied NO initial values
% get starting values using positive and negative y’s
ypos = find(y > 0); yneg = find(y <= 0);
res = ols(y(ypos,1),x1(ypos,:)); sig1 = res.sige; b1 = res.beta;
res = ols(y(yneg,1),x2(yneg,:)); sig2 = res.sige; b2 = res.beta;
% b3 starting values using ad-hockery
.::UdecomBooks::.
res = ols(y,x3); b3 = res.beta;

else % user supplied b1,b2,b3 we need to find sig1,sig2
sig1 = (y - x1*b1)’*(y - x1*b1)/nobs;
sig2 = (y - x2*b2)’*(y - x2*b2)/nobs;
end;
while converge > crit % start EM-loop
f1=norm_pdf(((y-x1*b1)/sig1)/sig1); f2=norm_pdf(((y-x2*b2)/sig2)/sig2);
lamda=norm_cdf(-x3*b3); h=(lamda.*f1)+((1-lamda).*f2);
w1=lamda.*f1./h; w2=ones(nobs,1)-w1;
ep=x3*b3+(f2-f1).*norm_pdf(-x3*b3./h); t1=sum(w1); t2=sum(w2);
if t1 <= 1; error(’switch_em: regime 1 weights all near zero’); end;
if t2 <= 1; error(’switch_em: regime 2 weights all near zero’); end;
w1 = sqrt(w1); w2 = sqrt(w2); xx1=matmul(x1,w1); xx2=matmul(x2,w2);
y1=y.*w1; y2=y.*w2; b01 = b1; b02 = b2; b03 = b3;
b1=inv(xx1’*xx1)*xx1’*y1; b2=inv(xx2’*xx2)*xx2’*y2; b3=inv(x3’*x3)*x3’*ep;
sig01 = sig1; sig02 = sig2; sig1=y1-xx1*b1; sig2=y2-xx2*b2;
sig1=sqrt(sig1’*sig1/t1); sig2=sqrt(sig2’*sig2/t2);
% check for convergence
c = max(abs(b1-b01)); c = max(c,max(abs(b2-b02)));
c = max(c,max(abs(b3-b03))); c = max(c,max(sig1-sig01));
converge = max(c,max(sig2-sig02));
iter = iter + 1;
if iter > maxit; warn(’switch_em: max # of iterations exceeded’);
break; end;
end; % end of EM-loop
% compute t-statistics
tmp = sig1*(inv(xx1’*xx1)); results.t1 = b1./sqrt(diag(tmp));
tmp = sig2*(inv(xx2’*xx2)); results.t2 = b2./sqrt(diag(tmp));
tmp = inv(x3’*x3);% sig3=1; results.t3 = b3./sqrt(diag(tmp));
results.meth = ’switch_em’; % return results
results.iter = iter; results.crit = converge;
results.beta1 = b1; results.beta2 = b2; results.beta3 = b3;
results.nobs = nobs; results.k1 = k1; results.k2 = k2; results.k3 = k3;
results.nvar = k1+k2+k3; results.y = y;
results.prob1 = w1.^2; results.prob2 = w2.^2;
results.yhat1 = xx1*b1; results.yhat2 = xx2*b2;
results.r1 = y1-xx1*b1; results.r2 = y2-xx2*b2;
results.sig1 = sig1; results.sig2 = sig2;
% compute R-squared based on two regimes
sigu1 = 0; sigu2 = 0; y1sum = 0; y2sum = 0; y1s = 0; y2s = 0;
prob1 = results.prob1; nobs1 = 0; nobs2 = 0;
for i=1:nobs
if prob1(i,1) > 0.5, nobs1 = nobs1+1;
sigu1 = sigu1 + results.r1(i,1)*results.r1(i,1);
y1sum = y1sum + y(i,1); y1s = y1s + y1sum*y1sum;
else, nobs2 = nobs2+1;
sigu2 = sigu2 + results.r2(i,1)*results.r2(i,1);
y2sum = y2sum + y(i,1); y2s = y2s + y2sum*y2sum;
end;
.::UdecomBooks::.
end;
results.rsqr1 = 1 - sigu1/y1s; results.rsqr2 = 1 - sigu2/y2s;
like = log((1-lamda).*f1 + lamda.*f2); results.like = sum(like);
An example program that generates regression data based on two regimes

and estimates the parameters using switch em is:
% --- Example 10.2 EM estimation of switching regime model
% generate data from switching regression model
nobs = 100; n1 = 3; n2 = 3; n3 = 3; sig1 = 1; sig2 = 1;
b1 = ones(n1,1); b2 = ones(n2,1)*5; b3 = ones(n3,1);
x1 = randn(nobs,n1); x2 = randn(nobs,n2); x3 = randn(nobs,n3);
for i=1:nobs;
if x3(i,:)*b3 <= 0, y(i,1) = x1(i,:)*b1 + randn(1,1)*2;
else, y(i,1) = x2(i,:)*b2 + randn(1,1)*2;
end;
end;
result = switch_em(y,x1,x2,x3,b1,b2,b3);
vnames1 = strvcat(’y1’,’x1_1’,’x1_2’,’x1_3’);
vnames2 = strvcat(’y2’,’x2_1’,’x2_2’,’x2_3’);
vnames3 = strvcat(’x3_1’,’x3_2’,’x3_3’);
vnames = strvcat(vnames1,vnames2,vnames3);
prt(result,vnames);
plt(result,vnames);
Example 10.2 generates a set of y−values based on X3 β3 ≤ 0 for regime

1 and otherwise regime 2. The regime 1 relation contains true parameters
(β1 ) equal to unity, whereas regime 2 sets the parameters (β2 ) equal to five.
The values used for β3 were unity, which should produce a roughly equal
number of observations in the two regimes, given that the values of X3 are
normally distributed around a mean of zero.
The printout of estimation results from example 10.2 and the plot pro-
duced by the plt function are shown below.
EM Estimates - Switching Regression model
Regime 1 equation
Dependent Variable = y1
R-squared = 0.9925
sigma^2 = 4.2517
***************************************
x1_1 0.92050544 3.40618422 0.00095997
x1_2 0.61875719 2.25506692 0.02637700
x1_3 1.15667358 4.21939446 0.00005514
Regime 2 equation
Dependent Variable = y2
.::UdecomBooks::.
R-squared = 0.9934
sigma^2 = 2.4245
***************************************
x2_1 4.78698864 19.75899413 0.00000000
x2_2 4.67524699 21.70327247 0.00000000
x2_3 5.26189140 24.83553712 0.00000000
Switching equation
Conv criterion = 0.00097950
# iterations = 53
# obs regime 1 = 51
# obs regime 2 = 49
log Likelihood = -136.7960
***************************************
x3_1 1.04653177 10.97136877 0.00000000
x3_2 0.96811598 8.97270638 0.00000000
x3_3 1.04817888 10.45553196 0.00000000
The estimated parameters are close to the true values of one and five
for the two regimes as are the parameters of unity used in the switching
equation. Figure 10.1 shows the actual y-values versus predictions. The plot
shows predictions classified into regimes 1 and 2 based on the probabilities
for each regime > 0.5. The results structure returned by switch em returns
predicted values for all n observations, but the graph only shows predicted
values based on classification by probabilities for the two regimes.
Summarizing the EM approach to optimization, we require a likelihood
function that can be maximized once some missing data values or parameters
are replaced by their expectations. We then loop through an expectation-
maximization sequence where the expected values of the parameters or miss-
ing sample data are substituted to produce a full-data likelihood function.
The remaining parameters or observations are computed based on maxi-
mizing the likelihood function and these estimates are used to produce new
expectations for the missing sample data or parameters. This process con-
tinues until convergence in the parameter estimates. A number of estimation
problems have been structured in a form amenable to EM optimization. For
example, Shumway and Stoffer (1982) provide an EM algorithm for estima-
tion time-series state space models, van Norden and Schaller (1993) provide
an EM algorithm for estimation of the Markov transition regime-switching
model of Hamilton (1989). McMillen (1992) sets forth an EM approach to
estimating spatial autoregressive logit/probit models.
An interesting point is that for most cases where an EM approach to es-
.::UdecomBooks::.
Regime 1 and Regime 2 Actual vs. Predicted

30
Actual
Regime1
20
Regime2
10
-10
-20
0 10 20 30 40 50 60 70 80 90 100
Estimated probabilities
1
0.8
0.6
regime1
0.4 regime2
0.2
0
0 10 20 30 40 50 60 70 80 90 100
Figure 10.1: Plots from switching regime regression
timation exists, a corresponding Gibbs sampling estimation procedure can

be devised. As an example, the function sarp g implements a Gibbs sam-
pling approach to estimating the spatial autoregressive logit/probit model
that corresponds to an EM approach set forth in McMillen (1992).
For the switching regression example illustrated in this section, a Gibbs
sampling approach would carry out multivariate normal draws for β1 , β2 , β3
using means and variances based on the least-squares and weighted least-
squares estimates presented. Chi-squared draws for σ1 , σ2 would be made
based on the sum of squared residuals from the weighted least-squares es-
timates and a new value of λt would be calculated based on these param-
eter values. This process of updating the estimates β1 , β2 , β3 , σ1 , σ2 and
the weights based on λt would be continued to produce a large sample of
draws for the parameters in the model. Albert and Chib (1993) and Kim
(1994) provide Gibbs sampling approaches to Hamilton’s dynamic Markov
switching model.
An EM approach to Hamilton’s Markov switching model can be found in
.::UdecomBooks::.
the functions hmarkov em and emhergott contributed by Matt Hergott.

One criticism levelled against EM algorithms is that they are slow to
converge. This seems less and less of a problem with the speed enhancements
occurring in computing technology. Nonetheless, some problems can be more
efficiently handled if the EM algorithm is used at the outset to steer the
optimization problem towards high quality initial parameter values. After
some initial EM loops, one can switch to a maximum likelihood approach
based on the type of gradient routines discussed in the next section. This
avoids the problem of slow convergence near the end of the EM algorithm,
while taking advantage of a desirable aspect of the EM algorithm — it is
guaranteed to produce a sequence of estimates that monotonically improve
the value of the likelihood function.
10.3 Multivariate gradient optimization

By far the most difficult optimization problems are those that cannot be
reduced to lower-dimensional optimization problems amenable to simplex
algorithms, or handled by the EM algorithm, or Gibbs sampling. As we saw
in Chapter 7 for the case of logit/probit models, this may not be a problem
if analytical gradients and second derivatives are available.
For problems where analytical gradients and second derivatives are not
available or convenient, there are functions frpr min, dfp min, pow min
that represent code from the Numerical Recipes book converted to MAT-
LAB, a function maxlik that attempts to handle general maximum likeli-
hood optimization problems, as well as the function minz by Mike Cliff and
optsolv. The motivation for providing a host of different general purpose
optimization functions is that in some problems one function works better
than others.
Some differences between these functions are:
1. The frpr min, dfp min, pow min functions begin by setting the
hessian matrix to the identity and building up the hessian estimate
with updating. This avoids inversion of the hessian matrix which can
become non-positive definite in some problems. The maxlik function
inverts the hessian matrix but relies on the function invpd to force
non-positive definite matrices to be positive definite prior to inver-
sion. (The small eigenvalues of the non-positive definite matrix are
arbitrarily augmented to accomplish this.)
2. The frpr min, dfp min, pow min functions use a function linmin
.::UdecomBooks::.
that can encounter problems, simply print an error message and quit.
3. The optsolv function is based on a public domain MATLAB function

solvopt that implements a modification of Shor’s r-algorithm. This
function is used here to provide an example of how we use a ‘wrapper
function’ optsolv to incorporate a public domain algorithm in the
Econometrics Toolbox so it conforms in terms of calling arguments
and documentation. The original solvopt algorithm was implemented
by faculty at the University of Graz in Austria. Knowledge about the
performance and quality of the algorithm in econometric maximum
likelihood applications is sketchy.
4. The minz function incorporated in the Econometrics Toolbox differs

slightly from that found on Mike Cliff’s Internet site in order to make
the calling arguments conformable with the other optimization algo-
rithms in the toolbox.
As an illustration of using all of these optimization functions consider

solving the tobit estimation problem. First, we create a MATLAB function
to evaluate the log likelihood function. This function requires two optional
input arguments ‘y,x’ that we use to pass the sample data to the likeli-
hood function. Our optimization functions allow for an arbitrary number
of arguments that can be passed, making their use perfectly general. The
log-likelihood function for the tobit model is shown below:
function like = to_liked(b,y,x);
% PURPOSE: evaluate tobit log-likelihood
% to demonstrate optimization routines
%-----------------------------------------------------
% USAGE: like = to_liked(b,y,x)
% x = explanatory variables matrix (n x m)
%-----------------------------------------------------
% NOTE: this function returns a scalar equal to the negative
% of the log-likelihood function
% k ~= m because we may have additional parameters
% in addition to the m bhat’s (e.g. sigma)
%-----------------------------------------------------
% error check
if nargin ~= 3,error(’wrong # of arguments to to_like’); end;
h = .000001; % avoid sigma = 0
[m junk] = size(b);
beta = b(1:m-1); % pull out bhat
sigma = max([b(m) h]); % pull out sigma
.::UdecomBooks::.
xb = x*beta;
llf1 = -(y-xb).^2./(2*sigma) - .5*log(2*pi*sigma);
xbs = xb./sqrt(sigma); cdf = .5*(1+erf(xbs./sqrt(2)));
llf2 = log(h+(1-cdf));
llf = (y > 0).*llf1 + (y <= 0).*llf2;
like = -sum(llf);% scalar result
Now we can create a function to read the data and call these optimization
functions to solve the problem. The documentation for all of the optimiza-
tion functions take a similar form, so they can be used interchangeably when
attempting to solve maximum likelihood problems. The documentation for
dfp min is shown below. There are four required input arguments, the
function name given as a string, a vector of starting parameter values, a
structure variable containing optimization options, and a variable argument
list of optional arguments that will be passed to the function. In this case,
the variable argument list contains the data vector for y and data matrix X.
Other optimization options can be input as fields in the structure variable,
and these differ depending on the optimization function being called. For
example, the dfp min function allows the user to provide optional argu-
ments in the structure fields for the maximum number of iterations, a flag
for printing intermediate results and a convergence tolerance.
% PURPOSE: DFP minimization routine to minimize func

% (Converted from Numerical Recipes book dfpmin routine)
%---------------------------------------------------
% USAGE: [pout,fout, hessn, niter] = dfp_min(func,b,info,varargin)
% Where: func = likelihood function to minimize
% b = parameter vector fed to func
% info = a structure with fields:
% info.maxit = maximum # of iterations (default = 500)
% info.tol = tolerance for convergence (default = 1e-7)
% info.pflag = 1 for printing iterations, 0 for no printing
% varargin = list of arguments passed to function
%---------------------------------------------------
% RETURNS: pout = (kx1) minimizing vector
% fout = value of func at solution values
% hessn = hessian evaluated at the solution values
% niter = # number of iterations
%---------------------------------------------------
% NOTE: func must take the form func(b,varargin)
% varargin = arguments passed to the function
Example 10.3 shows a program that calls five of the optimization func-
tions to solve the tobit estimation problem. A comparison of the time
.::UdecomBooks::.
taken by each optimization function, the estimates produced, and the log-
likelihood function values and the hessian matrices evaluated at the solution
are presented.
% --- Example 10.3 Maximum likelihood estimation of the Tobit model
n=200; k=5; randn(’seed’,20201); x = randn(n,k); beta = ones(k,1);
y = x*beta + randn(n,1); % generate uncensored data
for i=1:n, if y(i,1) < 0, y(i,1) = 0.0; end; end;
% use ols for starting values
res = ols(y,x); b = res.beta; sige = res.sige;
parm = [b
sige]; % starting values
info.maxit = 100;
% solve using frpr_min routine
tic; [parm1,like1,hess1,niter1] = frpr_min(’to_liked’,parm,info,y,x);
disp(’time taken by frpr routine’); toc;
% solve using dfp_min routine
tic; [parm2,like2,hess2,niter2] = dfp_min(’to_liked’,parm,info,y,x);
disp(’time taken by dfp routine’); toc;
% solve using pow_min routine
tic; [parm3,like3,hess3,niter3] = pow_min(’to_liked’,parm,info,y,x);
disp(’time taken by powell routine’); toc;
% solve using maxlik routine
tic; [parm4,like4,hess4,grad,niter4,fail] = maxlik(’to_liked’,parm,info,y,x);
disp(’time taken by maxlik routine’); toc;
% solve using minz routine
infoz.call = ’other’;
infoz.maxit = 500;
infoz.prt = 0;
tic;
[parm5,infoz,stat] = minz(’to_liked’,parm,infoz,y,x);
disp(’time taken by minz routine’); toc;
niter5 = stat.iter;
like5 = stat.f;
hess5 = stat.Hi;
% formatting information for mprint routine

in.cnames = strvcat(’fprf’,’dfp’,’powell’,’maxlik’,’minz’);
in.fmt = ’%8.4f’;
fprintf(1,’comparison of bhat estimates \n’);
mprint([parm1(1:k,1) parm2(1:k,1) parm3(1:k,1) parm4(1:k,1) parm5(1:k,1)],in);
fprintf(1,’comparison of sige estimates \n’);
mprint([parm1(k+1,1) parm2(k+1,1) parm3(k+1,1) parm4(k+1,1) parm5(k+1,1)],in);
fprintf(1,’comparison of likelihood functions \n’);
mprint([like1 like2 like3 like4 like5],in);
in.fmt = ’%4d’; fprintf(1,’comparison of # of iterations \n’);
mprint([niter1 niter2 niter3 niter4 niter5],in);
fprintf(1,’comparison of hessians \n’); in2.fmt = ’%8.2f’;
.::UdecomBooks::.
fprintf(1,’fprf hessian’); mprint(hess1,in2);

fprintf(1,’dfp hessian’); mprint(hess2,in2);
fprintf(1,’powell hessian’); mprint(hess3,in2);
fprintf(1,’maxlik hessian’); mprint(hess4,in2);
fprintf(1,’minz hessian’); mprint(hess5,in2);
The output from example 10.3 is shown below, where we see that all
four functions found nearly identical results. A primary difference was in
the time needed by the alternative optimization methods.
time taken by frpr routine = 3.9557

time taken by dfp routine = 3.6244
time taken by powell routine = 12.8823
time taken by maxlik routine = 0.9818
time taken by minz routine = 1.3015
comparison of bhat estimates

fprf dfp powell maxlik minz
0.8822 0.8822 0.8820 0.8822 0.8828
0.8509 0.8509 0.8508 0.8509 0.8515
1.0226 1.0226 1.0224 1.0226 1.0247
0.9057 0.9057 0.9055 0.9057 0.9067
0.9500 0.9500 0.9497 0.9500 0.9517
comparison of sige estimates

0.8808 0.8809 0.8769 0.8808 0.8844
comparison of likelihood functions

153.0692 153.0692 153.0697 153.0692 153.0700
comparison of # of iterations
8 7 5 10 11
comparison of hessians
fprf hessian
137.15 -6.66 -30.32 -2.11 2.97 -2.50
-6.66 136.03 -17.15 -4.10 -5.99 -2.13
-30.32 -17.15 126.68 -7.98 -10.56 -5.10
-2.11 -4.10 -7.98 126.65 -4.78 -5.57
2.97 -5.99 -10.56 -4.78 145.19 -8.04
-2.50 -2.13 -5.10 -5.57 -8.04 71.06
dfp hessian
137.12 -6.66 -30.31 -2.11 2.97 -2.52
-6.66 136.00 -17.15 -4.10 -5.98 -2.11
-30.31 -17.15 126.65 -7.97 -10.57 -5.14
.::UdecomBooks::.
-2.11 -4.10 -7.97 126.63 -4.77 -5.54

2.97 -5.98 -10.57 -4.77 145.16 -8.06
-2.52 -2.11 -5.14 -5.54 -8.06 71.01
powell hessian
137.69 -6.68 -30.43 -2.12 3.00 -2.38
-6.68 136.57 -17.24 -4.12 -6.00 -2.10
-30.43 -17.24 127.14 -8.00 -10.61 -5.27
-2.12 -4.12 -8.00 127.14 -4.79 -5.58
3.00 -6.00 -10.61 -4.79 145.77 -8.03
-2.38 -2.10 -5.27 -5.58 -8.03 72.27
maxlik hessian
137.14 -6.66 -30.31 -2.11 2.97 -2.50
-6.66 136.02 -17.15 -4.10 -5.99 -2.12
-30.31 -17.15 126.66 -7.98 -10.57 -5.12
-2.11 -4.10 -7.98 126.64 -4.77 -5.56
2.97 -5.99 -10.57 -4.77 145.18 -8.04
-2.50 -2.12 -5.12 -5.56 -8.04 71.04
minz hessian
136.60 -6.63 -30.19 -2.09 2.95 -2.52
-6.63 135.49 -17.08 -4.08 -5.96 -2.14
-30.19 -17.08 126.15 -7.94 -10.53 -5.31
-2.09 -4.08 -7.94 126.16 -4.76 -5.63
2.95 -5.96 -10.53 -4.76 144.61 -8.21
-2.52 -2.14 -5.31 -5.63 -8.21 70.00
A point to note about the approach taken to designing the optimization

functions is that we can easily add new algorithms from the Internet or
other sources to the optimization function library. As an illustration of this,
we demonstrate how to add a new MATLAB optimization function named
“solvopt” which was placed in the public domain on the Internet by Alexei
Kuntsevich and Franz Kappel.
We construct a ‘wrapper function’ that we name optsolv to call this
function with the appropriate input arguments.
The first step is to change the function input arguments and documen-
tation to conform to the others in our optimization function library. The
function solvopt was written to handle constrained as well as unconstrained
optimization problems, but we are only interested in using the function for
unconstrained problems. The new format for the optsolv function is shown
below and it takes the same form as our other functions with the excep-
tion of the optimization options. This function has the responsibility for
interpreting the input arguments and passing them along to the function
solvopt that does the real work of optimization.
function [x,f,hessn,gradn,niter]=optsolv(fun,x,info,varargin)
PURPOSE: a modified version of Shor’s r-algorithm to minimize func
.::UdecomBooks::.
---------------------------------------------------
USAGE: [x,f,hess,gradn,niter] = optsolv(func,b,info)
Where: func = likelihood function to minimize (<=36 characters long)
b = parameter vector fed to func
info = a structure with fields:
info.maxit = maximum # of iterations (default = 1000)
info.btol = b tolerance for convergence (default = 1e-7)
info.ftol = func tolerance for convergence (default = 1e-7)
info.pflag = 1 for printing iterations, 0 for no printing
varargain = arguments passed to the function
---------------------------------------------------
RETURNS: x = (kx1) minimizing vector
f = value of func at solution values
hessn = hessian evaluated at the solution values
gradn = gradient evaluated at the solution
niter = # number of iterations
---------------------------------------------------
NOTE: - func must take the form func(b,P0,P1,...)
where: b = parameter vector (k x 1)
P0,P1,... = arguments passed to the function
---------------------------------------------------
These modifications required that we parse our input option arguments

into a vector variable named ‘options’ that the solvopt function uses. This
was accomplished with the following code fragment:
if ~isstruct(info)
error(’optsolv: options should be in a structure variable’);
end;
% default options
options=[-1,1.e-4,1.e-6,1000,0,1.e-8,2.5,1e-11];
app = 1; % no user supplied gradients
constr = 0; % unconstrained problem
% parse options
fields = fieldnames(info);
nf = length(fields); xcheck = 0; ycheck = 0;
for i=1:nf
if strcmp(fields{i},’maxit’), options(4) = info.maxit;
elseif strcmp(fields{i},’btol’), options(2) = info.btol;
elseif strcmp(fields{i},’ftol’), options(3) = info.ftol;
elseif strcmp(fields{i},’pflag’), options(5) = info.pflag;
end;
end;
funfcn = fcnchk(funfcn,length(varargin));
We simply feed the user-input options directly into the options vector used
by the function.
.::UdecomBooks::.
A second step involves use of a MATLAB function fcnchk shown as the

last line in the code fragment above. This function sets up a string argument
for the function name so we can pass it to the gradient and hessian functions.
This MATLAB function is used by fmin and fmins along with the approach
we take to passing the function name on to other sub-functions used by the
optimization algorithm. A good way to discover programming tricks is to
examine the functions crafted by the MathWorks folks, which is what was
done here.
Given the function name string, ‘funfcn’, we can use this in a function
call to the function apprgrdn that solvopt uses to compute numerical
gradients. We have to change this function to rely on a function name
argument based on the string in ‘funfcn’ and our variable argument list that
will be passed to this function. This function will need to call our likelihood
function and properly pass the variable argument list. The original calls to
apprgrdn were in the form:
g=apprgrdn(b,fp,fun,deltax,1);
which do not supply the y and X data arguments required by our function.
The modified function apprgrdn was written to accept a call with a variable
arguments list.
g=apprgrdn(parm,f,funfcn,deltaparm,varargin{:});
The third step involves changing likelihood function calls from inside
the apprgrdn function The calls to the likelihood function in the original
apprgrdn function were in the form:
fi=feval(fun,b);
These were modified with the code shown below. First we modify the func-
tion to rely on the MATLAB fcnchk function to provide a string argument
for the function name. Next, we call the likelihood function using the string
‘funfcn’ containing its name, a vector of parameters ‘x0’, and the variable
argument list ‘varargin:’.
funfcn = fcnchk(funfcn,length(varargin));
fi=feval(funfcn,x0,varargin{:});
A fourth step relates to the fact that many optimization routines crafted
for non-econometric purposes do not provide a hessian argument. This can
be remedied with the function hessian from the optimization function li-
brary. It evaluates the hessian at a given parameter vector using a likelihood
function argument. The documentation for hessian is:
.::UdecomBooks::.
PURPOSE: computes hessian matrix of 2nd partial derivative

for the likelihood function f evaluated at b
---------------------------------------------------
USAGE: hessn = hessian(f,b,varargin)
where: f = a string containing the likelihood function
b = a parameter vector (k x 1)
varargin = arguments that will be passed to the function
---------------------------------------------------
NOTE: f must take the form f(b,P0,P1,...)
where: b = parameter vector (k x 1)
P0,P1,... = optional variables passed to likelihood
---------------------------------------------------
RETURNS: hess = hessian matrix (k x k)
---------------------------------------------------
Notice that it was designed to work with likelihood functions matching the
format of those in the optimization function library. We can produce a
hessian matrix evaluated at the solution vector with a call to hessian, for
optimization algorithms like solvopt that don’t provide this information.
An example of using this new optimization routine is shown below, where
we compare the solution times and results with the maxlik function demon-
strated previously.
% ----- Example 10.4 Using the optsolv() function

n=200; k=5; randn(’seed’,20211); x = randn(n,k); beta = ones(k,1);
y = x*beta + randn(n,1); % generate uncensored data
for i=1:n, if y(i,1) < 0,y(i,1) = 0.0; end; end;
res = ols(y,x); b = res.beta; sige = res.sige;
parm = [b
sige]; % starting values
info.maxit = 100;
% solve using optsolv routine
tic;
[parm1,like1,hess1,grad1,niter1] = optsolv(’to_liked’,parm,info,y,x);
disp(’time taken by optsolv routine’); toc;
% solve using maxlik routine
tic;
[parm2,like2,hess2,grad2,niter2,fail] = maxlik(’to_liked’,parm,info,y,x);
disp(’time taken by maxlik routine’); toc;
in.fmt = ’%9.3f’; in.cnames = strvcat(’optsolv’,’maxlik’);
fprintf(1,’comparison of estimates \n’);
mprint([parm1 parm2],in);
fprintf(1,’comparison of # of iterations \n’);
in.fmt = ’%8d’;
mprint([niter1 niter2],in);
fprintf(1,’comparison of hessians \n’); in2.fmt = ’%8.3f’;
mprint(hess1,in2); mprint(hess2,in2);
.::UdecomBooks::.
The results from example 10.4 are shown below, indicating that the
solvopt function might be a valuable addition to the optimization function
library.
time taken by optsolv routine = 7.4620
time taken by maxlik routine = 1.1209
comparison of estimates
optsolv maxlik
1.082 1.082
1.145 1.145
1.168 1.168
1.021 1.021
1.155 1.155
0.867 0.867
comparison of # of iterations
optsolv maxlik
23 10
comparison of hessians
optsolv
145.137 -5.641 -6.786 7.562 -14.985 -7.246
-5.641 125.331 -11.404 -4.000 -16.000 -1.483
-6.786 -11.404 145.101 31.777 -10.786 -0.818
7.562 -4.000 31.777 129.980 -28.289 -1.575
-14.985 -16.000 -10.786 -28.289 152.888 -1.994
-7.246 -1.483 -0.818 -1.575 -1.994 76.805
maxlik
145.145 -5.641 -6.787 7.562 -14.985 -7.248
-5.641 125.339 -11.405 -4.000 -16.001 -1.486
-6.787 -11.405 145.110 31.779 -10.787 -0.815
7.562 -4.000 31.779 129.989 -28.291 -1.575
-14.985 -16.001 -10.787 -28.291 152.898 -1.985
-7.248 -1.486 -0.815 -1.575 -1.985 76.824
10.4 Chapter summary

We provided some simple examples of maximum likelihood estimation for
models involving univariate and multivariate parameter vectors. Simplex
and gradient methods as well as the EM algorithm were demonstrated. Be-
yond providing examples, a design was set forth to allow new optimization
algorithms to be incorporated into the optimization function library in a
consistent, easy-to-use fashion. A demonstration of re-coding an existing
optimization algorithm from the Internet to work with the function library
was provided in the last section of the chapter.
.::UdecomBooks::.
Chapter 10 Appendix
Optimization functions discussed in this chapter discussed in this chapter

are in the subdirectory optimize.
Optimization functions library
--------------- optimization functions -----------------
dfp_min - Davidson-Fletcher-Powell
frpr_min - Fletcher-Reeves-Polak-Ribiere
maxlik - general all-purpose optimization routine
minz - general purpose optimziation routine
pow_min - Powell conjugate gradient
optsolv - yet another general purpose optimization routine
optim1_d - dfp, frpr, pow, maxlik demo

optim2_d - optsolv demo
optim3_d - fmins demo
optim4_d - minz demo
--------------- support programs -----------------------
apprgrdn - computes gradient for optsolv

box_like1 - used by optim3_d
gradnt - computes gradient
gradt - computes gradient
hessian - evaluates hessian
hessz - used by minz
linmin - line minimization routine (used by dfp, frpr, pow)
numz - used by minz
stepz - used by minz
tol_liked - used by optim1_d, optim2_d, optim4_d
--------------------------------------------------------
288
.::UdecomBooks::.
Chapter 11
Handling sparse matrices
This chapter provides demonstrations of using the MATLAB sparse matrix

algorithms to solve econometric problems. A sparse matrix is one that
contains a large proportion of zeros. The illustrations used involve spatial
econometric estimation problems where large but sparse matrices are often
encountered.
Section 11.1 discusses the computational savings associated with intelli-
gent handling of sparse matrices. In Section 11.2 we illustrate how the MAT-
LAB sparse matrix algorithms can be used to solve a spatial econometric
estimation problem involving a 3,107 by 3,107 matrix. The log-likelihood
function for this model requires that we evaluate the determinant of this
large but sparse matrix. Another illustration involving the use of sparse
matrix algorithms in a Gibbs sampling situation is provided in Section 11.3.
11.1 Computational savings with sparse matrices

As an example of a sparse matrix, consider the first-order contiguity matrix
for a sample of 3,107 U.S. counties used in Pace and Berry (1997). Re-
call, a first-order contiguity matrix has zeros on the main diagonal, rows
that contain zeros in positions associated with non-contiguous observational
units and ones in positions reflecting neighboring units that are (first-order)
contiguous. An example is shown in (11.1) for a sample of five areas.
289
.::UdecomBooks::.
CHAPTER 11. HANDLING SPARSE MATRICES 290
 
0 1 1 0 0
 
 1 0 1 0 0 
 
W = 1 1 0 1 0  (11.1)
 
 0 0 1 0 1 
0 0 0 1 0
Information regarding first-order contiguity is recorded for each obser-
vation as ones for areas that are neighbors (e.g., observations 2 and 3 are
neighbors to 1) and zeros for those that are not (e.g., observations 4 and
5 are not neighbors to 1). By convention, zeros are placed on the main
diagonal of the spatial weight matrix.
For the case of our 3,107 county sample, this matrix would be sparse
since the largest number of neighbors to any county is 8 and the average
number of neighbors is 4. A great many of the elements in the contiguity
matrix W are zero, meeting the definition of a sparse matrix.
To understand how sparse matrix algorithms conserve on storage space
and computer memory, consider that we need only record the non-zero ele-
ments of a sparse matrix for storage. Since these represent a small fraction
of the total 3107x3107 = 9,653,449 elements in the weight matrix, we save
a trememdous amount of computer memory. In fact for our example of the
3,107 U.S. counties, only 12,429 non-zero elements were found in the first-
order spatial contiguity matrix, representing a very small fraction (about
0.4 percent) of the total elements.
MATLAB provides a function sparse that can be used to construct a
large sparse matrix by simply indicating the row and column positions of
non-zero elements and the value of the matrix element for these non-zero
row and column elements. Continuing with our example, we can store the
first-order contiguity matrix in a single data file containing 12,429 rows with
3 columns that take the form:
row column value
This represents a considerable savings in computational space when com-

pared to storing a matrix containing 9,653,449 elements. A handy utility
function in MATLAB is spy which allows one to produce a specially for-
matted graph showing the sparsity structure associated with sparse matrices.
We demonstrate by executing spy(W) on our weight matrix W from the
Pace and Berry data set, which produced the graph shown in Figure 11.1.
As we can see from the figure, most of the non-zero elements reside near the
diagonal.
.::UdecomBooks::.
500
1000
1500
2000
2500
3000
0 500 1000 1500 2000 2500 3000

nz = 12428
Figure 11.1: Sparsity structure of W from Pace and Berry
As an example of storing a sparse first-order contiguity matrix, consider

example 11.1 below that reads data from the file ‘ford.dat’ in sparse for-
mat and uses the function sparse to construct a working spatial contiguity
matrix W . The example also produces a graphical display of the sparsity
structure using the MATLAB function spy.
% ----- Example 11.1 Using sparse matrix functions

load ford.dat; % 1st order contiguity matrix
% stored in sparse matrix form
ii = ford(:,1);
jj = ford(:,2);
.::UdecomBooks::.
ss = ford(:,3);
clear ford; % clear out the matrix to save RAM memory
W = sparse(ii,jj,ss,3107,3107);
clear ii; clear jj; clear ss; % clear out these vectors to save memory
spy(W);
MATLAB does not automatically treat a matrix as sparse, users must

provide a declaration using the sparse command. If we know a matrix
is sparse and want to use this to our advantage when executing matrix
operations, we convert the matrix to sparse form using: ‘b = sparse(a)’.
Subsequent MATLAB matrix operations such as multiplication, addition,
subtraction, division and the MATLAB ‘slash’ operator (discussed below)
involving ‘b’ and other sparse matrices will produce sparse matrix results.
That is, operating on two sparse matrices will create a new sparse matrix
as a result. Other operations that produce sparse matrix results when given
a sparse matrix argument are the cholesky decomposition chol, the house-
holder QR decomposition qr and the LU decomposition lu.
If one matrix in a binary operation is sparse and the other a full matrix,
the result will be a full matrix. When you are uncertain about the nature
of a matrix, the function issparse will provide an answer. This function
returns a value of 1 if the matrix is sparse and 0 otherwise. Another useful
function for dealing with sparse matrices is nnz, which returns the number
of non-zero elements in a matrix.
To illustrate working with sparse matrices, example 11.2 generates a
sparse matrix system which we solve using the MATLAB ‘slash’ operator
that carries out Gaussian elimination to solve the problem. A full matrix
solution is also provided, and we compare the time needed using the tic and
toc commands.
% ----- Example 11.2 Solving a sparse matrix system
n = 500; density = 0.1; knum=0.5;
for i=1:5;
density = density*0.5;
X = sprandsym(n,density,knum);
y = randn(n,1);
tic; b = X\y; time1 = toc;
xf = full(X);
tic; b = xf\y; time2 = toc;
fprintf(’# of non-zero elements in X = %6d \n’,nnz(X));
fprintf(’sparse matrix solution time = %8.3f \n’,time1);
fprintf(’full matrix solution time = %8.3f \n’,time2);
end;
The example relies on the sparse matrix function sprandsym to gener-
.::UdecomBooks::.
ate a symmetric matrix of dimension ‘n=500’ with random normal numbers

distributed with a density provided by an input argument to the function.
This ‘density’ argument represents the fraction of elements that will be non-
zero. We produce a sequence of five matrix systems with increasing spar-
sity by varying the ‘density’ input argument each time through the ‘for
i=1:5’ loop. A third argument allows us to specify the condition number
for the generated matrix using a fraction ‘knum’ between zero and one such
that the condition number will be 1/knum. We use this option to prevent
ill-conditioning in our matrix system as the number of non-zero elements
becomes smaller.
The results are shown below where we see that the full matrix operation
takes roughly 3.3 seconds to solve all five systems. In contrast, the sparse
matrix solution is around 10 times faster when we have a density of 0.05,
taking only .356 seconds. As the sparsity increases, the time required to
solve the problem decreases dramatically to 0.010 seconds, which is around
300 times faster. For the U.S counties the density is: 12,429/9,653,449 =
0.00128. By comparison, the smallest density in example 11.2 is 0.003125,
so it should be clear that one could spend weeks of computer time solving
large problems without sparse matrix algorithms or a few seconds with these
algorithms.
# of non-zero elements in X = 11878

sparse matrix solution time = 0.356
full matrix solution time = 3.260
The subject of sparse matrix computational efficiency is beyond the scope

of this text, but we can provide some insights that may be useful. Gaussian
elimination includes a special preordering of the columns of the matrix la-
beled “minimum degree ordering”, which reduces “fill-in” that occurs with
matrix operations. The term “fill-in” refers to the introducion of non-zero
elements as a result of matrix manipulation. As an illustration of fill-in,
.::UdecomBooks::.
we carry out matrix multiplication of the sparse first-order contiguity ma-

trix taken from Anselin’s (1988) Columbus neighborhood crime data set. A
comparison of the sparsity associated with the original first-order contiguity
matrix W and the matrix product W ∗ W ∗ W is shown in Figure 11.2.
10
15
20
25
30
35
40
45
W
W*W*W
50
0 5 10 15 20 25 30 35 40 45 50
nz = 1165
Figure 11.2: An illustration of fill-in from matrix multiplication
To illustrate the importance of preordering in matrix computations, we

carry out matrix multiplication on a symmetric sparse random matrix gen-
erated using sprandsym from example 11.2. We produce a symmetric min-
imum degree ordering, using the MATLAB function symmmd and carry
out matrix multiplication based on this ordering. We also rely on MATLAB
to automatically do this for us and carry out matrix multiplication involving
the sparse matrix. A comparison to the case of a full matrix multiplication
.::UdecomBooks::.
is provided in terms of both floating point operations (flops) and the time
required.
% ----- Example 11.3 Symmetric minium degree ordering operations

n = 500; density = 0.1; knum=0.5;
for i=1:5;
density = density*0.5;
X = sprandsym(n,density,knum);
% produce symmetric minimum degree ordering
mo = symmmd(X);
tic; flops(0); result = X(mo,mo)*X(mo,mo); time0 = toc; flops0 = flops;
tic; flops(0); result = X*X; time1 = toc; flops1 = flops;
xf = full(X);
tic; flops(0); result = xf*xf; time2 = toc; flops2 = flops;
fprintf(’# of non-zero elements in X = %6d \n’,nnz(X));
fprintf(’ordered matrix solution time = %8.3f flops = %16d \n’,time0,flops0);
fprintf(’sparse matrix solution time = %8.3f flops = %16d \n’,time1,flops1);
fprintf(’full matrix solution time = %8.3f flops = %16d \n’,time2,flops2);
end;
From the results presented below we see that the minimum degree or-
dering accounts for the greatly reduced number of floating point opera-
tions during sparse matrix multiplication. The time is slightly faster for the
straightforward sparse matrix multiplication because the internal ordering
is faster than our explicit ordering.

ordered matrix solution time = 0.179 flops = 1228634
sparse matrix solution time = 0.169 flops = 1228634
full matrix solution time = 8.320 flops = 250000018
.::UdecomBooks::.
As a graphical illustration of symmetric minimum degree ordering, Fig-

ure 11.3 shows the 3,107 county contiguity matrix sparsity structure in the
original un-ordered form along with this structure for a symmetric minimum
degree ordering of this matrix.
original un-ordered
0
500
1000
1500
2000
2500
3000
0 1000 2000 3000
nz = 12428
minimum degree ordering

0
500
1000
1500
2000
2500
3000
0 1000 2000 3000
nz = 12428
Figure 11.3: Minimum degree ordering versus unordered Pace and Barry
matrix
MATLAB provides other orderings for non-symmetric matrices using the

function colmmd and a Cuthill MacGee ordering can be generated using
the function symrcm.
As a final example of sparse matrix functionality, we examine some code
from a function normw in the spatial econometrics library that normalizes
contiguity matrices to have row-sums of unity. One might approach this
.::UdecomBooks::.
problem inefficiently using a loop over the rows as shown below.
rsum = sum(win’);
wout = zeros(n1,n1);
for i=1:n1
wout(i,:) = win(i,:)/rsum(1,i);
end;
This code is inefficient because we need to allocate space for wout, which
could be a very large but sparse matrix containing millions of elements, most
of which are zero. Rather than allocate this memory, we can rely on the find
command to extract only non-zero elements on which the standardization
operates as shown below.
[n1 n2] = size(win);

if n1 ~= n2,
error(’normw: W matrix must be square’);
end;
d = zeros(n1,1);
[junk,ind,val] = find(sum(win’));
d(ind) = val.^(-1);
D = diag(d);
wout = D*win;
We find a vector of non-zero row-sums whose index is ‘ind’. The index

variable ‘ind’ is then used to standardize only the non-zero values by oper-
ating on the entire matrix with the diagonal matrix containing the inverse
of non-zero row-sums.
11.2 Estimation using sparse matrix algorithms

MATLAB contains functions to carry out matrix operations on sparse ma-
trices. To illustrate some of these, we construct a function far that provides
maximum likelihood estimates for a first-order spatial autoregressive model
shown in (11.2).
y = ρW1 y + ε (11.2)
ε ∼ N (0, σ In ) 2
This model attempts to explain variation in y as a linear combination of

contiguous or neighboring units with no other explanatory variables. The
model is termed a first order spatial autoregression because its represents a
.::UdecomBooks::.
spatial analogy to the first order autoregressive model from time series anal-
ysis, yt = ρyt−1 + εt , where total reliance is on the past period observations
to explain variation in yt .
It is conventional to standardize the spatial weight matrix W so that
the row sums are unity and to put the vector of sample observations y
in deviations from the means form to eliminate the intercept term from
the model. Anselin (1988) shows that ordinary least-squares estimation of
this model will produce biased and inconsistent estimates. Beacuse of this,
a maximum likelihood approach can be used to find an estimate of the
parameter ρ using the likelihood function shown in (11.3).
1 1
L(y|ρ, σ 2 ) = |In − ρW | exp{− (y − ρW y)0 (y − ρW y)} (11.3)
2πσ 2 (n/2) 2σ 2
In order to simplify the maximization problem, we obtain a concen-

trated log likelihood function based on eliminating the parameter σ 2 for
the variance of the disturbances. This is accomplished by substituting
σ̂ 2 = (1/n)(y − ρW y)0 (y − ρW y) in the likelihood (11.3) and taking logs
which yields:
n
Ln(L) ∝ − ln(y − ρW y)0 (y − ρW y) + ln|In − ρW | (11.4)
2
This expression can be maximized with respect to ρ using a simplex univari-
ate optimization routine. The estimate for the parameter σ 2 can be obtained
using the value of ρ that maximizes the log-likelihood function (say, ρ̃) in:
σ̂ 2 = (1/n)(y − ρ̃W y)0 (y − ρ̃W y).
Note that the log-likelihood function in (11.4) contains an expression for
the determinant of (In −ρW ), which for our example data sample is 3,107 by
3,107. We demonstrate a sparse matrix algorithm approach to evaluating
this determinant that allows us to solve problems involving thousands of
observations quickly with small amounts of computer memory.
Two implementation details arise with this approach to solving for maxi-
mum likelihood estimates. First, there is a constraint that we need to impose
on the parameter ρ. This parameter can take on feasible values in the range
(Anselin and Florax, 1994):
where λmin represents the minimum eigenvalue of the standardized spatial

contiguity matrix W and λmax denotes the largest eigenvalue of this matrix.
.::UdecomBooks::.
This suggests that we need to constrain our optimization procedure search

over values of ρ within this range. Note that this requires we solve for the
maximum and minimum eigenvalues of the matrix W which is 3,107 by
3,107, not a trivial problem.
The second implementation issue is that for problems involving a small
number of observations, we used our knowledge of the theoretical infor-
mation matrix to produce estimates of dispersion. This approach is com-
putationally impossible when dealing with large scale problems involving
thousands of observations. In these cases we can evaluate the numerical
hessian matrix using the maximum likelihood estimates of ρ and σ 2 as well
as a sparse matrix non-concentrated version of the likelihood function.
Our first task is to construct a function to evaluate the concentrated
log likelihood based on the sparse matrix algorithms. This function named
f far is shown below.
function llike = f_far(rho,y,W)

% PURPOSE: evaluate the concentrated log-likelihood for the first-order
% spatial autoregressive model using sparse matrix algorithms
% ---------------------------------------------------
% USAGE:llike = f_far(rho,y,W)
% where: rho = spatial autoregressive parameter
% y = dependent variable vector
% W = spatial weight matrix
% ---------------------------------------------------
% RETURNS: a scalar equal to minus the log-likelihood
% function value at the parameter rho
% --------------------------------------------------
% SEE ALSO: far, f_sar, f_sac, f_sem
% ---------------------------------------------------
n = length(y); spparms(’tight’);
z = speye(n) - 0.1*sparse(W);
p = colmmd(z);
z = speye(n) - rho*sparse(W);
[l,u] = lu(z(:,p));
detval = sum(log(abs(diag(u))));
epe = y’*z’*z*y;
llike = (n/2)*log(pi) + (n/2)*log(epe) - detval;
The function solves the determinant of (I − ρW ) using the LU decompo-

sition implemented by the MATLAB lu command. This algorithm operates
on sparse matrices in an intelligent way. The command spparms sets op-
tions for operation of the sparse matrix algorithms. The option ‘tight’ sets
the minimum degree ordering parameters to particular settings, which lead
to sparse matrix orderings with less fill-in, but make the ordering functions
.::UdecomBooks::.
use more execution time. Some experimentation on my part with the vari-
ous options that can be set has led me to believe this is an optimal setting
for this type of model. The command sparse informs MATLAB that the
matrix W is sparse and the command speye creates an identity matrix
in sparse format. We set up an initial matrix based on (In − 0.1W ) from
which we construct a column vector of minimum degree permutations for
this sparse matrix. By executing the lu command with this vector, we man-
age to operate on a sparser set of LU factors than if we operated on the
matrix z = (I − ρW ).
Given this function to evaluate the log likelihood for very large spatial
weight matrices W , we can now rely on the same fmin simplex optimization
algorithm demonstrated in Chapter 10. Another place where we can rely
on sparse matrix functions is in determining the minimum and maximum
eigenvalues of the matrix W . We will use these values to set the feasible
range for ρ in our call to the simplex search function fmin. The code for
carrying this out is:
opt.tol = 1e-3; opt.disp = 0;

lambda = eigs(sparse(W),speye(n),2,’BE’,opt);
lmin = 1/lambda(2);
lmax = 1/lambda(1);
The MATLAB function eigs works to compute selected eigenvalues for

a sparse matrix, and we use the option ‘BE’ to compute only the maximum
and minimum eigenvalues which we need to determine the feasible range for
ρ. The options set as fields in the ‘opt’ structure variable indicate a tolerance
to be used when solving the eigenvalue problem and to prohibit display of
the iterative solution results. The default tolerance is 1e-10, but using the
tolerance of 1e-3 speeds up the solution by a factor of four times. Note that
we are not overly concerned about loosing a few decimal digits of accuracy,
since it is unlikely that the maximum likelihood values of ρ are near these
bounds on the feasible range. If we find a maximum likelihood estimate
near these limits, we can always change the tolerance option argument to
the eigs function.
Another point is that for the case of standardized weight matrices W , the
maximum eigenvalue will always take on a value of unity, so we could save
time by only computing one eigenvalue. However, not all researchers will
use a row-standardized matrix W , so we make the function more generally
useful by computing both eigenvalues.
The final issue we need to address is computing measures of dispersion
for the estimates ρ and σ 2 in our estimation problem. As already noted,
.::UdecomBooks::.
we cannot rely on the information matrix approach because this involves

matrix operations on very large matrices. An approach that we take to
produce measures of dispersion is to numerically evaluate the hessian matrix
using the maximum likelihood estimates of ρ and σ 2 . Our function hessian
from Chapter 10 can be used to compute the hessian matrix given a non-
concentrated log-likelihood function and maximum likelihood estimates for
ρ and σ. This non-concentrated version of the log-likelihood function is
shown below.
function llike = f2_far(parm,y,W)

% PURPOSE: evaluate the log-likelihood for ML rho,sigma values
% for the first-order spatial autoregressive model
% ---------------------------------------------------
% USAGE: llike = f2_far(parm,y,W)
% where: parm = 2x1 vector with rho,sigma values
% ---------------------------------------------------
% RETURNS: a scalar equal to minus the log-likelihood
% function value at the ML parameters rho,sigma
% --------------------------------------------------
% SEE ALSO: far, f2_sar, f2_sac, f2_sem
% ---------------------------------------------------
n = length(y); rho = parm(1,1); sige = parm(2,1);
spparms(’tight’); z = speye(n) - 0.1*sparse(W);
p = colmmd(z); z = speye(n) - rho*sparse(W);
[l,u] = lu(z(:,p));
detval = sum(log(abs(diag(u))));
epe = y’*z’*z*y;
llike = (n/2)*log(pi) + (n/2)*log(epe) +(n/2)*log(sige) - detval;
The function far that implements this approach to estimation of the

first-order autoregressive model is shown below.
function results = far(y,W,convg,maxit)

% PURPOSE: computes 1st-order spatial autoregressive estimates
% model: y = p*W*y + e, using sparse matrix algorithms
% ---------------------------------------------------
% USAGE: results = far(y,W,convg,maxit)
% where: y = dependent variable vector
% W = standardized contiguity matrix
% convg = (optional) convergence criterion (default = 1e-8)
% maxit = (optional) maximum # of iterations (default = 500)
% ---------------------------------------------------
% results.meth = ’far’
% results.rho = rho
.::UdecomBooks::.
% results.tstat = asymptotic t-stat

% results.sige = sige = (y-p*W*y)’*(y-p*W*y)/n
% results.lik = -log likelihood
% results.nvar = nvar = 1
% results.iter = # of iterations taken
% results.romax = 1/maximum eigenvalue of W
% results.romin = 1/minimum eigenvalue of W
% --------------------------------------------------
% SEE ALSO: prt(results), sar, sem, sac
% ---------------------------------------------------
options = zeros(1,18);
if nargin == 2 % set default optimization options
options(1,1) = 0; options(1,2) = 1.e-8; options(14) = 500;
elseif nargin == 3 % set user supplied convergence option
options(1,1) = 0; options(1,2) = convg; options(1,14) = 500;
elseif nargin == 4 % set user supplied convg and maxit options
options(1,1) = 0; options(1,2) = convg; options(1,14) = maxit;
else, error(’Wrong # of arguments to far’);
end;
[n junk] = size(y); results.y = y; results.nobs = n;
results.nvar = 1; results.meth = ’far’;
lmin = 1/lambda(2); lmax = 1/lambda(1);
results.romax = lmax; results.romin = lmin;
% step 1) maximize concentrated likelihood function;
[p options] = fmin(’f_far’,1/lmin,1/lmax,options,y,W);
if options(10) == options(14),
fprintf(1,’far: convergence not obtained in %4d iterations’,options(14));
else, results.iter = options(10);
end;
Wy = sparse(W)*y; epe = (y - p*Wy)’*(y-p*Wy); sige = epe/n;
results.rho = p; results.yhat = p*Wy;
results.resid = y - results.yhat; results.sige = sige;
% asymptotic t-stats using numerical hessian
parm = [p
sige];
hessn = hessian(’f2_far’,parm,y,W); xpxi = inv(hessn);
results.tstat = results.rho/sqrt(xpxi(1,1));
ym = y - mean(y); rsqr1 = results.resid’*results.resid;
rsqr2 = ym’*ym; results.rsqr = 1.0-rsqr1/rsqr2; % r-squared
results.lik = -f2_far(p,y,W); % -log likelihood
As an example of using this function, we solve for maximum likelihood
.::UdecomBooks::.
estimates using a large sample of 3,107 observations representing counties

in the continental U.S. from Pace and Berry (1997). They examine pres-
idential election results for this large sample of observations covering the
U.S. presidential election of 1980 between Carter and Reagan. The variable
we wish to explain using the first-order spatial autoregressive model is the
proportion of total possible votes cast for both candidates. Only persons
18 years and older are eligible to vote, so the proportion is based on those
voting for both candidates divided by the population over 18 years of age.
Example 11.4 shows a MATLAB program that reads in data from a
file ‘elect.dat’ that contains 3,107 rows with the county-level sample data.
Another file named ‘ford.dat’ holds the first-order contiguity matrix informa-
tion in sparse matrix storage form. This allows us to store the information
in three columns containing 12,429 rows with the non-zero values. We rely
on the MATLAB sparse command to construct the first-order contiguity
matrix from the compactly stored information.
% ----- Example 11.4 Using the far() function

% with very large data set from Pace and Berry
load elect.dat; % load data on votes

y = elect(:,7)./elect(:,8); % proportion of voters casting votes
ydev = y - mean(y); % deviations from the means form
clear y; % conserve on RAM memory
clear elect; % conserve on RAM memory
load ford.dat; % 1st order contiguity matrix stored in sparse matrix form
ii = ford(:,1); jj = ford(:,2); ss = ford(:,3);
n = 3107;
clear ford; % clear ford matrix to save RAM memory
W = sparse(ii,jj,ss,n,n);
clear ii; clear jj; clear ss; % conserve on RAM memory
tic; res = far(ydev,W); toc;
prt(res);
A breakdown of time needed to solve various aspects of the estimation

problem is shown along with the output below. Our use of the simplex
optimization algorithm required 13 function evaluations which took 10.6
seconds. The total time required to compute the estimates and measures of
dispersion for ρ and σ, the R−squared statistics and log likelihood function
value was around 100 seconds.
elapsed_time = 59.8226 % computing min,max eigenvalues

elapsed_time = 10.6622 % time required for simplex solution of rho
elapsed_time = 1.7681 % time required for hessian evaluation
elapsed_time = 1.7743 % time required for likelihood evaluation
.::UdecomBooks::.
First-order spatial autoregressive model Estimates

R-squared = 0.5375
sigma^2 = 0.0054
log-likelihood = 3506.3203
min and max rho = -1.0710, 1.0000
***************************************************************
rho 0.721474 59.567710 0.000000
11.3 Gibbs sampling and sparse matrices

We have already used the first-order spatial autoregressive model to illus-
trate Metropolis-within-Gibbs sampling in Chapter 6. Here we extend the
approach to include the use of sparse matrix algorithms that will allow us
to solve very large models of the type illustrated in the previous section.
We focus attention on implementation details concerned with construct-
ing a MATLAB function far g that will produce estimates for a Bayesian
variant of the first-order spatial autoregressive model. This function will
rely on a sparse matrix algorithm approach to handle problems involving
large data samples. It will also allow for diffuse or informative priors and
handle the case of heterogeneity in the disturbance variance.
The first thing we need to consider is that in order to produce a large
number of draws, say 1,000, we would need to evaluate the conditional
distribution of ρ given σ 2,000 times since we need to call this function
twice during metropolis sampling. Each evaluation would require that we
compute the determinant of the matrix (In − ρW ), which we have already
seen is a non-trivial task for large data samples. To avoid this, we rely on an
approach suggested by Pace and Berry (1997). They suggested evaluating
this determinant over a grid of values in the feasible range of ρ once at the
outset. Given that we have carried out this evaluation and stored the values
for the determinant and associated values of ρ, we can simply “look-up”
the appropriate determinant in our function that evaluates the conditional
distribution. Each call to the conditional distribution function will provide
a value of ρ for which we need to evaluate the conditional distribution. If
we already know the determinant for a grid of all feasible ρ values, we can
simply look up the determinant value closest to the ρ value and use it during
evaluation of the conditional distribution. This saves us the time involved
in computing the determinant twice for each draw of ρ.
.::UdecomBooks::.
The code that we execute at the outset in our function far g to compute
determinant values over a grid of ρ values is shown below. After finding the
minimum and maximum eigenvalues using our approach from the previous
section, we define a grid based on increments of 0.01 over these values and
evaluate the determinant over this grid.

lmin = 1/lambda(2); lmax = 1/lambda(1);
results.romax = lmax; results.romin = lmin;
% compute a detval vector based on Pace and Berry’s approach
rvec = lmin:.01:lmax;
spparms(’tight’);
z = speye(n) - 0.1*sparse(W);
p = colmmd(z);
niter = length(rvec);
detval = zeros(niter,2);
for i=1:niter;
rho = rvec(i); z = speye(n) - rho*sparse(W);
[l,u] = lu(z(:,p));
detval(i,1) = sum(log(abs(diag(u)))); detval(i,2) = rho;
end;
Note that we save the values of the determinant alongside the associated
values of ρ in a 2-column matrix named detval. We will simply pass this
matrix to the conditional distribution function c far which is shown below:
function cout = c_far(rho,y,W,detval,sige,c,T)

% PURPOSE: evaluate the conditional distribution of rho given sige
% 1st order spatial autoregressive model using sparse matrix algorithms
% ---------------------------------------------------
% USAGE:cout = c_far(rho,y,W,detval,c,T)
% where: rho = spatial autoregressive parameter
% detval = an (ngrid,2) matrix of values for det(I-rho*W)
% over a grid of rho values
% detval(:,1) = determinant values
% detval(:,2) = associated rho values
% sige = sige value
% c = optional prior mean for rho
% T = optional prior variance for rho
% ---------------------------------------------------
% RETURNS: a conditional used in Metropolis-Hastings sampling
% NOTE: called only by far_g
% --------------------------------------------------
% SEE ALSO: far_g, c_sar, c_sac, c_sem
.::UdecomBooks::.
% ---------------------------------------------------
i1 = find(detval(:,2) <= rho + 0.005);
i2 = find(detval(:,2) <= rho - 0.005);
i1 = max(i1); i2 = max(i2);
index = round((i1+i2)/2);
detm = detval(index,1); n = length(y);
z = speye(n) - rho*sparse(W);
if nargin == 5, % diffuse prior
epe = (n/2)*log(y’*z’*z*y);
elseif nargin == 7 % informative prior
epe = (n/2)*log(y’*z’*z*y + (rho-c)^2/T);
end;
cout = -epe -(n/2)*log(sige) + detm;
In the function c far, we find the determinant value that is closest to the
ρ value for which we are evaluating the conditional distribution. This is very
fast in comparison to calculating the determinant. Since we need to carry
out a large number of draws, this approach works better than computing
determinants for every draw.
Another point is that we allow for the case of a normally distributed
informative prior on the parameter ρ in the model, which changes the con-
ditional distribution slightly. The documentation for our function far g is
shown below.
PURPOSE: Gibbs sampling estimates of the 1st-order Spatial

model: y = rho*W*y + e, e = N(0,sige*V),
rho = N(c,T), sige = gamma(nu,d0)
----------------------------------------------------------------
USAGE: result = far_g(y,W,ndraw,nomit,prior,start)
W = nobs x nobs 1st-order contiguity matrix (standardized)
prior = a structure variable for prior information input
prior.rho, prior mean for rho, c above, default = diffuse
priov.rcov, prior rho variance , T above, default = diffuse
prior.rmin, (optional) min value of rho to use in sampling
prior.rmax, (optional) max value of rho to use in sampling
ndraw = # of draws
start = (optional) (2x1) vector of rho, sige starting values
.::UdecomBooks::.
(defaults, rho = 0.5, sige = 1.0)

---------------------------------------------------------------
results.meth = ’far_g’
results.bdraw = rho draws (ndraw-nomit x 1)
results.sdraw = sige draws (ndraw-nomit x 1)
results.vmean = mean of vi draws (nobs x 1)
results.rdraw = r-value draws (ndraw-nomit x 1)
results.pmean = rho prior mean (if prior input)
results.pstd = rho prior std dev (if prior input)
results.nu = prior nu-value for sige (if prior input)
results.d0 = prior d0-value for sige (if prior input)
results.r = value of hyperparameter r (if input)
results.m = m prior parameter (if input)
results.k = k prior parameter (if input)
results.ndraw = # of draws
results.nomit = # of initial draws omitted
results.yhat = mean of posterior for y-predicted (nobs x 1)
results.time = time taken for sampling
results.accept = acceptance rate
results.pflag = 1 for prior, 0 for no prior
results.rmax = 1/max eigenvalue of W (or rmax if input)
results.rmin = 1/min eigenvalue of W (or rmin if input)
----------------------------------------------------------------
----------------------------------------------------------------
As the documentation makes clear, there are a number of user options

to facilitate different models. We can use the function for homoscedastic
as well as heteroscedastic data samples as well as those containing outliers.
This was accomplished using the approach of Geweke (1993) illustrated in
Chapter 6. In addition, an informative prior can be used for the parameter
σ.
Example 11.5 illustrates using the function with various input options
on the large Pace and Berry data set. We set an r value of 5 which will
capture heterogeneity if it exists.
% ----- Example 11.5 Gibbs sampling with sparse matrices

% using a very large data set from Pace and Berry
load elect.dat; % load data on votes in 3,107 counties
y = (elect(:,7)./elect(:,8)); % convert to per capita variables
ydev = y - mean(y);
clear elect; % conserve on RAM memory
load ford.dat; % 1st order contiguity matrix stored in sparse matrix form
.::UdecomBooks::.
ii = ford(:,1); jj = ford(:,2); ss = ford(:,3);

n = 3107;
clear ford; % clear ford matrix to save RAM memory
W = sparse(ii,jj,ss,n,n);
clear ii; clear jj; clear ss; % conserve on RAM memory
prior.rval = 5; ndraw = 1100; nomit = 100;
res = far_g(ydev,W,ndraw,nomit,prior);
prt(res);
We also present the maximum likelihood results for comparison with the
Gibbs sampling results. If there is no substantial heterogeneity in the distur-
bance, the two sets of estimates should be similar, as we saw in Chapter 6.
From the results, we see that the estimates are similar, suggesting a lack of
heterogeneity that would lead to different estimated values for ρ and σ.
% Maximum likelihood results

First-order spatial autoregressive model Estimates
R-squared = 0.5375
sigma^2 = 0.0054
log-likelihood = 3506.3203
min and max rho = -1.0710, 1.0000
***************************************************************
rho 0.721474 59.567710 0.000000
% Gibbs sampling estimates
Gibbs sampling First-order spatial autoregressive model
R-squared = 0.5711
sigma^2 = 0.0050
r-value = 5
acceptance rate = 0.8773
min and max rho = -1.0710, 1.0000
***************************************************************
rho 0.718541 42.026284 0.000000
Note that the time needed to produce 1100 draws was around 378 sec-
onds, making this estimation method competitive with the maximum likeli-
hood approach which took around 100 seconds.
.::UdecomBooks::.
11.4 Chapter summary

We illustrated the use of sparse matrix functions that represent an important
feature of MATLAB. These functions allow the solution of problems involv-
ing large but sparse matrices with a minimum of both time and computer
memory.
Our illustrations demonstrated the use of these functions using a spatial
econometrics problem, but other examples where these functions would be
of use surely exists in econometrics. Cases where block diagonal matrices
are used would be one example.
Extensive use of the approach presented in this chapter has been made
to implement a host of maximum likelihood and Gibbs sampling implemen-
tations of spatial autoregressive estimation functions contained in a spatial
econometrics library.
.::UdecomBooks::.
Chapter 11 Appendix
The spatial econometrics library is in a subdirectory spatial.
spatial econometrics library

------- spatial econometrics functions -----------
gwr_logit - logit version of gwr
gwr_probit - probit version of gwr
bgwrv - robust geographically weighted regression
normw - normalizes a spatial weight matrix
sacp_g - Gibbs sampling Bayesian sac probit model
sarp_g - Gibbs sampling Bayesian sar probit model
sart_g - Gibbs sampling Bayesian sar tobit model
semp_g - Gibbs sampling Bayesian spatial error probit model
semt_g - Gibbs sampling Bayesian spatial error tobit model
sdm_g - Gibbs sampling Bayesian sdm model
sdmp_g - Gibbs sampling Bayesian sdm probit model
sdmt_g - Gibbs sampling Bayesian sdm tobit model
slag - creates spatial lags
310
.::UdecomBooks::.


sacp_gd - sac Gibbs probit demo
sact_gd - sac Gibbs tobit demo
sact_gd2 - sac tobit right-censoring demo
sarp_gd - sar probit Gibbs sampling demo
sart_gd - sar tobit model Gibbs sampling demo
sart_gd2 - sar tobit right-censoring demo
sdmp_g - sdm Gibbs probit demo
sdmt_g - sdm Gibbs tobit demo
semp_gd - sem Gibbs probit demo
semt_gd - sem Gibbs tobit demo
semt_gd2 - sem tobit right-censoring demo
slag_d - demo of slag function
.::UdecomBooks::.


anselin.dat - Anselin (1988) Columbus crime data
.::UdecomBooks::.
References
Albert, J. and S. Chib. 1993. “Bayes inference via Gibbs sampling

of autoregressive time series subject to Markov mean and variance
shifts,” Journal of Business and Economic Statistics, Vol. 11, pp. 1-
15.
Albert, James A. and Siddartha Chib. 1993. “Bayesian Analysis of
Binary and Polytomous Response Data”. Journal of the American
Statistical Association, Vol. 88, pp. 669-679.
Anselin, Luc. 1988. Spatial Econometrics: Methods and Models,
(Dordrecht: Kluwer Academic Publishers).
Belsley, D.A., R.E. Welsch and E. Kuh, 1980. Regression Diagnostics,
John Wiley & Sons, Inc. (New York: New York).
Best, N.G., M.K. Cowles, and S.K. Vines. 1995. CODA: Manual
version 0.30. Biostatistics Unit, Cambridge U.K. http://www.mrc-
bsu.cam.ac.uk
Box, George E.P., and George C. Tiao. 1992. Bayesian Inference in
Statistical Analysis, (John Wiley and Sons: New York).
Casella, G. and E.I. George. 1992. “Explaining the Gibbs Sampler”,
American Statistician, Vol. 46, pp. 167-174.
Chib, Siddhartha. 1993. “Bayes regression with autoregressive error:
A Gibbs sampling approach,” Journal of Econometrics, Vol. 58, pp.
275-294.
Chib, Siddhartha. 1992. “Bayes inference in the Tobit censored re-
gression model”, Journal of Econometrics. Vol. 51, pp. 79-100.
Cook, R.D. and S. Weisberg, 1982. Residuals and Influence in Regres-
sion, Chapman and Hall, (New York, London).
313
.::UdecomBooks::.
REFERENCES 314
Dempster, A.P. N.M. Laird and D.B. Rubin. 1977. “Maximum like-
lihood from incomplete data via the EM algorithm,” Journal of the
Royal Statistical Society, Series B, Vol. 39, pp. 1-38.
Dickey, David A., Dennis W. Jansen and Daniel L. Thornton. 1991.
“A primer on cointegration with an application to money and in-
come,” Federal Reserve Bulletin, Federal Reserve Bank of St. Louis,
March/April, pp. 58-78.
Doan, Thomas, Robert. B. Litterman, and Christopher A. Sims. 1984.
“Forecasting and conditional projections using realistic prior distribu-
tions,” Econometric Reviews, Vol. 3, pp. 1-100.
Engle, Robert F. and Clive W.J. Granger. 1987. “Co-integration
and error Correction: Representation, Estimation and Testing,” Eco-
noemtrica, Vol. 55, pp. 251-76.
Estrella, Artuto. 1998. “A new measure of fit for equations with di-
chotmous dependent variable”, Journal of Business & Economic Statis-
tics, Vol. 16, no. 2, pp. 198-205.
Fomby, Thomas B., R. Carter Hill and Stanley R. Johnson. 1984.
Advanced Econometric Methods, (Springer Verlag: New York).
Geman, S., and D. Geman. 1984. “Stochastic relaxation, Gibbs distri-
butions, and the Bayesian restoration of images,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, Vol. 6, pp. 721-741.
Gelfand, Alan E., and A.F.M Smith. 1990. “Sampling-Based Ap-
proaches to Calculating Marginal Densities”, Journal of the American
Gelfand, Alan E., Susan E. Hills, Amy Racine-Poon and Adrian F.M.
Smith. 1990. “Illustration of Bayesian Inference in Normal Data Mod-
els Using Gibbs Sampling”, Journal of the American Statistical Asso-
ciation, Vol. 85, pp. 972-985.
Gelman, Andrew, John B. Carlin, Hal S. Stern, and Donald B. Rubin.
1995. Bayesian Data Analysis, (London: Chapman & Hall).
Geweke, John. 1991. “Efficient Simulation from the Multivariate nor-
mal and Student-t Distributions Subject to Linear Constraints,” in
Computing Science and Statistics: Proceedings of the Twenty-Third
Symposium on the Interface.
.::UdecomBooks::.
REFERENCES 315
Geweke, John. 1992. “Evaluating the Accuracy of Sampling-Based

Approaches to the Calculation of Posterior Moments”, Bayesian Statis-
tics, 4, J.M Bernardo, J.0. Berger, A.P. Dawid, and A.F.M Smith,
(eds.), pp. 641-649. (Oxford: Clarendon Press).
Geweke, John. 1993. “Bayesian Treatment of the Independent Stu-

dent t Linear Model”, Journal of Applied Econometrics, Vol. 8, s19-
s40.
Geweke, John and Michael Keane. 1997. “Mixture of Normals Pro-

bit Models”, Research Department Staff Report 237, Federal Reserve
Bank of Minneapolis.
Gilks, W.R., S. Richardson and D.J. Spiegelhalter. 1996. Markov

Chain Monte Carlo in Practice, (London: Chapman & Hall).
Goldfeld S.M. and R. E. Quandt 1973. “A Markov model for switching

regressions”, Journal of Econometrics, Vol. 1, pp. 3-16.
Green, William H., 1997. Econometric Analysis, Third Edition, (Pren-

tice Hall: New Jersey).
Hamilton, James D. 1989. “A new approach to economic analysis of

nonstationary time series and the business cycle”, Econometrica, Vol.
57, number 2, pp. 357-384.
Hanselmann, D. adn B. Littlefield. 1997. The Student Edition of

MATLAB, Version 5 User’s Guide. (New Jersey: Prentice Hall).
Hastings, W. K. 1970. “Monte Carlo sampling methods using Markov

chains and their applications,” Biometrika, Vol. 57, pp. 97-109.
Hoerl A.E., and R.W. Kennard, 1970. “Ridge Regression: Biased

Estimation and Applications for Nonorthogonal Problems,” Techno-
metrics, Vol. 12, pp. 55-82.
Hoerl A.E., R.W. Kennard and K.F. Baldwin, 1975. “Ridge Regres-
sion: Some Simulations,” Communications in Statistics, A, Vol. 4, pp.
105-23.
Johansen, Soren. 1988. “Statistical Analysis of Co-integration vec-

tors,” Journal of Economic Dynamics and Control, Vol. 12, pp. 231-
254.
.::UdecomBooks::.
REFERENCES 316
Johansen, Soren. 1995. Likelihood-based Inference in Cointegrated

Vector autoregressive Models, Oxford: Oxford University Press.
Johansen, Soren and Katarina Juselius. 1990. ‘Maximum likelihood
estimation and inference on cointegration - with applications to the
demand for money”, Oxford Bulletin of Economics and Statistics, Vol.
52, pp. 169-210.
Judge G. C., R.C. Hill, W.E. Griffiths, H. Lutkepohl, T. Lee, 1988.
Introduction to the Theory and Practice of Econometrics, John Wiley
& Sons, Inc. (New York: New York).
Kim, Chaing-Jin. 1994. “Dynamic linear models with Markov switch-
ing”, Journal of Econometrics, Vol. 60, pp. 1-22.
Lange, K.L., R.J.A. Little, and J.M.G. Taylor. 1989. “Robust Sta-
tistical Modeling Using the t Distribution,” Journal of the American
Leamer, Edward E. 1983. “Model Choice and Specification Analysis”,
in Handbook of Econometrics, Volume 1, Zvi Griliches and Michael
D. Intriligator, eds. (North-Holland: Amsterdam).
LeSage, James P. 1990. “A Comparison of the Forecasting Ability of
ECM and VAR Models,” Review of Economics and Statistics, Vol. 72,
pp. 664-671.
LeSage, James P. 1997. “Bayesian Estimation of Spatial Autoregres-
sive Models”, International Regional Science Review, 1997 Vol. 20,
number 1&2, pp. 113-129. Also available at www.econ.utoledo.edu.
LeSage, James P. 1998. “Bayesian Estimation of Spatial Probit/Tobit
Models”, Available at www.econ.utoledo.edu.
LeSage, James P. and Anna Krivelyova. 1999. “A Spatial Prior for
Bayesian Vector Autoregressive Models,” Journal of Regional Science,
Vol. 39, number 2, pp. 297-317.
LeSage, James P. and Anna Krivelova. 1998. “A Random Walk Aver-
aging Prior for Bayesian Vector Autoregressive Models,” available at:
www.econ.utoledo.edu.
LeSage, James P. and Michael Magura. 1991. “Using interindustry
input-output relations as a Bayesian prior in employment forecasting
models”, International Journal of Forecasting, Vol. 7, pp. 231-238.
.::UdecomBooks::.
REFERENCES 317
LeSage, James P. and Zheng Pan. 1995. ‘Using Spatial Contiguity as

Bayesian Prior Information in Regional Forecasting Models”, Interna-
tional Regional Science Review, Vol. 18, no. 1, pp. 33-53.
Lindley, David V. 1971. “The estimation of many parameters,” in

Foundations of Statistical Science, V.P. Godambe and D.A. Sprout
(eds.) (Toronto: Holt, Rinehart, and Winston).
Litterman, Robert B. 1986. “Forecasting with Bayesian Vector Au-

toregressions — Five Years of Experience,” Journal of Business &
Economic Statistics, Vol. 4, pp. 25-38.
MacKinnon, J.G. 1994 “Approximate Asymptotic Distribution Func-

tions for unit-root and cointegration tests,” Journal of Business &
Economic Statistics, Vol. 12, pp. 167-176.
MacKinnon, J.G. 1996 “Numerical distribution functions for unit-root

and cointegration tests,” Journal of Applied Econometrics, Vol. 11,
pp. 601-618.
McFadden, Daniel. 1984. “Econometric Analysis of Qualitative Re-

sponse Models”, in Zvi Griliches and Michael D. Intriligator, eds.
Handbook of Econometrics, Volume 2, (North-Holland: Amsterdam).
McMillen, Daniel P. 1992. “Probit with Spatial Autocorrelation”,

Journal of Regional Science, Vol. 32, no. 3, pp. 335-348.
Metroplis, N., A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller and

E. Teller. 1953. “Equation of state calculations by fast computing
machines,” Journal of Chemical Physics, Vol. 21, pp. 1087-1092.
Pace, R. Kelley and Ronald Barry. 1997. “Quick Computation of Spa-

tial Autoregressive Estimators”, forthcoming in Geographical Analysis.
Raftery, Adrian E., and Steven M. Lewis. 1992a. “How many iter-
ations in the Gibbs sampler?”, in Bayesian Statistics, Vol. 4, J.M.
Bernardo, J.O. Berger, A.P. Dawid and A.F.M Smith, eds. (Oxford
University Press: Oxford, pp. 763-773.)
Raftery, Adrian E., and Steven M. Lewis. 1992b. “One long run
with diagnostics: Implementation strategies for Markov chain Monte
Carlo”, Statistical Science, Vol. 7, pp. 493-497.
.::UdecomBooks::.
REFERENCES 318
Raftery, Adrian E., and Steven M. Lewis. 1995. “The number of it-
erations, convergence diagnostics and generic Metropolis algorithms”,
forthcoming in Practical Markov Chain Monte Carlo, W.R. Gilks, D.J.
Spiegelhalter and S. Richardson, eds. (Chapman and Hall: London).
Raftery, Adrian E., David Madigan and Jennifer A. Hoeting. 1997.
“Bayesian model averaging for linear regression models,”, Journal of
the American Statistical Association, Vol. 92, pp. 179-191.
Shoesmith, Gary L. 1992. “Cointegration, error Correction and Im-
proved Regional VAR Forecasting,” Journal of Forecasting, Vol. 11,
pp. 91-109.
Shoesmith, Gary L. 1995. “Multiple Cointegrating Vectors, error Cor-
rection, and Litterman’s Model” International Journal of Forecasting,
Vol. 11, pp. 557-567.
Simon, S.D., and J.P. LeSage, 1988a. “The Impact of Collinearity In-
volving the Intercept Term on the Numerical Accuracy of Regression,”
Computational Statistics in Economics and Management Science, Vol.
1 no. 2, pp. 137-152.
Simon, S.D., and J.P. LeSage, 1988b. “Benchmarking Numerical Ac-
curacy of Statistical Algorithms,” with Stephen D. Simon, Computa-
tional Statistics and Data Analysis, Vol. 7, pp. 197-209.
Sims, Christopher A. 1980. “Macroeconomics and Reality,” Econo-
metrica Vol. 48, pp. 1-48.
Smith, A.F.M and G.O. Roberts (1992). “Bayesian Statistics without
Tears: A Sampling-Resampling Perspective”, The American Statisti-
cian, Vol. 46, pp. 84-88.
Spector, L., and M. Mazzeo. 1980. “Probit Analysis and Economic
Education.” Journal of Economic Education, Vol. 11, pp. 37-44.
Tanner, Martin A. 1991. Tools for Statistical Inference, (Springer-
Verlag: New York).
Theil, Henri., 1971. Principles of Econometrics, (John Wiley & Sons:
New York).
Theil, Henri and Arthur S. Goldberger. 1961. “On Pure and Mixed
Statistical Estimation in Economics,” International Economic Review,
Vol. 2, 65-78.
.::UdecomBooks::.
REFERENCES 319
van Norden, Simon and Huntley Schaller. 1993. “The predictability

of stock market regime: evidence from the Toronto Stock Exchange,”
Review of Economics and Statistics.
Wampler, R.H., 1980. “Test Procedures and Test Problems for Least-
Squares Algorithms,” Journal of Econometrics, Vol. 12, pp. 3-22.
Wecker, William E. 1979. “Predicting the Turning Points of a Time

Series,” Journal of Business, Vol. 55, pp. 57-85.
Zellner, Arnold and Chanisk Hong. 1988. “Bayesian Methods for

Forecasting Turning Points in Economic Time Series: Sensitivity of
Forecasts to Asymmetricy of Loss Structures,” in Leading Economic
Indicators: New Approaches and Forecasting Records, (K. Lahiri and
G. Moore, eds.), pp. 129-140. (England: Cambridge Univ. Press).
Zellner, Arnold, Chanisk Hong, and G. M. Gulati. 1990. “Turning

Points in Economic Time Series, Loss Structures and Bayesian Fore-
casting”, in Bayesian Likelihood Methods in Statistics and Economet-
rics: Essays in Honor of George A. Barnard, (S. Geisser, J. Hodges,
S.J. Press, and A. Zellner, eds.), pp. 371-393. (Amsterdam: North-
Holland Publ.).
.::UdecomBooks::.
Appendix: Toolbox functions
The Econometric Toolbox is organized in a set of directories, each containing

a different library of functions. When your Internet browser unpacks the
compressed file containing the Econometric Toolbox the files will be placed
in the appropriate directories.
To install the toolbox:
1. create a single subdirectory in the MATLAB toolbox directory:
C:\matlab\toolbox\econ
Where we have used the name econ for the directory.
2. Copy the system of directories to this subdirectory.
3. Use the graphical path tool in MATLAB to add these directories to

your path. (On a unix or linux system, you may need to edit your
environment variables that set the MATLAB path.) the graphical
path tool in MATLAB to add these directories to your path. (On a
unix or linux system, you may need to edit your environment variables
that set the MATLAB path.)
A listing of the contents file from each subdirectory is presented on the

following pages.
320
.::UdecomBooks::.
APPENDIX: TOOLBOX FUNCTIONS 321
The regression function library is in a subdirectory regress.
regression function library
ar_g - Gibbs sampling Bayesian autoregressive model

bma_g - Gibbs sampling Bayesian model averaging
boxcox - Box-Cox regression with 1 parameter
boxcox2 - Box-Cox regression with 2 parameters
egarchm - EGARCH(p,q)-in-Mean regression model
emhergott - EM estimates of Hamilton’s markov switching model
garchs - garchs(1,1,1) model using non-central t-distribution
hmarkov_em - Hamilton’s markov switching model
hwhite - Halbert White’s heteroscedastic consistent estimates
lad - least-absolute deviations regression
lm_test - LM-test for two regression models
nwest - Newey-West hetero/serial consistent estimates
ols - ordinary least-squares
ols_g - Gibbs sampling Bayesian linear model
olsar1 - Maximum Likelihood for AR(1) errors ols model
olsc - Cochrane-Orcutt AR(1) errors ols model
rtrace - ridge estimates vs parameters (plot)
switch_em - switching regime regression using EM-algorithm
waldf - Wald F-test
-------- Demonstration programs ------------------
ar_gd - demonstration of Gibbs sampling ar_g

bma_gd - demonstrates Bayesian model averaging
boxcox_d - demonstrates Box-Cox 1-parameter model
boxcox2_d - demonstrates Box-Cox 2-parmaeter model
demo_all - demos most regression functions
egarchm_d - demos egarchm function
garchs_d - demos garchs function
hmarkov_emd - demos Hamilton’s model
.::UdecomBooks::.
hmarkov_emd2 - another demo of Hamilton’s model

hwhite_d - H. White’s hetero consistent estimates demo
lad_d - demos lad regression
lm_test_d - demos lm_test
mlogit_d - demonstrates multinomial logit
nwest_d - demonstrates Newey-West estimates
ols_d - demonstrates ols regression
ols_d2 - Monte Carlo demo using ols regression
ols_gd - demo of Gibbs sampling ols_g
olsar1_d - Max Like AR(1) errors model demo
olsc_d - Cochrane-Orcutt demo
olst_d - olst demo
probit_gd - demo of Gibbs sampling Bayesian probit model
ridge_d - ridge regression demo
switch_emd - demonstrates switching regression
tobit_d2 - tobit right-censoring demo
tobit_gd - demo of Gibbs sampling Bayesian tobit model
tobit_gd2 - Bayesian tobit right-censoring demo
waldf_d - demo of using wald F-test function
-------- Support functions ------------------------
ar1_like - used by olsar1 (likelihood)

bmapost - used by bma_g
box_lik - used by box_cox (likelihood)
box_lik2 - used by box_cox2 (likelihood)
chis_prb - computes chi-squared probabilities
egarchm_lik - likelihood used by egarchm
garchs_llf - likelihood used by garchs
herg_llf - likelihood used by emhergott
herg_llf2 - likelihood used by emhergott
hmarkov_llf - likelihood used by hmarkov_em
hmarkov_llf2 - likelihood used by hmarkov_em
fdis_prb - computes F-statistic probabilities
find_new - used by bma_g
hessian - used by tobit to determine numerical hessian
mcov - used by hwhite
.::UdecomBooks::.

norm_pdf - used by prt_reg, probit
plt_reg - plots regressions
prt_gibbs - prints Gibbs sampling models
prt_reg - prints regressions
prt_swm - prints switching regression results
sample - used by bma_g
tdis_prb - computes t-statistic probabilities
to_like - used by tobit (likelihood)
The utility functions are in a subdirectory util.

utility function library
accumulate - accumulates column elements of a matrix

blockdiag - creates a block diagonal matrix
cal - associates obs # with time-series calendar
ccorr1 - correlation scaling to normal column length
ccorr2 - correlation scaling to unit column length
crlag - circular lag function
diagrv - replaces main diagonal of square matrix with vector
findnear - finds matrix element nearest a scalar value
fturns - finds turning-points in a time-series
growthr - converts time-series matrix to growth rates
ical - associates time-series dates with obs #
indicator - converts a matrix to indicator variables
invccorr - inverse for ccorr1, ccorr2
invpd - makes a matrix positive-definite, then inverts
kernel_n - normal kernal density estimates
lag - generates a lagged variable vector or matrix
levels - generates factor levels variable
lprint - prints a matrix in LaTeX table-formatted form
lprintf - enhanced lprint function
matadd - adds non-conforming matrices, row or col compatible.
.::UdecomBooks::.
matdiv - divides non-conforming matrices, row or col compatible.

matmul - multiplies non-conforming matrices, row or col compatible.
matsub - divides non-conforming matrices, row or col compatible.
mlag - generates a var-type matrix of lags
mode - calculates the mode of a distribution
mprint - prints a matrix
mprint3 - prints coefficient, t-statistics matrices
mth2qtr - converts monthly to quarterly data
nclag - generates a matrix of non-contiguous lags
plt - wrapper function, plots all result structures
prt - wrapper function, prints all result strucutres
recserar - recursive AR series (like Gauss)
recsercp - recursive series product (like Gauss)
roundoff - rounds matrix to fixed number of decimal digits
sacf - sample autocorrelation function estimates
sdiff - seasonal differencing
sdummy - generates seasonal dummy variables
seqa - a sequence of numbers with a beginning and increment
shist - plots spline smoothed histogram
spacf - sample partial autocorrelation estimates
stdc - std deviations of columns returned as a column vector
tally - computes frequencies of distinct levels
tdiff - time-series differencing
trimc - trims columns of a matrix (or vector) like Gauss
trimr - trims rows of a matrix (or vector) like Gauss
tsdates - time-series dates function
tsprint - print time-series matrix
unsort - unsorts a sorted vector or matrix
vec - turns a matrix into a stacked vector
vech - matrix from lower triangular columns of a matrix
xdiagonal - spreads x(nxk) out to X(n*n x n*k) diagonal matrix
yvector - repeats y(nx1) to form Y(n*n x 1)
-------- demonstration programs -------------
cal_d - demonstrates cal function

fturns_d - demonstrates fturns and plt
ical_d - demonstrates ical function
lprint_d - demonstrates lprint function
lprintf_d - demonstrates lprintf function
mprint_d - demonstrates mprint function
mprint3_d - demonstrates mprint3 function
sacf_d - demonstrates sacf
spacf_d - demonstrates spacf
tsdate_d - demonstrates tsdate function
.::UdecomBooks::.
tsprint_d - demonstrates tsprint function

util_d - demonstrated some of the utility functions
A set of graphing functions are in a subdirectory graphs.
graphing function library
-------- graphing programs ---------------------------
tsplot - time-series graphs

pltdens - density plots
pairs - scatter plot (uses histo)
plt - plots results from all functions
-------- Demonstration programs ----------------------
tsplot_d - demonstrates tsplot

pltdens_d - demonstrates pltdens
plt_turnsd - demonstrates plt_turns
pairs_d - demonstrates pairwise scatter
plt_d - demonstrates plt on results structures
------- support functions ----------------------------
histo - used by pairs

plt_turns - used by plt to plot turning points
A library of routines in the subdirectory diagn contain the regression

diagnostics functions.
regression diagnostics library
------- regression diagnostic programs -----------
arch - ARCH(p) test

bkw - BKW collinearity diagnostics
bpagan - Breusch-Pagan heteroscedasticity test
cusums - Brown,Durbin,Evans cusum squares test
dfbeta - BKW influential observation diagnostics
diagnose - compute diagnostic statistics
plt_cus - plots cusums
recresid - compute recursive residuals
rdiag - graphical residuals diagnostics
studentize - standarization transformation
unstudentize - reverses studentize transformation
qstat2 - Box-Ljung Q-statistic
.::UdecomBooks::.
------- demonstration programs --------------------
arch_d - demonstrates arch

bkw_d - demonstrates bkw
bpagan_d - demonstrates bpagan
cusums_d - demonstrates cusums
dfbeta_d - demonstrates dfbeta, plt_dfb, plt_dff
diagnose_d - demonstrates diagnose
recresid_d - demonstrates recresid
rdiag_d - demonstrates rdiag
unstudentize_d - demonstrates studentize, unstudentize
qstat2_d - demonstrates qstat2
------- support functions -------------------------
plt_cus - plots cusums test results

../util/plt - plots everything
../regress/ols.m - least-squares regression
The vector autoregressive library is in a subdirectory var bvar.
vector autoregressive function library
------- VAR/BVAR functions -------------------

becmf - Bayesian ECM model forecasts
bvar - BVAR model
bvarf - BVAR model forecasts
ecm - ECM (error correction) model estimates
ecmf - ECM model forecasts
irf - impulse response functions
lrratio - likelihood ratio tests for lag length
recm - ecm version of rvar
recmf - random-walk averaging ECM forecasts
rvar - Bayesian random-walk averaging prior model
rvarf - Bayesian RVAR model forecasts
var - VAR model
varf - VAR model forecasts
.::UdecomBooks::.
becm_d - BECM model demonstration

becm_g - Gibbs sampling BECM estimates demo
becmf_d - becmf demonstration
bvar_d - BVAR model demonstration
bvarf_d - bvarf demonstration
ecm_d - ECM model demonstration
ecmf_d - ecmf demonstration
irf_d - impulse response function demo
irf_d2 - another irf demo
lrratio_d - demonstrates lrratio
pftest_d - demo of pftest function
recm_d - RECM model demonstration
recmf_d - recmf demonstration
rvar_d - RVAR model demonstration
rvar_gd - Gibbs sampling rvar model demo
rvarf_d - rvarf demonstration
var_d - VAR model demonstration
varf_d - varf demonstration
------- support functions ------------------
johansen - used by ecm,ecmf,becm,becmf,recm,recmf

lag - does ordinary lags
mlag - does var-type lags
nclag - does contiguous lags (used by rvar,rvarf,recm,recmf)
ols - used for VAR estimation
pftest - prints Granger F-tests
pgranger - prints Granger causality probabilities
prt - prints results from all functions
prt_coint - used by prt_var for ecm,becm,recm
prt_var - prints results of all var/bvar models
rvarb - used for RVARF forecasts
scstd - does univariate AR for BVAR
theilbf - used for BVAR forecasts
theilbv - used for BVAR estimation
trimr - used by VARF,BVARF, johansen (in /util/trimr.m)
vare - used by lrratio (vare uses /regress/olse.m)
.::UdecomBooks::.
The co-integration library functions are in a subdirectory coint.
co-integration library
------ Co-integration testing routines --------
johansen - carries out Johansen’s co-integration tests

adf - carries out Augmented Dickey-Fuller unit root tests
cadf - carries out ADF tests for co-integration
phillips - carries out Phillips-Peron co-integration tests
prt_coint - prints results from adf,cadf,johansen
------ Demonstration programs -----------------
johansen_d - demonstrates johansen

adf_d - demonstrates adf
cadf_d - demonstrates cadf
phillips_d - demonstrates phillips
------ Support functions -----------------------
c_sja - returns critical values for SJ maximal eigenvalue test

c_sjt - returns critical values for SJ trace test
ztcrit - returns critical values for adf test
rztcrit - returns critical values for cadf test
detrend - used by johansen to detrend data series
ptrend - used by adf to create time polynomials
trimr - /util/trimr.m (like Gauss trimr)
cols - /util/cols.m (like Gauss cols)
rows - /util/rows.m (like Gauss rows)
tdiff - /util/tdiff.m differences
The Gibbs convergence diagnostic functions are in a subdirectory gibbs.
Gibbs sampling convergence diagnostics functions
--------- convergence testing functions ---------
apm - Geweke’s chi-squared test

coda - convergence diagnostics
momentg - Geweke’s NSE, RNE
raftery - Raftery and Lewis program Gibbsit for convergence
--------- demonstration programs ----------------
apm_d - demonstrates apm

coda_d - demonstrates coda
momentg_d - demonstrates momentg
raftery_d - demonstrates raftery
.::UdecomBooks::.
--------- support functions ---------------------
prt_coda - prints coda, raftery, momentg, apm output (use prt)

empquant - These were converted from:
indtest - Rafferty and Lewis FORTRAN program.
mcest - These function names follow the FORTRAN subroutines
mctest -
ppnd -
thin -
Distribution functions are in the subdirectory distrib.

------- pdf, cdf, inverse functions -----------
beta_cdf - beta(a,b) cdf

beta_inv - beta inverse (quantile)
beta_pdf - beta(a,b) pdf
bino_cdf - binomial(n,p) cdf
bino_inv - binomial inverse (quantile)
bino_pdf - binomial pdf
chis_cdf - chisquared(a,b) cdf
chis_inv - chi-inverse (quantile)
chis_pdf - chisquared(a,b) pdf
chis_prb - probability for chi-squared statistics
fdis_cdf - F(a,b) cdf
fdis_inv - F inverse (quantile)
fdis_pdf - F(a,b) pdf
fdis_prb - probabililty for F-statistics
gamm_cdf - gamma(a,b) cdf
gamm_inv - gamma inverse (quantile)
gamm_pdf - gamma(a,b) pdf
hypg_cdf - hypergeometric cdf
hypg_inv - hypergeometric inverse
hypg_pdf - hypergeometric pdf
logn_cdf - lognormal(m,v) cdf
logn_inv - lognormal inverse (quantile)
logn_pdf - lognormal(m,v) pdf
logt_cdf - logistic cdf
logt_inv - logistic inverse (quantile)
logt_pdf - logistic pdf
norm_cdf - normal(mean,var) cdf
norm_inv - normal inverse (quantile)
norm_pdf - normal(mean,var) pdf
pois_cdf - poisson cdf
pois_inv - poisson inverse
pois_pdf - poisson pdf
.::UdecomBooks::.
stdn_cdf - std normal cdf

stdn_inv - std normal inverse
stdn_pdf - std normal pdf
tdis_cdf - student t-distribution cdf
tdis_inv - student t inverse (quantile)
tdis_pdf - student t-distribution pdf
tdis_prb - probabililty for t-statistics
------- random samples -----------------------
beta_rnd - random beta(a,b) draws

bino_rnd - random binomial draws
chis_rnd - random chi-squared(n) draws
fdis_rnd - random F(a,b) draws
gamm_rnd - random gamma(a,b) draws
hypg_rnd - random hypergeometric draws
logn_rnd - random log-normal draws
logt_rnd - random logistic draws
nmlt_rnd - left-truncated normal draw
nmrt_rnd - right-truncated normal draw
norm_crnd - contaminated normal random draws
norm_rnd - multivariate normal draws
pois_rnd - poisson random draws
tdis_rnd - random student t-distribution draws
unif_rnd - random uniform draws (lr,rt) interval
wish_rnd - random Wishart draws
-------- demonstration and test programs ------
beta_d - demo of beta distribution functions

bino_d - demo of binomial distribution functions
chis_d - demo of chi-squared distribution functions
fdis_d - demo of F-distribution functions
gamm_d - demo of gamma distribution functions
hypg_d - demo of hypergeometric distribution functions
logn_d - demo of lognormal distribution functions
logt_d - demo of logistic distribution functions
pois_d - demo of poisson distribution functions
stdn_d - demo of std normal distribution functions
tdis_d - demo of student-t distribution functions
trunc_d - demo of truncated normal distribution function
unif_d - demo of uniform random distribution function
-------- support functions ---------------------
betacfj - used by fdis_prb

betai - used by fdis_prb
bincoef - binomial coefficients
com_size - test and converts to common size
.::UdecomBooks::.
gammalnj - used by fdis_prb

is_scalar - test for scalar argument
Optimization functions are in the subdirectory optimize.
Optimization functions library
--------------- optimization functions -----------------
dfp_min - Davidson-Fletcher-Powell
frpr_min - Fletcher-Reeves-Polak-Ribiere
maxlik - general all-purpose optimization routine
pow_min - Powell conjugate gradient
solvopt - yet another general purpose optimization routine
optim1_d - dfp, frpr, pow, maxlik demo

optim2_d - solvopt demo
optim3_d - fmins demo
--------------- support functions -----------------------
apprgrdn - computes gradient for solvopt

box_like1 - used by optim3_d
gradt - computes gradient
hessian - evaluates hessian
linmin - line minimization routine (used by dfp, frpr, pow)
stepsize - stepsize determination
tol_like1 - used by optim1_d, optim2_d
updateh - updates hessian
A library of spatial econometrics functions are in the subdirectory spa-

tial.
------- spatial econometrics functions -----------

.::UdecomBooks::.

sarp_g - Gibbs sampling Bayesian sar Probit model
sart_g - Gibbs sampling Bayesian sar Tobit model
sdm_g - Gibbs sampling Bayesian spatial Durbin model

sarp_gd - sar Probit Gibbs sampling demo
sart_gd - sar Tobit model Gibbs sampling demo
.::UdecomBooks::.

anselin.dat- Anselin (1988) Columbus crime data

.::UdecomBooks::.

Applied Econometrics Using Matlab

Uploaded by

Copyright:

Available Formats

Applied Econometrics Using Matlab

Uploaded by

Document Information

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

Applied Econometrics Using Matlab

Uploaded by

Copyright:

Available Formats

Applied Econometrics using MATLAB

This text describes a set of MATLAB functions that implement a host of

models that robustify against outliers and accommodate heteroscedastic dis-

cipher the information and provide a printed or graphical presentation of

Much of the econometric code I encounter on the internet is simply too

First I would like to acknowledge my students who suffered through previous

2 Regression using MATLAB 5

Chapter 4 Appendix 101

5 VAR and Error Correction Models 103

Chapter 5 Appendix 148

6 Markov Chain Monte Carlo Models 151

Chapter 6 Appendix 199

7 Limited Dependent Variable Models 204

Chapter 7 Appendix 228

8 Simultaneous Equation Models 230

8.4 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . 244

Chapter 8 Appendix 246

9 Distribution functions library 247

Chapter 9 Appendix 257

10 Optimization functions library 260

Chapter 10 Appendix 288

11 Handling sparse matrices 289

Chapter 11 Appendix 310

Appendix: Toolbox functions 320

1.1 Demonstrate regression using the ols() function . . . . . . . . 8

4.7 Using the robust() function . . . . . . . . . . . . . . . . . . . 97

10.1 Simplex maximum likelihood for Box-Cox model . . . . . . . 265

2.1 Histogram of β̂ outcomes . . . . . . . . . . . . . . . . . . . . 33

3.1 Output from tsplot() function . . . . . . . . . . . . . . . . . . 62

4.1 Ridge trace plot . . . . . . . . . . . . . . . . . . . . . . . . . . 91

5.1 Impulse response graphs . . . . . . . . . . . . . . . . . . . . . 113

6.1 Prior Vi distributions for various values of r . . . . . . . . . . 171

7.1 Cumulative distribution functions compared . . . . . . . . . . 206

9.1 Beta distribution demonstration program plots . . . . . . . . 250

10.1 Plots from switching regime regression . . . . . . . . . . . . . 277

11.1 Sparsity structure of W from Pace and Berry . . . . . . . . . 291

11.3 Minimum degree ordering versus unordered Pace and Barry

2.1 Timing (in seconds) for Cholesky and QR Least-squares . . . 13

4.1 Variance-decomposition proportions table . . . . . . . . . . . 84

6.1 A Comparison of FAR Estimators . . . . . . . . . . . . . . . 190

The Econometrics Toolbox contains around 50 functions that implement

of passing information to functions and returning results. These are used

Simultaneous equation systems are the subject of Chapter 8, where we

estimation and testing functions and most of the utility functions.

Regression using MATLAB

This chapter describes the design and implementation of a regression func-

2.1 Design of the regression library

2.2 The ols function

All functions in the econometrics toolbox present a unified documenta-

PURPOSE: computes Hoerl-Kennard Ridge Regression

An alternative solution based on the Cholesky decomposition is faster,

Given either of these solutions, we are in a position to use our estimates

2.3 Selecting a least-squares algorithm

% ----- Example 2.2 Least-squares timing information

Using this type of program, we explored the timing differences between

Table 2.1: Timing (in seconds) for Cholesky and QR Least-squares

We can quantify the conditioning of a problem by calculating a condition

increasingly ill-conditioned least-squares problem.

accj = −log10 (|1.0 − β̂ j |) (2.2)

Table 2.2: Digits of accuracy for Cholesky vs. QR decomposition