Multinomial Logit,

Discrete Choice Modeling

An Introduction to Designing Choice Experiments,

and Collecting, Processing, and Analyzing Choice Data
with SAS R

Warren F. Kuhfeld
SAS

August 31, 2002

70 TS-677E Multinomial Logit, Discrete Choice Modeling

Contents

Introduction 74

Preliminaries 76
Experimental Design Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Efficiency of an Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Efficiency of a Choice Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Customizing the Multinomial Logit Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Orthogonal Coding, Efficiency, Balance, and Orthogonality . . . . . . . . . . . . . . . . . . . . . . . 80

Candy Example 83
The Multinomial Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
The Input Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Fitting the Multinomial Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Multinomial Logit Model Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Fitting the Multinomial Logit Model, All Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Probability of Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Fabric Softener Example 94

Set Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Designing the Choice Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Examining the Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Randomizing the Design, Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Generating the Questionnaire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Entering the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Processing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Binary Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Fitting the Multinomial Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Multinomial Logit Model Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Probability of Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Custom Questionnaires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Processing the Data for Custom Questionnaires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

Vacation Example 116

Set Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Designing the Choice Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Contents 71

The %MktEx Macro Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

Examining the Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Generating the Questionnaire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Entering and Processing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Binary Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Quantitative Price Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Quadratic Price Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Effects Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Alternative-Specific Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

Vacation Example, with Alternative-Specific Attributes 152

Choosing the Number of Choice Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Designing the Choice Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Ensuring that Certain Key Interactions are Estimable . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Examining the Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Blocking an Existing Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Generating the Questionnaire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Generating Artificial Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Reading, Processing, and Analyzing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Aggregating the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

Brand Choice Example with Aggregate Data 173

Processing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Simple Price Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Alternative-Specific Price Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Mother Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Aggregating the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Choice and Breslow Likelihood Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

Food Product Example with Asymmetry and Availability Cross Effects 192
The Multinomial Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Set Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Designing the Choice Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
When You Have a Long Time to Search for an Efficient Design . . . . . . . . . . . . . . . . . . . . . 198
Examining the Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
Designing the Choice Experiment, More Choice Sets . . . . . . . . . . . . . . . . . . . . . . . . . . 202
72 TS-677E Multinomial Logit, Discrete Choice Modeling

Examining the Subdesigns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

Examining the Aliasing Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Blocking the Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
The Final Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Testing the Design Before Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Generating Artificial Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Processing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Cross Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Multinomial Logit Model Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Modeling Subject Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

Allocation of Prescription Drugs 237

Designing the Allocation Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Processing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
Coding and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Multinomial Logit Model Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Analyzing Proportions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249

Chair Design with Generic Attributes 252

Generic Attributes, Alternative Swapping, Large Candidate Set . . . . . . . . . . . . . . . . . . . . . 253
Generic Attributes, Alternative Swapping, Small Candidate Set . . . . . . . . . . . . . . . . . . . . . 258
Generic Attributes, a Constant Alternative, and Alternative Swapping . . . . . . . . . . . . . . . . . . 261
Generic Attributes, a Constant Alternative, and Choice Set Swapping . . . . . . . . . . . . . . . . . . 264
Design Algorithm Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

Initial Designs 268

Improving an Existing Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
When Some Choice Sets are Fixed in Advance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

Partial Profiles and Restrictions 274

Pair-wise Partial Profile Choice Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Linear Partial Profile Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
Choice from Triples; Partial Profiles Constructed Using Restrictions . . . . . . . . . . . . . . . . . . 280
Advanced Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285

The Macros 287

%ChoicEff Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
Contents 73

%MktAllo Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303

%MktBal Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
%MktBlock Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
%MktDes Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
%MktDups Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
%MktEval Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
%MktEx Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
%MktKey Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
%MktLab Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
%MktMerge Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
%MktOrth Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
%MktRoll Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
%MktRuns Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
%PhChoice Macro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364

Concluding Remarks 369

References 370

Multinomial Logit Models (SUGI Paper) 372

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Modeling Discrete Choice Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Fitting Discrete Choice Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
Cross-Alternative Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Final Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385

Index 386
74 TS-677E Multinomial Logit, Discrete Choice Modeling

Multinomial Logit, Discrete Choice Modeling

This report shows you how to use the multinomial logit model (Manski and McFadden, 1981; Louviere and
Woodworth, 1983) to investigate consumer’s stated choices. The multinomial logit model is an alternative to
full-profile conjoint analysis and is extremely popular in marketing research (Louviere, 1991; Carson et. al.,
1994). We will discuss designing a choice experiment, preparing the questionnaire, inputting and processing the
data, performing the analysis, and interpreting the results. Discrete choice, using the multinomial logit model,
is sometimes referred to as “choice-based conjoint.” However, discrete choice uses a different model from full-
profile conjoint analysis. Discrete choice applies a nonlinear model to aggregate choice data, whereas full-profile
conjoint analysis applies a linear model to individual-level rating or ranking data.
This report is TS-677E, the August 31, 2002 edition for SAS Version 9.0. It is a revision of the April 1, 2001
report for Version 8.2 and other previous editions. This edition uses SAS macros and features that are new in
Version 9, including the new experimental design macro, %MktEx . This report heavily relies on a number of
macros and procedures.

 We use the %MktRuns autocall macro to suggest design sizes. See page 360 for documentation. The
%MktRuns macro has been revised since the 2001 book.
 We use the %MktEx autocall macro to generate most of our experimental designs. It is easier to use and
usually produces better results than the methods suggested in earlier reports. See page 327 for documen-
tation. The %MktEx macro is new with the 2002 book.
 We use the %MktEval autocall macro to evaluate our designs. See page 325 for documentation. The
%MktEval macro has been revised since the 2001 book.
 We use the %ChoicEff autocall macro to generate certain specialized choice designs. See page 288 for
documentation.
 We use the autocall macros %MktRoll, %MktMerge, and %MktAllo to prepare the data and design for
analysis. See pages 356, 353, and 303 for documentation.
 We use PROC TRANSREG to do all of our design coding.
 We use the %PhChoice autocall macro to customize our printed output. This macro uses PROC TEM-
PLATE and ODS (Output Delivery System) to customize the output from PROC PHREG, which fits the
multinomial logit model. See page 364 for documentation.
 The %MktBal macro can be used to make perfectly balanced designs. See page 305 for documentation.
The %MktBal macro is new with the 2002 book.
 The %MktBlock macro can be used to block a linear or choice design. See page 307 for documentation.
The %MktBlock macro is new with the 2002 book.
 The %MktDes experimental design macro, which was heavily used in previous reports, is called by the
%MktEx macro, and it can still be called directly. See page 307 for documentation.
 The %MktDups macro can be used to search for duplicate runs or choice sets. See page 319 for documen-
tation. The %MktDups macro is new with the 2002 book.
 The %MktLab macro can be used to assign different variable names, labels and levels to experimental
designs and to add an intercept. See page 345 for documentation. The %MktLab macro is new with the
2002 book.
 The %MktOrth macro can be used to list orthogonal experimental designs that the %MktEx macro can
produce. See page 354 for documentation. The %MktOrth macro is new with the 2002 book.
Introduction 75

All of these macros are distributed with Version 9.0 of SAS as autocall macros (see page 287 for more information
on autocall macros). Note however, that Version 9.0 was finished before the macros were finalized and this book
finished. Hence there are a few differences between the macros used in this book and those shipped with Version
9.0 of SAS. If you are running version 9.0 or any earlier version of SAS, get the latest macros from the web or by
writing [Link]@[Link]. This report and the macros are available from the Technical Support web site
at [Link] [Link] . This information is provided by SAS as a service
to its users. It is provided “as is.” There are no warranties, expressed or implied, as to merchantability or fitness
for a particular purpose regarding the accuracy of the materials or code contained herein.
Several examples are discussed.

 The candy example is a first, very simple example that discusses the multinomial logit model, the input
data, analysis, results, and computing probability of choice.
 The fabric softener example is a small, somewhat more realistic example that discusses designing the
choice experiment, randomization, generating the questionnaire, entering and processing the data, analysis,
results, probability of choice, and custom questionnaires.
 The first vacation example is a larger, symmetric example that discusses designing the choice experi-
ment, blocks, randomization, generating the questionnaire, entering and processing the data, coding, and
alternative-specific effects.
 The second vacation example is a larger, asymmetric example that discusses designing the choice experi-
ment, blocks, blocking an existing design, interactions, generating the questionnaire, generating artificial
data, reading, processing, and analyzing the data, aggregating the data to save time and memory.
 The brand choice example is a small example that discusses the processing of aggregate data, the mother
logit model, and the likelihood function.
 The food product example is a medium sized example that discusses asymmetry, coding, checking the
design to ensure that all effects are estimable, availability cross effects, interactions, overnight design
searches, modeling subject attributes, and designs when balance is of primary importance.
 The drug allocation example is a small example that discusses data processing for studies where respon-
dents potentially make multiple choices.
 The chair example is a purely generic-attributes study, and it uses the %ChoicEff macro to create exper-
imental designs.
 The last example sections contains miscellaneous examples including improving an existing design, aug-
menting a design with some choice sets are fixed in advance, and partial profiles.

This document would not be possible without the help of Randy Tobias who contributed to the discussion of
experimental design and Ying So who contributed to the discussion of analysis. Randy Tobias wrote PROC
FACTEX and PROC OPTEX. Ying So wrote PROC PHREG. Warren F. Kuhfeld wrote PROC TRANSREG and
the macros.
 All of the sample data sets are artificially generated.
76 TS-677E Multinomial Logit, Discrete Choice Modeling

Preliminaries
This section defines some design terms that we will use later and shows how to customize the multinomial logit
output listing. Impatient readers may skip ahead to the candy example on page 83 and refer back to this section
as needed.

Experimental Design Terminology

An experimental design is a plan for running an experiment. The factors of an experimental design are the
columns or variables that have two or more fixed values, or levels. The rows of a design are called runs and cor-
respond to product profiles in a full-profile conjoint study or choice sets in a discrete choice study. Experiments
are performed to study the effects of the factor levels on the dependent variable. In a discrete-choice study, the
factors are the attributes of the hypothetical products or services, and the response is choice. For example, the
following table contains an experimental design in 8 runs with three factors, Brand 1 price, Brand 2 price, and
Brand 3 price. Each factor has two levels, $1.99 and $2.99.
Linear Design
For a Choice Model
Brand 1 Brand 2 Brand3
Price Price Price
1.99 1.99 1.99
1.99 1.99 2.99
1.99 2.99 1.99
1.99 2.99 2.99
2.99 1.99 1.99
2.99 1.99 2.99
2.99 2.99 1.99
2.99 2.99 2.99
This is an example of a full-factorial design. It consists of all possible combinations of the levels of the factors.
Factorial designs allow you to estimate main effects and interactions. A main effect is a simple effect, such as a
price or brand effect. In a main-effects model, for example, the brand effect is the same at the different prices and
the price effect is the same for the different brands. Interactions involve two or more factors, such as a brand by
price interaction. In a model with interactions, for example, brand preference is different at the different prices
and the price effect is different for the different brands.
In a full-factorial design, all main effects, all two-way interactions, and all higher-order interactions are estimable
and uncorrelated. The problem with a full-factorial design is that, for most practical situations, it is too cost-
prohibitive and tedious to have subjects consider all possible combinations. For example, with five factors, two
at four levels and three at five levels (denoted 42 53 ), there are 4  4  5  5  5 = 2000 combinations in the
full-factorial design. For this reason, researchers often use fractional-factorial designs, which have fewer runs
than full-factorial designs. The price of having fewer runs is that some effects become confounded. Two effects
are confounded or aliased when they are not distinguishable from each other.
A special type of fractional-factorial design is the orthogonal array. An orthogonal array or orthogonal design
is one in which all estimable effects are uncorrelated. Orthogonal arrays are categorized by their resolution. The
resolution identifies which effects, possibly including interactions, are estimable. For example, for resolution
III designs, all main effects are estimable free of each other, but some of them are confounded with two-factor
interactions. For resolution V designs, all main effects and two-factor interactions are estimable free of each
other. More generally, if resolution (r) is odd, then effects of order e = (r 1)=2 or less are estimable free of
each other. However, at least some of the effects of order e are confounded with interactions of order e + 1. If
r is even, then effects of order e = (r 2)=2 are estimable free of each other and are also free of interactions
of order e + 1. Higher resolutions require larger designs. Orthogonal arrays come in specific numbers of runs
(such as 16, 18, 20, 24, 27, 28, ...) for specific numbers of factors with specific numbers of levels. Resolution III
orthogonal arrays are frequently used in marketing research.
Preliminaries 77

The term “orthogonal array,” as it is sometimes used in practice, is imprecise. It is correctly used to refer to
designs that are both orthogonal and balanced, and hence optimal. The term is sometimes also used to refer to
designs that are orthogonal but not balanced, and hence not 100% efficient and sometimes not even optimal. A
design is balanced when each level occurs equally often within each factor, which that means the intercept is
orthogonal to each effect. Imbalance is a generalized form of nonorthogonality, hence it increases the variances
of the parameter estimates and decreases efficiency.

Efficiency of an Experimental Design

The goodness or efficiency of an experimental design can be quantified. Common measures of the efficiency of
an (ND  p) design matrix X are based on the information matrix X0 X. The variance-covariance matrix of the
vector of parameter estimates ^ in a least-squares analysis is proportional to (X0 X) 1 . An efficient design will
have a “small” variance matrix, and the eigenvalues of (X0 X) 1 provide measures of its “size.” The two most
prominent efficiency measures are based on quantifying the idea of matrix size by averaging (in some sense)
the eigenvalues or variances. A-efficiency is a function of the arithmetic mean of the eigenvalues, which is also
the arithmetic mean of the variances and is given by trace ((X0 X) 1 )=p. (The trace is the sum of the diagonal
elements of a matrix, which is the sum of the eigenvalues.) D-efficiency is a function of the geometric mean of the
eigenvalues, which is given by j(X0 X) 1 j1=p . (The determinant, j(X0 X) 1 j, is the product of the eigenvalues of
(X0 X) 1 .) A third common efficiency measure, G-efficiency, is based on M , the maximum standard error for
prediction over the candidate set. All three of these criteria are convex functions of the eigenvalues of (X0 X) 1
and hence are usually highly correlated.
For all three criteria, if a balanced and orthogonal design exists, then it has optimum efficiency; conversely,
the more efficient a design is, the more it tends toward balance and orthogonality. A design is balanced and
orthogonal when (X0 X) 1 is diagonal, N1D I, for a suitably coded X. A design is orthogonal when the submatrix
of (X0 X) 1 , excluding the row and column for the intercept, is diagonal; there may be off-diagonal nonzeros
for the intercept. A design is balanced when all off-diagonal elements in the intercept row and column are zero.
These measures of efficiency can be scaled to range from 0 to 100 (see page 80 for the orthogonal coding of X
that must be used with these formulas):

= 100  N 1
trace ((X0 X) 1 )=p
A-efficiency
D

D-efficiency = 100  N j(X01X) 1 j1 =p

D

p
G-efficiency = 100  p=N

D

These efficiencies measure the goodness of the design relative to hypothetical orthogonal designs that may not
exist, so they are not useful as absolute measures of design efficiency. Instead, they should be used relatively,
to compare one design to another for the same situation. Efficiencies that are not near 100 may be perfectly
satisfactory. Throughout this report, we will use the %MktEx macro to find good, efficient experimental designs.

Efficiency of a Choice Design

All of the theory in the preceding section concerned linear models. In linear models, the parameter estimates ^
have variances proportional to (X0 X) 1 . In contrast, the variances of the parameter estimates in the multinomial
logit model are given by
  " " ##
@ 2 `( ) 1
 exp(x0 )x x0 ( =1 exp(x0 )x )( =1 exp(x0 )x )0 1
V ( ^) =
m m m

@ 2
=  =1 N =1 exp(x0 )
n
k
j
m
j j j j j j

( =1 exp(x0 ))2
m
j j j

=1 j j j j

where
78 TS-677E Multinomial Logit, Discrete Choice Modeling

exp(( =1 f x0 ) )
m

`( ) =  n j j j
=1
k
( =1 exp(x0 ))
m
j j
N

m– brands
n – choice sets
N– people

We will often create experimental designs for choice models using efficiency criteria for linear models. Consider
an extremely simple example of three brands and two prices. We could use linear model theory to create a design
for a full-profile conjoint study. The full-profile conjoint design has two factors, one for brand and one for price.
For the same brands and prices, we could instead use linear model theory to create a linear design from which
we will construct a choice design to use in a discrete choice study. The linear design for a pricing study with
three brands has three factors (Brand 1 Price, Brand 2 Price, and Brand 3 Price) and one row for each choice set.
More generally, the linear design has one factor for each attribute of each alternative (or brand), and brand is not
a factor in the linear design. Each brand is a “bin” into which its factors are collected.

Full-Profile
Linear Design
Conjoint Design
Used to Make a Choice Design
Brand Price
Brand 1 Brand 2 Brand3
1 1.99
Price Price Price
1 2.99
1.99 1.99 1.99
2 1.99
1.99 2.99 2.99
2 2.99
2.99 1.99 2.99
3 1.99
2.99 2.99 1.99
3 2.99

Before we fit the choice model, we will construct a choice design from the linear design and code the choice
design. See the three tables below.
Linear Design Choice Design Choice Design Coding
Brand 1 Brand 2 Brand 3
1 2 3 Brand Price Brand 1 Brand 2 Brand 3 Price Price Price
1.99 1.99 1.99 1 1.99 1 0 0 1.99 0 0
2 1.99 0 1 0 0 1.99 0
3 1.99 0 0 1 0 0 1.99
1.99 2.99 2.99 1 1.99 1 0 0 1.99 0 0
2 2.99 0 1 0 0 2.99 0
3 2.99 0 0 1 0 0 2.99
2.99 1.99 2.99 1 2.99 1 0 0 2.99 0 0
2 1.99 0 1 0 0 1.99 0
3 2.99 0 0 1 0 0 2.99
2.99 2.99 1.99 1 2.99 1 0 0 2.99 0 0
2 2.99 0 1 0 0 2.99 0
3 1.99 0 0 1 0 0 1.99
The linear design has one row per choice set. The choice design has three rows for each choice set. The linear
design and the choice design contain different arrangements of the exact same information. In the linear design,
brand is a bin into which its factors are collected (in this case one factor per brand). In the choice design, brand
and price are both factors, because the design has been rearranged from one row per choice set to one row per
alternative per choice set. For this problem, with only one attribute per brand, the first row of the choice design
matrix corresponds to the first value in the linear design matrix, Brand 1 at $1.99. The second row of the choice
design matrix corresponds to the second value in the linear design matrix, Brand 2 at $1.99. The third row of the
choice design matrix corresponds to the third value in the linear design matrix, Brand 3 at $1.99, and so on.
Preliminaries 79

We will go through how to construct linear and choice designs many times in the examples. For now, just notice
that the conjoint design is different from the linear design, which is different from the choice design. They aren’t
even the same size! Also note that we cannot use linear efficiency criteria to directly construct the choice design
bypassing the linear design step.
We make a good design for a linear model by picking x’s that minimize functions of (X0 X) 1. In the choice
model, ideally we would like to minimize functions of

" " ##
 =1 exp(x0 )x x0 ( =1 exp(x0 )x )( =1 exp(x0 )x )0 1
V ( ^) =  =1 N
m m m
n j j j j j j j j j j
k
 =1 exp(x0 )
m
j j
( =1 exp(x0 ))2
m
j j

We cannot do this unless we know , and if we knew , we would not need to do the experiment. (However, in
the chair example on pages 252 267, we will see how to make an efficient choice design when we are willing to
make assumptions about .)
Certain assumptions must be made before applying ordinary general-linear-model theory to problems in mar-
keting research. The usual goal in linear modeling is to estimate parameters and test hypotheses about those
parameters. Typically, independence and normality are assumed. In full-profile conjoint analysis, each subject
rates all products and separate ordinary-least-squares analyses are run for each subject. This is not a standard
general linear model; in particular, observations are not independent and normality cannot be assumed. Discrete
choice models, which are nonlinear, are even more removed from the general linear model.
Marketing researchers have always made the critical assumption that designs that are good for general linear
models are also good designs for conjoint analysis and discrete choice models. We also make this assumption.
We will assume that an efficient design for a linear model is a good design for the multinomial logit model
used in discrete choice studies. We assume that if we create the linear design (one row per choice set and all
of the attributes of all of the alternatives comprise that row), and if we strive for linear-model efficiency (near
balance and orthogonality), then we will have a good design for measuring the utility of each alternative and the
contributions of the factors to that utility. When we construct choice designs in this way, our designs will have
two nice properties. 1) Each attribute level will occur equally often (or at least nearly equally often) for each
attribute of each alternative across all choice sets. 2) Each attribute will be independent of every other attribute
(or at least nearly independent), both those in the current alternative and those in all of the other alternatives. The
design techniques discussed in this book that are based on the assumption that linear design efficiency is a good
surrogate for choice design goodness have been used quite successfully in the field for many years.
In most of the examples, we will use the %MktEx macro to create a good linear design, from which we will
construct our choice design. This seems to be a good safe strategy. It is safe in the sense that you have enough
choice sets and collect enough information so that very complex models, including models with alternative-
specific effects, availability effects, and cross effects, can be fit. However, it is good to remember that when you
run the %MktEx macro and you get an efficiency value, it corresponds to the linear design, not the choice design.
It is a surrogate for the criterion of interest, the efficiency of the choice design, which is unknowable unless you
know the parameters.

Customizing the Multinomial Logit Output

The multinomial logit model for discrete choice experiments is fit using the SAS/STAT R procedure PHREG
(proportional hazards regression), with the ties=breslow option. The likelihood function of the multinomial
logit model has the same form as a survival analysis model fit by PROC PHREG. The output from PROC PHREG
is primarily designed for survival analysis studies. Before we fit the multinomial logit model with PROC PHREG,
we can customize the output to make it more appropriate for choice experiments. We will use the autocall macro
%PhChoice macro. See page 287 for information on autocall macros. You can run the following macro to
customize PROC PHREG output.
%phchoice(on)
80 TS-677E Multinomial Logit, Discrete Choice Modeling

The macro uses PROC TEMPLATE and ODS (Output Delivery System) to customize the output from PROC
PHREG. Running this code edits the templates and stores copies in sasuser. These changes will remain in
effect until you delete them, so typically, you only have to run this macro once. Note that these changes assume
that each effect in the choice model has a variable label associated with it, so there is no need to print variable
names. If you are coding with PROC TRANSREG, this will usually be the case. To return to the default output
from PROC PHREG, run the following macro.
%phchoice(off)

See page 364 for more information on the %PhChoice macro.

Orthogonal Coding, Efficiency, Balance, and Orthogonality

We mentioned on page 77 that we use a special orthogonal coding of X when computing design efficiency. This
section shows that coding. All but the most dedicated readers may skip ahead to the candy example on page 83.
Recall that our measures of design efficiency are scaled to range from 0 to 100.

= 100  N 1
trace ((X0 X) )=p
A-efficiency 1
D

D-efficiency = 100  N j(X01X) 1 j1 =p

D

When computing D-efficiency or A-efficiency, we code X so that when the design is orthogonal and balanced.
X0 X = ND I where I is a p  p identity matrix. When our design is orthogonal and balanced, (X0 X) 1 = N1 I,
D
and trace ((X0 X) 1 )=p = j(X0 X) 1 j1=p = 1=ND . In this case, the two denominator terms cancel and efficiency
is 100%. As the average variance increases, efficiency decreases.
Here are the orthogonal codes for two-level through five-level factors.
Two-Level Three-Level Four-Level Five-Level
a 1.00 a 1.22 -0.71 a 1.41 -0.82 -0.58 a 1.58 -0.91 -0.65 -0.50
b -1.00 b 0 1.41 b 0 1.63 -0.58 b 0 1.83 -0.65 -0.50
c -1.22 -0.71 c 0 0 1.73 c 0 0 1.94 -0.50
d -1.41 -0.82 -0.58 d 0 0 0 2.00
e -1.58 -0.91 -0.65 -0.50
Notice that the sum of squares for the coding of the two-level factor is 2; for all of the columns of the three-level
factor, the sums of squares are 3; for the four-level factor, the sums of squares are all 4; and for the five-level
factor, the sums of squares are all 5. Also notice that each column within a factor is orthogonal to all of the
other columns the sum of cross products is zero. For example, in the last two columns of the five-level factor,
0:65  0:5 + 0:65  0:5 + 1:94  :05 + 0  2 + 0:65  0:5 = 0. Finally notice that the codings for
each level form a contrast the ith level versus all of the preceding levels and the last level.
This example shows the coding of a 2  6 full-factorial design in 12 runs using a coding function that requires
that the factors levels are consecutive positive integers beginning with one and ending with m for an m-level
factor. Note that the IML operator # performs ordinary (scalar) multiplication, and ## performs exponentiation.
Preliminaries 81

proc iml; /* orthogonal coding, levels must be 1, 2, ..., m */

reset fuzz;

start orthogcode(x);
levels = max(x);
xstar = shape(x, levels - 1, nrow(x))‘;
j = shape(1 : (levels - 1), nrow(x), levels - 1);
r = sqrt(levels # (x / (x + 1))) # (j = xstar) -
sqrt(levels / (j # (j + 1))) # (j > xstar | xstar = levels);
return(r);
finish;
design = (1:2)‘ @ j(6, 1, 1) || {1, 1} @ (1:6)‘;
x = j(12, 1, 1) || orthogcode(design[,1]) || orthogcode(design[,2]);
print design[format=1.] ’ ’ x[format=5.2 colname={’Int’ ’Two’ ’Six’}];

xpx = x‘ * x; print xpx[format=best5.];

inv = inv(xpx); print inv[format=best5.];
d_eff = 100 / (nrow(x) # det(inv) ## (1 / ncol(inv)));
a_eff = 100 / (nrow(x) # trace(inv) / ncol(inv));
print ’D-efficiency =’ d_eff[format=6.2]
’ A-efficiency =’ a_eff[format=6.2];

X
DESIGN Int Two Six

1 1 1.00 1.00 1.73 -1.00 -0.71 -0.55 -0.45

1 2 1.00 1.00 0.00 2.00 -0.71 -0.55 -0.45
1 3 1.00 1.00 0.00 0.00 2.12 -0.55 -0.45
1 4 1.00 1.00 0.00 0.00 0.00 2.19 -0.45
1 5 1.00 1.00 0.00 0.00 0.00 0.00 2.24
1 6 1.00 1.00 -1.73 -1.00 -0.71 -0.55 -0.45
2 1 1.00 -1.00 1.73 -1.00 -0.71 -0.55 -0.45
2 2 1.00 -1.00 0.00 2.00 -0.71 -0.55 -0.45
2 3 1.00 -1.00 0.00 0.00 2.12 -0.55 -0.45
2 4 1.00 -1.00 0.00 0.00 0.00 2.19 -0.45
2 5 1.00 -1.00 0.00 0.00 0.00 0.00 2.24
2 6 1.00 -1.00 -1.73 -1.00 -0.71 -0.55 -0.45

XPX

12 0 0 0 0 0 0
0 12 0 0 0 0 0
0 0 12 0 0 0 0
0 0 0 12 0 0 0
0 0 0 0 12 0 0
0 0 0 0 0 12 0
0 0 0 0 0 0 12

INV

0.083 0 0 0 0 0 0
0 0.083 0 0 0 0 0
0 0 0.083 0 0 0 0
0 0 0 0.083 0 0 0
0 0 0 0 0.083 0 0
0 0 0 0 0 0.083 0
0 0 0 0 0 0 0.083
82 TS-677E Multinomial Logit, Discrete Choice Modeling

D_EFF A_EFF

D-efficiency = 100.00 A-efficiency = 100.00

With this orthogonal and balanced design, X0 X =N D I = 12I, which means (X0 X) 1 = N
1
DI = 1
12 I, and
D-efficiency = 100%.
With a nonorthogonal design, for example with the first 10 rows of the 2  6 full-factorial design, D-efficiency
and A-efficiency are less than 100%.
design = design[1:10,];
x = j(10, 1, 1) || orthogcode(design[,1]) || orthogcode(design[,2]);
inv = inv(x‘ * x);
d_eff = 100 / (nrow(x) # det(inv) ## (1 / ncol(inv)));
a_eff = 100 / (nrow(x) # trace(inv) / ncol(inv));
print ’D-efficiency =’ d_eff[format=6.2]
’ A-efficiency =’ a_eff[format=6.2];
quit;

D_EFF A_EFF

D-efficiency = 92.90 A-efficiency = 84.00

In this case, j(X0 X) 1 j1=p and trace ((X0 X) 1 )=p are multiplied in the denominator of the efficiency formulas
1 . If an orthogonal and balanced design were available for this problem, then (X0 X) 1 would equal
by N1D = 10
1 I = 1 I. Since an orthogonal and balanced design is not possible (6 does not divide 10), both D-efficiency
ND 10
and A-efficiency will be less than 100%, even with the optimal design. A main-effects, orthogonal and balanced
design, with a variance matrix equal to N1D I, is the standard by which 100% efficiency is gauged, even when
we know such a design cannot exist. The standard is the average variance for the maximally efficient potentially
hypothetical design, which is knowable, not the average variance for the optimal design, which for many practical
problems we have no way of knowing.
For our purposes in this report, we will never consider an experimental design with fewer runs than a saturated
design. A saturated design has as many runs as there are parameters. The number of parameters in a main-
effects model is 1 (for the intercept) plus the sum of the numbers of levels of all of the factors, minus the number
of factors. Equivalently, since there are m 1 parameters in an m-level factor, the number of parameters is
1 + Pkj=1 (mj 1) for k factors, each with mj levels.
If a main-effects design is orthogonal and balanced, then the design must be at least as large as the saturated
design and the number of runs must be divisible by the number of levels of all the factors and by the products of
the number of levels of all pairs of factors. For example, a 2  2  3  3  3 design cannot be orthogonal and
balanced unless the number of runs is divisible by 2 (twice because there are two 2’s), 3 (three times because
there are three 3’s), 2  2 = 4 (once, because there is one pair of 2’s), 2  3 = 6 (six times, two 2’s times three
3’s), and 3  3 = 9 (three times, three pairs of 3’s). If the design is orthogonal and balanced, then all of the
divisions will work without a remainder. However, all of the divisions working is a necessary but not sufficient
condition for the existence of an orthogonal and balanced design. For example, 45 is divisible by 3 and 3  3 = 9,
but an orthogonal and balanced saturated design 322 (22 three-level factors) in 45 runs does not exist.
Candy Example 83

Candy Example
We begin with a very simple example. In this example, we will discuss the multinomial logit model, data input
and processing, analysis, results, interpretation, and probability of choice. In this example, each of ten subjects
was presented with eight different chocolate candies and asked to choose one. The eight candies consist of the
23 combinations of dark or milk chocolate, soft or chewy center, and nuts or no nuts. Each subject saw all eight
candies and made one choice. Experimental choice data such as these are typically analyzed with a multinomial
logit model.

The Multinomial Logit Model

The multinomial logit model assumes that the probability that an individual will choose one of the m alternatives,
ci , from choice set C is

p(c jC ) = P
exp(U (c )) = P exp(x )
i i

=1 exp(U (c )) =1 exp(x )
i m m
j j j j

where xi is a vector of alternative attributes and is a vector of unknown parameters. U (ci ) = xi is the utility
for alternative ci , which is a linear function of the attributes. The probability that an individual will choose one
of the m alternatives, ci , from choice set C is the exponential of the utility of the alternative divided by the sum
of all of the exponentiated utilities.
There are m = 8 attribute vectors in this example, one for each alternative. Let x = (Dark/Milk, Soft/Chewy,
Nuts/No Nuts) where Dark/Milk = (1 = Dark, 0 = Milk), Soft/Chewy = (1 = Soft, 0 = Chewy), Nuts/No Nuts =
(1 = Nuts, 0 = No Nuts). The eight attribute vectors are

x1 = (0 0 0) (Milk, Chewy, No Nuts)

x2 = (0 0 1) (Milk, Chewy, Nuts )
x3 = (0 1 0) (Milk, Soft, No Nuts)
x4 = (0 1 1) (Milk, Soft, Nuts )
x5 = (1 0 0) (Dark, Chewy, No Nuts)
x6 = (1 0 1) (Dark, Chewy, Nuts )
x7 = (1 1 0) (Dark, Soft, No Nuts)
x8 = (1 1 1) (Dark, Soft, Nuts )

Say, hypothetically that 0 = (4 2 1): That is, the part-worth utility for dark chocolate is 4, the part-
worth utility for soft center is -2, and the part-worth utility for nuts is 1. The utility for each of the combinations,
xi , would be as follows.

U(Milk, Chewy, No Nuts) = 04 + 0 2 + 01 = 0

U(Milk, Chewy, Nuts ) = 04 + 0 2 + 11 = 1
U(Milk, Soft, No Nuts) = 04 + 1 2 + 01 = -2
U(Milk, Soft, Nuts ) = 04 + 1 2 + 11 = -1
U(Dark, Chewy, No Nuts) = 14 + 0 2 + 01 = 4
U(Dark, Chewy, Nuts ) = 14 + 0 2 + 11 = 5
U(Dark, Soft, No Nuts) = 14 + 1 2 + 01 = 2
U(Dark, Soft, Nuts ) = 14 + 1 2 + 11 = 3
84 TS-677E Multinomial Logit, Discrete Choice Modeling

P
=1 exp(x ), is exp(0) + exp(1) + exp( 2) + exp( 1) +
m
The denominator of the probability formula, j j

exp(4) + exp(5)
P + exp(2) + exp(3) = 234 :707. The probability that each alternative is chosen,
exp(x )= =1 exp(x ), is
i
m
j j

p(Milk, Chewy, No Nuts) = exp(0) / 234.707 = 0.004

p(Milk, Chewy, Nuts ) = exp(1) / 234.707 = 0.012
p(Milk, Soft, No Nuts) = exp(-2) / 234.707 = 0.001
p(Milk, Soft, Nuts ) = exp(-1) / 234.707 = 0.002
p(Dark, Chewy, No Nuts) = exp(4) / 234.707 = 0.233
p(Dark, Chewy, Nuts ) = exp(5) / 234.707 = 0.632
p(Dark, Soft, No Nuts) = exp(2) / 234.707 = 0.031
p(Dark, Soft, Nuts ) = exp(3) / 234.707 = 0.086
Note that even combinations with a negative or zero utility have a nonzero probability of choice. Also note that
adding a constant to the utilities will not change the probability of choice, however multiplying by a constant
will.
Probability of choice is a nonlinear and increasing function of utility. The following plot shows the relationship
between utility and probability of choice for this hypothetical situation.
data x;
do u = -2 to 5 by 0.1;
p = exp(u) / 234.707;
output;
end;
run;
proc gplot;
title h=1 ’Probability of Choice as a Function of Utility’;
plot p * u;
symbol1 i=join;
run; quit;
Candy Example 85

This plot shows the function exp( 2) to exp(5), scaled into the range zero to one, the range of probability values.
For the small negative utilities, the probability of choice is essentially zero. As utility increases beyond two, the
function starts rapidly increasing.
In this example, the chosen alternatives are x5 , x6 , x7 , x5 , x2 , x6 , x2 , x6 , x6 , x6 . Alternative x2 was chosen 2
times, x5 was chosen 2 times, x6 was chosen 5 times, and x7 was chosen 1 time. The choice model likelihood
for these data is the product of ten terms, one for each choice set for each subject. Each term consists of the
probability that the chosen alternative is chosen. For each choice set, the utilities for all of the alternatives enter
into the denominator, and the utility for the chosen alternative enters into the numerator. The choice model
likelihood for these data is

L = hP8exp(x5 ) i  hP8exp(x6 ) i  hP exp(x7) i  h exp(x5 )

P8 i
=1 exp(x ) =1 exp(x ) exp(x ) =1 exp(x )
C
8
j j j j j =1 j j j

hP exp(x2 ) i  hP exp(x6 ) i  hP exp(x2 ) i  hP exp(x6 ) i

=1 exp(x ) =1 exp(x ) =1 exp(x ) =1 exp(x )
8 8 8 8
j j j j j j j j

hP exp(x6 ) i  hP exp(x6 ) i
=1 exp(x ) =1 exp(x )
8 8
j j j j

= exp((2hxP2 + 2x5 + 5x6 i+10x7 ) )

=1 exp(x )
8
j j

The Input Data

The data set consists of one observation for each alternative of each choice set for each subject. (A typical choice
study has more than one choice set per person. This first example only has one choice set to help keep it simple.)
All of the chosen and unchosen alternatives must appear in the data set. The data set must contain variables that
identify the subject, the choice set, which alternative was chosen, and the set of alternatives from which it was
chosen. In this example, the data set contains 10  1  8 = 80 observations: 10 subjects each saw 1 choice set
with 8 alternatives.
Typically, two variables are used to identify the choice sets, subject ID and choice set within subject. In this
simple case where each subject only made one choice, the choice set variable is not necessary. However, we use
it here to illustrate the general case. The variable Subj is the subject number, and Set identifies the choice set
within subject. The chosen alternative is indicated by c=1, which means first choice. All second and subsequent
choices are unobserved, so the unchosen alternatives are indicated by c=2, which means that all we know is
that they would have been chosen after the first choice. Both the chosen and unchosen alternatives must appear
in the input data set since both are needed to construct the likelihood function. The c=2 observations enter
into the denominator of the likelihood function, and the c=1 observations enter into both the numerator and the
denominator of the likelihood function. In this input DATA step, the data for four alternatives appear on one
line, and all of the data for a choice set of eight alternatives appear on two lines. The DATA step shows data
entry in the way that requires the fewest programming statements. Each execution of the input statement reads
information about one alternative. The @@ in the input statement specifies that SAS should not automatically
go to a new input data set line when it reads the next row of data. This specification is needed here because each
line in the input data set contains the data for four output data set rows. The data from the first two subjects is
printed.
86 TS-677E Multinomial Logit, Discrete Choice Modeling

title ’Choice of Chocolate Candies’;

data chocs;
input Subj c Dark Soft Nuts @@;
Set = 1;
datalines;
1 2 0 0 0 1 2 0 0 1 1 2 0 1 0 1 2 0 1 1
1 1 1 0 0 1 2 1 0 1 1 2 1 1 0 1 2 1 1 1
2 2 0 0 0 2 2 0 0 1 2 2 0 1 0 2 2 0 1 1
2 2 1 0 0 2 1 1 0 1 2 2 1 1 0 2 2 1 1 1
3 2 0 0 0 3 2 0 0 1 3 2 0 1 0 3 2 0 1 1
3 2 1 0 0 3 2 1 0 1 3 1 1 1 0 3 2 1 1 1
4 2 0 0 0 4 2 0 0 1 4 2 0 1 0 4 2 0 1 1
4 1 1 0 0 4 2 1 0 1 4 2 1 1 0 4 2 1 1 1
5 2 0 0 0 5 1 0 0 1 5 2 0 1 0 5 2 0 1 1
5 2 1 0 0 5 2 1 0 1 5 2 1 1 0 5 2 1 1 1
6 2 0 0 0 6 2 0 0 1 6 2 0 1 0 6 2 0 1 1
6 2 1 0 0 6 1 1 0 1 6 2 1 1 0 6 2 1 1 1
7 2 0 0 0 7 1 0 0 1 7 2 0 1 0 7 2 0 1 1
7 2 1 0 0 7 2 1 0 1 7 2 1 1 0 7 2 1 1 1
8 2 0 0 0 8 2 0 0 1 8 2 0 1 0 8 2 0 1 1
8 2 1 0 0 8 1 1 0 1 8 2 1 1 0 8 2 1 1 1
9 2 0 0 0 9 2 0 0 1 9 2 0 1 0 9 2 0 1 1
9 2 1 0 0 9 1 1 0 1 9 2 1 1 0 9 2 1 1 1
10 2 0 0 0 10 2 0 0 1 10 2 0 1 0 10 2 0 1 1
10 2 1 0 0 10 1 1 0 1 10 2 1 1 0 10 2 1 1 1
;
proc print data=chocs noobs;
where subj <= 2;
var subj set c dark soft nuts;
run;

Choice of Chocolate Candies

Subj Set c Dark Soft Nuts

1 1 2 0 0 0
1 1 2 0 0 1
1 1 2 0 1 0
1 1 2 0 1 1
1 1 1 1 0 0
1 1 2 1 0 1
1 1 2 1 1 0
1 1 2 1 1 1
2 1 2 0 0 0
2 1 2 0 0 1
2 1 2 0 1 0
2 1 2 0 1 1
2 1 2 1 0 0
2 1 1 1 0 1
2 1 2 1 1 0
2 1 2 1 1 1
Candy Example 87

These next steps illustrate a more typical form of data entry. The experimental design is stored in a separate data
set from the choices and is merged with the choices as the data are read, which produces the same results as the
preceding steps.
title ’Choice of Chocolate Candies’;

* Alternative Form of Data Entry;

data combos; /* Read the design matrix. */

input Dark Soft Nuts;
datalines;
0 0 0
0 0 1
0 1 0
0 1 1
1 0 0
1 0 1
1 1 0
1 1 1
;
data chocs; /* Create the data set. */
input Choice @@; drop choice; /* Read the chosen combo num. */
Subj = _n_; Set = 1; /* Store subj, choice set num. */
do i = 1 to 8; /* Loop over alternatives. */
c = 2 - (i eq choice); /* Designate chosen alt. */
set combos point=i; /* Read design matrix. */
output; /* Output the results. */
end;
datalines;
5 6 7 5 2 6 2 6 6 6
;
The variable Choice is the number of the chosen alternative. For each choice set, each of the eight observations
in the experimental design is read. The point= option on the set statement is used to read the ith observation
of the data set COMBOS. When i (the alternative index) equals Choice (the number of the chosen alterna-
tive), the logical expression (i eq choice) equals 1; otherwise it is 0. The statement c = 2 - (i eq
choice) sets c to 1 (two minus one) when the alternative is chosen and 2 (two minus zero) otherwise. All
eight observations in the COMBOS data set are read 10 times, once per subject. The resulting data set is the
same as the one we created previously. In all of the remaining examples, we will simplify this process by using
the %MktMerge macro to merge the design and data. The basic logic underlying this macro is shown in the
preceding step. The number of a chosen alternative is read, then each alternative of the choice set is read, the
chosen alternative is flagged (c = 1), and the unchosen alternatives are flagged (c = 2). One observation
per choice set per subject is read from the input data stream, and one observation per alternative per choice set
per subject is written.

Fitting the Multinomial Logit Model

The data are now in the right form for analysis. In SAS, the multinomial logit model is fit with the SAS/STAT
procedure PHREG (proportional hazards regression), with the ties=breslow option. The likelihood function
of the multinomial logit model has the same form as a survival analysis model fit by PROC PHREG.
In a discrete choice study, subjects are presented with sets of alternatives and asked to choose the most preferred
alternative. The data for one choice set consist of one alternative that was chosen and m 1 alternatives that were
not chosen. First choice was observed. Second and subsequent choices were not observed; it is only known that
the other alternatives would have been chosen after the first choice. In survival analysis, subjects (rats, people,
light bulbs, machines, and so on) are followed until a specific event occurs (such as failure or death) or until the
experiment ends. The data are event times. The data for subjects who have not experienced the event (such as
88 TS-677E Multinomial Logit, Discrete Choice Modeling

those who survive past the end of a medical experiment) are censored. The exact event time is not known, but
it is known to have occurred after the censored time. In a discrete choice study, first choice occurs at time one,
and all subsequent choices (second choice, third choice, and so on) are unobserved or censored. The survival and
choice models are the same. To fit the multinomial logit model, use PROC PHREG as follows.
proc phreg data=chocs outest=betas;
strata subj set;
model c*c(2) = dark soft nuts / ties=breslow;
label dark = ’Dark Chocolate’ soft = ’Soft Center’
nuts = ’With Nuts’;
run;
The data= option specifies the input data set. The outest= option requests an output data set called BETAS
with the parameter estimates. The strata statement specifies that each combination of the variables Set and
Subj forms a set from which a choice was made. Each term in the likelihood function is a stratum. There is
one term or stratum per choice set per subject, and each is composed of information about the chosen and all the
unchosen alternatives.
In the left side of the model statement, you specify the variables that indicate which alternatives were chosen
and unchosen. While this could be two different variables, we will use one variable c to provide both pieces of in-
formation. The response variable c has values 1 (chosen or first choice) and 2 (unchosen or subsequent choices).
The first c of the c*c(2) in the model statement specifies that c indicates which alternative was chosen. The
second c specifies that c indicates which alternatives were not chosen, and (2) means that observations with
values of 2 were not chosen. When c is set up such that 1 indicates the chosen alternative and 2 indicates the un-
chosen alternatives, always specify c*c(2) on the left of the equal sign in the model statement. The attribute
variables are specified after the equal sign. Specify ties=breslow after a slash to explicitly specify the like-
lihood function for the multinomial logit model. (Do not specify any other ties= options; ties=breslow
specifies the most efficient and always appropriate way to fit the multinomial logit model.) The label statement
is added since we are using a template that assumes each variable has a label.
Note that the c*c(n) syntax allows second choice (c=2) and subsequent choices (c=3, c=4, ...) to be entered.
Just enter in parentheses one plus the number of choices actually made. For example, with first and second choice
data specify c*c(3). Note however that some experts believe that second and subsequent choice data are much
less reliable than first choice data.

Multinomial Logit Model Results

The output is shown next. Recall that we used %phchoice(on)on page 79 to customize the output from PROC
PHREG.

Choice of Chocolate Candies

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW
Candy Example 89

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Subj Set Alternatives Alternatives Chosen

1 1 1 8 1 7
2 2 1 8 1 7
3 3 1 8 1 7
4 4 1 8 1 7
5 5 1 8 1 7
6 6 1 8 1 7
7 7 1 8 1 7
8 8 1 8 1 7
9 9 1 8 1 7
10 10 1 8 1 7
---------------------------------------------------------------------------
Total 80 10 70

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 41.589 28.727

AIC 41.589 34.727
SBC 41.589 35.635

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 12.8618 3 0.0049

Score 11.6000 3 0.0089
Wald 8.9275 3 0.0303

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Dark Chocolate 1 1.38629 0.79057 3.0749 0.0795

Soft Center 1 -2.19722 1.05409 4.3450 0.0371
With Nuts 1 0.84730 0.69007 1.5076 0.2195

The first table, ’Model Information’, contains the input data set name, dependent variable name, censoring infor-
mation, and tie handling option.
The ’Summary of Subjects, Sets, and Chosen and Unchosen Alternatives’ table is printed by default and should
be used to check the data entry. In general, there are as many strata as there are combinations of the Subj
and Set variables. In this case, there are ten strata. Each stratum must be composed of m alternatives. In this
case, there are eight alternatives. The number of chosen alternatives should be 1, and the number of unchosen
alternatives is m 1 (in this case 7). Always check the summary table to ensure that the data are arrayed
correctly.
90 TS-677E Multinomial Logit, Discrete Choice Modeling

The next table, ’Convergence Status’, shows that the iterative algorithm successfully converged. The next tables,
’Model Fit Statistics’ and ’Testing Global Null Hypothesis: BETA=0’ contain the overall fit of the model. The -2
LOG L statistic under ’With Covariates’ is 28.727 and the Chi-Square statistic is 12.8618 with 3 df (p=0.0049),
which is used to test the null hypothesis that the attributes do not influence choice. At common alpha levels
such as 0.05 and 0.01, we would reject the null hypothesis of no relationship between choice and the attributes.
Note that 41.589 (-2 LOG L Without Covariates, which is -2 LOG L for a model with no explanatory variables)
minus 28.727 (-2 LOG L With Covariates, which is -2 LOG L for a model with all explanatory variables) equals
12.8618 (Model Chi-Square, which is used to test the effects of the explanatory variables).
Next is the ’Multinomial Logit Parameter Estimates’ table. For each effect, it contains the maximum likelihood
parameter estimate, its estimated standard error (the square root of the corresponding diagonal element of the
estimated covariance matrix), the Wald Chi-Square statistic (the square of the parameter estimate divided by its
standard error), the df of the Wald Chi-Square statistic (1 unless the corresponding parameter is redundant or
infinite, in which case the value is 0), and the p-value of the Chi-Squared statistic with respect to a chi-squared
distribution with one df. The parameter estimate with the smallest p-value is for soft center. Since the parameter
estimate is negative, chewy is the more preferred level. Dark is preferred over milk, and nuts over no nuts,
however only the p-value for Soft is less than 0.05.

Fitting the Multinomial Logit Model, All Levels

It is instructive to perform some manipulations on the data set and analyze it again. These steps will perform the
same analysis as before, only now, coefficients for both levels of the three attributes are printed. Binary variables
for the missing levels are created by subtracting the existing binary variables from 1.
data chocs2;
set chocs;
Milk = 1 - dark; Chewy = 1 - Soft; NoNuts = 1 - nuts;
label dark = ’Dark Chocolate’ milk = ’Milk Chocolate’
soft = ’Soft Center’ chewy = ’Chewy Center’
nuts = ’With Nuts’ nonuts = ’No Nuts’;
run;

proc phreg data=chocs2;

strata subj set;
model c*c(2) = dark milk soft chewy nuts nonuts / ties=breslow;
run;

Choice of Chocolate Candies

The PHREG Procedure

Model Information

Data Set WORK.CHOCS2

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW
Candy Example 91

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Subj Set Alternatives Alternatives Chosen

1 1 1 8 1 7
2 2 1 8 1 7
3 3 1 8 1 7
4 4 1 8 1 7
5 5 1 8 1 7
6 6 1 8 1 7
7 7 1 8 1 7
8 8 1 8 1 7
9 9 1 8 1 7
10 10 1 8 1 7
---------------------------------------------------------------------------
Total 80 10 70

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 41.589 28.727

AIC 41.589 34.727
SBC 41.589 35.635

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 12.8618 3 0.0049

Score 11.6000 3 0.0089
Wald 8.9275 3 0.0303

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Dark Chocolate 1 1.38629 0.79057 3.0749 0.0795

Milk Chocolate 0 0 . . .
Soft Center 1 -2.19722 1.05409 4.3450 0.0371
Chewy Center 0 0 . . .
With Nuts 1 0.84730 0.69007 1.5076 0.2195
No Nuts 0 0 . . .

Now the zero coefficients for the reference levels, milk, chewy, and no nuts are printed. The part-worth utility for
Milk Chocolate is a structural zero, and the part-worth utility for Dark Chocolate is larger at 1.38629. Similarly,
the part-worth utility for Chewy Center is a structural zero, and the part-worth utility for Soft Center is smaller
at -2.19722. Finally, the part-worth utility for No Nuts is a structural zero, and the part-worth utility for Nuts is
larger at 0.84730.
92 TS-677E Multinomial Logit, Discrete Choice Modeling

Probability of Choice
The parameter estimates are used next to construct the estimated probability that each alternative will be chosen.
The DATA step program uses the following formula to create the choice probabilities.

p(c jC ) = P
exp(x ) i

=1 exp(x )
i m
j j

* Estimate the probability that each alternative will be chosen;

data p;
retain sum 0;
set combos end=eof;

* On the first pass through the DATA step (_n_ is the pass
number), get the regression coefficients in B1-B3.
Note that they are automatically retained so that they
can be used in all passes through the DATA step.;

if _n_ = 1 then
set betas(rename=(dark=b1 soft=b2 nuts=b3));
keep dark soft nuts p;
array x[3] dark soft nuts;
array b[3] b1-b3;
* For each combination, create x * b;
p = 0;
do j = 1 to 3;
p = p + x[j] * b[j];
end;
* Exponentiate x * b and sum them up;
p = exp(p);
sum = sum + p;

* Output sum exp(x * b) in the macro variable ’&sum’;

if eof then call symput(’sum’,put(sum,best12.));
run;

proc format;
value df 1 = ’Dark’ 0 = ’Milk’;
value sf 1 = ’Soft’ 0 = ’Chewy’;
value nf 1 = ’Nuts’ 0 = ’No Nuts’;
run;
* Divide each exp(x * b) by sum exp(x * b);
data p;
set p;
p = p / (&sum);
format dark df. soft sf. nuts nf.;
run;

proc sort;
by descending p;
run;
proc print;
run;
Candy Example 93

Choice of Chocolate Candies

Obs Dark Soft Nuts p

1 Dark Chewy Nuts 0.50400

2 Dark Chewy No Nuts 0.21600
3 Milk Chewy Nuts 0.12600
4 Dark Soft Nuts 0.05600
5 Milk Chewy No Nuts 0.05400
6 Dark Soft No Nuts 0.02400
7 Milk Soft Nuts 0.01400
8 Milk Soft No Nuts 0.00600

The three most preferred alternatives are Dark/Chewy/Nuts, Dark/Chewy/No Nuts, and Milk/Chewy/Nuts.
94 TS-677E Multinomial Logit, Discrete Choice Modeling

Fabric Softener Example

In this example, subjects are asked to choose among fabric softeners. This example shows all of the steps
in a discrete choice study, including experimental design creation and evaluation, creating the questionnaire,
inputting the raw data, creating the data set for analysis, coding, fitting the discrete choice model, interpretation,
and probability of choice. In addition, custom questionnaires are discussed. We assume the reader is familiar
with the experimental design issues discussed in Kuhfeld, Tobias, and Garratt (1994), on page 25. Some of these
concepts are reviewed starting on page 76.

Set Up
The study involves four fictitious fabric softeners Sploosh, Plumbbob, Platter, and Moosey. Each choice set
consists of each of these four brands and a constant alternative Another. Each of the brands is available at three
prices, $1.49, $1.99, and $2.49. Another is only offered at $1.99. There are 50 subjects, each of which will see
the same choice sets. We can use the %MktRuns autocall macro to help us choose the number of choice sets. All
of the autocall macros used in this report are documented starting on page 287. To use this macro, you specify
the number of levels for each of the factors. With four brands each with three prices, you specify four 3’s (or
3 ** 4).
title ’Choice of Fabric Softener’;

%mktruns( 3 3 3 3 )
The output first tells us that we specified a design with four factors, each with three levels. The next table reports
the size of the saturated design, which is the number of parameters in the linear design, and suggests design sizes.

Choice of Fabric Softener

Design Summary

Number of
Levels Frequency

3 4

Choice of Fabric Softener

Saturated = 9
Full Factorial = 81

Some Reasonable Cannot Be

Design Sizes Violations Divided By

9 * 0
18 * 0
12 6 9
15 6 9
10 10 3 9
11 10 3 9
13 10 3 9
14 10 3 9
16 10 3 9
17 10 3 9

* - 100% Efficient Design can be made with the MktEx Macro.

 Of course real studies would use real brands. Since we have not collected real data, we cannot use real brand names. We picked these
silly names so no one would confuse our artificial data with real data.
Fabric Softener Example 95

Choice of Fabric Softener

n Design Reference

9 3 ** 4 Fractional-factorial
18 2 ** 1 3 ** 7 Taguchi, 1987
18 3 ** 6 6 ** 1 Taguchi, 1987

The output from this macro tells us that the saturated design has nine runs and the full-factorial design has 81
runs. It also tells us that 9 and 18 are optimal design sizes with zero violations. The macro tells us that in nine
runs, an orthogonal design with 4 three-level factors is available, and in 18 runs, two orthogonal and balanced
designs are available: one with a two-level factor and 7 three-level factors, and one with 6 three-level factors
and a six-level factor. There are zero violations with these designs because these sizes can be divided by 3 and
3  3 = 9. Twelve and 15 are also reported as potential design sizes, but each has 6 violations. Six times
(the 4(4 1)=2 = 6 pairs of the four threes) 12 and 15 cannot be divided by 3  3 = 9. Ideally, we would
like to have a manageable number of choice sets for people to evaluate and a design that is both orthogonal and
balanced. When violations are reported, orthogonal and balanced designs are not possible. While orthogonality
and balance are not required, they are nice properties to have. With 4 three-level factors, the number of choice
sets in all orthogonal and balanced designs must be divisible by 3  3 = 9.
Nine choice sets is a bit small. Furthermore, there are no error df. We set the number of choice sets to 18 since it
is small enough for each person to see all choice sets, large enough to have reasonable error df, and an orthogonal
and balanced design is available. It is important to remember however that the concept of number of parameters
and error df discussed here applies to the linear design and not to the choice design. We could use the nine-run
design for a discrete choice model and have error df in the choice model. If we were to instead use this design
for a full-profile conjoint (not recommended), there would be no error df.
To make the code easier to modify for future use, the number of choice sets and alternatives are stored in macro
variables and the prices are put into a format. Our design, in raw form, will have values for price of 1, 2, and
3. We will use a format to assign the actual prices: $1.49, $1.99, and $2.49. The format also creates a price of
$1.99 for missing, which will be used for the constant alternative.
%let n = 18; /* n choice sets */
%let m = 5; /* m alternative including constant */
%let mm1 = %eval(&m - 1); /* m - 1 */

proc format; /* create a format for the price */

value price 1 = ’$1.49’ 2 = ’$1.99’ 3 = ’$2.49’ . = ’$1.99’;
run;

Designing the Choice Experiment

In the next steps, an efficient experimental design is created. We will use an autocall macro %MktEx to create
most of our designs. (All of the autocall macros used in this report are documented starting on page 287.) When
you invoke the %MktEx macro for a simple problem, you only need to specify the numbers of levels, and number
of runs. The macro does the rest. Here is the %MktEx macro usage for this example:
%mktex(3 ** 4, n=&n)

The syntax ’n  m’ means m factors each at n levels. This example has four factors, x1 through x4, all with
three levels. A design with 18 runs is requested. The n= option specifies the number of runs. These are all the
options that are needed for a simple problem such as this one. However, throughout this report, random number
seeds are explicitly specified with the seed= option so that the results will be reproducible. Here is the macro
 By specifying a random number seed, results should be reproducible within a SAS release for a particular operating system. However,
due to machine differences, some results may not be exactly reproducible on other machines. For most orthogonal and balanced designs,
the results should be reproducible. When computerized searches are done, it is likely that you will not get the same design as the one in the
book, although you would expect the efficiency differences to be slight.
96 TS-677E Multinomial Logit, Discrete Choice Modeling

call with the random number seed specified:

%mktex(3 ** 4, n=&n, seed=7654321)

proc print; run;

Here are the results.

Choice of Fabric Softener

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 100.0000 100.0000 Tab
1 End 100.0000

Choice of Fabric Softener

The OPTEX Procedure

Class Level Information

Class Levels -Values-

x1 3 1 2 3
x2 3 1 2 3
x3 3 1 2 3
x4 3 1 2 3

Choice of Fabric Softener

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 100.0000 100.0000 100.0000 0.7071 p

Choice of Fabric Softener

Obs x1 x2 x3 x4

1 1 1 1 1
2 1 1 2 3
3 1 2 1 3
4 1 2 3 2
5 1 3 2 2
6 1 3 3 1
7 2 1 1 2
8 2 1 3 3
9 2 2 2 2
Fabric Softener Example 97

10 2 2 3 1
11 2 3 1 3
12 2 3 2 1
13 3 1 2 1
14 3 1 3 2
15 3 2 1 1
16 3 2 2 3
17 3 3 1 2
18 3 3 3 3

The macro found a perfect, orthogonal and balanced, 100% efficient design consisting of 4 three-level factors,
x1-x4. The levels are the integers 1 to 3. For this problem, the macro generated the design directly. For other
problems, the macro may have to use a computerized search. See page 123 for more information on how the
%MktEx macro works.

Examining the Design

It is good to run basic checks on all designs. You can use the %MktEval macro to display information about
the design. The macro first prints a matrix of canonical correlations between the factors. We hope to see an
identity matrix (a matrix of ones on the diagonal and zeros everywhere else). Next, the macro prints all one-
way frequencies for all attributes, all two-way frequencies, and all n-way frequencies (in this case four-way)
frequencies. We hope to see equal or at least nearly equal one-way and two-way frequencies, and we want to see
that each combination occurs only once.
%mkteval;

Choice of Fabric Softener

Canonical Correlations Between the Factors
There are 0 Canonical Correlations Greater Than 0.316

x1 x2 x3 x4

x1 1 0 0 0
x2 0 1 0 0
x3 0 0 1 0
x4 0 0 0 1

Choice of Fabric Softener

Summary of Frequencies
There are 0 Canonical Correlations Greater Than 0.316

Frequencies

x1 6 6 6
x2 6 6 6
x3 6 6 6
x4 6 6 6
x1 x2 2 2 2 2 2 2 2 2 2
x1 x3 2 2 2 2 2 2 2 2 2
x1 x4 2 2 2 2 2 2 2 2 2
x2 x3 2 2 2 2 2 2 2 2 2
x2 x4 2 2 2 2 2 2 2 2 2
x3 x4 2 2 2 2 2 2 2 2 2
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
98 TS-677E Multinomial Logit, Discrete Choice Modeling

The first table shows the canonical correlations between pairs of coded factors. A canonical correlation is the
maximum correlation between linear combinations of the coded factors. All zeros off the diagonal show that
this design is orthogonal for main effects. If any off-diagonal canonical correlations had been greater than 0.316
(r2 > 0:1), the macro would have listed them in a separate table. The last title line tells you that none of them
was this large. For nonorthogonal designs and designs with interactions, the canonical-correlation matrix is not
a substitute for looking at the variance matrix (with examine=v, discussed on pages 126, 162, and 336). It
just provides a quick and more-compact picture of the correlations between the factors. The variance matrix is
sensitive to the actual model specified and the coding. The canonical-correlation matrix just tells you if there is
some correlation between the main effects. In this case, there are no correlations.
The macro also prints one-way, two-way and n-way frequencies. The equal one-way frequencies show you that
this design is balanced. The equal two-way frequencies show you that this design is orthogonal. The n-way
frequencies, all equal to one, show you that there are no duplicate profiles. This is a perfect design for a main-
effects model. However, there are other 100% efficient designs for this problem with duplicate observations.
In the last part of the output, the N-Way frequencies may contain some 2’s for those designs. You can specify
options=nodups to ensure that there are no duplicates.
The %MktEval macro produces a very compact summary of the design, hence some information, for example
the levels to which the frequencies correspond, is not shown. You can use the print=freqs option to get a
less compact and more detailed display.
%mkteval(data=design, print=freqs);

Here are some of the results.

Choice of Fabric Softener

Frequencies
There are 0 Canonical Correlations Greater Than 0.316

Effects Frequency x1 x2 x3 x4

x1 6 1 . . .
6 2 . . .
6 3 . . .

x2 6 . 1 . .
6 . 2 . .
6 . 3 . .
.
.
.

x1 x2 2 1 1 . .
2 1 2 . .
2 1 3 . .
2 2 1 . .
2 2 2 . .
2 2 3 . .
2 3 1 . .
2 3 2 . .
2 3 3 . .
.
.
.
Fabric Softener Example 99

x3 x4 2 . . 1 1
2 . . 1 2
2 . . 1 3
2 . . 2 1
2 . . 2 2
2 . . 2 3
2 . . 3 1
2 . . 3 2
2 . . 3 3

N-Way 1 1 1 1 1
1 1 1 2 3
1 1 2 1 3
1 1 2 3 2
1 1 3 2 2
1 1 3 3 1
1 2 1 1 2
1 2 1 3 3
1 2 2 2 2
1 2 2 3 1
1 2 3 1 3
1 2 3 2 1
1 3 1 2 1
1 3 1 3 2
1 3 2 1 1
1 3 2 2 3
1 3 3 1 2
1 3 3 3 3

Randomizing the Design, Postprocessing

The design we just looked at and examined was in the default output data set DESIGN. The DESIGN data
set is sorted and often has a first row consisting entirely of ones. For these reasons, you should actually use the
randomized design. In the randomized design, the choice sets are presented in a random order and the levels have
been randomly reassigned. Neither of these operations affects the design efficiency, balance, or orthogonality.
The macro automatically randomizes the design and stores the results in a data set called RANDOMIZED. The
next steps assign formats and labels, and store the results in a SAS data set [Link] so that it will still be
available after the data are collected.
proc print data=randomized; run;
data [Link];
set randomized;
format x1-x&mm1 price.;
label x1 = ’Sploosh’ x2 = ’Plumbbob’ x3 = ’Platter’ x4 = ’Moosey’;
run;
100 TS-677E Multinomial Logit, Discrete Choice Modeling

This is the final design.

proc print data=[Link] label; /* print final design */
title2 ’Efficient Design’;
run;

title2;

Choice of Fabric Softener

Efficient Design

Obs Sploosh Plumbbob Platter Moosey

1 $2.49 $1.99 $1.99 $2.49

2 $1.49 $2.49 $1.99 $1.99
3 $1.49 $1.99 $1.99 $1.49
4 $2.49 $2.49 $2.49 $1.49
5 $1.99 $1.49 $1.99 $2.49
6 $1.49 $1.49 $1.49 $2.49
7 $1.49 $1.99 $1.49 $1.99
8 $1.49 $1.49 $2.49 $1.49
9 $2.49 $1.49 $1.49 $1.49
10 $1.99 $2.49 $1.99 $1.49
11 $2.49 $1.49 $1.99 $1.99
12 $1.99 $1.99 $2.49 $2.49
13 $1.99 $2.49 $1.49 $1.99
14 $2.49 $1.99 $2.49 $1.99
15 $2.49 $2.49 $1.49 $2.49
16 $1.99 $1.49 $2.49 $1.99
17 $1.49 $2.49 $2.49 $2.49
18 $1.99 $1.99 $1.49 $1.49

Generating the Questionnaire

A questionnaire based on the design is printed using the DATA step. The statement array brands[&m]
$ - temporary- (’Sploosh’ ’Plumbbob’ ’Platter’ ’Moosey’ ’Another’) creates a con-
stant array so that brands[1] accesses the string ’Sploosh’, brands[2] accesses the string ’Plumb-
bob’, and so on. The - temporary- specification means that no output data set variables are created for
this array. The linesleft= specification in the file statement creates the variable ll, which contains the
number of lines left on a page. This ensures that each choice set is not split over two pages.
Fabric Softener Example 101

options ls=80 ps=60 nonumber nodate;

title;

data _null_; /* print questionnaire */

array brands[&m] $ _temporary_ (’Sploosh’ ’Plumbbob’ ’Platter’
’Moosey’ ’Another’);
array x[&m] x1-x&m;
file print linesleft=ll;
set [Link];

x&m = 2; /* constant alternative */

format x&m price.;

if _n_ = 1 or ll < 12 then do;

put _page_;
put @60 ’Subject: _________’ //;
end;
put _n_ 2. ’) Circle your choice of ’
’one of the following fabric softeners:’ /;
do brnds = 1 to &m;
put ’ ’ brnds 1. ’) ’ brands[brnds] ’brand at ’
x[brnds] +(-1) ’.’ /;
end;
run;
In the interest of space, only the first two choice sets are printed. The questionnaire is printed, copied, and the
data are collected.

Subject: _________

1) Circle your choice of one of the following fabric softeners:

1) Sploosh brand at $2.49.

2) Plumbbob brand at $1.99.

3) Platter brand at $1.99.

4) Moosey brand at $2.49.

5) Another brand at $1.99.

2) Circle your choice of one of the following fabric softeners:

1) Sploosh brand at $1.49.

2) Plumbbob brand at $2.49.

3) Platter brand at $1.99.

4) Moosey brand at $1.99.

5) Another brand at $1.99.

102 TS-677E Multinomial Logit, Discrete Choice Modeling

In practice, data collection may be much more elaborate than this. It may involve art work, photographs, and the
choice sets may be presented and data may be collected over the web. However the choice sets are presented and
the data are collected, the essential ingredients remain the same. Subjects are shown sets of alternatives and are
asked to make a choice, then they go on to the next set.

Entering the Data

The data consist of a subject number followed by 18 integers in the range 1 to 5. These are the alternatives that
were chosen for each choice set. For example, the first subject chose alternative 3 (Platter brand at $1.99) in the
first choice set, alternative 3 (Platter brand at $1.99) in the second choice set, and so on. In the interest of space,
data from three subjects appear on one line.
title ’Choice of Fabric Softener’;

data results; /* read choice data set */

input Subj (choose1-choose&n) (1.) @@;
datalines;
1 334523313433343413 2 314423343433333215 3 333423323323333233
4 314432343444343254 5 315423313421543213 6 334433344423323233
7 334433343432343413 8 334423323434343214 9 333433313351343213
10 345423325322343233 11 333433343433323213 12 314423343423333253
13 534423313322343213 14 214421144423323214 15 334423345433343235
16 333433345333335313 17 534423343452343413 18 344435544432343513
19 334433345432343433 20 311423343422325513 21 334453543423543213
22 314423353323333513 23 333423343333333433 24 545422323322323213
25 354433343433333313 26 314425525422343214 27 334353523423353213
28 334433313333333233 29 333423543335353234 30 334453343533343433
31 354423344322333413 32 354422343323333213 33 314423343352343215
34 334423343443333213 35 314553344453343215 36 333433544433343233
37 314423343424543214 38 353433324423353533 39 333453323333323513
40 314433343422333214 41 334423344442343444 42 334433323422323213
43 333423354433343213 44 314423323422333213 45 314452544422343214
46 414423345423543214 47 544423544442343414 48 335453343423323453
49 314523344424333214 50 334423343332343413
;

Processing the Data

Next, we prepare the experimental design for analysis. Our design, stored in the data set [Link], is
stored with one row per choice set. We call this the linear design (see page 78). The linear design, which came
directly from the %MktEx macro, is conveniently arrayed for generating the questionnaire, however it is not in
the right form for analysis. For analysis, we need a choice design with one row for each alternative of each choice
set. We will use the macro %MktRoll to “roll out” the linear design into the choice design, which is in proper
form for analysis. First, we must create a data set that describes how the design will be processed. We call this
data set the design key.
In this example, we want a choice design with two factors, Brand and Price. Brand has levels ’Sploosh’,
’Plumbbob’, ’Platter’, ’Moosey’, and ’Another’. Price has levels $1.49, $1.99, and $2.49. Brand and Price
are created by different processes. The Price factor will be constructed from the factors of the linear design
matrix. In contrast, there is no Brand factor in the linear design. Each brand is a bin into which its factors are
collected. The variable Brand will be named on the alt= option of the %MktRoll macro as the alternative
variable, so its values will be read directly out of the KEY data set. Price will not be named on the alt=
macro option, so its values in the KEY data set are variable names from the linear design data set. The values of
Price in the final choice design will be read from the named variables in the linear design data set. The Price
factor in the choice design is created from the four linear design factors (x1 for Sploosh, x2 for Plumbbob, x3
for Platter, x4 for Moosey, and no attribute for Another, the constant alternative).
Fabric Softener Example 103

Here is how the KEY data set is created. The Brand factor levels and the Price linear design factors are stored
in the KEY data set.
title2 ’Key Data Set’;

data key;
input Brand $ Price $;
datalines;
Sploosh x1
Plumbbob x2
Platter x3
Moosey x4
Another .
;

proc print; run;

title2;

Choice of Fabric Softener

Key Data Set

Obs Brand Price

1 Sploosh x1
2 Plumbbob x2
3 Platter x3
4 Moosey x4
5 Another

Note that the value of Price for alternative Another in the KEY data set is blank (character missing). The period
in the in-stream data set is simply a placeholder, used with list input to read both character and numeric missing
data. A period is not stored with the data. Next, we use the %MktRoll macro to process the design.
%mktroll(design=[Link], key=key, alt=brand, out=rolled)

The %MktRoll step processes the design=[Link] linear design data set using the rules specified in
the key=key data set, naming the alt=brand variable as the alternative name variable, and creating an output
SAS data set called ROLLED, which contains the choice design. The input design=[Link] data set
has 18 observations, one per choice set, and the output out=rolled data set has 5  18 = 90 observations,
one for each alternative of each choice set. Here are the first three observations of the linear design data set.
title2 ’Linear Design (First 3 Sets)’;

proc print data=[Link](obs=3); run;

title2;

Choice of Fabric Softener

Linear Design (First 3 Sets)

Obs x1 x2 x3 x4

1 $2.49 $1.99 $1.99 $2.49

2 $1.49 $2.49 $1.99 $1.99
3 $1.49 $1.99 $1.99 $1.49
104 TS-677E Multinomial Logit, Discrete Choice Modeling

These observations define the first three choice sets. Here are those same observations, arrayed for analysis in
the choice design data set.
title2 ’Choice Design (First 3 Sets)’;

proc print data=rolled(obs=15); format price price.; run;

title2;

Choice of Fabric Softener

Choice Design (First 3 Sets)

Obs Set Brand Price

1 1 Sploosh $2.49
2 1 Plumbbob $1.99
3 1 Platter $1.99
4 1 Moosey $2.49
5 1 Another $1.99
6 2 Sploosh $1.49
7 2 Plumbbob $2.49
8 2 Platter $1.99
9 2 Moosey $1.99
10 2 Another $1.99
11 3 Sploosh $1.49
12 3 Plumbbob $1.99
13 3 Platter $1.99
14 3 Moosey $1.49
15 3 Another $1.99

The choice design data set has a choice set variable Set, an alternative name variable Brand, and a price
variable Price. The prices come from the linear design, and the price for Another is a constant $1.99. Recall
that the prices are assigned by the following format.
proc format; /* create a format for the price */
value price 1 = ’$1.49’ 2 = ’$1.99’ 3 = ’$2.49’ . = ’$1.99’;
run;
The next step merges the choice data with the choice design using the %MktMerge macro.
%mktmerge(design=rolled, data=results, out=res2,
nsets=&n, nalts=&m, setvars=choose1-choose&n)
This step reads the design=rolled choice design and the data=results data set and creates the
out=res2 output data set. The data are from an experiment with nsets=&n choice sets, nalts=&m al-
ternatives, with variables setvars=choose1-choose&n containing the numbers of the chosen alternatives.
Here are the first 15 observations.
title2 ’Choice Design and Data (First 3 Sets)’;

proc print data=res2(obs=15); run;

title2;
Fabric Softener Example 105

Choice of Fabric Softener

Choice Design and Data (First 3 Sets)

Obs Subj Set Brand Price c

1 1 1 Sploosh 3 2
2 1 1 Plumbbob 2 2
3 1 1 Platter 2 1
4 1 1 Moosey 3 2
5 1 1 Another . 2
6 1 2 Sploosh 1 2
7 1 2 Plumbbob 3 2
8 1 2 Platter 2 1
9 1 2 Moosey 2 2
10 1 2 Another . 2
11 1 3 Sploosh 1 2
12 1 3 Plumbbob 2 2
13 1 3 Platter 2 2
14 1 3 Moosey 1 1
15 1 3 Another . 2

The data set contains the subject ID variable Subj from the data=results data set, the Set, Brand, and
Price variables from the design=rolled data set, and the variable c, which indicates which alternative was
chosen. The variable c indicates the chosen alternatives: 1 for first choice and 2 for second or subsequent choice.
This subject chose the third alternative, Platter, for each of the first two choice sets, and Moosey for the third.
This data set has 4500 observations: 50 subjects times 18 choice sets times 5 alternatives.
Since we did not specify a format, we see in the design the raw design values for Price: 1, 2, 3 and missing
for the constant alternative. If we were going to treat Price as a categorical variable for analysis, this would
be fine. We would simply assign our price format to Price and designate it as a class variable. However, in
this analysis we are going to treat price as quantitative and use the actual prices in the analysis. Hence, we must
convert our design values of 1, 2, 3, and . to 1.49, 1.99, 2.49, and 1.99. We cannot do this by simply assigning
a format. Formats create character strings that are printed in place of the original value. We need to convert a
numeric variable from one set of numbers to another. We could use if and assignment statements. We could
also use the %MktLab macro, which is used in later examples. However, instead we will use the put function
to write the formatted value into a character string, then we read it back using a dollar format and the input
function. For example, the expression put(price, price.) converts a number, say 2, into a string (in this
case ’$1.99’), then the input function reads the string and converts it to a numeric 1.99. This step also assigns
a label to the variable Price.
data res3; /* Create a numeric actual price */
set res2;
price = input(put(price, price.), dollar5.);
label price = ’Price’;
run;

Binary Coding
One more thing must be done to these data before they can be analyzed. A binary or zero-one design matrix must
be coded for the brand effect. This can be done with PROC TRANSREG.
proc transreg design=5000 data=res3 nozeroconstant norestoremissing;
model class(brand / zero=none order=data)
identity(price) / lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
id subj set c;
run;
106 TS-677E Multinomial Logit, Discrete Choice Modeling

The design option specifies that no model is fit; the procedure is just being used to code a design. When
design is specified, dependent variables are not required. Optionally, design can be followed by “=n” where
n is the number of observations to process at one time. By default, PROC TRANSREG codes all observations
in one big group. For very large data sets, this can consume large amounts of memory and time. Processing
blocks of smaller numbers of observations is more efficient. The option design=5000 processes observations
in blocks of 5000. For smaller computers, try something like design=1000.
The nozeroconstant and norestoremissing options are not necessary for this example but are included
here because sometimes they are very helpful in coding choice models. The nozeroconstant option specifies
that if the coding creates a constant variable, it should not be zeroed. The nozeroconstant option should
always be specified when you specify design=n because the last group of observations may be small and may
contain constant variables. The nozeroconstant option is also important if you do something like coding by
subj set because sometimes an attribute is constant within a choice set. The norestoremissing option
specifies that missing values should not be restored when the out= data set is created. By default, the coded
class variable contains a row of missing values for observations in which the class variable is missing. When
you specify the norestoremissing option, these observations contain a row of zeros instead. This option
is useful when there is a constant alternative indicated by missing values. Both of these options, like almost all
options in PROC TRANSREG, can be abbreviated to three characters (noz and nor).
The model statement names the variables to code and provides information about how they should be coded.
The specification class(brand / ...) specifies that the variable Brand is a classification variable and
requests a binary coding. The zero=none option creates binary variables for all categories. In contrast, by
default, a binary variable is not created for the last category the parameter for the last category is a structural
zero. The zero=none option is used when there are no structural zeros or when you want to see the structural
zeros in the multinomial logit parameter estimates table. The order=data option sorts the levels into the order
they were first encountered in the data set. The specification identity(price) specifies that Price is a
quantitative factor that should be analyzed as is (not expanded into dummy variables).
The lprefix=0 option specifies that when labels are created for the binary variables, zero characters of the
original variable name should be used as a prefix. This means that the labels are created only from the level
values. For example, ’Sploosh’ and ’Plumbbob’ are created as labels not ’Brand Sploosh’ and
’Brand Plumbbob’.
An output statement names the output data set and drops variables that are not needed. These variables do
not have to be dropped. However, since they are variable names that are often found in special data set types,
PROC PHREG prints warnings when it finds them. Dropping the variables prevents the warnings. Finally, the
id statement names the additional variables that we want copied from the input to the output data set. The next
steps print the first three coded choice sets.
proc print data=coded(obs=15) label;
title2 ’First 15 Observations of Analysis Data Set’;
id subj set c;
run;

title2;

Choice of Fabric Softener

First 15 Observations of Analysis Data Set

Subj Set c Sploosh Plumbbob Platter Moosey Another Price Brand

1 1 2 1 0 0 0 0 2.49 Sploosh
1 1 2 0 1 0 0 0 1.99 Plumbbob
1 1 1 0 0 1 0 0 1.99 Platter
1 1 2 0 0 0 1 0 2.49 Moosey
1 1 2 0 0 0 0 1 1.99 Another
Fabric Softener Example 107

1 2 2 1 0 0 0 0 1.49 Sploosh
1 2 2 0 1 0 0 0 2.49 Plumbbob
1 2 1 0 0 1 0 0 1.99 Platter
1 2 2 0 0 0 1 0 1.99 Moosey
1 2 2 0 0 0 0 1 1.99 Another
1 3 2 1 0 0 0 0 1.49 Sploosh
1 3 2 0 1 0 0 0 1.99 Plumbbob
1 3 2 0 0 1 0 0 1.99 Platter
1 3 1 0 0 0 1 0 1.49 Moosey
1 3 2 0 0 0 0 1 1.99 Another

Fitting the Multinomial Logit Model

The next step fits the discrete choice, multinomial logit model.
proc phreg data=coded outest=betas brief;
title2 ’Discrete Choice Model’;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;

title2;
The brief option requests a brief summary for the strata. As with the candy example, c*c(2) designates the
chosen and unchosen alternatives in the model statement. We specify the &- trgind macro variable for the
model statement independent variable list. PROC TRANSREG automatically creates this macro variable. It
contains the list of coded independent variables generated by the procedure. This is so you do not have to figure
out what names TRANSREG created and specify them. In this case, PROC TRANSREG sets &- trgind to
contain the following list.
BrandSploosh BrandPlumbbob BrandPlatter BrandMoosey BrandAnother Price

The ties=breslow option specifies a PROC PHREG model that has the same likelihood as the multinomial
logit model for discrete choice. The strata statement specifies that the combinations of Set and Subj indicate
the choice sets. This data set has 4500 observations consisting of 18  50 = 900 strata and five observations per
stratum.
Each subject rated 18 choice sets, but the multinomial logit model assumes each stratum is independent. That
is, the multinomial logit model assumes each person makes only one choice. The option of collecting only one
datum from each subject is too expensive to consider for many problems, so multiple choices are collected from
each subject, and the repeated measures aspect of the problem is ignored. This practice is typical, and it usually
works well.

Multinomial Logit Model Results

The output is shown next. (Recall that we used %phchoice(on)on page 79 to customize the output from
PROC PHREG.)
108 TS-677E Multinomial Logit, Discrete Choice Modeling

Choice of Fabric Softener

Discrete Choice Model

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 900 5 1 4

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 2896.988 1567.162

AIC 2896.988 1577.162
SBC 2896.988 1601.174

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 1329.8263 5 <.0001

Score 1197.0568 5 <.0001
Wald 608.1816 5 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Sploosh 1 -1.11167 0.20114 30.5462 <.0001

Plumbbob 1 -0.16186 0.16951 0.9117 0.3397
Platter 1 1.95389 0.14609 178.8686 <.0001
Moosey 1 0.78650 0.15517 25.6909 <.0001
Another 0 0 . . .
Price 1 -4.25545 0.19663 468.3659 <.0001
Fabric Softener Example 109

The procedure output begins with information about the data set, variables, and options. This is followed by
information about the 900 strata. Since the brief option was specified, this table contains one row for each
stratum pattern. In contrast, the default table would have 900 rows, one for each choice set and subject com-
bination. Each subject and choice set combination consists of a total of five observations, one that was chosen
and four that were not chosen. This pattern was observed 900 times. This table provides a check on data entry.
Unless we have an availability or allocation study (page 226) or a nonconstant number of alternatives in different
choice sets, we would expect to see one pattern of results where one of the m alternatives was chosen for each
choice set. If you do not observe this for a study like this, there was probably a mistake in the data entry or
processing.
The most to least preferred brands are: Platter, Moosey, Another, Plumbbob, and Sploosh. Increases in
price have a negative utility. For example, the predicted utility of Platter brand at $1.99 is xi which is
(0 0 1 0 0 $1:99) ( 1:11 0:16 1:95 0:79 0 4:26)0 = 1:95 + 1:99  4:26 = 6:53.
Since Price was analyzed as a quantitative factor, we can see for example that the utility of Platter at $1.89,
which was not in any choice set, is 1:95 + 1:89  4:26 = 6:10, which is a $0:10  4:26 = 0:43 increase in
utility.

Probability of Choice
These next steps compute the expected probability that each alternative is chosen within each choice set. This
code could easily be modified to compute expected market share for hypothetical marketplaces that do not di-
rectly correspond to the choice sets. Note however, that a term like “expected market share,” while widely used,
is a misnomer. Without purchase volume data, it is unlikely that these numbers would mirror true market share.
First, PROC SCORE is used to compute the predicted utility for each alternative.
proc score data=coded(where=(subj=1) drop=c)
score=betas type=parms out=p;
var &_trgind;
run;
The data set to be scored is named with the data= option, and the coefficients are specified in the option
score=beta. Note that we only need to read all of the choice sets once, since the parameter estimates were
computed in an aggregate analysis. This is why we specified where=(subj=1). We do not need xj ^ for each
of the different subjects. We dropped the variable c from the CODED data set since this name will be used by
PROC SCORE for the results (xj ^ ). The option type=parms specifies that the score= data set contains the
parameters in - TYPE- = ’PARMS’ observations. The output data set with the predicted utilities is named
P. Scoring is based on the coded variables from PROC TRANSREG, whose names are contained in the macro
variable &- trgind. The next step exponentiates xj ^ .
data p2;
set p;
p = exp(c);
run;

Next, exp(xj ^ ) is summed for each choice set.

proc means data=p2 noprint;
output out=s sum(p) = sp;
by set;
run;
P
Finally, each xj ^ is divided by ^.
m
j =1 xj
data p;
merge p2 s(keep=set sp);
by set;
p = p / sp;
keep brand set price p;
run;
110 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are the results for the first three choice sets.
proc print data=p(obs=15);
title2 ’Choice Probabilities for the First 3 Choice Sets’;
run;

title2;

Choice of Fabric Softener

Choice Probabilities for the First 3 Choice Sets

Obs Price Brand Set p

1 2.49 Sploosh 1 0.00426

2 1.99 Plumbbob 1 0.09238
3 1.99 Platter 1 0.76635
4 2.49 Moosey 1 0.02840
5 1.99 Another 1 0.10861
6 1.49 Sploosh 2 0.21061
7 2.49 Plumbbob 2 0.00772
8 1.99 Platter 2 0.53800
9 1.99 Moosey 2 0.16741
10 1.99 Another 2 0.07625
11 1.49 Sploosh 3 0.09176
12 1.99 Plumbbob 3 0.02825
13 1.99 Platter 3 0.23439
14 1.49 Moosey 3 0.61237
15 1.99 Another 3 0.03322

Custom Questionnaires
In this part of the example, a custom questionnaire is printed for each person. Previously, each subject saw the
same questionnaire, with the same choice sets, each containing the same alternatives, with everything in the same
order. In this example, the order of the choice sets and all alternatives within choice sets are randomized for each
subject. Randomizing avoids any systematic effects due to the order of the alternatives and choice sets. The
constant alternative is always printed last. If you have no interest in custom questionnaires, you can skip ahead
to page 116.
First, the macro variable &forms is created. It contains the number of separate questionnaires (or forms or
subjects, in this case 50). We can use the %MktEx macro to create a data set with one observation for each
alternative of each choice set for each person. The specification %mktex(&forms &n &mm1, n=&forms
* &n * &mm1) is %mktex(50 18 4, n=50 * 18 * 4) and creates a 50  18  4 full-factorial design.
Note that the n= specification allows expressions. The macro %MktLab is then used to assign the variable names
Form, Set, and Alt instead of the default x1 - x3. The data set is sorted by Form. Within Form, the choice
sets are sorted into a random order, and within choice set, the alternatives are sorted into a random order. The 72
observations for each choice set contain 18 blocks of 4 observations one block per choice set in a random order
and the 4 alternatives within each choice set, again in a random order. Note that we store these in a permanent
SAS data set so they will be available after the data are collected.
Fabric Softener Example 111

%let forms = 50;

title2 ’Create 50 Custom Questionnaires’;

*---Make the design---;

%mktex(&forms &n &mm1, n=&forms * &n * &mm1)

*---Assign Factor Names---;

%mktlab(data=design, vars=Form Set Alt)

*---Set up for Random Ordering---;

data [Link];
set final;
by form set;
retain r1;
if [Link] then r1 = uniform(7);
r2 = uniform(7);
run;

*---Random Sort---;
proc sort out=[Link](drop=r:); by form r1 r2; run;
proc print data=[Link](obs=16); run;

The first 16 observations in this data set are shown next.

Choice of Fabric Softener

Create 50 Custom Questionnaires

Obs Form Set Alt

1 1 4 4
2 1 4 2
3 1 4 3
4 1 4 1
5 1 8 4
6 1 8 2
7 1 8 1
8 1 8 3
9 1 7 4
10 1 7 3
11 1 7 2
12 1 7 1
13 1 14 4
14 1 14 1
15 1 14 3
16 1 14 2
112 TS-677E Multinomial Logit, Discrete Choice Modeling

The data set is transposed, so the resulting data set contains 50  18 = 900 observations, one per subject per
choice set. The alternatives are in the variables Col1-Col4. The first 18 observations, which contain the
ordering of the choice sets for the first subject, are shown next.
proc transpose data=[Link] out=[Link](drop=_name_);
by form notsorted set;
run;
proc print data=[Link](obs=18);
run;

Choice of Fabric Softener

Create 50 Custom Questionnaires

Obs Form Set COL1 COL2 COL3 COL4

1 1 4 4 2 3 1
2 1 8 4 2 1 3
3 1 7 4 3 2 1
4 1 14 4 1 3 2
5 1 9 2 1 4 3
6 1 11 4 1 2 3
7 1 17 3 2 4 1
8 1 1 4 1 2 3
9 1 12 1 3 2 4
10 1 13 2 1 4 3
11 1 18 3 1 4 2
12 1 3 3 1 4 2
13 1 2 3 4 2 1
14 1 5 3 4 1 2
15 1 15 1 3 4 2
16 1 6 3 2 1 4
17 1 16 2 1 4 3
18 1 10 4 1 3 2

The following DATA step prints the 50 custom questionnaires.

options ls=80 ps=60 nodate nonumber;
title;

data _null_;
array brands[&mm1] $ _temporary_
(’Sploosh’ ’Plumbbob’ ’Platter’ ’Moosey’);
array x[&mm1] x1-x&mm1;
array c[&mm1] col1-col&mm1;
format x1-x&mm1 price.;
file print linesleft=ll;
Fabric Softener Example 113

do frms = 1 to &forms;
do choice = 1 to &n;
if choice = 1 or ll < 12 then do;
put _page_;
put @60 ’Subject: ’ frms //;
end;
put choice 2. ’) Circle your choice of ’
’one of the following fabric softeners:’ /;
set [Link];
set [Link] point=set;
do brnds = 1 to &mm1;
put ’ ’ brnds 1. ’) ’ brands[c[brnds]] ’brand at ’
x[c[brnds]] +(-1) ’.’ /;
end;
put ’ 5) Another brand at $1.99.’ /;
end;
end;
stop;
run;
The loop do frms = 1 to &forms creates the 50 questionnaires. The loop do choice = 1 to &n
creates the alternatives within each choice set. On the first choice set and when there is not enough room
for the next choice set, we skip to a new page (put - page- ) and print the subject (forms) number. The
data set [Link] is read and the Set variable is used to read the relevant observation from
[Link] using the point= option in the set statement. The order of the alternatives is in the
c array and variables col1-col&mm1 from the [Link] data set. In the first observation of
[Link], Set=16, Col1=2, Col2=3, Col3=1, and Col4=4. The first brand, is c[brnds] =
c[1] = col1 = 2, so brands[c[brnds]] = brands[c[1]] = brands[2] = ’Plumbbob’,
and the price, from observation Set=16 of [Link], is x[c[brnds]] = x[2] = $1.99. The
second brand, is c[brnds] = c[2] = col2 = 3, so brands[c[brnds]] = brands[c[2]] =
brands[3] = ’Platter’, and the price, from observation Set=16 of [Link], is x[c[brnds]]
= x[3] = $1.49.
In the interest of space, only the first two choice sets are printed. Note that the subject number is printed on the
form. This information is needed to restore all data to the original order.

Subject: 1

1) Circle your choice of one of the following fabric softeners:

1) Moosey brand at $1.49.

2) Plumbbob brand at $2.49.

3) Platter brand at $2.49.

4) Sploosh brand at $2.49.

5) Another brand at $1.99.

114 TS-677E Multinomial Logit, Discrete Choice Modeling

2) Circle your choice of one of the following fabric softeners:

1) Moosey brand at $1.49.

2) Plumbbob brand at $1.49.

3) Sploosh brand at $1.49.

4) Platter brand at $2.49.

5) Another brand at $1.99.

Processing the Data for Custom Questionnaires

Here are the data. (Actually, these are the data that would have been collected if the same people as in the
previous situation made the same choices, without error and uninfluenced by order effects.) Before these data
are analyzed, the original order must be restored.
title ’Choice of Fabric Softener’;

data results; /* read choice data set */

input Subj (choose1-choose&n) (1.) @@;
datalines;
1 532144442413142131 2 512111233124212431 3 122111234233443322
4 141153221441433422 5 114112352113441523 6 243134342341213233
7 324132423331441234 8 233333444114324111 9 132244342211532121
10 123313551442431124 11 432413112334322232 12 344144441133222325
13 241112415231331332 14 234311233124243122 15 431114535444121411
16 145431412414225122 17 352214223453322244 18 153214343455431432
19 311424145334343233 20 322242435132322453 21 422313351152215321
22 144434352124221354 23 431322144242341234 24 234313521411125323
25 222341432223153413 26 452315344524342414 27 512514432245223241
28 223133334224321424 29 134244525425322423 30 442434215232335231
31 425133234223121332 32 125322244221121142 33 341335431241533111
34 423142124231241323 35 524415342341154511 36 431241423531214311
37 151222121421141413 38 342143531454133541 39 342124325132141235
40 442421422321313223 41 411341134114122434 42 124333224232434133
43 342222124412221521 44 331422224311344343 45 523222441214141453
46 332325444322322225 47 442244212532412252 48 235451235314313322
49 133111422252312124 50 242434413113342131
;
The data set is transposed, and the original order is restored.
proc transpose data=results /* create one obs per choice set */
out=res2(rename=(col1=choose) drop=_name_);
by subj;
run;
data res3(keep=subj set choose);
array c[&mm1] col1-col&mm1;
merge [Link] res2;
if choose < 5 then choose = c[choose];
run;

proc sort;
by subj set;
run;
Fabric Softener Example 115

The actual choice number, stored in Choose, indexes the alternative numbers from [Link] to
restore the original alternative orders. For example, for the first subject, the first choice was 5, which is the
Another alternative. The second choice was 3. The data set [Link] shows in the second observation
that this choice of 3 corresponds to the first alternative (in the third column variable, Col3 = 1) of choice set
Set= 8. This DATA step writes out the data after the original order has been restored. It matches the data on
page 102.
data _null_;
set res3;
by subj;
if [Link] then do;
if mod(subj, 3) eq 1 then put;
put subj 4. +1 @@;
end;
put choose 1. @@;
run;

1 334523313433343413 2 314423343433333215 3 333423323323333233

4 314432343444343254 5 315423313421543213 6 334433344423323233
7 334433343432343413 8 334423323434343214 9 333433313351343213
10 345423325322343233 11 333433343433323213 12 314423343423333253
13 534423313322343213 14 214421144423323214 15 334423345433343235
16 333433345333335313 17 534423343452343413 18 344435544432343513
19 334433345432343433 20 311423343422325513 21 334453543423543213
22 314423353323333513 23 333423343333333433 24 545422323322323213
25 354433343433333313 26 314425525422343214 27 334353523423353213
28 334433313333333233 29 333423543335353234 30 334453343533343433
31 354423344322333413 32 354422343323333213 33 314423343352343215
34 334423343443333213 35 314553344453343215 36 333433544433343233
37 314423343424543214 38 353433324423353533 39 333453323333323513
40 314433343422333214 41 334423344442343444 42 334433323422323213
43 333423354433343213 44 314423323422333213 45 314452544422343214
46 414423345423543214 47 544423544442343414 48 335453343423323453
49 314523344424333214 50 334423343332343413

The data can be combined with the design and analyzed as in the previous example.
116 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example
This example illustrates the design and analysis for a larger problem. We will discuss designing a choice experi-
ment, evaluating the design, generating the questionnaire, processing the data, binary coding, generic attributes,
quantitative price effects, quadratic price effects, effects coding, alternative-specific effects, analysis, and inter-
pretation of the results.
A researcher is interested in studying choice of vacation destinations. There are five destinations (alternatives)
of interest: Hawaii, Alaska, Mexico, California, and Maine. Here are two summaries of the design, with factors
grouped by attribute and grouped by destination.

Factor Destination Attribute Levels

X1 Hawaii Accommodations Cabin, Bed & Breakfast, Hotel
X2 Alaska Accommodations Cabin, Bed & Breakfast, Hotel
X3 Mexico Accommodations Cabin, Bed & Breakfast, Hotel
X4 California Accommodations Cabin, Bed & Breakfast, Hotel
X5 Maine Accommodations Cabin, Bed & Breakfast, Hotel

X6 Hawaii Scenery Mountains, Lake, Beach

X7 Alaska Scenery Mountains, Lake, Beach
X8 Mexico Scenery Mountains, Lake, Beach
X9 California Scenery Mountains, Lake, Beach
X10 Maine Scenery Mountains, Lake, Beach

X11 Hawaii Price $999, $1249, $1499

X12 Alaska Price $999, $1249, $1499
X13 Mexico Price $999, $1249, $1499
X14 California Price $999, $1249, $1499
X15 Maine Price $999, $1249, $1499

Factor Destination Attribute Levels

X1 Hawaii Accommodations Cabin, Bed & Breakfast, Hotel
X6 Scenery Mountains, Lake, Beach
X11 Price $999, $1249, $1499

X2 Alaska Accommodations Cabin, Bed & Breakfast, Hotel

X7 Scenery Mountains, Lake, Beach
X12 Price $999, $1249, $1499

X3 Mexico Accommodations Cabin, Bed & Breakfast, Hotel

X8 Scenery Mountains, Lake, Beach
X13 Price $999, $1249, $1499

X4 California Accommodations Cabin, Bed & Breakfast, Hotel

X9 Scenery Mountains, Lake, Beach
X14 Price $999, $1249, $1499

X5 Maine Accommodations Cabin, Bed & Breakfast, Hotel

X10 Scenery Mountains, Lake, Beach
X15 Price $999, $1249, $1499
Vacation Example 117

Each alternative is composed of three factors: package cost ($999, $1,249, $1,499), scenery (mountains, lake,
beach), and accommodations (cabin, bed & breakfast, and hotel). There are five destinations, each with three
attributes, for a total of 15 factors. This problem requires a design with 15 three-level factors, denoted 315 . Each
row of the design matrix contains the description of the five alternatives in one choice set. Note that the levels do
not have to be the same for all destinations. For example, the cost for Hawaii and Alaska could be different from
the other destinations. However, for this example, each destination will have the same attributes.

Set Up
We can use the %MktRuns autocall macro to suggest design sizes. (All of the autocall macros used in this report
are documented starting on page 287.) To use this macro, you specify the number of levels for each of the factors.
With 15 attributes each with three prices, you specify fifteen 3’s (3 3 3 3 3 3 3 3 3 3 3 3 3 3 3) or
you can use the more compact syntax of 3 ** 15.
title ’Vacation Example’;

%mktruns( 3 ** 15 )
The output tells us the size of the saturated design, which is the number of parameters in the linear design, and
suggests design sizes.

Vacation Example

Design Summary

Number of
Levels Frequency

3 15

Vacation Example

Saturated = 31
Full Factorial = 14,348,907

Some Reasonable Cannot Be

Design Sizes Violations Divided By

36 0
45 0
54 * 0
63 0
72 * 0
33 105 9
39 105 9
42 105 9
48 105 9
51 105 9

* - 100% Efficient Design can be made with the MktEx Macro.

Vacation Example

n Design Reference

54 2 ** 1 3 ** 25 Taguchi, 1987
54 3 ** 24 6 ** 1 Hedayat, Sloane, and Stufken, 1999
54 3 ** 18 18 ** 1 Hedayat, Sloane, and Stufken, 1999
118 TS-677E Multinomial Logit, Discrete Choice Modeling

72 2 ** 23 3 ** 24 Dey, 1985
72 2 ** 20 3 ** 24 4 ** 1 Wang, 1996
72 2 ** 16 3 ** 25 Wang, 1996
72 2 ** 14 3 ** 24 6 ** 1 Wang, 1996
72 2 ** 13 3 ** 25 4 ** 1 Wang, 1996
72 2 ** 12 3 ** 24 12 ** 1 Hedayat, Sloane, and Stufken, 1999
72 2 ** 11 3 ** 24 4 ** 1 6 ** 1 Wang, 1996
72 3 ** 25 8 ** 1 Hedayat, Sloane, and Stufken, 1999
72 3 ** 24 24 ** 1 Hedayat, Sloane, and Stufken, 1999

In this design, there are 15  (3 1) + 1 = 31 parameters, so at least 31 choice sets must be created. With
all three-level factors, the number of choice sets in all orthogonal and balanced designs must be divisible by
3  3 = 9. Hence, optimal designs for this problem have at least 36 choice sets (the smallest number  31 and
divisible by 9) and the number of choice sets must be a multiple of 9. Note however, that zero violations does
not imply that a 100% efficient design exists. It just means that 100% efficiency is not precluded by unequal
frequencies. In fact. the %MktEx orthogonal design catalogue does not include orthogonal designs for this
problem in 36, 45, and 63 runs (because they do not exist).
Thirty-six would be a good design size (2 blocks of size 18) as would 54 (3 blocks of size 18). Fifty-four would
probably be the best choice, and that is what we would recommend for this study. However, we will instead
create an efficient experimental design with 36 choice sets using the %MktEx macro. In practice, with more
difficult designs, an orthogonal design is not available, and using 36 choice sets will allow us to see an example
of using the %mkt family of macros to get nonorthogonal designs.
We can see what orthogonal designs with three-level factors are available in 36 runs as follows. The macro
%MktOrth creates a data set with information about the orthogonal designs that the %MktEx macro knows how
to make. This macro produces a data set called MKTDESLEV that contains variables n, the number of runs;
Design, a description of the design; and Reference, one of the (sometimes many) references for the design.
In addition, there are variables: x1, the number of 1-level factors (which is always zero); x2, the number of
2-level factors; x3, the number of 3-level factors; and so on. We can sort this data set, excluding all but the
36-run designs, such that designs with the most three-level factors are printed first.
%mktorth;

proc sort data=mktdeslev out=list(drop=x:);

by descending x3;
where n = 36;
run;

proc print; run;

Vacation Example 119

Vacation Example

Obs n Design Reference

1 36 2 ** 4 3 ** 13 Taguchi, 1987
2 36 3 ** 13 4 ** 1 Dey, 1985
3 36 2 ** 11 3 ** 12 Taguchi, 1987
4 36 2 ** 2 3 ** 12 6 ** 1 Wang and Wu, 1991
5 36 3 ** 12 12 ** 1 Wang and Wu, 1991
6 36 2 ** 1 3 ** 8 6 ** 2 Zhang, Lu, and Pang, 1999
7 36 3 ** 7 6 ** 3 Finney, 1982
8 36 2 ** 13 3 ** 4 Suen, 1989
9 36 2 ** 4 3 ** 3 6 ** 1 Hedayat, Sloane, and Stufken, 1999
10 36 2 ** 20 3 ** 2 Hedayat, Sloane, and Stufken, 1999
11 36 2 ** 11 3 ** 2 6 ** 1 Hedayat, Sloane, and Stufken, 1999
12 36 2 ** 2 3 ** 2 6 ** 2 Hedayat, Sloane, and Stufken, 1999
13 36 2 ** 27 3 ** 1 Hedayat, Sloane, and Stufken, 1999
14 36 2 ** 18 3 ** 1 6 ** 1 Hedayat, Sloane, and Stufken, 1999
15 36 2 ** 9 3 ** 1 6 ** 2 Hedayat, Sloane, and Stufken, 1999
16 36 2 ** 35 Hadamard
17 36 2 ** 13 9 ** 1 Suen, 1989
18 36 2 ** 2 18 ** 1 Hedayat, Sloane, and Stufken, 1999
19 36 2 ** 1 6 ** 3 SAS Procedure OPTEX

There are 13 two-level factors available in 36 runs, and we need 15, so we would expect to make a pretty good
nonorthogonal design.

Designing the Choice Experiment

The following code creates a design.
%let m = 6; /* m alts including constant */
%let mm1 = %eval(&m - 1); /* m - 1 */
%let n = 18; /* number of choice sets per person */
%let blocks = 2; /* number of blocks */

%mktex(3 ** 15 2, n=&n * &blocks, seed=7654321)

The specification 3 ** 15 requests a design with 15 factors, x1 x15, each with three levels. This specifica-
tion also requests a two-level factor (the 2 following the 3 ** 15). This is because 36 choice sets may be too
many for one person to rate, so we may want to block the design into two blocks, and we can use a two-level
factor to do this. A design with 18  2 = 36 runs is requested, which will mean 36 choice sets. A random number
seed is explicitly specified so we will be able to reproduce these exact results.
Here are some of the log messages. The macro searches a fractional-factorial candidate set of 81 runs, and it also
generates a tabled design in 36 runs to try as part of the design. This will be explained in more detail on page
123.
NOTE: Generating the candidate set.
NOTE: Performing 20 searches of 81 candidates, full-factorial=28,697,814.
NOTE: Generating the tabled design, n=36.
120 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are some of the results from the %MktEx macro.

Vacation Example

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 82.2172 82.2172 Can
1 End 82.2172

2 Start 78.5039 Tab,Ran

2 5 14 83.2098 83.2098
2 6 14 83.3917 83.3917
2 6 15 83.5655 83.5655
2 7 14 83.7278 83.7278
2 7 15 84.0318 84.0318
2 7 15 84.3370 84.3370
2 8 14 85.1449 85.1449
.
.
.
2 End 98.0624

3 Start 79.1390 Tab,Ran

3 19 15 98.1101 98.1101
3 19 15 98.6368 98.6368
3 End 98.6368

.
.
.

5 Start 79.6983 Tab,Ran

5 30 14 98.8933 98.8933
5 End 98.8933

.
.
.

11 Start 80.5274 Tab,Ran

11 End 98.4043

12 Start 51.8915 Ran,Mut,Ann

12 End 93.0214

.
.
.

21 Start 47.6990 Ran,Mut,Ann

21 End 93.6096
Vacation Example 121

Vacation Example

Design Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 98.8933 98.8933 Ini

1 Start 80.4296 Tab,Ran

1 End 98.8567

2 Start 82.0528 Tab,Ran

2 End 98.7672

.
.
.

43 Start 80.4332 Tab,Ran

43 35 14 98.9438 98.9438
43 End 98.9438

.
.
.

62 Start 74.7712 Tab,Ran

62 16 15 98.9438 98.9438
62 End 98.9438

.
.
.

84 Start 81.9302 Tab,Ran

84 9 15 98.9438 98.9438
84 End 98.9438

.
.
.

124 Start 85.2024 Tab,Ran

124 21 14 98.9438 98.9438
124 End 98.9438

.
.
.

149 Start 78.9082 Tab,Ran

149 31 14 98.9438 98.9438
149 End 98.9438

NOTE: Stopping since it appears that no improvement is possible.

122 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example

Design Refinement History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 98.9438 98.9438 Ini

1 Start 94.7490 Pre,Mut,Ann

1 End 92.1336

.
.
.

10 Start 90.7390 Pre,Mut,Ann

10 End 92.3775

Vacation Example

The OPTEX Procedure

Class Level Information

Class Levels -Values-

x1 3 1 2 3
x2 3 1 2 3
x3 3 1 2 3
x4 3 1 2 3
x5 3 1 2 3
x6 3 1 2 3
x7 3 1 2 3
x8 3 1 2 3
x9 3 1 2 3
x10 3 1 2 3
x11 3 1 2 3
x12 3 1 2 3
x13 3 1 2 3
x14 3 1 2 3
x15 3 1 2 3

Vacation Example

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 98.9437 97.9592 98.9743 0.9428

The %MktEx macro used 5.24 minutes and found a design that is almost 99% efficient. (Differences in the fourth
decimal place between the iteration history and the final table, in this case 98.9438 versus 98.9437, are due to
rounding error and differences in ridging strategies between the macro and PROC OPTEX and are nothing to
worry about.)
Vacation Example 123

The %MktEx Macro Algorithm

The %MktEx macro creates efficient linear experimental designs using several approaches. The macro will try
to create a tabled design, it will search a set of candidate runs (rows of the design), and it will use a coordinate-
exchange algorithm using both random initial designs and also a partial tabled design initialization. The macro
stops if at any time it finds a perfect, 100% efficient, orthogonal and balanced design. This first phase is the
algorithm search phase. In it, the macro determines which approach is working best for this problem. At the
end of this phase, the macro chooses the method that has produced the best design and performs another set of
iterations using exclusively the chosen approach. Finally, the macro performs a third set of iterations where it
takes the best design it found so far and tries to improve it.
The %MktEx macro can directly generate, without iterations, hundreds of different 100% D-efficient, orthogonal
and balanced, tabled designs. It does this using its design catalogue and many different general and ad hoc
algorithms. See page 328 for a list of some of the orthogonal designs that the %MktEx macro knows how to
make. The closest design that the macro knows how to make for this problem is 21 313 in 36 runs.
The candidate-set search has two parts. First, either PROC PLAN is run to create a full-factorial design for
small problems, or PROC FACTEX is run to create a fractional-factorial design for large problems. Either
way, this design is a candidate set that in the second part is searched by PROC OPTEX using the modified
Federov algorithm. A design is built from a selection of the rows of the candidate set (Federov, 1972; Cook and
Nachtsheim, 1980). The modified Federov algorithm considers each run in the design and each candidate run.
Candidate runs are swapped in and design runs are swapped out if the swap improves D-efficiency. In this case,
since the full-factorial design is large (over 28 million runs), the candidate-set search step calls PROC FACTEX
to make the candidate set and then PROC OPTEX to do the search. The Can line of the iteration history shows
that this step found a design that was 82.2172% efficient.
Next, the %MktEx macro uses the coordinate-exchange algorithm, based on Meyer and Nachtsheim (1995). The
coordinate-exchange algorithm considers each level of each factor, and considers the effect on D-efficiency of
changing a level (1 ! 2, or 1 ! 3, or 2 ! 1, or 2 ! 3, or 3 ! 1, or 3 ! 2, and so on). Exchanges that increase
efficiency are performed. In this step, the macro first tries to initialize the design with a tabled design (Tab) and
a random design (Ran) both. In this case, 14 of the 16 columns can be initialized with 14 columns of 24 313 , and
the other two columns are randomly initialized. Levels that are not orthogonally initialized may be exchanged
for other levels if the exchange increases efficiency. For example, the iteration history for this example shows
that the macro exchanged levels in row 5 column 14, row 6 column 14, row 6 column 15, and so on.
The initialization may be more complicated in other problems. Say you asked for the design 41 51 34 in 18 runs.
The macro would use the tabled design 36 61 in 18 runs to initialize the three-level factors orthogonally, and
the five-level factor with the six-level factor coded down to five levels (and hence unbalanced). The four-level
factor would be randomly initialized. The macro would also try the same initialization but with a random rather
than unbalanced initialization of the five-level factor, as a minor variation on the first initialization. In the next
initialization variation, the macro would use a fully random initialization. If the number of runs requested were
smaller than the number or runs in the initial tabled design, the macro would initialize the design with just the
first n rows of the tabled design. Similarly, if the number of runs requested were larger than the number or runs
in the initial tabled design, the macro would initialize part of the design with the orthogonal tabled design and the
remaining rows and columns randomly. The coordinate-exchange algorithm considers each level of each factor
that is not orthogonally initialized, and it exchanges a level if the exchange improves D-efficiency. When the
number or runs in the tabled design does not match the number of runs desired, none of the design is initialized
orthogonally.
The coordinate-exchange algorithm is not restricted by having a candidate set and hence can potentially consider
every possible design. In practice, however, both the candidate-set-based and coordinate-exchange algorithms
consider only a tiny fraction of the possible designs. When the number of runs in the full-factorial design is small
(up to several thousand), the modified Federov algorithm is usually superior to coordinate exchange. When the
full-factorial design is larger, coordinate exchange is usually the superior approach. However, heuristics like
this are sometimes wrong, which is why the macro tries both methods to see which one is really best for each
problem.
124 TS-677E Multinomial Logit, Discrete Choice Modeling

In the first attempt at coordinate exchange (Design 2), the macro found a design that is 98.0624% efficient
(Design 2, End). In design 3 and subsequent designs, the macro uses this same approach, but different random
initializations of the remaining two columns. In design 5, the %MktEx macro finds a design that is 98.8933%
efficient. Designs 12 through 21 use a purely random initialization and simulated annealing and are not as good
as previous designs. During these iterations, the macro is considering exchanging every level of every factor with
all of the other levels, one level of one factor at a time.
At this point, the %MktEx macro determines that the combination of tabled and random initialization is working
best and tries more iterations using that approach. It starts by printing the initial (Ini) best efficiency of 98.8933.
In designs 43, 62, 84, 124, and 149, it finds a design that is 98.9438% efficient. After iteration 149, the macro
stops since it keeps finding the same design over and over. This does not necessarily mean the macro found the
optimal design; it means it found a very attractive (perhaps local) optimum, and it is unlikely it will do better
using this approach.
Next, the %MktEx macro tries to improve the best design it found previously. Using the previous best design
as an initialization (Pre), and random mutations of the initialization Mut) and simulated annealing (Ann), the
macro uses the coordinate-exchange algorithm to try to find a better design. This step is important because the
best design that the macro found may be an intermediate design and not be the final design at the end of an
iteration. Sometimes the iterations deliberately make the designs less efficient, and sometimes, the macro never
finds a design as efficient or more efficient again. Hence it is worthwhile to see if the best design found so far
can be improved. In this case the macro fails to improve the design. At the end, PROC OPTEX is called to print
the levels of each factor and the final D-efficiency.
Random mutations add random noise to the initial design before iterations start (levels are randomly changed).
This may eliminate the perfect balance that will often be in the initial design. By default, random mutations are
used with designs with fully random initializations and in the design refinement step; orthogonal initial designs
are not mutated.
Simulated annealing allows the design to get worse occasionally but with decreasing probability as the number
of swaps increases. For design 1, for the first level of the first factor, by default, the macro may execute a swap
(say change a 2 to a 1), that makes the design worse, with probability 0.05. As more and more swaps occur,
this probability decreases so at the end of the processing of design 1, swaps that decrease efficiency are hardly
ever done. For design 2, this same process is repeated, again starting by default with an annealing probability of
0.05. This often helps the algorithm overcome local efficiency maxima. To envision, this, imagine that you are
standing on a molehill next to a mountain. The only way you can start going up the mountain is to first step down
off the molehill. Once you are on the mountain, you may occasionally hit a dead end, where all you can do is step
down and look for a better place to continue going up. Simulated annealing, by occasionally stepping down the
efficiency function, often allows the macro to go farther up it than it would otherwise. The simulated annealing is
why you will sometimes see designs getting worse in the iteration history. Recall however, that the macro keeps
track of the best design, not the final design in each step. By default, annealing is used with designs with fully
random initializations and in the design refinement step; simulated annealing is not used with orthogonal initial
designs.
For this example, the %MktEx macro ran in less than 6 minutes. If an orthogonal design had been available, run
time would have been a few seconds. If the fully random initialization method had been the best method, run
time might have been on the order of 20 to 45 minutes. Since the tabled initialization worked best, run time was
on the order of several minutes. While it is possible to construct huge problems that will take much longer, for
any design that most marketing researchers are likely to encounter, run time should be less than one hour. One
of the macro options, maxtime=, ensures this.

Examining the Design

Before you use a design, you should always look at its characteristics. We will use the %MktEval macro.
%mkteval;
Vacation Example 125

Here are some of the results.

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15 x16

x1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x2 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x3 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
x4 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
x5 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
x6 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
x7 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
x8 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
x9 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
x10 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
x11 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
x12 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
x13 0 0 0 0 0 0 0 0 0 0 0 0 1 0.25 0.25 0
x14 0 0 0 0 0 0 0 0 0 0 0 0 0.25 1 0.25 0
x15 0 0 0 0 0 0 0 0 0 0 0 0 0.25 0.25 1 0
x16 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1
Vacation Example
Summary of Frequencies
There are 0 Canonical Correlations Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

x1 12 12 12
x2 12 12 12
x3 12 12 12
x4 12 12 12
x5 12 12 12
x6 12 12 12
x7 12 12 12
x8 12 12 12
x9 12 12 12
x10 12 12 12
x11 12 12 12
x12 12 12 12
x13 12 12 12
x14 12 12 12
x15 12 12 12
x16 18 18
x1 x2 4 4 4 4 4 4 4 4 4
x1 x3 4 4 4 4 4 4 4 4 4
x1 x4 4 4 4 4 4 4 4 4 4
x1 x5 4 4 4 4 4 4 4 4 4
x1 x6 4 4 4 4 4 4 4 4 4
x1 x7 4 4 4 4 4 4 4 4 4
x1 x8 4 4 4 4 4 4 4 4 4
x1 x9 4 4 4 4 4 4 4 4 4
x1 x10 4 4 4 4 4 4 4 4 4
x1 x11 4 4 4 4 4 4 4 4 4
x1 x12 4 4 4 4 4 4 4 4 4
x1 x13 4 4 4 4 4 4 4 4 4
x1 x14 4 4 4 4 4 4 4 4 4
x1 x14 4 4 4 4 4 4 4 4 4
x1 x15 4 4 4 4 4 4 4 4 4
x1 x16 6 6 6 6 6 6
126 TS-677E Multinomial Logit, Discrete Choice Modeling

.
.
.
* x13 x14 3 6 3 6 3 3 3 3 6
* x13 x15 6 3 3 3 6 3 3 3 6
x13 x16 6 6 6 6 6 6
* x14 x15 3 6 3 6 3 3 3 3 6
x14 x16 6 6 6 6 6 6
x15 x16 6 6 6 6 6 6
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

This design looks great! The factors x1-x13 form an orthogonal design, x14 and x15 are slightly correlated
with each other and with x13. The blocking factor x16 is orthogonal to all the other factors. All of the factors
are perfectly balanced. The N-Way frequencies show that each choice set appears once.
What if there had been some larger canonical correlations? Would this be a problem? That depends. You have
to decide this for yourself based on your particular study. You do not want large correlations between your most
important factors. If you have high correlations between the wrong factors, you can swap them with other factors
with the same number of levels, or try to make a new design with a different seed, or change the number of
choice sets, and so on. While this design looks great, we should make one minor adjustment based on these
results. Since our correlations are in the factors we originally planned to make price factors, we should change
our plans slightly and use those factors for less important attributes like scenery.
You can run the %MktEx macro to provide additional information about a design, for example asking to examine
the information matrix (i) and its inverse (v), which is the covariance matrix of the parameter estimates. You
hope to see that all of the off-diagonal elements of the variance matrix, the covariances, are small relative to the
variances on the diagonal. When options=check is specified, the macro evaluates an initial design instead of
generating a design. The option init=randomized names the design to evaluate, and the examine= option
displays the information and variance matrices. The blocking variable was dropped.
%mktex(3 ** 15, n=&n * &blocks, init=randomized(drop=x16),
options=check, examine=i v)
Here is a small part of the output.

Vacation Example

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 98.9099 97.8947 98.9418 0.9280
Vacation Example 127

Information Matrix

Intercept x11 x12 x21 x22 x31 x32 x41 x42 x51 x52

Intercept 36 0 0 0 0 0 0 0 0 0 0
x11 0 36 -0 -0 0 0 -0 0 -0 0 0
x12 0 -0 36 0 -0 -0 0 0 -0 0 -0
x21 0 -0 0 36 0 -0 -0 0 0 0 -0
x22 0 0 -0 0 36 0 -0 -0 -0 0 -0
x31 0 0 -0 -0 0 36 -0 -0 0 0 0
x32 0 -0 0 -0 -0 -0 36 0 0 0 -0
x41 0 0 0 0 -0 -0 0 36 -0 -0 -0
x42 0 -0 -0 0 -0 0 0 -0 36 0 -0
x51 0 0 0 0 0 0 0 -0 0 36 -0
.
.
.

Information Matrix

x112 x121 x122 x131 x132 x141 x142 x151 x152

x112 36 -0 -0 0 0 0 -0 -0 -0
x121 -0 36 -0 -0 -0 -0 -0 0 0
x122 -0 -0 36 -0 0 -0 -0 -0 -0
x131 0 -0 -0 36 -0 -5 -8 9 -0
x132 0 -0 0 -0 36 8 -5 -0 -9
x141 0 -0 -0 -5 8 36 -0 -5 -8
x142 -0 -0 -0 -8 -5 -0 36 -8 5
x151 -0 0 -0 9 -0 -5 -8 36 -0
x152 -0 0 -0 -0 -9 -8 5 -0 36

Variance Matrix

Intercept x11 x12 x21 x22 x31

Intercept 0.0278 -0.0000 -0.0000 -0.0000 -0.0000 -0.0000

x11 -0.0000 0.0278 0.0000 0.0000 -0.0000 -0.0000
x12 -0.0000 0.0000 0.0278 -0.0000 0.0000 0.0000
x21 -0.0000 0.0000 -0.0000 0.0278 -0.0000 0.0000
x22 -0.0000 -0.0000 0.0000 -0.0000 0.0278 -0.0000
x31 -0.0000 -0.0000 0.0000 0.0000 -0.0000 0.0278
.
.
.

Variance Matrix

x131 x132 x141 x142 x151 x152

x131 0.0309 0.0000 0.0031 0.0053 -0.0062 0.0000

x132 0.0000 0.0309 -0.0053 0.0031 0.0000 0.0062
x141 0.0031 -0.0053 0.0309 0.0000 0.0031 0.0053
x142 0.0053 0.0031 0.0000 0.0309 0.0053 -0.0031
x151 -0.0062 0.0000 0.0031 0.0053 0.0309 0.0000
x152 0.0000 0.0062 0.0053 -0.0031 0.0000 0.0309
128 TS-677E Multinomial Logit, Discrete Choice Modeling

This design still looks good. The D-efficiency for the design excluding the blocking factor is 98.9099%. We
can see that the nonorthogonality between x13-x15 make their variances larger than the other factors (0.0309
versus 0.0278).
This variance matrix is a little hard to look at. All of the 0.0000 and -0.0000’s tend to obscure the nonzeros. We
can use ODS along with PROC FORMAT and PROC PRINT to make a better display. The variance matrix is
excluded from the printed output and instead is output to a SAS data set. The persist options are used since
the ODS statements need to persist through the macro steps until the macro reaches the PROC OPTEX step. In
Version 9.0 and previous SAS versions, the match- all option must be specified with persist on the ods
output statement. PROC FORMAT is used to construct a format so that the values within rounding error of
zero print as ’0’. PROC PRINT is called to print the results. The label statement gives the row ID variable,
rowname a null header.
ods exclude ’variance matrix’(persist);
ods output ’variance matrix’(persist match_all)=v;
%mktex(3 ** 15, n=&n * &blocks, init=randomized(drop=x16),
options=check, examine=v)

proc format;
value zer -1e-8 - 1e-8 = ’ 0 ’;
run;

proc print label data=v(drop=_:);

format _numeric_ zer7.4;
label rowname = ’00’x;
id rowname;
run;

Vacation Example

Intercept x11 x12 x21 x22 x31 x32 x41

Intercept 0.0278 0 0 0 0 0 0 0
x11 0 0.0278 0 0 0 0 0 0
x12 0 0 0.0278 0 0 0 0 0
x21 0 0 0 0.0278 0 0 0 0
x22 0 0 0 0 0.0278 0 0 0
x31 0 0 0 0 0 0.0278 0 0
x32 0 0 0 0 0 0 0.0278 0
x41 0 0 0 0 0 0 0 0.0278
.
.
.

x122 x131 x132 x141 x142 x151 x152

x112 0 0 0 0 0 0 0
x121 0 0 0 0 0 0 0
x122 0.0278 0 0 0 0 0 0
x131 0 0.0309 0 0.0031 0.0053 -0.0062 0
x132 0 0 0.0309 -0.0053 0.0031 0 0.0062
x141 0 0.0031 -0.0053 0.0309 0 0.0031 0.0053
x142 0 0.0053 0.0031 0 0.0309 0.0053 -0.0031
x151 0 -0.0062 0 0.0031 0.0053 0.0309 0
x152 0 0 0.0062 0.0053 -0.0031 0 0.0309
Vacation Example 129

These next steps use the %MktLab macro to reassign the variable names, store the design in a permanent SAS
data set, [Link], and then use the %MktEx macro to check the results. The vars= option
provides the new variable names: the first variable (originally x1) becomes x1 (still), ..., the fifth variable (orig-
inally x5) becomes x5 (still), the sixth variable (originally x6) becomes x11, ... the tenth variable (originally
x10) becomes x15, the eleventh through fifteenth original variables become x6, x9, x7, x8, x10, and finally
the last variable becomes Block. We made the correlated variables correspond to the least important attributes
in different alternatives (in this case the scenery factors for Alaska, Mexico, and Maine).
%mktlab(data=randomized, vars=x1-x5 x11-x15 x6 x9 x7 x8 x10 Block,
out=[Link])

%mkteval(blocks=block)
Here is the output from the %MktLab macro, which shows the correspondence between the original and new
variable names.

Variable Mapping:
x1 : x1
x2 : x2
x3 : x3
x4 : x4
x5 : x5
x6 : x11
x7 : x12
x8 : x13
x9 : x14
x10 : x15
x11 : x6
x12 : x9
x13 : x7
x14 : x8
x15 : x10
x16 : Block

Here is some of the output from the %MktEval macro.

Vacation Example
Canonical Correlations Between the Factors
There are 0 Canonical Correlations Greater Than 0.316

Block x1 x2 x3 x4 x5 x11 x12 x13 x14 x15 x6 x9 x7 x8 x10

Block 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x2 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
x3 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
x4 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
x5 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
x11 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
x12 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
x13 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
x14 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
x15 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
x6 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
x9 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
x7 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0.25 0.25
x8 0 0 0 0 0 0 0 0 0 0 0 0 0 0.25 1 0.25
x10 0 0 0 0 0 0 0 0 0 0 0 0 0 0.25 0.25 1
130 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example
Summary of Frequencies
There are 0 Canonical Correlations Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

Block 18 18
x1 12 12 12
x2 12 12 12
x3 12 12 12
x4 12 12 12
x5 12 12 12
x11 12 12 12
x12 12 12 12
x13 12 12 12
x14 12 12 12
x15 12 12 12
x6 12 12 12
x9 12 12 12
x7 12 12 12
x8 12 12 12
x10 12 12 12

Block x1 6 6 6 6 6 6
Block x2 6 6 6 6 6 6
Block x3 6 6 6 6 6 6
Block x4 6 6 6 6 6 6
Block x5 6 6 6 6 6 6
Block x11 6 6 6 6 6 6
Block x12 6 6 6 6 6 6
Block x13 6 6 6 6 6 6
Block x14 6 6 6 6 6 6
Block x15 6 6 6 6 6 6
Block x6 6 6 6 6 6 6
Block x9 6 6 6 6 6 6
Block x7 6 6 6 6 6 6
Block x8 6 6 6 6 6 6
Block x10 6 6 6 6 6 6

x1 x2 4 4 4 4 4 4 4 4 4
x1 x3 4 4 4 4 4 4 4 4 4
x1 x4 4 4 4 4 4 4 4 4 4
.
.
.
x9 x7 4 4 4 4 4 4 4 4 4
x9 x8 4 4 4 4 4 4 4 4 4
x9 x10 4 4 4 4 4 4 4 4 4
* x7 x8 3 3 6 6 3 3 3 6 3
* x7 x10 6 3 3 3 3 6 3 6 3
* x8 x10 3 3 6 3 6 3 6 3 3
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Vacation Example 131

Generating the Questionnaire

This next DATA step prints the questionnaires. They are then copied and the data are collected.
title;
proc sort data=[Link]; by block; run;

options ls=80 ps=60 nodate nonumber;

data _null_;
array dests[&mm1] $ 10 _temporary_ (’Hawaii’ ’Alaska’ ’Mexico’
’California’ ’Maine’);
array prices[3] $ 5 _temporary_ (’$999’ ’$1249’ ’$1499’);
array scenes[3] $ 13 _temporary_
(’the Mountains’ ’a Lake’ ’the Beach’);
array lodging[3] $ 15 _temporary_
(’Cabin’ ’Bed & Breakfast’ ’Hotel’);
array x[15];
file print linesleft=ll;
set [Link];
by block;

if [Link] then do;

choice = 0;
put _page_;
put @50 ’Form: ’ block ’ Subject: ________’ //;
end;
choice + 1;
if ll < 19 then put _page_;
put choice 2. ’) Circle your choice of ’
’vacation destinations:’ /;
do dest = 1 to &mm1;
put ’ ’ dest 1. ’) ’ dests[dest]
+(-1) ’, staying in a ’ lodging[x[dest]]
’near ’ scenes[x[&mm1 + dest]] +(-1) ’,’ /
’ with a package cost of ’
prices[x[2 * &mm1 + dest]] +(-1) ’.’ /;
end;
put " &m) Stay at home this year." /;
run;
In this design, there are five destinations, and each destination has three attributes. Each destination name is
accessed from the array dests. Note that destination is not a factor in the design; it is a bin into which the
attributes are grouped. The factors in the design are named in the statement array x[15], which is a short-
hand notation for array x[15] x1-x15. The first five factors are used for the lodging attribute of the five
destinations. The actual descriptions of lodging are accessed by lodging[x[dest]]. The variable Dest
varies from 1 to 5 destinations, so x[dest] extracts the levels for the Dest destination. Similarly for scenery,
scenes[x[&mm1 + dest]] extracts the descriptions of the scenery. The index &mm1 + dest accesses
factors 6 through 10, and x[&mm1 + dest] indexes the scenes array. For prices, prices[x[2 * &mm1
+ dest]], the index 2 * &mm1 + dest accesses the factors 11 through 15. Here are the first two choice
sets.
132 TS-677E Multinomial Logit, Discrete Choice Modeling

Form: 1 Subject: ________

1) Circle your choice of vacation destinations:

1) Hawaii, staying in a Bed & Breakfast near a Lake,

with a package cost of $1249.

2) Alaska, staying in a Hotel near a Lake,

with a package cost of $1499.

3) Mexico, staying in a Cabin near the Mountains,

with a package cost of $1249.

4) California, staying in a Hotel near the Beach,

with a package cost of $1249.

5) Maine, staying in a Cabin near the Beach,

with a package cost of $1249.

6) Stay at home this year.

2) Circle your choice of vacation destinations:

1) Hawaii, staying in a Hotel near the Mountains,

with a package cost of $999.

2) Alaska, staying in a Bed & Breakfast near a Lake,

with a package cost of $1249.

3) Mexico, staying in a Hotel near the Mountains,

with a package cost of $999.

4) California, staying in a Cabin near the Mountains,

with a package cost of $999.

5) Maine, staying in a Hotel near the Beach,

with a package cost of $999.

6) Stay at home this year.

In practice, data collection may be much more elaborate than this. It may involve art work, photographs, and the
choice sets may be presented and data may be collected over the web. However the choice sets are presented and
the data collected, the essential ingredients remain the same. Subjects are shown sets of alternatives and asked to
make a choice, then they go on to the next set.
Vacation Example 133

Entering and Processing the Data

Here are some of the input data. Data from a total of 200 subjects were collected, 100 per form.
data results;
input Subj Form (choose1-choose&n) (1.) @@;
datalines;
1 1 151312311511353143 2 2 414435155112312411 3 1 131311331154453453
4 2 424434451134315111 5 1 131311111112442443 6 2 354434253152315111
7 1 141543311511154143 8 2 414415211434353511 9 1 111512331151412153
10 2 424333133154311511 11 1 151511331543123143 12 2 424155153114312511
13 1 151542331412114143 14 2 323414141152313111 15 1 151511311511413113
16 2 313435133154353113 17 1 154511311554152143 18 2 423434153121353111
19 1 111311131315144413 20 2 413435331112413111 21 1 131341331533413143
.
.
.
;
These next steps prepare the design for analysis. We need to create a data set KEY that describes how the factors
in our design will be used for analysis. It will contain all of the factor names, x1, x2, x3, ... x15. We can run
the %MktKey macro to get these names in the SAS log, for cutting and pasting into the program without typing
them.
%mktkey(x1-x15)

The %MktKey macro produced the following line.

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15
This code makes the KEY data set and processes the design.
title ’Vacation Example’;

data key;
input Place $ 1-10 (Lodge Scene Price) ($);
datalines;
Hawaii x1 x6 x11
Alaska x2 x7 x12
Mexico x3 x8 x13
California x4 x9 x14
Maine x5 x10 x15
Home . . .
;

%mktroll(design=[Link], key=key, alt=place, out=rolled)

For analysis, the design will have four factors as shown by the variables in the data set KEY. Place is the alter-
native name; its values are directly read from the KEY in-stream data. Lodge is an attribute whose values will
be constructed from the [Link] data set. Lodge is created from x1 for Hawaii, x2 for Alaska,
..., x5 for Maine, and no attribute for Home. Similarly, Scene is created from x6-x10, and Price is created
from x11-x15. The macro %MktRoll is used to create the data set ROLLED from [Link]
using the mapping in KEY and using the variable Place as the alternative ID variable. The macro warns us:
WARNING: The variable block is in the DESIGN= data set but not
the KEY= data set.
While this message could indicate a problem, in this case it does not. The variable Block in the de-
sign=[Link] data set will not appear in the final design. The purpose of the variable Block
(sorting the design into blocks) has already been achieved. These next steps show the results for the first two
choice sets. The data set is converted from a design matrix with one row per choice set to a design matrix with
one row per alternative per choice set.
134 TS-677E Multinomial Logit, Discrete Choice Modeling

proc print data=[Link](obs=2);

id Block;
var x1-x15;
run;

proc print data=rolled(obs=12); run;

Vacation Example

Block x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15

1 2 3 1 3 1 2 2 1 3 3 2 3 2 2 2
1 3 2 3 1 3 1 2 1 1 3 1 2 1 1 1

Vacation Example

Obs Set Place Lodge Scene Price

1 1 Hawaii 2 2 2
2 1 Alaska 3 2 3
3 1 Mexico 1 1 2
4 1 California 3 3 2
5 1 Maine 1 3 2
6 1 Home . . .
7 2 Hawaii 3 1 1
8 2 Alaska 2 2 2
9 2 Mexico 3 1 1
10 2 California 1 1 1
11 2 Maine 3 3 1
12 2 Home . . .

The next steps assign formats, convert the variable Price to contain actual prices, and recode the constant
alternative.
proc format;
value price 1 = ’ 999’ 2 = ’1249’
3 = ’1499’ 0 = ’ 0’;
value scene 1 = ’Mountains’ 2 = ’Lake’
3 = ’Beach’ 0 = ’Home’;
value lodge 1 = ’Cabin’ 2 = ’Bed & Breakfast’
3 = ’Hotel’ 0 = ’Home’;
run;

data rolled2;
set rolled;
if place = ’Home’ then do; lodge = 0; scene = 0; price = 0; end;
price = input(put(price, price.), 5.);
format scene scene. lodge lodge.;
run;

proc print data=rolled2(obs=12); run;

Vacation Example 135

Vacation Example

Obs Set Place Lodge Scene Price

1 1 Hawaii Bed & Breakfast Lake 1249

2 1 Alaska Hotel Lake 1499
3 1 Mexico Cabin Mountains 1249
4 1 California Hotel Beach 1249
5 1 Maine Cabin Beach 1249
6 1 Home Home Home 0
7 2 Hawaii Hotel Mountains 999
8 2 Alaska Bed & Breakfast Lake 1249
9 2 Mexico Hotel Mountains 999
10 2 California Cabin Mountains 999
11 2 Maine Hotel Beach 999
12 2 Home Home Home 0

It is not necessary to recode the missing values for the constant alternative. In practice, we usually will not do
this step. However, for this first analysis, we will want all nonmissing values of the attributes so we can see all
levels in the final printed output. We also recode Price so that for a later analysis, we can analyze Price as
a quantitative effect. For example, the expression put(price, price.) converts a number, say 2, into a
string (in this case ’1249’), then the input function reads the string and converts it to a numeric 1249. Next, we
use the macro %MktMerge to combine the data and design and create the variable c, indicating whether each
alternative was a first choice or a subsequent choice.
%mktmerge(design=rolled2, data=results, out=res2, blocks=form,
nsets=&n, nalts=&m, setvars=choose1-choose&n)

proc print data=res2(obs=12); run;

This macro takes the design=rolled2 experimental design, merges it with the data=result data set,
creating the out=res2 output data set. The RESULTS data set contains the variable Form that contains the
block number. Since there are two blocks, this variable must have values of 1 and 2. This variable must be
specified in the blocks= option. The experiment has nsets=&n choice sets, nalts=6 alternatives, and the
variables setvars=choose1-choose&n contain the numbers of the chosen alternatives. The output data
set RES2 has 21600 observations (200 subjects who each saw 18 choice sets with 6 alternatives). Here are the
first two choice sets.

Vacation Example

Obs Subj Form Set Place Lodge Scene Price c

1 1 1 1 Hawaii Bed & Breakfast Lake 1249 1

2 1 1 1 Alaska Hotel Lake 1499 2
3 1 1 1 Mexico Cabin Mountains 1249 2
4 1 1 1 California Hotel Beach 1249 2
5 1 1 1 Maine Cabin Beach 1249 2
6 1 1 1 Home Home Home 0 2
7 1 1 2 Hawaii Hotel Mountains 999 2
8 1 1 2 Alaska Bed & Breakfast Lake 1249 2
9 1 1 2 Mexico Hotel Mountains 999 2
10 1 1 2 California Cabin Mountains 999 2
11 1 1 2 Maine Hotel Beach 999 1
12 1 1 2 Home Home Home 0 2
136 TS-677E Multinomial Logit, Discrete Choice Modeling

Binary Coding
One more thing must be done to these data before they can be analyzed. The binary design matrix is coded for
each effect. This can be done with PROC TRANSREG.
proc transreg design=5000 data=res2 nozeroconstant norestoremissing;
model class(place / zero=none order=data)
class(price scene lodge / zero=none order=formatted) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
id subj set form c;
run;
The design option specifies that no model is fit; the procedure is just being used to code a design. When
design is specified, dependent variables are not required. Optionally, design can be followed by “=n” where
n is the number of observations to process at one time. By default, PROC TRANSREG codes all observations
in one big group. For very large data sets, this can consume large amounts of memory and time. Processing
blocks of smaller numbers of observations is more efficient. The option design=5000 processes observations
in blocks of 5000. For smaller computers, try something like design=1000.
The nozeroconstant and norestoremissing options are not necessary for this example but are included
here because sometimes they are very helpful in coding choice models. The nozeroconstant option specifies
that if the coding creates a constant variable, it should not be zeroed. The nozeroconstant option should
always be specified when you specify design=n because the last group of observations may be small and may
contain constant variables. The nozeroconstant option is also important if you do something like coding by
subj set because sometimes an attribute is constant within a choice set. The norestoremissing option
specifies that missing values should not be restored when the out= data set is created. By default, the coded
class variable contains a row of missing values for observations in which the class variable is missing. When
you specify the norestoremissing option, these observations contain a row of zeros instead. This option
is useful when there is a constant alternative indicated by missing values. Both of these options, like almost all
options in PROC TRANSREG, can be abbreviated to three characters (noz and nor).
The model statement names the variables to code and provides information about how they should be coded.
The specification class(place / ...) specifies that the variable Place is a classification variable and
requests a binary coding. The zero=none option creates binary variables for all categories. The order=data
option sorts the levels into the order they were first encountered in the data set. It is specified so ’Home’
will be the last destination in the analysis, as it is in the data set. The class(price scene lodge /
...) specification names the variables Price, Scene, and Lodge as categorical variables and creates binary
variables for all of the levels of all of the variables. The levels are sorted into order based on their formatted
values. The lprefix=0 option specifies that when labels are created for the binary variables, zero characters
of the original variable name should be used as a prefix. This means that the labels are created only from the
level values. For example, ’Mountains’ and ’Bed & Breakfast’ are created as labels not ’Scene
Mountains’ and ’Lodge Bed & Breakfast’.
An output statement names the output data set and drops variables that are not needed. These variables do
not have to be dropped. However, since they are variable names that are often found in special data set types,
PROC PHREG prints warnings when it finds them. Dropping the variables prevents the warnings. Finally, the
id statement names the additional variables that we want copied from the input to the output data set. The next
steps print the first coded choice set.
proc print data=coded(obs=6);
id place;
var subj set form c price scene lodge;
run;
proc print data=coded(obs=6) label;
var pl:;
run;
Vacation Example 137

proc print data=coded(obs=6) label;

id place;
var sc:;
run;
proc print data=coded(obs=6) label;
id place;
var lo: pr:;
run;

Vacation Example

Place Subj Set Form c Price Scene Lodge

Hawaii 1 1 1 1 1249 Lake Bed & Breakfast

Alaska 1 1 1 2 1499 Lake Hotel
Mexico 1 1 1 2 1249 Mountains Cabin
California 1 1 1 2 1249 Beach Hotel
Maine 1 1 1 2 1249 Beach Cabin
Home 1 1 1 2 0 Home Home

Vacation Example

Obs Hawaii Alaska Mexico California Maine Home Place

1 1 0 0 0 0 0 Hawaii
2 0 1 0 0 0 0 Alaska
3 0 0 1 0 0 0 Mexico
4 0 0 0 1 0 0 California
5 0 0 0 0 1 0 Maine
6 0 0 0 0 0 1 Home

Vacation Example

Place Beach Home Lake Mountains Scene

Hawaii 0 0 1 0 Lake
Alaska 0 0 1 0 Lake
Mexico 0 0 0 1 Mountains
California 1 0 0 0 Beach
Maine 1 0 0 0 Beach
Home 0 1 0 0 Home

Vacation Example

Bed &
Place Breakfast Cabin Home Hotel Lodge 0 999 1249 1499 Price

Hawaii 1 0 0 0 Bed & Breakfast 0 0 1 0 1249

Alaska 0 0 0 1 Hotel 0 0 0 1 1499
Mexico 0 1 0 0 Cabin 0 0 1 0 1249
California 0 0 0 1 Hotel 0 0 1 0 1249
Maine 0 1 0 0 Cabin 0 0 1 0 1249
Home 0 0 1 0 Home 1 0 0 0 0
138 TS-677E Multinomial Logit, Discrete Choice Modeling

The coded design consists of binary variables for destinations Hawaii Home, scenery Beach Mountains,
lodging Bed & Breakfast Hotel, and price 0 1499. For example, in the last printed panel of the first choice
set, the Bed & Breakfast column has a 1 for Hawaii since Hawaii has B & B lodging in this choice set. The Bed
& Breakfast column has a 0 for Alaska since Alaska does not have B & B lodging in this choice set. These binary
variables will form the independent variables in the analysis.
Note that we are fitting a model with generic attributes. Generic attributes are assumed to be the same for all
alternatives. For example, our model is structured so that the part-worth utility for being on a lake will be the same
for Hawaii, Alaska, and all of the other destinations. Similarly, the part-worth utilities for the different prices will
not depend on the destinations. In contrast, on page 146, using the same data, we will code alternative-specific
effects where the part-worth utilities are allowed by the model to be different for each of the destinations.
PROC PHREG is run in the usual way to fit the choice model.
proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
We specify the &- trgind macro variable for the model statement independent variable list. PROC TRANS-
REG automatically creates this macro variable. It contains the list of coded independent variables generated by
the procedure. This is so you do not have to figure out what names TRANSREG created and specify them. In
this case, PROC TRANSREG sets &- trgind to contain the following list.
PlaceHawaii PlaceAlaska PlaceMexico PlaceCalifornia PlaceMaine PlaceHome
Price0 Price999 Price1249 Price1499 SceneBeach SceneHome SceneLake
SceneMountains LodgeBed___Breakfast LodgeCabin LodgeHome LodgeHotel
The analysis is stratified by subject and choice set. Each stratum consists of a set of alternatives from which a
subject made one choice. In this example, each stratum consists of six alternatives, one of which was chosen
and five of which were not chosen. (Recall that we used %phchoice(on)on page 79 to customize the output
from PROC PHREG.) The full table of the strata would be quite large with one line for each of the 3600 strata,
so the brief option was specified on the PROC PHREG statement. This option produces a brief summary of
the strata. In this case, we see there were 3600 choice sets that all fit one response pattern. Each consisted of 6
alternatives, 1 of which was chosen and 5 of which were not chosen. There should be one pattern for all choice
sets in an example like this one the number of alternatives, number of chosen alternatives, and the number not
chosen should be constant.

Vacation Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 3600 6 1 5

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Vacation Example 139

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 12900.668 6365.337

AIC 12900.668 6387.337
SBC 12900.668 6455.412

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 6535.3316 11 <.0001

Score 5858.0723 11 <.0001
Wald 2158.0226 11 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Hawaii 1 3.59233 0.41717 74.1524 <.0001

Alaska 1 0.73317 0.42650 2.9552 0.0856
Mexico 1 2.73979 0.41843 42.8735 <.0001
California 1 2.04754 0.42037 23.7249 <.0001
Maine 1 1.34308 0.42439 10.0153 0.0016
Home 0 0 . . .
0 0 0 . . .
999 1 3.51899 0.08910 1559.9275 <.0001
1249 1 1.23293 0.07617 261.9979 <.0001
1499 0 0 . . .
Beach 1 1.41301 0.07173 388.0235 <.0001
Home 0 0 . . .
Lake 1 0.71264 0.06292 128.2840 <.0001
Mountains 0 0 . . .
Bed & Breakfast 1 0.66322 0.05778 131.7652 <.0001
Cabin 1 -1.53715 0.07178 458.5709 <.0001
Home 0 0 . . .
Hotel 0 0 . . .

The destinations, from most preferred to least preferred, are Hawaii, Mexico, California, Maine, Alaska, and then
stay at home. The utility for lower price is greater than the utility for higher price. The beach is preferred over a
lake, which is preferred over the mountains. A bed & breakfast is preferred over a hotel, which is preferred over a
cabin. Notice that the coefficients for the constant alternative (home and zero price) are all zero. Also notice that
for each factor, destination, price, scenery and accommodations, the coefficient for the last level is always zero.
This will always occur when we code with zero=none. The last level of each factor is a reference level, and
the other coefficients will have values relative to this zero. For example, all of the coefficients for the destination
are positive relative to the zero for staying at home. For scenery, all of the coefficients are positive relative to the
zero for the mountains. For accommodations, the coefficient for cabin is less than the zero for hotel, which is less
than the coefficient for bed & breakfast. In some sense, each class variable in a choice model with a constant
alternative has two reference levels or two levels that will always have a zero coefficient: the level corresponding
to the constant alternative and the level corresponding to the last level. At first, it is reassuring to run the model
with all levels represented to see that all the right levels get zeroed. Later, we will see ways to eliminate these
levels from the output.
140 TS-677E Multinomial Logit, Discrete Choice Modeling

Quantitative Price Effect

These data can also be analyzed in a different way. The Price variable can be specified directly as a quantitative
variable, instead of with indicator variables for a qualitative price effect. You could print the independent variable
list and copy and edit it, removing the Price indicator variables and adding Price.
%put &_trgind;

Alternatively, you could run PROC TRANSREG again with the new coding. We use this latter approach, because
it is easier, and it will allow us to illustrate other options. In the previous analysis, there were a number of
structural-zero parameter estimates in the results due to the usage of the zero=none option in the PROC
TRANSREG coding. This is a good thing, particularly for a first attempt at the analysis. It is good to specify
zero=none and check the results and make sure you have the right pattern of zeros and nonzeros. Later, you
can run again excluding some of the structural zeros. This time, we will explicitly specify the ’Home’ level in
the zero= option as the reference level so it will be omitted from the &- trgind variable list. The first class
specification specifies zero=’Home’ since there is one variable. The second class specification specifies
zero=’Home’ ’Home’ specifying the reference level for each of the two variables. The variable Price is
designated as an identity variable. The identity transformation is the no-transformation option, which
is used for variables that need to enter the model with no further manipulations. The identity variables are
simply copied into the output data set and added to the &- trgind variable list. The statement label price
= ’Price’ is specified to explicitly set a label for the identity variable price. This is because we explicitly
modified PROC PHREG output using %phchoice(on)so that the rows of the parameter estimate table would
be labeled only with variable labels not variable names. A label for Price must be explicitly specified in order
for the output to contain a label for the price effect.
proc transreg design data=res2 nozeroconstant norestoremissing;
model class(place / zero=’Home’ order=data) identity(price)
class(scene lodge / zero=’Home’ ’Home’ order=formatted) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’;
id subj set form c;
run;

proc phreg data=coded brief;

model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
Here are the results.

Vacation Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 3600 6 1 5
Vacation Example 141

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 12900.668 6445.698

AIC 12900.668 6465.698
SBC 12900.668 6527.585

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 6454.9702 10 <.0001

Score 5633.7761 10 <.0001
Wald 2243.9574 10 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Hawaii 1 14.21266 0.46894 918.5650 <.0001

Alaska 1 11.46095 0.45555 632.9626 <.0001
Mexico 1 13.38866 0.46367 833.7971 <.0001
California 1 12.70294 0.46109 758.9939 <.0001
Maine 1 12.12909 0.45754 702.7502 <.0001
Price 1 -0.00729 0.0001785 1670.5323 <.0001
Beach 1 1.31418 0.06946 357.9800 <.0001
Lake 1 0.68484 0.06252 120.0016 <.0001
Mountains 0 0 . . .
Bed & Breakfast 1 0.64743 0.05770 125.9040 <.0001
Cabin 1 -1.47644 0.06929 453.9843 <.0001
Hotel 0 0 . . .

The results of the two different analyses are similar. The coefficients for the destinations all increase by a non-
constant amount (approximately 10.65) but the pattern is the same. There is still a negative effect for price. Also,
the fit of this model is slightly worse, Chi-Square = 6454.9702 , compared to the previous value of 6535.3316
(bigger values mean better fit), because price has one fewer parameter.

Quadratic Price Effect

Previously, we saw price treated as a qualitative factor with two parameters and two df, then we saw price treated
as a quantitative factor with one parameter and one df. Alternatively, we could treat price as quantitative and
add a quadratic price effect (price squared). Like treating price as a qualitative factor, there are two parameters
and two df for price. First, we create PriceL, the linear price term by centering the original price and dividing
by the price increment (250). This maps (999, 1249, 1499) to (-1, 0, 1). Next, we run PROC TRANSREG and
PROC PHREG with the new price variables.
142 TS-677E Multinomial Logit, Discrete Choice Modeling

data res3;
set res2;
PriceL = price;
if price then pricel = (price - 1249) / 250;
run;

proc transreg design=5000 data=res3 nozeroconstant norestoremissing;

model class(place / zero=’Home’ order=data)
pspline(pricel / degree=2)
class(scene lodge / zero=’Home’ ’Home’ order=formatted) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label pricel = ’Price’;
id subj set form c;
run;
The pspline or polynomial spline expansion with the degree=2 option replaces the variable PriceL with
two coded variables, PriceL- 1 (which is the same as the original PriceL) and PriceL- 2 (which is Pri-
ceL squared). A degree=2 spline with no knots (neither knots= nor nknots= were specified) simply
expands the variable into a quadratic polynomial.
proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
This step produced the following results.

Vacation Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 3600 6 1 5

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 12900.668 6365.337

AIC 12900.668 6387.337
SBC 12900.668 6455.412
Vacation Example 143

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 6535.3316 11 <.0001

Score 5858.0723 11 <.0001
Wald 2158.0226 11 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Hawaii 1 4.82525 0.41472 135.3715 <.0001

Alaska 1 1.96610 0.42218 21.6877 <.0001
Mexico 1 3.97272 0.41540 91.4641 <.0001
California 1 3.28047 0.41666 61.9889 <.0001
Maine 1 2.57601 0.42133 37.3804 <.0001
Price 1 1 -1.75950 0.04455 1559.9275 <.0001
Price 2 1 0.52657 0.05879 80.2229 <.0001
Beach 1 1.41301 0.07173 388.0235 <.0001
Lake 1 0.71264 0.06292 128.2840 <.0001
Mountains 0 0 . . .
Bed & Breakfast 1 0.66322 0.05778 131.7652 <.0001
Cabin 1 -1.53715 0.07178 458.5709 <.0001
Hotel 0 0 . . .

The fit is exactly the same as when price was treated as qualitative, Chi-Square = 6535.3316. This is because
both models are the same except for the different but equivalent 2 df codings of price. The coefficients for
the destinations in the two models differ by a constant 1.23293. The coefficients for the factors after price are
unchanged. The part-worth utility for $999 is 1:75950  (999 1249)=250+0:52657  ((999 1249)=250)2 =
2:28607, the part-worth utility for $1249 is 1:75950  (1249 1249)=250+0:52657  ((1249 1249)=250)2 =
0, and the part-worth utility for $1499 is 1:75950  (1499 1249)=250+0:52657  ((1499 1249)=250)2 =
1:23293, which differ from the coefficients when price was treated as qualitative, by a constant -1.23293.

Effects Coding
In the previous analyses, binary (1, 0) codings were used for the variables. The next analysis illustrates effects
(1, 0, -1) coding. The two codings differ in how the final reference level is coded. In binary coding, the reference
level is coded with zeros. In effects coding, the reference level is coded with minus ones.

Binary Coding Effects Coding

Levels One Two One Two

1 1 0 1 0
2 0 1 0 1
3 0 0 -1 -1

In this example, we will use a binary coding for the destinations and effects codings for the attributes.
PROC TRANSREG can be used for effects coding. The effects option used inside the parentheses after
class asks for a (0, 1, -1) coding. The zero= option specifies the levels that receive the -1’s. PROC PHREG
is run with almost the same variable list as before, except now the variables for the reference levels, those whose
parameters are structural zeros are omitted. Refer back to the parameter estimates table on page 139, a few select
lines of which are reproduced next:
144 TS-677E Multinomial Logit, Discrete Choice Modeling

(Some Lines in the)

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Home 0 0 . . .
0 0 0 . . .
1499 0 0 . . .
Home 0 0 . . .
Mountains 0 0 . . .
Home 0 0 . . .
Hotel 0 0 . . .

Notice that the coefficients for the constant alternative (home and zero price) are all zero. Also notice that for
each factor, destination, price, scenery and accommodations, the coefficient for the last level is always zero. In
some sense, each class variable in a choice model with a constant alternative has two reference levels or two
levels that will always have a zero coefficient: the level corresponding to the constant alternative and the level
corresponding to the last level. In some of the preceding examples, we eliminated the ’Home’ levels by specifying
zero=Home. Now we will see how to eliminate all of the structural zeros from the parameter estimate table.
First, for each classification variable, we change the level for the constant alternative to missing. (Recall that
they were originally missing and we only made them nonmissing to deliberately produce the zero coefficients.)
This will cause PROC TRANSREG to ignore those levels when constructing dummy variables. When you use
this strategy, you must specify the norestoremissing option in the PROC TRANSREG statement. During
the first stage of design matrix creation, PROC TRANSREG puts zeros in the dummy variables for observations
with missing class levels. At the end, it replaces the zeros with missings, “restoring the missing values.” When
the norestoremissing option is specified, missing values are not restored and we get zeros in the dummy
variables for missing class levels. The DATA step if statements recode the constant levels to missing. Next, in
PROC TRANSREG, the reference levels ’Mountains’ and ’Hotel’ are listed in the zero= option in the class
specification.
data res4;
set res3;
if scene = 0 then scene = .;
if lodge = 0 then lodge = .;
run;

proc transreg design=5000 data=res4 nozeroconstant norestoremissing;

model class(place / zero=’Home’ order=data)
pspline(pricel / degree=2)
class(scene lodge /
effects zero=’Mountains’ ’Hotel’ order=formatted) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label pricel = ’Price’;
id subj set form c;
run;
Vacation Example 145

The coded data and design matrix are printed for the first choice set. The coded design matrix begins with five
binary columns for the destinations, ’Hawaii’ through ’Maine’. There is not a column for the stay-at-home
destination and the row for stay at home has all zeros in the coded variables. Next is the linear price effect,
’Price 1’, consisting of 0, 1, and -1. It is followed by the quadratic price effect, ’Price 2’, which is
’Price 1’ squared. Next are the scenery terms, effects coded. ’Beach’ and ’Lake’ have values of 0 and
1; -1’s in the fourth row for the reference level, ’Mountains’; and zeros in the last row for the stay-at-home
alternative. Next are the lodging terms, effects coded. ’Bed & Breakfast’ and ’Cabin’ have values of 0
and 1; -1’s in the first, third and fourth row for the reference level, ’Hotel’; and zeros in the last row for the
stay-at-home alternative.
proc print data=coded(obs=6) label;
run;

Vacation Example

Price Price Bed &

Obs Hawaii Alaska Mexico California Maine 1 2 Beach Lake Breakfast

1 1 0 0 0 0 0 0 0 1 1
2 0 1 0 0 0 1 1 0 1 -1
3 0 0 1 0 0 0 0 -1 -1 0
4 0 0 0 1 0 0 0 1 0 -1
5 0 0 0 0 1 0 0 1 0 0
6 0 0 0 0 0 0 0 0 0 0

Obs Cabin Place Price Scene Lodge Subj Set Form c

1 0 Hawaii 0 Lake Bed & Breakfast 1 1 1 1

2 -1 Alaska 1 Lake Hotel 1 1 1 2
3 1 Mexico 0 Mountains Cabin 1 1 1 2
4 -1 California 0 Beach Hotel 1 1 1 2
5 1 Maine 0 Beach Cabin 1 1 1 2
6 0 Home 0 . . 1 1 1 2

proc phreg data=coded brief;

model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;

Vacation Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 3600 6 1 5
146 TS-677E Multinomial Logit, Discrete Choice Modeling

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 12900.668 6365.337

AIC 12900.668 6387.337
SBC 12900.668 6455.412

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 6535.3316 11 <.0001

Score 5858.0723 11 <.0001
Wald 2158.0226 11 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Hawaii 1 5.24249 0.41160 162.2273 <.0001

Alaska 1 2.38334 0.41768 32.5603 <.0001
Mexico 1 4.38996 0.41216 113.4464 <.0001
California 1 3.69771 0.41325 80.0633 <.0001
Maine 1 2.99325 0.41631 51.6947 <.0001
Price 1 1 -1.75950 0.04455 1559.9275 <.0001
Price 2 1 0.52657 0.05879 80.2229 <.0001
Beach 1 0.70446 0.03931 321.2158 <.0001
Lake 1 0.00409 0.03390 0.0146 0.9040
Bed & Breakfast 1 0.95453 0.04069 550.3877 <.0001
Cabin 1 -1.24584 0.04754 686.7095 <.0001

It is instructive to compare the results of this analysis to the previous analysis on page 142. First, the model fit and
chi-square statistics are the same indicating the models are equivalent. The coefficients for the destinations differ
by a constant 0.41724, the price coefficients are the same, the scenery coefficients differ by a constant -0.70855,
and the lodging coefficients differ by a constant 0.29131. Notice that 0:41724 + 0 + 0:70855 + 0:29131 = 0,
so the utility for each alternative is unchanged by the different but equivalent codings.

Alternative-Specific Effects
In all of the analyses presented so far in this example, we have assumed that the effects for price, scenery,
and accommodations are generic or constant across the different destinations. Equivalently, we assumed that
destination does not interact with the attributes. Next, we show a model with alternative-specific effects that
does not make this assumption. Our new model allows for different price, scenery and lodging effects for each
destination. The coding can be done with PROC TRANSREG and its syntax for interactions. Before we do the
coding, let’s go back to the design preparation stage and redo it in a more normal fashion so reference levels will
be omitted from the analysis.
Vacation Example 147

We start by creating the data set KEY. This step differs from the one we saw on page 133 only in that now we
have a missing value for Place for the constant alternative.
data key;
input Place $ 1-10 (Lodge Scene Price) ($);
datalines;
Hawaii x1 x6 x11
Alaska x2 x7 x12
Mexico x3 x8 x13
California x4 x9 x14
Maine x5 x10 x15
. . . .
;
Next, we use the %MktRoll macro to process the design and the %MktMerge macro to merge the design and
data.
%mktroll(design=[Link], key=key, alt=place, out=rolled)

%mktmerge(design=rolled, data=results, out=res2, blocks=form,

nsets=&n, nalts=&m, setvars=choose1-choose&n,
stmts=%str(price = input(put(price, price.), 5.);
format scene scene. lodge lodge.;))

proc print data=res2(obs=12); run;

The usage of the %MktRoll macro is exactly the same as we saw on page 133. The %MktMerge macro usage
differs from page 135 in that instead of assigning labels and recoding price in a separate DATA step, we now
do it directly in the macro. The stmts= option is used to add a price = assignment statement and format
statement to the DATA step that merges the two data sets. The statements were included in a %str( ) macro
since they contain semicolons. Here are the first two choice sets.

Vacation Example

Obs Subj Form Set Place Lodge Scene Price c

1 1 1 1 Hawaii Bed & Breakfast Lake 1249 1

2 1 1 1 Alaska Hotel Lake 1499 2
3 1 1 1 Mexico Cabin Mountains 1249 2
4 1 1 1 California Hotel Beach 1249 2
5 1 1 1 Maine Cabin Beach 1249 2
6 1 1 1 . . . 2
7 1 1 2 Hawaii Hotel Mountains 999 2
8 1 1 2 Alaska Bed & Breakfast Lake 1249 2
9 1 1 2 Mexico Hotel Mountains 999 2
10 1 1 2 California Cabin Mountains 999 2
11 1 1 2 Maine Hotel Beach 999 1
12 1 1 2 . . . 2
148 TS-677E Multinomial Logit, Discrete Choice Modeling

Notice that the attributes for the constant alternative are all missing. Next, we code with PROC TRANSREG.
Since we are using missing values for the constant alternative, we must specify the norestoremissing
option in the PROC TRANSREG statement. First, we specify the variable Place as a class variable. Next,
we interact Place with all of the attributes, Price, Scene, and Lodge, to create the alternative-specific
effects.
proc transreg design=5000 data=res2 nozeroconstant norestoremissing;
model class(place / zero=none order=data)
class(place * price place * scene place * lodge /
zero=none order=formatted) / lprefix=0 sep=’ ’ ’, ’;
output out=coded(drop=_type_ _name_ intercept);
id subj set form c;
run;
proc print data=coded(obs=6) label noobs;
run;
The coded design matrix consists of:

 five binary columns, ’Hawaii’ through ’Maine’, for the five destinations,

 fifteen binary columns (5 destinations times 3 prices), ’Alaska, 999’ through ’Mexico, 1499’,
for the alternative-specific price effects,
 fifteen binary columns (5 destinations times 3 sceneries), ’Alaska, Beach’ through ’Mexico,
Mountains’, for the alternative-specific scenery effects,
 fifteen binary columns (5 destinations times 3 lodgings), ’Alaska, Bed & Breakfast’ through
’Mexico, Hotel’, for the alternative-specific lodging effects.

The entire sixth row of the coded design matrix, the stay-at-home alternative, consists of zeros.

Vacation Example

Alaska, Alaska, Alaska,

Hawaii Alaska Mexico California Maine 999 1249 1499

1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 1
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0

California, California, California, Hawaii, Hawaii, Hawaii, Maine, Maine, Maine,

999 1249 1499 999 1249 1499 999 1249 1499

0 0 0 0 1 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0
Vacation Example 149

Mexico, Mexico, Mexico, Alaska, Alaska, Alaska, California, California, California,

999 1249 1499 Beach Lake Mountains Beach Lake Mountains

0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0
0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0

Hawaii, Hawaii, Hawaii, Maine, Maine, Maine, Mexico, Mexico, Mexico,

Beach Lake Mountains Beach Lake Mountains Beach Lake Mountains

0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 0 0

Alaska, California,
Bed & Alaska, Alaska, Bed & California, California,
Breakfast Cabin Hotel Breakfast Cabin Hotel

0 0 0 0 0 0
0 0 1 0 0 0
0 0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 0 0
0 0 0 0 0 0

Hawaii, Maine, Bed Mexico,

Bed & Hawaii, Hawaii, & Maine, Maine, Bed & Mexico, Mexico,
Breakfast Cabin Hotel Breakfast Cabin Hotel Breakfast Cabin Hotel

1 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 0 0 0 0 0

Price Scene Lodge Subj Set Form c

Place
1249 Lake Bed & Breakfast 1 1 1 1
Hawaii 1499 Lake Hotel 1 1 1 2
Alaska 1249 Mountains Cabin 1 1 1 2
Mexico 1249 Beach Hotel 1 1 1 2
California 1249 Beach Cabin 1 1 1 2
Maine . . . 1 1 1 2

Analysis proceeds by running PROC PHREG as before.

proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
150 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 3600 6 1 5

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 12900.668 6340.136

AIC 12900.668 6410.136
SBC 12900.668 6626.741

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 6560.5318 35 <.0001

Score 6461.9024 35 <.0001
Wald 2047.3514 35 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Hawaii 1 3.48997 0.42964 65.9821 <.0001

Alaska 1 -0.14055 0.67506 0.0434 0.8351
Mexico 1 2.84951 0.44410 41.1695 <.0001
California 1 2.08314 0.46841 19.7780 <.0001
Maine 1 1.24023 0.50709 5.9818 0.0145
Alaska, 999 1 4.14782 0.46236 80.4767 <.0001
Alaska, 1249 1 1.93389 0.49549 15.2336 <.0001
Alaska, 1499 0 0 . . .
California, 999 1 3.59002 0.19604 335.3481 <.0001
California, 1249 1 1.04103 0.22174 22.0424 <.0001
California, 1499 0 0 . . .
Vacation Example 151

Hawaii, 999 1 3.61719 0.13710 696.0668 <.0001

Hawaii, 1249 1 1.24098 0.13098 89.7622 <.0001
Hawaii, 1499 0 0 . . .
Maine, 999 1 3.86257 0.25424 230.8154 <.0001
Maine, 1249 1 1.72816 0.28715 36.2202 <.0001
Maine, 1499 0 0 . . .
Mexico, 999 1 3.54171 0.15019 556.0958 <.0001
Mexico, 1249 1 1.22636 0.15974 58.9413 <.0001
Mexico, 1499 0 0 . . .
Alaska, Beach 1 1.67842 0.25920 41.9321 <.0001
Alaska, Lake 1 0.94543 0.24197 15.2667 <.0001
Alaska, Mountains 0 0 . . .
California, Beach 1 1.43286 0.16370 76.6131 <.0001
California, Lake 1 0.63691 0.16172 15.5103 <.0001
California, Mountains 0 0 . . .
Hawaii, Beach 1 1.60035 0.12782 156.7505 <.0001
Hawaii, Lake 1 1.02525 0.12842 63.7366 <.0001
Hawaii, Mountains 0 0 . . .
Maine, Beach 1 1.40565 0.18525 57.5772 <.0001
Maine, Lake 1 0.34761 0.19683 3.1189 0.0774
Maine, Mountains 0 0 . . .
Mexico, Beach 1 1.32202 0.14113 87.7431 <.0001
Mexico, Lake 1 0.63658 0.13770 21.3708 <.0001
Mexico, Mountains 0 0 . . .
Alaska, Bed & Breakfast 1 0.87845 0.21590 16.5553 <.0001
Alaska, Cabin 1 -1.72057 0.25842 44.3295 <.0001
Alaska, Hotel 0 0 . . .
California, Bed & Breakfast 1 0.64269 0.13784 21.7398 <.0001
California, Cabin 1 -1.32720 0.16997 60.9734 <.0001
California, Hotel 0 0 . . .
Hawaii, Bed & Breakfast 1 0.55551 0.12773 18.9151 <.0001
Hawaii, Cabin 1 -1.63720 0.12169 181.0058 <.0001
Hawaii, Hotel 0 0 . . .
Maine, Bed & Breakfast 1 0.60884 0.15443 15.5423 <.0001
Maine, Cabin 1 -1.76896 0.21614 66.9816 <.0001
Maine, Hotel 0 0 . . .
Mexico, Bed & Breakfast 1 0.64489 0.12357 27.2366 <.0001
Mexico, Cabin 1 -1.75349 0.17161 104.4079 <.0001
Mexico, Hotel 0 0 . . .

There are zero coefficients for the reference alternative. Do we need this more complicated model instead of the
simpler model? To answer this, first look at the coefficients. Are they similar across different destinations? In
this case, they seem to be. This suggests that the simpler model may be sufficient.
More formally, the two models can be statistically compared. You can test the null hypothesis that the two
models are not significantly different by comparing their likelihoods. The difference between two 2 log(LC )’s
(the number reported under ’With Covariates’ in the output) has a chi-square distribution. We can get the df for
the test by subtracting the two df for the two likelihoods. The difference 6560:5318 6535:3316 = 25:2002 is
distributed 2 with 35 11 = 24 df (p < 0:395). This more complicated model does not account for significantly
more variance than the simpler model.
152 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example, with Alternative-Specific

Attributes
A researcher is interested in studying choice of vacation destinations. This example discusses choosing the
number of choice sets, designing the choice experiment, ensuring that certain key interactions are estimable, ex-
amining the design, blocking an existing design, generating the questionnaire, generating artificial data, reading,
processing, and analyzing the data, binary coding, generic attributes, alternative-specific effects, aggregating the
data, analysis, and interpretation of the results. Here are two summaries of the design, with factors grouped by
attribute and grouped by destination.
Factor Destination Attribute Levels
X1 Hawaii Accommodations Cabin, Bed & Breakfast, Hotel
X2 Alaska Accommodations Cabin, Bed & Breakfast, Hotel
X3 Mexico Accommodations Cabin, Bed & Breakfast, Hotel
X4 California Accommodations Cabin, Bed & Breakfast, Hotel
X5 Maine Accommodations Cabin, Bed & Breakfast, Hotel

X6 Hawaii Scenery Mountains, Lake, Beach

X7 Alaska Scenery Mountains, Lake, Beach
X8 Mexico Scenery Mountains, Lake, Beach
X9 California Scenery Mountains, Lake, Beach
X10 Maine Scenery Mountains, Lake, Beach

X11 Hawaii Price $1249, $1499, $1749

X12 Alaska Price $1249, $1499, $1749
X13 Mexico Price $999, $1249, $1499
X14 California Price $999, $1249, $1499, $1749
X15 Maine Price $999, $1249, $1499

X16 Hawaii Side Trip Yes, No

X17 Mexico Side Trip Yes, No

Factor Destination Attribute Levels

X1 Hawaii Accommodations Cabin, Bed & Breakfast, Hotel
X6 Scenery Mountains, Lake, Beach
X11 Price $1249, $1499, $1749
X16 Side Trip Yes, No

X2 Alaska Accommodations Cabin, Bed & Breakfast, Hotel

X7 Scenery Mountains, Lake, Beach
X12 Price $1249, $1499, $1749

X3 Mexico Accommodations Cabin, Bed & Breakfast, Hotel

X8 Scenery Mountains, Lake, Beach
X13 Price $999, $1249, $1499
X17 Side Trip Yes, No

X4 California Accommodations Cabin, Bed & Breakfast, Hotel

X9 Scenery Mountains, Lake, Beach
X14 Price $999, $1249, $1499, $1749

X5 Maine Accommodations Cabin, Bed & Breakfast, Hotel

X10 Scenery Mountains, Lake, Beach
X15 Price $999, $1249, $1499
Vacation Example, with Alternative-Specific Attributes 153

This example is a modification of the previous example. Now, all alternatives do not have the same factors, and
all factors do not have the same numbers of levels. There are still five destinations of interest: Hawaii, Alaska,
Mexico, California, and Maine. Each alternative is composed of three factors like before: package cost, scenery,
and accommodations, only now they do not all have the same levels, and the Hawaii and Mexico alternatives
are composed of one additional attribute. For Hawaii and Alaska, the costs are $1,249, $1,499, and $1,749; for
California, the prices are $999, $1,249, $1,499, and $1,749; and for Mexico and Maine, the prices are $999,
$1,249, and $1,499. Scenery (mountains, lake, beach) and accommodations (cabin, bed & breakfast, and hotel)
are the same as before. The Mexico trip now has the option of a side trip to sites of archaeological significance,
via bus, for an additional cost of $100. The Hawaii trip has the option of a side trip to an active volcano, via
helicopter, for an additional cost of $200. This is typical of the problems that marketing researchers face. We
have lots of factors and asymmetry each alternative is not composed of the same factors, and the factors do not
all have the same numbers of levels.

Choosing the Number of Choice Sets

We can use the %MktRuns autocall macro to suggest experimental design sizes. (All of the autocall macros used
in this report are documented starting on page 287.) As before, we specify a list containing the number of levels
of each factor.
title ’Vacation Example with Asymmetry’;

%mktruns( 3 ** 14 4 2 2 )
The output tells us the size of the saturated design, which is the number of parameters in the linear design, and
suggests design sizes.

Vacation Example with Asymmetry

Design Summary

Number of
Levels Frequency

2 2
3 14
4 1

Vacation Example with Asymmetry

Saturated = 34
Full Factorial = 76,527,504

Some Reasonable Cannot Be

Design Sizes Violations Divided By

72 * 0
144 0
36 2 8
108 2 8
54 18 4 8 12
90 18 4 8 12
126 18 4 8 12
45 48 2 4 6 8 12
63 48 2 4 6 8 12
81 48 2 4 6 8 12

* - 100% Efficient Design can be made with the MktEx Macro.

154 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example with Asymmetry

n Design Reference

72 2 ** 20 3 ** 24 4 ** 1 Wang, 1996
72 2 ** 13 3 ** 25 4 ** 1 Wang, 1996
72 2 ** 11 3 ** 24 4 ** 1 6 ** 1 Wang, 1996

We need at least 34 choice sets, as shown by ’(Saturated=34)’ in the listing. Any size that is a multiple of 72
would be optimal. We would recommend 72 choice sets, four blocks of size 18. However, like the previous
vacation example, we will use fewer choice sets so that we can illustrate getting an efficient but nonorthogonal
design. A design with 36 choice sets is pretty good. Thirty-six is not divisible by 8 = 2  4, so we cannot have
equal frequencies in the California price and Mexico and Hawaii side trip combinations. This should not pose
any problem. This leaves only 2 error df for the linear model, but in the choice model, we will have adequate
error df.

Designing the Choice Experiment

This problem requires a design with 1 four-level factor for price and 4 three-level factors for price. There are
10 three-level factors for scenery and accommodations as before. Also, we need 2 two-level factors for the two
side trips. Note that we do not need a factor for the price or mode of transportation of the side trips since they
are constant within each trip. With the %MktEx macro, making an asymmetric design is no more difficult than
making a symmetric design.
%mktex(3 ** 13 4 3 2 2, n=36, seed=7654321)
%mkteval;
Here is the last part of the results.

Vacation Example with Asymmetry

The OPTEX Procedure

Class Level Information

Class Levels -Values-

x1 3 1 2 3
x2 3 1 2 3
x3 3 1 2 3
x4 3 1 2 3
x5 3 1 2 3
x6 3 1 2 3
x7 3 1 2 3
x8 3 1 2 3
x9 3 1 2 3
x10 3 1 2 3
x11 3 1 2 3
x12 3 1 2 3
x13 3 1 2 3
x14 4 1 2 3 4
x15 3 1 2 3
x16 2 1 2
x17 2 1 2
Vacation Example, with Alternative-Specific Attributes 155

Vacation Example with Asymmetry

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 98.8874 97.5943 97.4925 0.9718

Vacation Example with Asymmetry

Canonical Correlations Between the Factors
There are 2 Canonical Correlations Greater Than 0.316

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15 x16 x17

x1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x2 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x3 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
x4 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
x5 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
x6 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
x7 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
x8 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
x9 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
x10 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
x11 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
x12 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
x13 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0.25 0 0
x14 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0.33 0.33
x15 0 0 0 0 0 0 0 0 0 0 0 0 0.25 0 1 0 0
x16 0 0 0 0 0 0 0 0 0 0 0 0 0 0.33 0 1 0
x17 0 0 0 0 0 0 0 0 0 0 0 0 0 0.33 0 0 1

The macro found a very nice, almost orthogonal and almost 99% efficient design in 5.5 minutes. However, we
will not use this design. Instead, we will make a larger design with interactions.

Ensuring that Certain Key Interactions are Estimable

Next, we will ensure that certain key interactions are estimable. Specifically, it would be good if in the aggregate,
the interactions between price and accommodations were estimable for each destination. We would like the
following interactions to be estimable: x1*x11 x2*x12 x3*x13 x4*x15 x5*x15. We will again use
the %MktEx macro.
%mktex(3 ** 13 4 3 2 2, n=36,
interact=x1*x11 x2*x12 x3*x13 x5*x15,
seed=7654321)
We immediately get this message.

ERROR: More parameters than runs.

If you really want to do this, specify RIDGE=.
ERROR: The MKTEX macro ended abnormally.
156 TS-677E Multinomial Logit, Discrete Choice Modeling

If we want interactions to be estimable, we will need more choice sets. The number of parameters is 1 for the
intercept, 14  (3 1)+(4 1)+2  (2 1) = 33 for main effects, and 4  (3 1)  (3 1)+(4 1)  (3 1) = 22
for interactions for a total of 1 + 33 + 22 = 56 parameters. This means we need at least 56 choice sets, and
ideally for this design with 2, 3, and 4 level factors, we would like the number of sets to be divisible by 2  2,
2  3, 2  4, 3  3, and 3  4. Sixty is divisible by 2, 3, 4, 6, and 12 so is a reasonable design size. Sixty choice
sets could be divided into three blocks of size 20, four blocks of size 15, or five blocks of size 12. Seventy-two
choice sets would be better, since unlike 60, 72 can be divided by 9. Unfortunately, 72 would require one more
block.
We can also run the %MktRuns macro to help us choose the number of choice sets. However, the %MktRuns
does not have a special syntax for interactions, you have to specify the main effects and interactions of two factors
as if it were a single factor. For example, for the interaction of 2 three-level factors, you specify 9 in the list.
For the interaction of a three-level factor and a four-level factor, you specify 12 in the list. Do not specify ’3 3
9’ or ’3 4 12’; just specify ’3’ and ’12’. In this example, we specify four 9’s for the four accommodation/price
interactions involving only three-level factors, one 12 for the California accommodation/price interaction, five
3’s for scenery, and two 2’s for the side trips. We also specified a keyword option max= to consider only the 45
design sizes from the minimum of 56 up to 100.
title ’Vacation Example with Asymmetry’;

%mktruns(9 9 9 9 12 3 3 3 3 3 2 2, max=45)

Vacation Example with Asymmetry

Design Summary

Number of
Levels Frequency

2 2
3 5
9 4
12 1

Vacation Example with Asymmetry

Saturated = 56
Full Factorial = 76,527,504

Some Reasonable Cannot Be

Design Sizes Violations Divided By

72 30 27 81 108
81 33 2 4 6 12 18 24 36 108
90 39 4 12 24 27 36 81 108
96 57 9 18 27 36 81 108
60 59 9 18 24 27 36 81 108
63 59 2 4 6 12 18 24 27 36 81 108
84 59 9 18 24 27 36 81 108
99 59 2 4 6 12 18 24 27 36 81 108
66 61 4 9 12 18 24 27 36 81 108
78 61 4 9 12 18 24 27 36 81 108
Vacation Example, with Alternative-Specific Attributes 157

We see that 72 cannot be divided by 27 = 9  3 so for example the Maine accommodation/price combinations
cannot occur with equal frequency with each of the three-level factors. We see that 72 cannot be divided by
81 = 9  9 so for example the Mexico accommodation/price combinations cannot occur with equal frequency
with each of the Hawaii accommodation/price combinations. We see that 72 cannot be divided by 108 = 9  12
so for example the California accommodation/price combinations cannot occur with equal frequency with each
of the Maine accommodation/price combinations. With interactions, there are many higher-order opportunities
for nonorthogonality. However, usually we will not be overly concerned about potential unequal frequencies on
combinations of attributes in different alternatives.
The smallest number of runs in the table is 60. While 72 is better in that it can be divided by more numbers,
either 72 or 60 should work fine. We will pick the larger number and run the %MktEx macro again with n=72
specified.
%mktex(3 ** 13 4 3 2 2, n=72, seed=7654321,
interact=x1*x11 x2*x12 x3*x13 x4*x14 x5*x15)
The macro printed these notes to the log.
NOTE: Performing 20 searches of 243 candidates, full-factorial=76,527,504.
NOTE: Generating the tabled design, n=72.
The candidate-set search is using a fractional-factorial candidate set with 35 = 243 candidates. The two-level
factors in the candidate set are made from three-level factors by coding down. Coding down replaces an m-level
factor with a factor with fewer than m levels, for example a two-level factor could be created from a three-level
factor: ((123) ) (121)). The four-level factor in the candidate set is made from 2 three-level factors and coding
down. ((1 2 3)  (1 2 3) ) (1 2 3 4 5 6 7 8 9) ) (1 2 3 4 1 2 3 4 1)). The tabled design used for the partial
initialization in the coordinate-exchange steps has 72 runs. Here are some of the results.

Vacation Example with Asymmetry

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 85.1188 85.1188 Can
1 End 85.1188

2 Start 65.6743 Tab

2 46 11 85.1337 85.1337
2 59 13 85.1617 85.1617
2 64 1 85.1630 85.1630
2 69 3 85.1663 85.1663
2 1 2 85.3358 85.3358
2 2 2 85.3706 85.3706
2 2 12 85.4221 85.4221
2 3 13 85.4374 85.4374
2 14 1 85.4819 85.4819
2 25 15 85.4824 85.4824
2 28 5 85.5177 85.5177
2 32 3 85.5350 85.5350
2 42 5 85.5801 85.5801
2 51 4 85.5826 85.5826
2 59 13 85.6058 85.6058
2 66 1 85.6455 85.6455
2 7 15 85.6703 85.6703
2 12 2 85.7176 85.7176
2 40 1 85.7682 85.7682
2 End 85.7682
158 TS-677E Multinomial Logit, Discrete Choice Modeling

3 Start 65.6743 Tab

3 40 1 85.7682 85.7682
3 End 85.7682

4 Start 65.6743 Tab

4 40 1 85.7682 85.7682
4 End 85.7682
.
.
.

11 Start 65.6743 Tab

11 40 1 85.7682 85.7682
11 End 85.7682

12 Start 56.4650 Ran,Mut,Ann

12 48 14 85.7817 85.7817
12 48 16 85.7888 85.7888
12 49 6 85.7888 85.7888
.
.
.
12 40 17 89.5461 89.5461
12 58 17 89.5503 89.5503
12 End 89.5503

13 Start 56.9380 Ran,Mut,Ann

13 12 16 89.5646 89.5646
13 15 16 89.5674 89.5674
.
.
.
13 62 17 89.8765 89.8765
13 End 89.8765
.
.
.

16 Start 59.8285 Ran,Mut,Ann

16 End 89.0880

NOTE: Quitting the algorithm search step after 10.40 minutes and 16 designs.

Vacation Example with Asymmetry

Design Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 89.8765 89.8765 Ini

1 Start 57.5919 Ran,Mut,Ann

1 End 89.2836
.
.
.
Vacation Example, with Alternative-Specific Attributes 159

14 Start 58.7160 Ran,Mut,Ann

14 End 88.7760

NOTE: Quitting the design search step after 20.95 minutes and 14 designs.

Vacation Example with Asymmetry

Design Refinement History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 89.8765 89.8765 Ini

1 Start 81.3493 Pre,Mut,Ann

1 End 89.7066
.
.
.

4 Start 81.5586 Pre,Mut,Ann

4 End 89.1031

NOTE: Quitting the refinenent step after 6.33 minutes and 4 designs.

Vacation Example with Asymmetry

The OPTEX Procedure

Class Level Information

Class Levels -Values-

x1 3 1 2 3
x2 3 1 2 3
x3 3 1 2 3
x4 3 1 2 3
x5 3 1 2 3
x6 3 1 2 3
x7 3 1 2 3
x8 3 1 2 3
x9 3 1 2 3
x10 3 1 2 3
x11 3 1 2 3
x12 3 1 2 3
x13 3 1 2 3
x14 4 1 2 3 4
x15 3 1 2 3
x16 2 1 2
x17 2 1 2
160 TS-677E Multinomial Logit, Discrete Choice Modeling

Vacation Example with Asymmetry

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 89.8765 80.0119 95.5854 0.8819

The algorithm search history shows that the candidate-set approach (Can) used in design 1 found a design that
was 85.1188% efficient. The macro makes no attempt to improve on this design, unless there are restriction on
the design, until the end in the design refinement step, and only if it is the best design found.
Designs 2 through 111 used the coordinate-exchange algorithm with a tabled design initialization (Tab). This
process found a design that was 85.7682% efficient. For this problem, the tabled design initialization initializes
all 72 rows; For other problems, when the number of runs in the design is greater than the number of runs in
the nearest tabled design, the remaining rows would be randomly initialized. The tabled design initialization
usually works very well when all but at most a very few rows and columns are left uninitialized and there are
no interactions or restrictions. That is not the case in this problem, and when the algorithm switches to a fully
random initialization in design 12, it immediately does better. In design 13, the macro finds a design with
89.8765% efficiency. After 16 iterations, the macro quit because the run time for the algorithm search exceeded
10 minutes (which is the default first value of the maxtime= option). The macro only checks the elapsed time
after it finishes making a design. This is why huge problems with restrictions can take much longer.
The algorithm search phase picked the coordinate-exchange algorithm with a random initialization and random
mutations and simulated annealing as the algorithm to use in the next step, the design search step. The design
search history is initialized with the best design (D-efficiency = 89.8765) found so far. The design search phase
starts out with the initial design (Ini) found in the algorithm search phase. Usually, you will see improvement
in the design search phase, however, in this case you do not. After 14 iterations, the macro quit because the run
time for the algorithm search exceeded 20 minutes (which is the default second value of the maxtime= option).
The final set of iterations tries to improve the best design found so far. Random mutations (Ran), simulated
annealing (Ann), and level exchanges are used on the previous best (Pre) design. The random mutations are
responsible for making the efficiency of the starting design worse than the previous best efficiency. After 4
iterations, the macro quit because the run time for the algorithm search exceeded 5 minutes (which is the default
third value of the maxtime= option).
All together, the macro used 37.78 minutes, which is more than the sum of the three reported times, because not
all of the macro’s calculations (most notably time spent in PROC FACTEX and OPTEX) are timed. Run time
was just slightly more than the 10 + 20 + 5 = 35 maximum time specified by maxtime=. Recall that the macro
stated that it ran 20 OPTEX iterations on 243 candidates. This will be very fast. If the full-factorial design had
been smaller (a few thousand runs) the macro may have done more iterations using PROC OPTEX. This could
have brought the run time up closer to an hour. When the full-factorial design is too big to search, and the macro
uses a fractional-factorial design, it does not spend much time using OPTEX, because PROC OPTEX is probably
not going to be the best approach. When the full-factorial design is manageable, the macro will spend more time
in OPTEX, because there is a good chance that it will be the best approach. The macro picks the number of
OPTEX iterations based on the size of the candidate set and the value of the maxtime= option.

Examining the Design

We can use the %MktEval macro to evaluate the goodness of this design.
%mkteval(data=design);
Vacation Example, with Alternative-Specific Attributes 161

Here are some of the results.

Vacation Example with Asymmetry

Canonical Correlations Between the Factors
There are 0 Canonical Correlations Greater Than 0.316

x1 x2 x3 x4 x5 x6 x7 x8 x9

x1 1 0.11 0.13 0.18 0.11 0.07 0.15 0.07 0.09

x2 0.11 1 0.19 0.11 0.05 0.11 0.08 0.12 0.11
x3 0.13 0.19 1 0.17 0.09 0.06 0.12 0.09 0.07
x4 0.18 0.11 0.17 1 0.05 0.09 0.05 0.06 0.11
x5 0.11 0.05 0.09 0.05 1 0.09 0.11 0.05 0.08
x6 0.07 0.11 0.06 0.09 0.09 1 0.07 0.09 0.09
x7 0.15 0.08 0.12 0.05 0.11 0.07 1 0.11 0.09
x8 0.07 0.12 0.09 0.06 0.05 0.09 0.11 1 0.07
x9 0.09 0.11 0.07 0.11 0.08 0.09 0.09 0.07 1
.
.
.

Vacation Example with Asymmetry

There are 0 Canonical Correlations Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

* x1 25 23 24
* x2 22 24 26
* x3 21 26 25
* x4 24 22 26
x5 24 24 24
* x6 22 24 26
* x7 26 23 23
* x8 24 25 23
* x9 26 23 23
* x10 23 25 24
x11 24 24 24
* x12 23 24 25
* x13 23 24 25
* x14 19 16 18 19
* x15 22 25 25
* x16 34 38
* x17 38 34
162 TS-677E Multinomial Logit, Discrete Choice Modeling

* x1 x2 7 8 10 6 9 8 9 7 8
* x1 x3 8 8 9 8 8 7 5 10 9
* x1 x4 10 5 10 6 9 8 8 8 8
* x1 x5 10 7 8 7 8 8 7 9 8
* x1 x6 8 8 9 7 7 9 7 9 8
* x1 x7 9 10 6 8 6 9 9 7 8
* x1 x8 8 9 8 8 7 8 8 9 7
* x1 x9 9 8 8 7 8 8 10 7 7
* x1 x10 8 8 9 9 8 6 6 9 9
* x1 x11 8 8 9 7 8 8 9 8 7
* x1 x12 7 8 10 8 8 7 8 8 8
* x1 x13 8 9 8 7 7 9 8 8 8
* x1 x14 7 6 6 6 5 4 7 7 7 6 5 6
* x1 x15 6 9 10 8 6 9 8 10 6
* x1 x16 12 13 10 13 12 12
* x1 x17 13 12 12 11 13 11
* x12 x13 8 7 8 6 9 9 9 8 8
* x12 x14 7 6 5 5 6 4 7 7 6 6 6 7
* x12 x15 7 7 9 6 10 8 9 8 8
* x12 x16 10 13 11 13 13 12
* x12 x17 12 11 12 12 14 11
* x13 x14 7 4 5 7 7 4 7 6 5 8 6 6
* x13 x15 8 7 8 9 7 8 5 11 9
* x13 x16 10 13 13 11 11 14
* x13 x17 12 11 13 11 13 12
* x14 x15 5 7 7 5 4 7 4 7 7 8 7 4
* x14 x16 9 10 7 9 9 9 9 10
* x14 x17 10 9 9 7 9 9 10 9
* x15 x16 11 11 13 12 10 15
* x15 x17 11 11 13 12 14 11
* x16 x17 19 15 19 19
.
.
.
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

We can the %MktEx macro to check the design and print the information matrix and variance matrix.
%mktex(3 ** 13 4 3 2 2, n=72, examine=i v, options=check, init=randomized,
interact=x1*x11 x2*x12 x3*x13 x4*x14 x5*x15)
In the interest of space, the results from this step are not shown.

Blocking an Existing Design

An existing design is blocked using the %MktBlock macro. The macro takes the observations in an existing
design and optimally sorts them into blocks. Here we are seeing how to block the linear version of the choice
design, but the macro can also be used directly on the choice design.
%mktblock(data=randomized, nblocks=4, out=[Link], seed=114)
Vacation Example, with Alternative-Specific Attributes 163

This step took 8.45 seconds. Here are some of the results including the one-way frequencies within blocks. They
should be examined to ensure that each level is well represented in each block. The design is nearly balanced in
most of the factors and blocks. Perfect balance is impossible for the three level factors.

Vacation Example with Asymmetry

Canonical Correlations Between the Factors
There are 0 Canonical Correlations Greater Than 0.316

Block x1 x2 x3 x4 x5 x6 x7 x8

Block 1 0.06 0.13 0.11 0.07 0.10 0.08 0.13 0.08

x1 0.06 1 0.11 0.13 0.18 0.11 0.07 0.15 0.07
x2 0.13 0.11 1 0.19 0.11 0.05 0.11 0.08 0.12
x3 0.11 0.13 0.19 1 0.17 0.09 0.06 0.12 0.09
x4 0.07 0.18 0.11 0.17 1 0.05 0.09 0.05 0.06
x5 0.10 0.11 0.05 0.09 0.05 1 0.09 0.11 0.05
x6 0.08 0.07 0.11 0.06 0.09 0.09 1 0.07 0.09
x7 0.13 0.15 0.08 0.12 0.05 0.11 0.07 1 0.11
x8 0.08 0.07 0.12 0.09 0.06 0.05 0.09 0.11 1

Vacation Example with Asymmetry

Summary of Frequencies
There are 0 Canonical Correlations Greater Than 0.316
* - Indicates Unequal Frequencies

Block 18 18 18 18
* x1 25 24 23
* x2 22 24 26
* x3 25 26 21
* x4 22 26 24
x5 24 24 24
* x6 24 26 22
* x7 23 26 23
* x8 24 25 23
* x9 23 23 26
* x10 23 25 24
x11 24 24 24
* x12 24 23 25
* x13 23 24 25
* x14 18 16 19 19
* x15 22 25 25
* x16 34 38
* x17 34 38
164 TS-677E Multinomial Logit, Discrete Choice Modeling

* Block x1 6 6 6 6 6 6 7 6 5 6 6 6
* Block x2 7 6 5 5 6 7 5 7 6 5 5 8
* Block x3 7 6 5 6 6 6 7 7 4 5 7 6
* Block x4 6 6 6 5 7 6 5 7 6 6 6 6
* Block x5 7 5 6 5 7 6 6 6 6 6 6 6
* Block x6 6 7 5 5 7 6 6 6 6 7 6 5
* Block x7 4 7 7 6 6 6 7 6 5 6 7 5
* Block x8 6 6 6 6 6 6 5 7 6 7 6 5
* Block x9 6 6 6 5 6 7 6 5 7 6 6 6
* Block x10 6 6 6 5 7 6 5 6 7 7 6 5
* Block x11 6 6 6 6 6 6 5 6 7 7 6 5
* Block x12 6 6 6 7 5 6 5 7 6 6 5 7
* Block x13 7 5 6 5 7 6 5 6 7 6 6 6
* Block x14 4 4 6 4 4 4 4 6 5 4 5 4 5 4 4 5
* Block x15 5 7 6 6 5 7 6 6 6 5 7 6
* Block x16 7 11 8 10 9 9 10 8
* Block x17 8 10 9 9 9 9 8 10
.
.
.

N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

Generating the Questionnaire

These next steps print the questionnaire.
%let m = 6; /* m alts including constant */
%let mm1 = %eval(&m - 1); /* m - 1 */
%let n = 18; /* number of choice sets */
%let blocks = 4; /* number of blocks */

title;
options ls=80 ps=60 nonumber nodate;

data _null_;
array dests[&mm1] $ 10 _temporary_ (’Hawaii’ ’Alaska’ ’Mexico’
’California’ ’Maine’);
array scenes[3] $ 13 _temporary_
(’the Mountains’ ’a Lake’ ’the Beach’);
array lodging[3] $ 15 _temporary_
(’Cabin’ ’Bed & Breakfast’ ’Hotel’);
array x[15];
array p[&mm1];
length price $ 6;
file print linesleft=ll;
set [Link];
by block;

p1 = 1499 + (x[11] - 2) * 250;

p2 = 1499 + (x[12] - 2) * 250;
p3 = 1249 + (x[13] - 2) * 250;
p4 = 1374 + (x[14] - 2.5) * 250;
p5 = 1249 + (x[15] - 2) * 250;
Vacation Example, with Alternative-Specific Attributes 165

if [Link] then do;

choice = 0;
put _page_;
put @50 ’Form: ’ block ’ Subject: ________’ //;
end;
choice + 1;
if ll < (19 + (x16 = 1) + (x17 = 1)) then put _page_;
put choice 2. ’) Circle your choice of ’
’vacation destinations:’ /;
do dest = 1 to &mm1;
price = left(put(p[dest], dollar6.));
put ’ ’ dest 1. ’) ’ dests[dest]
+(-1) ’, staying in a ’ lodging[x[dest]]
’near ’ scenes[x[&mm1 + dest]] +(-1) ’,’ /
+7 ’with a package cost of ’ price +(-1) @@;
if dest = 3 and x16 = 1 then
put ’, and an optional visit’ / +7
’to archaeological sites for an additional $100’ @@;
else if dest = 1 and x17 = 1 then
put ’, and an optional helicopter’ / +7
’flight to an active volcano for an additional $200’ @@;
put ’.’ /;
end;
put " &m) Stay at home this year." /;
run;
Here are the first two choice sets for the first subject.

Form: 1 Subject: ________

1) Circle your choice of vacation destinations:

1) Hawaii, staying in a Cabin near the Beach,

with a package cost of $1,749, and an optional helicopter
flight to an active volcano for an additional $200.

2) Alaska, staying in a Cabin near a Lake,

with a package cost of $1,249.

3) Mexico, staying in a Cabin near a Lake,

with a package cost of $1,249.

4) California, staying in a Bed & Breakfast near the Beach,

with a package cost of $999.

5) Maine, staying in a Cabin near the Beach,

with a package cost of $1,499.

6) Stay at home this year.

166 TS-677E Multinomial Logit, Discrete Choice Modeling

2) Circle your choice of vacation destinations:

1) Hawaii, staying in a Bed & Breakfast near the Mountains,

with a package cost of $1,249, and an optional helicopter
flight to an active volcano for an additional $200.

2) Alaska, staying in a Bed & Breakfast near a Lake,

with a package cost of $1,249.

3) Mexico, staying in a Bed & Breakfast near the Mountains,

with a package cost of $999.

4) California, staying in a Hotel near a Lake,

with a package cost of $1,499.

5) Maine, staying in a Hotel near the Mountains,

with a package cost of $1,499.

6) Stay at home this year.

In practice, data collection may be much more elaborate than this. It may involve art work, photographs, and the
choice sets may be presented and data may be collected over the web. However the choice sets are presented and
the data collected, the essential ingredients remain the same. Subjects are shown sets of alternatives and asked to
make a choice, and then they go on to the next set.

Generating Artificial Data

This next step generates an artificial set of data. Collecting data is time consuming and expensive. Generating
some artificial data before the data are collected to test your code and make sure the analysis will run is a good
idea. It helps avoid the “How am I going to analyze this?” question from occuring after the data have already
been collected. See page 215 for an alternative method of testing your design. This step generates data for 300
subjects, 100 per block.
data _null_;
array dests[&mm1] _temporary_ (5 -1 4 3 2);
array scenes[3] _temporary_ (-1 0 1);
array lodging[3] _temporary_ (0 3 2);
array u[&m];
array x[15];
do rep = 1 to 100;
n = 0;
do i = 1 to &blocks;
k + 1;
if mod(k,3) = 1 then put;
put k 3. +1 i 1. +2 @@;
do j = 1 to &n; n + 1;
set [Link] point=n;
do dest = 1 to &mm1;
u[dest] = dests[dest] + lodging[x[dest]] +
scenes[x[&mm1 + dest]] -
x[2 * &mm1 + dest] +
2 * normal(7);
end;
Vacation Example, with Alternative-Specific Attributes 167

u[1] = u[1] + (x16 = 1);

u[3] = u[3] + (x17 = 1);
u&m = -3 + 3 * normal(7);
m = max(of u1-u&m);
if abs(u1 - m) < 1e-4 then c = 1;
else if abs(u2 - m) < 1e-4 then c = 2;
else if abs(u3 - m) < 1e-4 then c = 3;
else if abs(u4 - m) < 1e-4 then c = 4;
else if abs(u5 - m) < 1e-4 then c = 5;
else c = 6;
put +(-1) c @@;
end;
end;
end;
stop;
run;
The dests, scenes, and lodging arrays are initialized with part-worth utilities for each level. The utilities
for each of the destinations are computed and stored in the array u in the statement u[dest] = ..., which
includes an error term 2 * normal(7). The utilities for the side trips are added in separately with u[1]
= u[1] + (x16 = 1) and u[3] = u[3] + (x17 = 1). The utility for the stay-at-home alternative
is -3 + 3 * normal(7). The maximum utility is computed, m = max(of u1-u&m) and the alternative
with the maximum utility is chosen. The put statement writes out the results to the log.

Reading, Processing, and Analyzing the Data

The results from the previous step are pasted into a DATA step and run to mimic reading real input data.
title ’Vacation Example with Asymmetry’;

data results;
input Subj Form (choose1-choose&n) (1.) @@;
datalines;
1 1 431313344341313133 2 2 314113115514415413 3 3 331313331143431411
4 4 431133311134311143 5 1 431151131141311333 6 2 111113113514335111
7 3 341513331311451134 8 4 431113211334311511 9 1 411121131141311153
.
.
.
;
The analysis proceeds in a fashion similar to before. Formats and the key to processing the design are created.
proc format;
value price 1 = ’ 999’ 2 = ’1249’ 3 = ’1499’ 4 = ’1749’;
value scene 1 = ’Mountains’ 2 = ’Lake’ 3 = ’Beach’;
value lodge 1 = ’Cabin’ 2 = ’Bed & Breakfast’ 3 = ’Hotel’;
value side 1 = ’Side Trip’ 2 = ’No’;
run;

data key;
input Place $ 1-10 (Lodge Scene Price Side) ($);
datalines;
Hawaii x1 x6 x11 x16
Alaska x2 x7 x12 .
Mexico x3 x8 x13 x17
California x4 x9 x14 .
Maine x5 x10 x15 .
. . . . .
;
168 TS-677E Multinomial Logit, Discrete Choice Modeling

For analysis, the design will have five attributes. Place is the alternative name. Lodge, Scene, Price and
Side are created from the design using the indicated factors. See page 133 for more information on creating
the design key. Notice that Side only applies to some of the alternatives and hence has missing values for the
others. Processing the design and merging it with the data are similar to what was done on pages 133 and 135.
One difference is now there are asymmetries in Price. For Hawaii’s price, x11, we need to change 1, 2, and
3 to $1249, $1499, and $1749. For Alaska’s price, x12, we need to change 1, 2, and 3 to $1249, $1499, and
$1749. For Mexico’s price, x13, we need to change 1, 2, and 3 to $999, $1249, and $1499. For California’s
price, x14, we need to change 1, 2, 3, and 4 to $999, $1249, $1499, and $1749. For Maine’s price, x11, we
need to change 1, 2, and 3 to $999, $1249, and $1499. We can simplify the problem by adding 1 to x11 and
x12, which are the factors that start at $1249 instead of $999. This will allow us to use a common format to set
the price. See page 211 for an example of handling more complicated asymmetries.
data temp;
set [Link];
x11 + 1;
x12 + 1;
run;

%mktroll(design=temp, key=key, alt=place, out=rolled)

%mktmerge(design=rolled, data=results, out=res2, blocks=form,

nsets=&n, nalts=&m, setvars=choose1-choose&n,
stmts=%str(price = input(put(price, price.), 5.);
format scene scene. lodge lodge. side side.;))

proc print data=res2(Obs=18); run;

Here are the first three choice sets.

Vacation Example with Asymmetry

Obs Subj Form Set Place Lodge Scene Price Side c

1 1 1 1 Hawaii Cabin Beach 1749 No 2

2 1 1 1 Alaska Cabin Lake 1249 . 2
3 1 1 1 Mexico Cabin Lake 1249 Side Trip 2
4 1 1 1 California Bed & Breakfast Beach 999 . 1
5 1 1 1 Maine Cabin Beach 1499 . 2
6 1 1 1 . . . . 2
7 1 1 2 Hawaii Bed & Breakfast Mountains 1249 No 2
8 1 1 2 Alaska Bed & Breakfast Lake 1249 . 2
9 1 1 2 Mexico Bed & Breakfast Mountains 999 Side Trip 1
10 1 1 2 California Hotel Lake 1499 . 2
11 1 1 2 Maine Hotel Mountains 1499 . 2
12 1 1 2 . . . . 2
13 1 1 3 Hawaii Hotel Lake 1499 Side Trip 1
14 1 1 3 Alaska Bed & Breakfast Mountains 1249 . 2
15 1 1 3 Mexico Cabin Mountains 1499 No 2
16 1 1 3 California Cabin Beach 1499 . 2
17 1 1 3 Maine Cabin Beach 1249 . 2
18 1 1 3 . . . . 2
Vacation Example, with Alternative-Specific Attributes 169

Indicator variables and labels are created using PROC TRANSREG like before.
proc transreg design=5000 data=res2 nozeroconstant norestoremissing;
model class(place / zero=none order=data)
class(price scene lodge / zero=none order=formatted)
class(place * side / zero=’ ’ ’No’ separators=’’ ’ ’) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
id subj set form c;
run;
proc print data=coded(obs=6) label;
run;
The design=5000 option specifies that no model is fit; the procedure is just being used to code a design
in blocks of 5000 observations at a time. The nozeroconstant option specifies that if the coding creates
a constant variable, it should not be zeroed. The norestoremissing option specifies that missing values
should not be restored when the out= data set is created. The model statement names the variables to code and
provides information about how they should be coded. The specification class(place / ...) specifies
that the variable Place is a classification variable and requests a binary coding. The zero=none option
creates binary variables for all categories. The order=data option sorts the levels into the order they were
first encountered in the data set. Similarly, the variables Price, Scene, and Lodge are classification variables.
The specification class(place * side / ...) creates alternative-specific side trip effects. The option
zero=’ ’ ’No’ specifies that dummy variables should be created for all levels of Place except blank, and
all levels of Side except ’No’. The specification zero=’ ’ is almost the same as zero=none. The zero=’
’ specification names a missing level as the reference level creating dummy variables for all nonmissing levels
of the class variables, just like zero=none. The difference is zero=none applies to all of the variables
named in the class specification. When you want zero=none to apply to only some variables, then you must
use zero=’ ’, as in zero=’ ’ ’No’ instead. In this case, zero=none applies to the first variable and
zero=’No’ applies to the second. With zero=’ ’, TRANSREG prints the following warning, which can be
safely ignored.
WARNING: Reference level ZERO=’’ was not found for variable Place.
The separators=” ’ ’ option (separators= quote quote space quote space quote) allows you to specify
two label component separators for the main effect and interaction terms, respectively. By specifying a blank for
the second value, we request labels for the side trip effects like ’Mexico Side Trip’ instead of the default
’Mexico * Side Trip’. This option is explained in more detail on page 177.
The lprefix=0 option specifies that when labels are created for the binary variables, zero characters of the
original variable name should be used as a prefix. This means that the labels are created only from the level
values. An output statement names the output data set and drops variables that are not needed. Finally, the id
statement names the additional variables that we want copied from the input to the output data set.

Vacation Example with Asymmetry

Obs Hawaii Alaska Mexico California Maine 999 1249 1499 1749 Beach Lake

1 1 0 0 0 0 0 0 0 1 1 0
2 0 1 0 0 0 0 1 0 0 0 1
3 0 0 1 0 0 0 1 0 0 0 1
4 0 0 0 1 0 1 0 0 0 1 0
5 0 0 0 0 1 0 0 1 0 1 0
6 0 0 0 0 0 0 0 0 0 0 0
170 TS-677E Multinomial Logit, Discrete Choice Modeling

Alaska Hawaii Maine Mexico

Bed & Side California Side Side Side
Obs Mountains Breakfast Cabin Hotel Trip Side Trip Trip Trip Trip

1 0 0 1 0 0 0 0 0 0
2 0 0 1 0 0 0 0 0 0
3 0 0 1 0 0 0 0 0 1
4 0 1 0 0 0 0 0 0 0
5 0 0 1 0 0 0 0 0 0
6 0 0 0 0 0 0 0 0 0

Obs Place Price Scene Lodge Side Subj Set Form c

1 Hawaii 1749 Beach Cabin No 1 1 1 2

2 Alaska 1249 Lake Cabin . 1 1 1 2
3 Mexico 1249 Lake Cabin Side Trip 1 1 1 2
4 California 999 Beach Bed & Breakfast . 1 1 1 1
5 Maine 1499 Beach Cabin . 1 1 1 2
6 . . . . 1 1 1 2

The PROC PHREG specification is the same as we have used before. (Recall that we used %phchoice(on)on
page 79 to customize the output from PROC PHREG.)
proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
Here are the results.

Vacation Example with Asymmetry

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 7200 6 1 5

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Vacation Example, with Alternative-Specific Attributes 171

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 25801.336 12829.273

AIC 25801.336 12857.273
SBC 25801.336 12953.618

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 12972.0636 14 <.0001

Score 12091.5568 14 <.0001
Wald 5019.1086 14 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Hawaii 1 3.42197 0.21382 256.1254 <.0001

Alaska 1 -1.11090 0.25638 18.7745 <.0001
Mexico 1 2.09610 0.22046 90.4029 <.0001
California 1 1.32260 0.21993 36.1656 <.0001
Maine 1 0.61641 0.22344 7.6105 0.0058
999 1 2.08485 0.07072 868.9811 <.0001
1249 1 1.43088 0.06177 536.5227 <.0001
1499 1 0.72192 0.05905 149.4626 <.0001
1749 0 0 . . .
Beach 1 1.41944 0.04504 993.3905 <.0001
Lake 1 0.81476 0.04675 303.7324 <.0001
Mountains 0 0 . . .
Bed & Breakfast 1 0.74857 0.04146 325.9274 <.0001
Cabin 1 -1.38428 0.04908 795.6207 <.0001
Hotel 0 0 . . .
Alaska Side Trip 0 0 . . .
California Side Trip 0 0 . . .
Hawaii Side Trip 1 0.69748 0.05815 143.8627 <.0001
Maine Side Trip 0 0 . . .
Mexico Side Trip 1 0.58787 0.06192 90.1210 <.0001

You would not expect the part-worth utilities to match those that were used to generate the data, but you would
expect a similar ordering within each factor, and in fact that does occur. These data can also be analyzed with
quantitative price effects and destination by attribute interactions, as in the previous vacation example.

Aggregating the Data

This data set is rather large with 43,200 observations. You can make the analysis run faster and with less memory
by aggregating. Instead of stratifying on each choice set and subject combination, you can stratify just on choice
set and specify the number of times each alternative was chosen or unchosen. First, use PROC SUMMARY to
count the number of times each observation occurs. Specify all the analysis variables, and in this example, also
specify Form. The variable Form was added to the list because Set designates choice set within form. It is the
Form and Set combinations that identify the choice sets. (In the previous PROC PHREG step, since the Subj
* Set combinations uniquely identified each stratum, Form was not needed.) PROC SUMMARY stores the
number of times each unique observation appears in the variable - freq- . PROC PHREG is then run with a
172 TS-677E Multinomial Logit, Discrete Choice Modeling

freq statement. Now, instead of analyzing a data set with 43,200 observations and 7200 strata, we analyze a
data set with at most 2  6  72 = 864 observations and 72 strata. For each of the 6 alternatives and 72 choice
sets, there are typically 2 observations in the aggregate data set: one that contains the number of times it was
chosen and one that contains the number of times it was not chosen. When one of those counts is zero, there will
be one observation. In this case, the aggregate data set has 726 observations.
proc summary data=coded nway;
class form set c &_trgind;
output out=agg(drop=_type_);
run;

proc phreg data=agg;

model c*c(2) = &_trgind / ties=breslow;
freq _freq_;
strata form set;
run;
PROC SUMMARY ran in three seconds, and PROC PHREG ran in less than one second. The parameter estimates
and Chi-Square statistics (not shown) are the same as before. The summary table shows the results of the
aggregation, 100 out of 600 alternatives were chosen in each stratum. The log likelihood statistics are different,
but that does not matter since the Chi-Square statistics are the same. The next example provides more information
about this.

Vacation Example with Asymmetry

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Frequency Variable _FREQ_
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Form Set Alternatives Alternatives Chosen

1 1 1 600 100 500

2 1 2 600 100 500
3 1 3 600 100 500
.
.
.
71 4 71 600 100 500
72 4 72 600 100 500
---------------------------------------------------------------------------
Total 36000 6000 30000
Brand Choice Example with Aggregate Data 173

Brand Choice Example with Aggregate Data

In this next example, subjects were presented with brands of a product at different prices. There were four brands
and a constant alternative, eight choice sets, and 100 subjects. This example shows how to handle data that come
to you already aggregated. It also illustrates comparing the fits of two competing models, the mother logit model,
cross effects, IIA, and techniques for handling large data sets. The choice sets, with the price of each alternative
and the number of times it was chosen in parentheses, are shown next.
Set Brand 1 Brand 2 Brand 3 Brand 4 Other
1 $3.99 (4) $5.99 (29) $3.99 (16) $5.99 (42) $4.99 (9)
2 $5.99 (12) $5.99 (19) $5.99 (22) $5.99 (33) $4.99 (14)
3 $5.99 (34) $5.99 (26) $3.99 (8) $3.99 (27) $4.99 (5)
4 $5.99 (13) $3.99 (37) $5.99 (15) $3.99 (27) $4.99 (8)
5 $5.99 (49) $3.99 (1) $3.99 (9) $5.99 (37) $4.99 (4)
6 $3.99 (31) $5.99 (12) $5.99 (6) $3.99 (18) $4.99 (33)
7 $3.99 (37) $3.99 (10) $5.99 (5) $5.99 (35) $4.99 (13)
8 $3.99 (16) $3.99 (14) $3.99 (5) $3.99 (51) $4.99 (14)
The first choice set consists of Brand 1 at $3.99, Brand 2 at $5.99, Brand 3 at $3.99, Brand 4 at $5.99, and Other
at $4.99. From this choice set, Brand 1 was chosen 4 times, Brand 2 was chosen 29 times, Brand 3 was chosen
16 times, Brand 4 was chosen 42 times, and Other was chosen 9 times.

Processing the Data

As in the previous examples, we will process the data to create a data set with one stratum for each choice set
within each subject and m alternatives per stratum. This example will have 100 people times 5 alternatives times
8 choice sets equals 4000 observations. The first five observations are for the first subject and the first choice
set, the next five observations are for the second subject and the first choice set, ..., the next five observations are
for the one-hundredth subject and the first choice set, the next five observations are for the first subject and the
second choice set, and so on. Subject 1 in the first choice set is almost certainly not the same as subject 1 in
subsequent choice sets since we were given aggregate data. However, that is not important. What is important is
that we have a subject and choice set variable whose unique combinations identify each choice set within each
subject. In previous examples, we specified strata Subj Set with PROC PHREG, and our data were sorted
by choice set within subject. We can still use the same specification even though our data are now sorted by
subject within choice set. This next step reads and prepares the data.
%let m = 5; /* Number of Brands in Each Choice Set */
/* (including Other) */

title ’Brand Choice Example, Multinomial Logit Model’;

proc format;
value brand 1 = ’Brand 1’ 2 = ’Brand 2’ 3 = ’Brand 3’
4 = ’Brand 4’ 5 = ’Other’;
run;
data price;
array p[&m] p1-p&m; /* Prices for the Brands */
array f[&m] f1-f&m; /* Frequency of Choice */

input p1-p&m f1-f&m;

keep subj set brand price c p1-p&m;
* Store choice set and subject number to stratify;
Set = _n_; Subj = 0;
174 TS-677E Multinomial Logit, Discrete Choice Modeling

do i = 1 to &m; /* Loop over the &m frequencies */

do ci = 1 to f[i]; /* Loop frequency of choice times */
subj + 1; /* Subject within choice set */
do Brand = 1 to &m; /* Alternatives within choice set */

Price = p[brand];

* Output first choice: c=1, unchosen: c=2;

c = 2 - (i eq brand); output;
end;
end;
end;

format brand brand.;

datalines;
3.99 5.99 3.99 5.99 4.99 4 29 16 42 9
5.99 5.99 5.99 5.99 4.99 12 19 22 33 14
5.99 5.99 3.99 3.99 4.99 34 26 8 27 5
5.99 3.99 5.99 3.99 4.99 13 37 15 27 8
5.99 3.99 3.99 5.99 4.99 49 1 9 37 4
3.99 5.99 5.99 3.99 4.99 31 12 6 18 33
3.99 3.99 5.99 5.99 4.99 37 10 5 35 13
3.99 3.99 3.99 3.99 4.99 16 14 5 51 14
;

proc print data=price(obs=15);

var subj set c price brand;
run;
The inner loop do Brand = 1 to &m creates all of the observations for the m alternatives within a per-
son/choice set combination. Within a choice set (row of input data), the outer two loops, do i = 1 to &m
and do ci = 1 to f[i] execute the code inside 100 times, the variable Subj goes from 1 to 100. In the
first choice set, they first create the data for the four subjects that chose Brand 1, then the data for the 29 subjects
that chose Brand 2, and so on. Here are the first 15 observations of the data set.

Brand Choice Example, Multinomial Logit Model

Obs Subj Set c Price Brand

1 1 1 1 3.99 Brand 1
2 1 1 2 5.99 Brand 2
3 1 1 2 3.99 Brand 3
4 1 1 2 5.99 Brand 4
5 1 1 2 4.99 Other
6 2 1 1 3.99 Brand 1
7 2 1 2 5.99 Brand 2
8 2 1 2 3.99 Brand 3
9 2 1 2 5.99 Brand 4
10 2 1 2 4.99 Other
11 3 1 1 3.99 Brand 1
12 3 1 2 5.99 Brand 2
13 3 1 2 3.99 Brand 3
14 3 1 2 5.99 Brand 4
15 3 1 2 4.99 Other
Brand Choice Example with Aggregate Data 175

Note that the data set also contains the variables p1-p5 which contain the prices of each of the alternatives.
These variables, which are used in constructing the cross effects, will be discussed in more detail on page 179.
proc print data=price(obs=5);
run;

Brand Choice Example, Multinomial Logit Model

Obs p1 p2 p3 p4 p5 Set Subj Brand Price c

1 3.99 5.99 3.99 5.99 4.99 1 1 Brand 1 3.99 1

2 3.99 5.99 3.99 5.99 4.99 1 1 Brand 2 5.99 2
3 3.99 5.99 3.99 5.99 4.99 1 1 Brand 3 3.99 2
4 3.99 5.99 3.99 5.99 4.99 1 1 Brand 4 5.99 2
5 3.99 5.99 3.99 5.99 4.99 1 1 Other 4.99 2

Simple Price Effects

The data are coded using PROC TRANSREG.
proc transreg design data=price nozeroconstant norestoremissing;
model class(brand / zero=none) identity(price) / lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’;
id subj set c;
run;
The design option specifies that no model is fit; the procedure is just being used to code a design. The
nozeroconstant option specifies that if the coding creates a constant variable, it should not be zeroed. The
norestoremissing option specifies that missing values should not be restored when the out= data set is
created. The model statement names the variables to code and provides information about how they should be
coded. The specification class(brand / zero=none) specifies that the variable Brand is a classification
variable and requests a binary coding. The zero=none option creates binary variables for all categories. The
specification identity(price) specifies that the variable Price is quantitative and hence should directly
enter the model without coding. The lprefix=0 option specifies that when labels are created for the binary
variables, zero characters of the original variable name should be used as a prefix. This means that the labels are
created only from the level values. An output statement names the output data set and drops variables that are
not needed. Finally, the id statement names the additional variables that we want copied from the input to the
output data set.
proc phreg data=coded brief;
title2 ’Discrete Choice with Common Price Effect’;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;

title2;
176 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are the results. (Recall that we used %phchoice(on)on page 79 to customize the output from PROC
PHREG.)

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Common Price Effect

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 800 5 1 4

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 2575.101 2425.214

AIC 2575.101 2435.214
SBC 2575.101 2458.637

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 149.8868 5 <.0001

Score 153.2328 5 <.0001
Wald 142.9002 5 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 0.66727 0.12305 29.4065 <.0001

Brand 2 1 0.38503 0.12962 8.8235 0.0030
Brand 3 1 -0.15955 0.14725 1.1740 0.2786
Brand 4 1 0.98964 0.11720 71.2993 <.0001
Other 0 0 . . .
Price 1 0.14966 0.04406 11.5379 0.0007
Brand Choice Example with Aggregate Data 177

Alternative-Specific Price Effects

In the next step, the data are coded for fitting a multinomial logit model with brand by price effects.
proc transreg design data=price nozeroconstant norestoremissing;
model class(brand / zero=none separators=’’ ’ ’) |
identity(price) / lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’;
id subj set c;
run;
The PROC TRANSREG model statement has a vertical bar, “|”, between the class specification and the
identity specification. Since the zero=none option is specified with class, the vertical bar creates two
sets of variables: five dummy variables for the brand effects and five more variables for the brand by price
interactions. The separators= option allows you to specify two label component separators as quoted strings.
The specification separators=” ’ ’ (separators= quote quote space quote space quote) specifies a
null string (quote quote) and a blank (quote space quote). The separators=” ’ ’ option in the class
specification specifies the separators that are used to construct the labels for the main effect and interaction
terms, respectively. By default, the alternative-specific price effects the brand by price interactions would
have labels like ’Brand 1 * Price’ since the default second value for separators= is ’ * ’ (a quoted
space asterisk space). Specifying ’ ’ (a quoted space) as the second value creates labels of the form ’Brand
1 Price’. Since lprefix=0, the main-effects separator, which is the first separators= value, ” (quote
quote), is ignored. Zero name or input variable label characters are used to construct the label. The label is simply
the formatted value of the class variable. The next steps print the first two coded choice sets and perform the
analysis.
proc print data=coded(obs=10) label;
title2 ’Discrete Choice with Brand by Price Effects’;
var subj set c brand price &_trgind;
run;
proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;

title2;

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Brand by Price Effects

Brand Brand Brand Brand

Obs Subj Set c Brand Price 1 2 3 4

1 1 1 1 Brand 1 3.99 1 0 0 0
2 1 1 2 Brand 2 5.99 0 1 0 0
3 1 1 2 Brand 3 3.99 0 0 1 0
4 1 1 2 Brand 4 5.99 0 0 0 1
5 1 1 2 Other 4.99 0 0 0 0
6 2 1 1 Brand 1 3.99 1 0 0 0
7 2 1 2 Brand 2 5.99 0 1 0 0
8 2 1 2 Brand 3 3.99 0 0 1 0
9 2 1 2 Brand 4 5.99 0 0 0 1
10 2 1 2 Other 4.99 0 0 0 0
178 TS-677E Multinomial Logit, Discrete Choice Modeling

Brand 1 Brand 2 Brand 3 Brand 4 Other

Obs Other Price Price Price Price Price

1 0 3.99 0.00 0.00 0.00 0.00

2 0 0.00 5.99 0.00 0.00 0.00
3 0 0.00 0.00 3.99 0.00 0.00
4 0 0.00 0.00 0.00 5.99 0.00
5 1 0.00 0.00 0.00 0.00 4.99
6 0 3.99 0.00 0.00 0.00 0.00
7 0 0.00 5.99 0.00 0.00 0.00
8 0 0.00 0.00 3.99 0.00 0.00
9 0 0.00 0.00 0.00 5.99 0.00
10 1 0.00 0.00 0.00 0.00 4.99

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Brand by Price Effects

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 800 5 1 4

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 2575.101 2424.812

AIC 2575.101 2440.812
SBC 2575.101 2478.288

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 150.2891 8 <.0001

Score 154.2563 8 <.0001
Wald 143.1425 8 <.0001
Brand Choice Example with Aggregate Data 179

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 -0.00972 0.43555 0.0005 0.9822

Brand 2 1 -0.62230 0.48866 1.6217 0.2028
Brand 3 1 -0.81250 0.60318 1.8145 0.1780
Brand 4 1 0.31778 0.39549 0.6456 0.4217
Other 0 0 . . .
Brand 1 Price 1 0.13587 0.08259 2.7063 0.1000
Brand 2 Price 1 0.20074 0.09210 4.7512 0.0293
Brand 3 Price 1 0.13126 0.11487 1.3057 0.2532
Brand 4 Price 1 0.13478 0.07504 3.2255 0.0725
Other Price 0 0 . . .

The likelihood for this model is essentially the same as for the simpler, common-price-slope model fit previously,
2 log(LC ) = 2425:214 compared to 2424.812. You can test the null hypothesis that the two models are not
significantly different by comparing their likelihoods. The difference between two 2 log(LC )’s (the number
reported under ’With Covariates’ in the output) has a chi-square distribution. We can get the df for the test by
subtracting the two df for the two likelihoods. The difference 2425:214 2424:812 = 0:402 is distributed 2
with 8 5 = 3 df and is not statistically significant.

Mother Logit Model

This next step fits the so-called “mother logit” model. This step creates the full design matrix, including the
brand, price, and cross effects. A cross effect represents the effect of one alternative on the utility of another
alternative. First, let’s look at the input data set for the first choice set.
proc print data=price(obs=5) label;
run;

Brand Choice Example, Multinomial Logit Model

Obs p1 p2 p3 p4 p5 Set Subj Brand Price c

1 3.99 5.99 3.99 5.99 4.99 1 1 Brand 1 3.99 1

2 3.99 5.99 3.99 5.99 4.99 1 1 Brand 2 5.99 2
3 3.99 5.99 3.99 5.99 4.99 1 1 Brand 3 3.99 2
4 3.99 5.99 3.99 5.99 4.99 1 1 Brand 4 5.99 2
5 3.99 5.99 3.99 5.99 4.99 1 1 Other 4.99 2

The input consists of Set, Subj, Brand, Price, and a choice time variable c. In addition, it contains five
variables p1 through p5. The first observation of the Price variable shows us that the first alternative costs
$3.99; p1 contains the cost of alternative 1, $3.99, which is the same for all alternatives. It does not matter which
alternative you are looking at, p1 shows that alternative 1 costs $3.99. Similarly, the second observation of the
Price variable shows us that the second alternative costs $5.99; p2 contains the cost of alternative 2, $5.99,
which is the same for all alternatives. There is one price variable, p1 through p5, for each of the five alternatives.
In all of the previous examples, we have used models that were coded so that the utility of an alternative only
depended on the attributes of that alternative. For example, the utility of Brand 1 would only depend on the Brand
1 name and its price. In contrast, p1-p5 contain information about each of the other alternatives’ attributes. We
will construct cross effects using the interaction of p1-p5 and the Brand variable. In a model with cross effects,
the utility for an alternative depends on both that alternative’s attributes and the other alternatives’ attributes. The
IIA (independence from irrelevant alternatives) property states that utility only depends on an alternative’s own
attributes. Cross effects add other alternative’s attributes to the model, so they can be used to test for violations
of IIA. (See pages 185, 192, 379, and 383 for other discussions of IIA.) Here is the PROC TRANSREG code for
180 TS-677E Multinomial Logit, Discrete Choice Modeling

the cross-effects model.

proc transreg design data=price nozeroconstant norestoremissing;
model class(brand / zero=none separators=’’ ’ ’) | identity(price)
identity(p1-p&m) *
class(brand / zero=none lprefix=0 separators=’’ ’ on ’) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’
p1 = ’Brand 1’ p2 = ’Brand 2’ p3 = ’Brand 3’
p4 = ’Brand 4’ p5 = ’Other’;
id subj set c;
run;
The class(brand / ...) | identity(price) specification in the model statement is the same as
the previous analysis. The additional terms, identity(p1-p&m) * class(brand / ...) create the
cross effects. The second value of the separators= option, ’ on ’ is used to create labels like ’Brand 1
on Brand 2’ instead of the default ’Brand 1 * Brand 2’. It is important to note that you must specify
the cross effect by specifying identity with the price factors, followed by the asterisk, followed by class
and the brand effect, in that order. The order of the specification determines the order in which brand names are
added to the labels. Do not specify the brand variable first; doing so will create incorrect labels.
With m alternatives, there are m  m cross effects, but as we will see, many of them are zero. The first coded
choice set is printed with the following PROC PRINT steps. Multiple steps are used to facilitate explaining the
coding.
title2 ’Discrete Choice with Cross Effects, Mother Logit’;
proc print data=coded(obs=5) label; var subj set c brand price; run;
proc print data=coded(obs=5) label; var Brand:; run;
proc print data=coded(obs=5) label; var p1B:; id brand; run;
proc print data=coded(obs=5) label; var p2B:; id brand; run;
proc print data=coded(obs=5) label; var p3B:; id brand; run;
proc print data=coded(obs=5) label; var p4B:; id brand; run;
proc print data=coded(obs=5) label; var p5B:; id brand; run;
The coded data set contains the strata variable Subj and Set, choice time variable c, and Brand and Price.
Brand and Price were used to create the coded independent variables but they are not used in the analysis
with PROC PHREG.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Obs Subj Set c Brand Price

1 1 1 1 Brand 1 3.99
2 1 1 2 Brand 2 5.99
3 1 1 2 Brand 3 3.99
4 1 1 2 Brand 4 5.99
5 1 1 2 Other 4.99

The effects ’Brand 1’ through ’Other’ in the next output are the binary brand effect variables. They indi-
cate the brand for each alternative. The effects ’Brand 1 Price’ through ’Other Price’ are alternative-
specific price effects. They indicate the price for each alternative. All ten of these variables are independent
variables in the analysis, and their names are part of the &- trgind macro variable list, as are all of the cross
effects that are described next.
Brand Choice Example with Aggregate Data 181

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Brand Brand Brand Brand Brand 1 Brand 2 Brand 3 Brand 4 Other

Obs 1 2 3 4 Other Price Price Price Price Price Brand

1 1 0 0 0 0 3.99 0.00 0.00 0.00 0.00 Brand 1

2 0 1 0 0 0 0.00 5.99 0.00 0.00 0.00 Brand 2
3 0 0 1 0 0 0.00 0.00 3.99 0.00 0.00 Brand 3
4 0 0 0 1 0 0.00 0.00 0.00 5.99 0.00 Brand 4
5 0 0 0 0 1 0.00 0.00 0.00 0.00 4.99 Other

The effects ’Brand 1 on Brand 1’ through ’Brand 1 on Other’ in the next output are the first five
cross effects.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Brand Brand Brand Brand

1 on 1 on 1 on 1 on Brand 1
Brand Brand 1 Brand 2 Brand 3 Brand 4 on Other

Brand 1 3.99 0.00 0.00 0.00 0.00

Brand 2 0.00 3.99 0.00 0.00 0.00
Brand 3 0.00 0.00 3.99 0.00 0.00
Brand 4 0.00 0.00 0.00 3.99 0.00
Other 0.00 0.00 0.00 0.00 3.99

They represent the effect of Brand 1 at its price on the utility of each alternative. The label ’Brand n on
Brand m’ is read as ’the effect of Brand n at its price on the utility of Brand m.’ For the first choice set, these
first five cross effects consist entirely of zeros and $3.99’s, where $3.99 is the price of Brand 1 in this choice set.
The nonzero value is constant across all of the alternatives in each choice set since Brand 1 has only one price
in each choice set. Notice the ’Brand 1 on Brand 1’ term, which is the effect of Brand 1 at its price on
the utility of Brand 1. Also notice the ’Brand 1 Price’ effect, which is shown in the previous output. The
description ’the effect of Brand 1 at its price on the utility of Brand 1’ is just a convoluted way of describing
the Brand 1 price effect. The ’Brand 1 on Brand 1’ cross effect is the same as the Brand 1 price effect,
hence when we do the analysis, we will see that the coefficient for the ’Brand 1 on Brand 1’ cross effect
is zero.
The effects ’Brand 2 on Brand 1’ through ’Brand 2 on Other’ in the next output are the next five
cross effects.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Brand Brand Brand Brand

2 on 2 on 2 on 2 on Brand 2
Brand Brand 1 Brand 2 Brand 3 Brand 4 on Other

Brand 1 5.99 0.00 0.00 0.00 0.00

Brand 2 0.00 5.99 0.00 0.00 0.00
Brand 3 0.00 0.00 5.99 0.00 0.00
Brand 4 0.00 0.00 0.00 5.99 0.00
Other 0.00 0.00 0.00 0.00 5.99
182 TS-677E Multinomial Logit, Discrete Choice Modeling

They represent the effect of Brand 2 at its price on the utility of each alternative. For the first choice set, these
five cross effects consist entirely of zeros and $5.99’s, where $5.99 is the price of Brand 2 in this choice set. The
nonzero value is constant across all of the alternatives in each choice set since Brand 2 has only one price in each
choice set. Notice the ’Brand 2 on Brand 2’ term, which is the effect of Brand 2 at its price on the utility
of Brand 2. The description “the effect of Brand 2 at its price on the utility of Brand 2” is just a convoluted way
of describing the Brand 2 price effect. The ’Brand 2 on Brand 2’ cross effect is the same as the Brand
2 price effect, hence when we do the analysis, we will see that the coefficient for the ’Brand 2 on Brand
2’ cross effect is zero.
The effects ’Brand 3 on Brand 1’ through ’Brand 3 on Other’ in the next output are the next five
cross effects.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Brand Brand Brand Brand

3 on 3 on 3 on 3 on Brand 3
Brand Brand 1 Brand 2 Brand 3 Brand 4 on Other

Brand 1 3.99 0.00 0.00 0.00 0.00

Brand 2 0.00 3.99 0.00 0.00 0.00
Brand 3 0.00 0.00 3.99 0.00 0.00
Brand 4 0.00 0.00 0.00 3.99 0.00
Other 0.00 0.00 0.00 0.00 3.99

They represent the effect of Brand 3 at its price on the utility of each alternative. For the first choice set, these
five cross effects consist entirely of zeros and $3.99’s, where $3.99 is the price of Brand 3 in this choice set.
Notice that the ’Brand 3 on Brand 3’ term is the same as the Brand 3 price effect, hence when we do the
analysis, we will see that the coefficient for the ’Brand 3 on Brand 3’ cross effect is zero.
Here are the remaining cross effects. They follow the same pattern that was described for the previous cross
effects.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Brand Brand Brand Brand

4 on 4 on 4 on 4 on Brand 4
Brand Brand 1 Brand 2 Brand 3 Brand 4 on Other

Brand 1 5.99 0.00 0.00 0.00 0.00

Brand 2 0.00 5.99 0.00 0.00 0.00
Brand 3 0.00 0.00 5.99 0.00 0.00
Brand 4 0.00 0.00 0.00 5.99 0.00
Other 0.00 0.00 0.00 0.00 5.99

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Cross Effects, Mother Logit

Other on Other on Other on Other on Other on

Brand Brand 1 Brand 2 Brand 3 Brand 4 Other

Brand 1 4.99 0.00 0.00 0.00 0.00

Brand 2 0.00 4.99 0.00 0.00 0.00
Brand 3 0.00 0.00 4.99 0.00 0.00
Brand 4 0.00 0.00 0.00 4.99 0.00
Other 0.00 0.00 0.00 0.00 4.99
Brand Choice Example with Aggregate Data 183

We have been describing variables by their labels. While it is not necessary to look at it, the &- trgind macro
variable name list that PROC TRANSREG creates for this problem is as follows:
%put &_trgind;
BrandBrand_1 BrandBrand_2 BrandBrand_3 BrandBrand_4 BrandOther
BrandBrand_1Price BrandBrand_2Price BrandBrand_3Price BrandBrand_4Price
BrandOtherPrice p1BrandBrand_1 p1BrandBrand_2 p1BrandBrand_3 p1BrandBrand_4
p1BrandOther p2BrandBrand_1 p2BrandBrand_2 p2BrandBrand_3 p2BrandBrand_4
p2BrandOther p3BrandBrand_1 p3BrandBrand_2 p3BrandBrand_3 p3BrandBrand_4
p3BrandOther p4BrandBrand_1 p4BrandBrand_2 p4BrandBrand_3 p4BrandBrand_4
p4BrandOther p5BrandBrand_1 p5BrandBrand_2 p5BrandBrand_3 p5BrandBrand_4
p5BrandOther
The analysis proceeds in exactly the same manner as before.
proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
Brand Choice Example, Multinomial Logit Model
Discrete Choice with Cross Effects, Mother Logit

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 800 5 1 4

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 2575.101 2349.325

AIC 2575.101 2389.325
SBC 2575.101 2483.018

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 225.7752 20 <.0001

Score 218.4500 20 <.0001
Wald 190.0257 20 <.0001
184 TS-677E Multinomial Logit, Discrete Choice Modeling

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 1.24963 1.31259 0.9064 0.3411

Brand 2 1 -0.16269 1.38579 0.0138 0.9065
Brand 3 1 -3.90179 1.56511 6.2150 0.0127
Brand 4 1 2.49435 1.25537 3.9480 0.0469
Other 0 0 . . .
Brand 1 Price 1 0.51056 0.13178 15.0096 0.0001
Brand 2 Price 1 -0.04920 0.13411 0.1346 0.7137
Brand 3 Price 1 -0.27594 0.15517 3.1623 0.0754
Brand 4 Price 1 0.28951 0.12192 5.6389 0.0176
Other Price 0 0 . . .
Brand 1 on Brand 1 0 0 . . .
Brand 1 on Brand 2 1 0.51651 0.13675 14.2653 0.0002
Brand 1 on Brand 3 1 0.66122 0.15655 17.8397 <.0001
Brand 1 on Brand 4 1 0.32806 0.12664 6.7105 0.0096
Brand 1 on Other 0 0 . . .
Brand 2 on Brand 1 1 -0.39876 0.12832 9.6561 0.0019
Brand 2 on Brand 2 0 0 . . .
Brand 2 on Brand 3 1 -0.01755 0.15349 0.0131 0.9090
Brand 2 on Brand 4 1 -0.33802 0.12220 7.6512 0.0057
Brand 2 on Other 0 0 . . .
Brand 3 on Brand 1 1 -0.43868 0.13119 11.1823 0.0008
Brand 3 on Brand 2 1 -0.31541 0.13655 5.3356 0.0209
Brand 3 on Brand 3 0 0 . . .
Brand 3 on Brand 4 1 -0.54854 0.12528 19.1723 <.0001
Brand 3 on Other 0 0 . . .
Brand 4 on Brand 1 1 0.24398 0.12781 3.6443 0.0563
Brand 4 on Brand 2 1 -0.01214 0.13416 0.0082 0.9279
Brand 4 on Brand 3 1 0.40500 0.15285 7.0211 0.0081
Brand 4 on Brand 4 0 0 . . .
Brand 4 on Other 0 0 . . .
Other on Brand 1 0 0 . . .
Other on Brand 2 0 0 . . .
Other on Brand 3 0 0 . . .
Other on Brand 4 0 0 . . .
Other on Other 0 0 . . .

The results consist of:

 four nonzero brand effects and a zero for the constant alternative
 four nonzero alternative-specific price effects and a zero for the constant alternative
 5  5 = 25 cross effects, the number of alternatives squared, but only (5 1)  (5 2) = 12 of them are
nonzero (four brands not counting Other affecting each of the remaining three brands).
 There are three cross effects for the effect of Brand 1 on Brands 2, 3, and 4.
 There are three cross effects for the effect of Brand 2 on Brands 1, 3, and 4.
 There are three cross effects for the effect of Brand 3 on Brands 1, 2, and 4.
 There are three cross effects for the effect of Brand 4 on Brands 1, 2, and 3.

All coefficients for the constant (other) alternative are zero as are the cross effects of a brand on itself.
Brand Choice Example with Aggregate Data 185

The mother logit model is used to test for violations of IIA (independence from irrelevant alternatives). IIA means
the odds of choosing alternative ci over cj do not depend on the other alternatives in the choice set. Ideally, this
more general model will not significantly explain more variation in choice than the restricted models. Also, if IIA
is satisfied, few if any of the cross-effect terms should be significantly different from zero. (See pages 179, 192,
379, and 383 for other discussions of IIA.) In this case, it appears that IIA is not satisfied (the data are artificial),
so the more general mother logit model is needed. The chi-square statistic is 2424:812 2349:325 = 75:487
with 20 8 = 12 df (p < 0:0001).
You could eliminate some of the zero parameters by changing zero=none to zero=’Other’ and eliminating
p5 (p&m) from the model.
proc transreg design data=price nozeroconstant norestoremissing;
model class(brand / zero=’Other’ separators=’’ ’ ’) | identity(price)
identity(p1-p4) * class(brand / zero=’Other’ separators=’’ ’ on ’) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’
p1 = ’Brand 1’ p2 = ’Brand 2’ p3 = ’Brand 3’
p4 = ’Brand 4’;
id subj set c;
run;
You could also eliminate the brand by price effects and instead capture brand by price effects as the cross effect
of a variable on itself.
proc transreg design data=price nozeroconstant norestoremissing;
model class(brand / zero=’Other’ separators=’’ ’ ’)
identity(p1-p4) * class(brand / zero=’Other’ separators=’’ ’ on ’) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’
p1 = ’Brand 1’ p2 = ’Brand 2’ p3 = ’Brand 3’
p4 = ’Brand 4’;
id subj set c;
run;
In both cases, the analysis (not shown) would be run in the usual manner. Except for the elimination of zero
terms, and in the second case, the change to capture the price effects in the cross effects, the results are identical.

Aggregating the Data

In all examples so far (except the last part of the last vacation example), the data set has been created for analysis
with one stratum for each choice set and subject combination. Such data sets can be large. The data can also be
arrayed with a frequency variable and each choice set forming a separate stratum. This example illustrates how.
title ’Brand Choice Example, Multinomial Logit Model’;
title2 ’Aggregate Data’;

%let m = 5; /* Number of Brands in Each Choice Set */

/* (including Other) */

proc format;
value brand 1 = ’Brand 1’ 2 = ’Brand 2’ 3 = ’Brand 3’
4 = ’Brand 4’ 5 = ’Other’;
run;

data price2;
array p[&m] p1-p&m; /* Prices for the Brands */
array f[&m] f1-f&m; /* Frequency of Choice */
186 TS-677E Multinomial Logit, Discrete Choice Modeling

input p1-p&m f1-f&m;

keep set price brand freq c p1-p&m;

* Store choice set number to stratify;

Set = _n_;
do Brand = 1 to &m;

Price = p[brand];

* Output first choice: c=1, unchosen: c=2;

Freq = f[brand]; c = 1; output;

* Output number of times brand was not chosen.;

freq = sum(of f1-f&m) - freq; c = 2; output;

end;

format brand brand.;

datalines;
3.99 5.99 3.99 5.99 4.99 4 29 16 42 9
5.99 5.99 5.99 5.99 4.99 12 19 22 33 14
5.99 5.99 3.99 3.99 4.99 34 26 8 27 5
5.99 3.99 5.99 3.99 4.99 13 37 15 27 8
5.99 3.99 3.99 5.99 4.99 49 1 9 37 4
3.99 5.99 5.99 3.99 4.99 31 12 6 18 33
3.99 3.99 5.99 5.99 4.99 37 10 5 35 13
3.99 3.99 3.99 3.99 4.99 16 14 5 51 14
;
proc print data=price2(obs=10);
var set c freq price brand;
run;

Brand Choice Example, Multinomial Logit Model

Aggregate Data

Obs Set c Freq Price Brand

1 1 1 4 3.99 Brand 1
2 1 2 96 3.99 Brand 1
3 1 1 29 5.99 Brand 2
4 1 2 71 5.99 Brand 2
5 1 1 16 3.99 Brand 3
6 1 2 84 3.99 Brand 3
7 1 1 42 5.99 Brand 4
8 1 2 58 5.99 Brand 4
9 1 1 9 4.99 Other
10 1 2 91 4.99 Other

This data set has 5 brands times 2 observations times 8 choice sets for a total of 80 observations, compared to
100  5  8 = 4000 using the standard method. Two observations are created for each alternative within each
choice set. The first contains the number of people who chose the alternative, and the second contains the number
of people who did not choose the alternative.
Brand Choice Example with Aggregate Data 187

To analyze the data, specify strata Set and freq Freq.

proc transreg design data=price2 nozeroconstant norestoremissing;
model class(brand / zero=none) identity(price) / lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’;
id freq set c;
run;

proc phreg data=coded;

title2 ’Discrete Choice with Common Price Effect, Aggregate Data’;
model c*c(2) = &_trgind / ties=breslow;
strata set;
freq freq;
run;

title2;
These steps produced the following results.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Common Price Effect, Aggregate Data

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Frequency Variable Freq
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Set Alternatives Alternatives Chosen

1 1 500 100 400

2 2 500 100 400
3 3 500 100 400
4 4 500 100 400
5 5 500 100 400
6 6 500 100 400
7 7 500 100 400
8 8 500 100 400
---------------------------------------------------------------
Total 4000 800 3200

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

188 TS-677E Multinomial Logit, Discrete Choice Modeling

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 9943.373 9793.486

AIC 9943.373 9803.486
SBC 9943.373 9826.909

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 149.8868 5 <.0001

Score 153.2328 5 <.0001
Wald 142.9002 5 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 0.66727 0.12305 29.4065 <.0001

Brand 2 1 0.38503 0.12962 8.8235 0.0030
Brand 3 1 -0.15955 0.14725 1.1740 0.2786
Brand 4 1 0.98964 0.11720 71.2993 <.0001
Other 0 0 . . .
Price 1 0.14966 0.04406 11.5379 0.0007

The summary table is small with eight rows, one row per choice set. Each row represents 100 chosen alternatives
and 400 unchosen. The ’Analysis of Maximum Likelihood Estimates’ table exactly matches the one produced
by the standard analysis. The -2 LOG L statistics are different than before: 9793.486 now compared to 2425.214
previously. This is because the data are arrayed in this example so that the partial likelihood of the proportional
hazards model fit by PROC PHREG with the ties=breslow option is now proportional to not identical
to the likelihood for the choice model. However, the Model Chi-Square statistics, df, and p-values are the
same as before. The two corresponding pairs of -2 LOG L’s differ by a constant 9943:373 2575:101 =
9793:486 2425:214 = 7368:272 = 2  800  log(100). Since the 2 is the -2 LOG L without covariates minus
-2 LOG L with covariates, the constants cancel and the 2 test is correct for both methods.
The technique of aggregating the data and using a frequency variable can be used for other models as well, for
example with brand by price effects.
proc transreg design data=price2 nozeroconstant norestoremissing;
model class(brand / zero=none separators=’’ ’ ’) |
identity(price) / lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
label price = ’Price’;
id freq set c;
run;

proc phreg data=coded;

title2 ’Discrete Choice with Brand by Price Effects, Aggregate Data’;
model c*c(2) = &_trgind / ties=breslow;
strata set;
freq freq;
run;
Brand Choice Example with Aggregate Data 189

This step produced the following results. The only thing that changes from the analysis with one stratum for each
subject and choice set combination is the likelihood.

Brand Choice Example, Multinomial Logit Model

Discrete Choice with Brand by Price Effects, Aggregate Data

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Frequency Variable Freq
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Set Alternatives Alternatives Chosen

1 1 500 100 400

2 2 500 100 400
3 3 500 100 400
4 4 500 100 400
5 5 500 100 400
6 6 500 100 400
7 7 500 100 400
8 8 500 100 400
---------------------------------------------------------------
Total 4000 800 3200

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 9943.373 9793.084

AIC 9943.373 9809.084
SBC 9943.373 9846.561

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 150.2891 8 <.0001

Score 154.2562 8 <.0001
Wald 143.1425 8 <.0001
190 TS-677E Multinomial Logit, Discrete Choice Modeling

The PHREG Procedure

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 -0.00972 0.43555 0.0005 0.9822

Brand 2 1 -0.62230 0.48866 1.6217 0.2028
Brand 3 1 -0.81250 0.60318 1.8145 0.1780
Brand 4 1 0.31778 0.39549 0.6456 0.4217
Other 0 0 . . .
Brand 1 Price 1 0.13587 0.08259 2.7063 0.1000
Brand 2 Price 1 0.20074 0.09210 4.7512 0.0293
Brand 3 Price 1 0.13126 0.11487 1.3057 0.2532
Brand 4 Price 1 0.13478 0.07504 3.2255 0.0725
Other Price 0 0 . . .

Previously, with one stratum per choice set within subject, we compared these models as follows: “The difference
2425:214 2424:812 = 0:402 is distributed 2 with 8 5 = 3 df and is not statistically significant.” The
difference between two 2 log(LC )’s equals the difference between two 2 log(LB )’s, since the constant terms
(800  log(100)) cancel, 9793:486 9793:084 = 2425:214 2424:812 = 0:402.

Choice and Breslow Likelihood Comparison

This section explains why the -2 LOG L values differ by a constant with aggregate data versus individual data.
It may be skipped by all but the most dedicated readers.
Consider the choice model with a common price slope. Let x0 represent the price of the brand. Let x1 , x2 , x3 ,
and x4 be indicator variables representing the choice of brands. Let x = (x0 x1 x2 x3 x4 ) be the vector of
alternative attributes. (A sixth element for ’Other’ is omitted, since its parameter is always zero given the other
brands.)
Consider the first choice set. There are five distinct vectors of alternative attributes
x1 = (3:99 1 0 0 0) x2 = (5:99 0 1 0 0) x3 = (3:99 0 0 1 0) x4 = (5:99 0 0 0 1)
x5 = (4:99 0 0 0 0)
The vector x2 , for example, represents choice of Brand 2, and x5 represents the choice of Other. One hundred
individuals were asked to choose one of the m = 5 brands from each of the eight sets. Let f1 , f2 , f3 , f4 , and f5
P set, f1 = 4, f2 = 29, f3 = 16, f4 = 42, and
be the number of times each brand was chosen. For the first choice
f5 = 9. Let N be the total frequency for each choice set, N = 5j=1 fj = 100. The likelihood LC1 for the first
choice set data is
P  
exp 5
=1 f xj
L1 = hP
j j
C
i
exp(x )
N
5
j =1 j

The joint likelihood for all eight choice sets is the product of the likelihoods

Y
8
L =
C LC
k
k =1
Brand Choice Example with Aggregate Data 191

The Breslow likelihood for this example, LB k

, for the kth choice set, is the same as the likelihood for the choice
model, except for a multiplicative constant.

L = N L = 100100L
C
k
N B
k
B
k

Therefore, the Breslow likelihood for all eight choice sets is

Y
8
L =
B L =N
B
k
8N
L = 100
C
800
L
C

k =1

The two likelihoods are not exactly the same, because each choice set is designated as a separate stratum, instead
of each choice set within each subject.
The log likelihood for the choice model is

log(L ) = 800  log(100) + log(L );

C B

log(L ) = 800  log(100) + ( 0:5)  9793:486;

C

log(L ) = 1212:607
C

and 2 log(LC ) = 2425:214, which matches the earlier output. However, it is usually not necessary to obtain
this value.
192 TS-677E Multinomial Logit, Discrete Choice Modeling

Food Product Example with Asymmetry and

Availability Cross Effects
This is the choice example from Kuhfeld, Tobias, and Garratt (1994), on page 25. This example discusses the
multinomial logit model, number of parameters, choosing the number of choice sets, designing the choice exper-
iment, long design searches, examining the design, examining the subdesigns, examining the aliasing structure,
blocking the design, testing the design before data collection, generating artificial data, processing the data,
coding, cross effects, availability, multinomial logit model results, modeling subject attributes, results, and inter-
pretation.
Consider the problem of using a discrete choice model to study the effect of introducing a retail food product.
This may be useful, for instance, to refine a marketing plan or to optimize a product prior to test market. A typical
brand team will have several concerns such as knowing the potential market share for the product, examining
the source of volume, and providing guidance for pricing and promotions. The brand team may also want to
know what brand attributes have competitive clout and want to identify competitive attributes to which they are
vulnerable.
To develop this further, assume our client wishes to introduce a line extension in the category of frozen entrées.
The client has one nationally branded competitor, a regional competitor in each of three regions, and a profusion
of private label products at the grocery chain level. The product may come in two different forms: stove-top or
microwaveable. The client believes that the private labels are very likely to mimic this line extension and to sell
it at a lower price. The client suspects that this strategy on the part of private labels may work for the stove-top
version but not for the microwaveable, where they have the edge on perceived quality. They also want to test the
effect of a shelf-talker that will draw attention to their product.

The Multinomial Logit Model

This problem can be set up as a discrete choice model in which a respondent’s choice among brands, given choice
set Ca of available brands, will correspond to the brand with the highest utility. For each brand i, the utility Ui
is the sum of a systematic component Vi and a random component ei . The probability of choosing brand i from
choice set Ca is therefore:

P (ijC ) = P (U > max(U )) = P (V + e > max(V + e ))

a i j i i j j 8 (j 6= i) 2 C
a

Assuming that the ei follow an extreme value type I distribution, the conditional probabilities P (ijCa ) can be
found using the multinomial logit (MNL) formulation of McFadden (1974)

P
P (ijC ) = exp(V )=
a i j 2 a exp(V )
C j

One of the consequences of the MNL formulation is the property of independence from irrelevant alternatives
(IIA). Under the assumption of IIA, all cross effects are assumed to be equal, so that if a brand gains in utility, it
draws share from all other brands in proportion to their current shares. Departures from IIA exist when certain
subsets of brands are in more direct competition and tend to draw a disproportionate amount of share from each
other than from other members in the category.
IIA is frequently described using a transportation example. Say you have three alternatives for getting to work:
bicycle, car, or a blue bus. If a fourth alternative became available, a red bus, then according to IIA the red bus
should draw riders from the other alternatives in proportion to their current usage. However, in this case, IIA
would be violated, and instead the red bus would draw more riders from the blue bus than from car drivers and
bicycle riders.
Food Product Example with Asymmetry and Availability Cross Effects 193

The mother logit formulation of McFadden (1974) can be used to capture departures from IIA. In a mother logit
model, the utility for brand i is a function of both the attributes of brand i and the attributes of other brands. The
effect of one brand’s attributes on another is termed a cross effect. In the case of designs in which only subsets
Ca of the full shelf set C appear, the effect of the presence/absence of one brand on the utility of another is
termed an availability cross effect. (See pages 179, 185, 379, and 383 for other discussions of IIA.)

Set Up
In the frozen entrée example, there are five alternatives: the client’s brand, the client’s line extension, a national
branded competitor, a regional brand and a private label brand. Several regional and private labels can be tested
in each market, then aggregated for the final model. Note that the line extension is treated as a separate alternative
rather than as a level of the client brand. This enables us to model the source of volume for the new entry and
to quantify any cannibalization that occurs. Each brand is shown at either two or three price points. Additional
price points are included so that quadratic models of price elasticity can be tested. The indicator for the presence
or absence of a brand in the shelf set is coded using one level of the Price variable. The layout of factors and
levels is given in the following table.

Factors and Levels

Alternative Factor Levels Brand Description

1 X1 4 Client 1.29, 1.69, 2.09 + absent

2 X2 4 Client Line Extension 1.39, 1.89, 2.39, + absent

X3 2 microwave/stove-top
X4 2 shelf-talker yes/no

3 X5 3 Regional 1.99, 2.49 + absent

4 X6 3 Private Label 1.49, 2.29 absent

X7 2 microwave/stove-top

5 X8 3 National 1.99 + 2.39 + absent

In addition to intercepts and main effects, we also require that all two-way interactions within alternatives be
estimable: x2*x3, x2*x4, x3*x4 for the line extension and x6*x7 for private labels. This will enable us
to test for different price elasticities by form (stove-top versus microwaveable) and to see if the promotion works
better combined with a low price or with different forms. Using a linear model for x1-x8, the total number of
parameters including the intercept, all main effects, and two-way interactions with brand is 25. This assumes that
price is treated as qualitative. The actual number of parameters in the choice model is larger than this because of
the inclusion of cross effects. Using indicator variables to code availability, the systematic component of utility
for brand i can be expressed as:

P P P
V =a +
i i k
(b  x ) +
ik ik j 6= z (d +
i j ij l
(g  x ))
ijl jl

where

ai = intercept for brand i

bik = effect of attribute k for brand i, where k = 1; ::; Ki
x ik = level of attribute k for brand i
dij
 of brand j on brand i
= availability cross effect
z 1 if j 2 Ca ;
j = availability code =
0 otherwise
gijl = cross effect of attribute l for brand j on brand i, where l = 1; ::; Lj
xjl = level of attribute l for brand j .
194 TS-677E Multinomial Logit, Discrete Choice Modeling

The xik and xjl could be expanded to include interaction and polynomial terms. In an availability-cross-effects
design, each brand is present in only a fraction of the choice sets. The size of this fraction or subdesign is
a function of the number of levels of the alternative-specific variable that is used to code availability (usually
price). For instance, if price has three valid levels and a fourth zero level to indicate absence, then the brand
will appear in only three out of four runs. Following Lazari and Anderson (1994), the size of each subdesign
determines how many model equations can be written for each brand in the discrete choice model. If Xi is
the subdesign matrix corresponding to Vi , then each Xi must be full rank to ensure that the choice set design
provides estimates for all parameters.
To create the design, a full-factorial candidate set is generated consisting of 3456 runs. It is then reduced to 2776
runs that contain between two and four brands so that the respondent is never required to compare more than four
brands at a time. In the model specification, we designate all variables as classification variables and require that
all main effects and two-way interactions within brands be estimable. The number of runs calculations are based
on the number of parameters that we wish to estimate in the various subdesigns Xi of X. Assuming that there is a
None alternative used as a reference level, the numbers of parameters required for various alternatives are shown
in the next table along with the sizes of the subdesigns (rounded down) for various numbers of runs. Parameters
for quadratic price models are given in parentheses. Note that the effect of private label being in a microwaveable
or stove-top form (stove/micro cross effect) is an explicit parameter under the client line extension.

Parameters
Client Private
Effect Client Line Extension Regional Label Competitor
intercept 1 1 1 1 1
availability cross effects 4 4 4 4 4
direct price effect 1 (2) 1 (2) 1 1 1
price cross effects 4 (8) 4 (8) 4 4 4
stove versus microwave - 1 - 1 -
stove/micro cross effects - 1 - - -
shelf-talker - 1 - - -
price*stove/microwave - 1 (2) - 1 -
price*shelf-talker - 1 (2) - - -
stove/micro*shelf-talker - 1 - - -

Total 10 (15) 16 (23) 10 12 10

Subdesign size

22 runs 16 16 14 14 14
26 runs 19 19 17 17 17
32 runs 24 24 21 21 21

The subdesign sizes are computed by taking the floor of the number of runs from the marginal times the expected
proportion of runs in which the alternative will appear. For example, for the client brand which has three prices
and not available and 22 runs, oor(22  3=4) = 16; for the competitor and 32 runs, oor(32  2=3) = 21. The
number of runs chosen was n=26. This number provides adequate degrees of freedom for the linear price model
and will also allow estimation of direct quadratic price effects. To estimate quadratic cross effects for price would
require 32 runs at the very least. Although the technique of using two-way interactions between nominal level
variables will usually guarantee that all direct and cross effects are estimable, it is sometimes necessary and good
practice to check the ranks of the subdesigns for more complex models (Lazari and Anderson 1994).

Designing the Choice Experiment

This example originated with Kuhfeld, Tobias, and Garratt (1994), long before the %MktRuns macro was pro-
grammed. At least for now, we will skip the customary step of running the %MktRuns macro to suggest a design
size and instead use the original size of 26 choice sets.
Food Product Example with Asymmetry and Availability Cross Effects 195

We will use the %MktEx autocall macro to create the design. (All of the autocall macros used in this report are
documented starting on page 287.) To recap, we want to make the design 23 33 42 in 26 runs, and we want the
following interactions to be estimable: x2*x3 x2*x4 x3*x4 x6*x7. Furthermore, there are restrictions
on the design. Each of the price variables, x1, x2, x5, x6, and x8, has one level the maximum level
that indicates the alternative is not available in the choice set. We use this to create choice sets with 2, 3, or
4 alternatives available. If (x1 < 4) then the first alternative is available, if (x2 < 4) then the second
alternative is available, if (x5 < 3) then the third alternative is available, and so on. A Boolean term such as
(x1 < 4) is one when true and zero otherwise. Hence,
((x1 < 4) + (x2 < 4) + (x5 < 3) + (x6 < 3) + (x8 < 3))
is the number of available alternatives. This is simply the sum of some 1’s if available and 0’s if not available.
We impose restrictions with the %MktEx macro by writing a macro, with IML statements, that quantifies the
badness of each run (or in this case, each choice set). We do this so bad = 0 is good and values larger than zero
are increasingly worse. We write our restrictions using an IML row vector x that contains the levels (integers
beginning with 1) of each of the factors in the ith choice set, the one the macro is currently seeking to improve.
The jth factor is x[j]. or we may also use the factor names (for example, x1, x2). (See page 280 for other
examples of restrictions.)
We must use IML logical operators, which are not as rich as DATA step operators:

= equals not: EQ
^ = or : = not equals not: NE
< less than not: LT
<= less than or equal to not: LE
> greater than not: GT
>= greater than or equal to not: GE
& and not: AND
j or not: OR
^ or : not not: NOT

To restrict the design, we must specify restrictions=macro-name, in this case restrictions=bad, that
names the macro that quantifies badness. The first statement counts up the number of available alternatives. The
second sets the actual badness values. If bad (the number available) is less than two or greater than 4, then the
Boolean expression ((bad < 2) | (bad > 4)) is true or 1. When the expression is true, then bad gets set
to the absolute difference between the number available and 3. Hence, zero available corresponds to bad = 3,
one available corresponds to bad = 2, two through four available corresponds to bad = 0, and five available
corresponds to bad = 2. We could just set bad to zero when everything is fine and one otherwise, but it is
better to help the macro by letting it know that when it switches from zero available to one available, it is going
in the right direction. Here is the code.
title ’Consumer Food Product Example’;

%macro bad;
bad = (x1 < 4) + (x2 < 4) + (x5 < 3) + (x6 < 3) + (x8 < 3);
bad = abs(bad - 3) * ((bad < 2) | (bad > 4));
%mend;

%mktex( 4 4 2 2 3 3 2 3, n=26, interact=x2*x3 x2*x4 x3*x4 x6*x7,

restrictions=bad, seed=377, outr=[Link] )
Here are the initial messages the macro prints.
NOTE: Generating the fractional-factorial design, n=27.
NOTE: Generating the candidate set.
NOTE: Performing 60 searches of 2,776 candidates, full-factorial=3,456.
196 TS-677E Multinomial Logit, Discrete Choice Modeling

The tabled design initialization part of the coordinate-exchange algorithm iterations will be initialized with the
first 26 rows of a 27 run fractional-factorial design. This design has 13 three-level factors, ten of which are used
to make 23 33 42 . The initial design will be unbalanced and one row short of orthogonal, so we would expect
that other methods would be better for this problem. The macro also tells us that it is performing 60 PROC
OPTEX searches of 2776 candidates, and that the full-factorial design has 3456 runs. The macro is searching
the full-factorial design minus the excluded choice sets. Since the full-factorial design is not too large (less than
5000), and since there is not tabled design that is very good for this problem, this is the kind of problem where
we would expect the PROC OPTEX algorithm to work best. The macro chose 60 OPTEX iterations. In the
fabric softener example, the macro did not try any OPTEX iterations, because it knew it could directly make a
100% efficient design. In the vacation examples, it ran the default minimum of 20 OPTEX iterations because the
macro’s heuristics concluded that OPTEX would probably not be the best approach for those problems. In this
example, the macro’s heuristics tried more iterations since this is the kind of example where OPTEX works best.
Here is some of the output.

Consumer Food Product Example

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 84.3176 Can
1 2 1 84.3176 84.3176 Conforms
1 End 84.3176

2 Start 49.1839 Tab,Unb,Ran

2 1 1 77.7982 Conforms
2 End 79.6170

.
.
.

11 Start 48.5995 Tab,Ran

11 23 1 76.9197 Conforms
11 End 79.8631

12 Start 25.4775 Ran,Mut,Ann

12 18 1 65.5812 Conforms
12 End 81.4087

.
.
.

21 Start 36.9535 Ran,Mut,Ann

21 1 1 73.6168 Conforms
21 End 83.1951

NOTE: Performing 600 searches of 2,776 candidates.

Food Product Example with Asymmetry and Availability Cross Effects 197

Consumer Food Product Example

Design Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 84.3176 84.3176 Ini

1 Start 85.4271 Can

1 2 1 85.4271 85.4271 Conforms
1 End 85.4271

Consumer Food Product Example

Design Refinement History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 85.4271 85.4271 Ini

1 Start 73.6369 Pre,Mut,Ann

1 2 1 71.6187 Conforms
1 End 82.1633

.
.
.

10 Start 62.6495 Pre,Mut,Ann

10 2 1 69.0658 Conforms
10 End 82.0098

Consumer Food Product Example

The OPTEX Procedure

Class Level Information

Class Levels -Values-

x1 4 1 2 3 4
x2 4 1 2 3 4
x3 2 1 2
x4 2 1 2
x5 3 1 2 3
x6 3 1 2 3
x7 2 1 2
x8 3 1 2 3
198 TS-677E Multinomial Logit, Discrete Choice Modeling

Consumer Food Product Example

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 85.4271 72.0480 98.0611 0.9806

Design 1 (Can), which was created by the candidate-set search (using PROC OPTEX), had D-efficiency or
84.3176, and the macro confirms that the design conforms to our restrictions. The tabled, unbalanced, and
random initializations do not work as well. For each design, the macro iteration history states the D-efficiency
for the initial design (49.1839 in design 2), the D-efficiency when the restrictions are met (77.7982, Conforms),
and the D-efficiency for the final design (79.6170). The fully random initialization tends to work a little better
than the tabled initialization for this problem, but not as well as PROC OPTEX. At the end of the algorithm
search phase, the macro decides to use PROC OPTEX and performs 600 more searches, and it finds a design
with 85.4271% D-efficiency. The design refinement step fails to improve on the best design. This step took 15
minutes.
As we will see in the next section, it turns out that this is a very good design for this example. Usually, for this
problem, we see final D-efficiencies less than 85.3. How did we happen to find a design this good? We just got
lucky. The random number seed that happened to go into the second OPTEX run, which was a function of all of
the random numbers generated previously by both OPTEX and the coordinate-exchange algorithm, was a really
good one.

When You Have a Long Time to Search for an Efficient Design

With a moderate sized candidate set such as this one (2000 to 6000 runs), we might be able to do better with more
iterations. To test this, PROC OPTEX was run 10,000 times over the winter holiday vacation, from December 22
through January 2, creating a total of 200,000 designs, 20 designs on each try. Here is a summary of the results.
PROC
OPTEX Percent
Run D-Efficiency Improvement
1 83.8959
2 83.9890 0.11%
3 84.3763 0.46%
6 84.7548 0.45%
84 85.1561 0.47%
1535 85.3298 0.20%
9576 85.3985 0.08%
This example is interesting, because it shows the diminishing value of increasing the number of iterations. Six
minutes into the search, in the first six passes through PROC OPTEX (6  20 = 120 total iterations), we found
a design with reasonably good D-efficiency=84.7548. Over an hour into the search, with (84 6)  20 = 1560
more iterations, we get a small 0.47% increase in efficiency to 85.1561. About one day into the search, with
(1535 84)  20 = 29; 020 more iterations, we get another small 0.20% increase in efficiency, 85.3298. Finally,
almost a week into the search, with (9576 1535)  20 = 160; 820 more iterations, we get another small 0.08%
increase in efficiency to 85.3985. Our overall improvement over the best design found in 120 iterations was
0.75952%, about three-quarters of a percent. These numbers will change with other problems and other seeds.
However, as these results show, usually the first few iterations will give you a good, efficient design, and usually,
 Author’s note: I do not know if this design is the optimal design. I know it is the best I have seen, and I frequently use this example as a
test case. If I had to guess, I would guess it is not the optimal design, but it is really close.
Food Product Example with Asymmetry and Availability Cross Effects 199

subsequent iterations will give you slight improvements but with a cost of much greater run times. Next, we will
construct a plot of this table.
data; input n e; datalines;
1 83.8959
2 83.9890
3 84.3763
6 84.7548
84 85.1561
1535 85.3298
9576 85.3985
;
proc gplot;
title h=1 ’Consumer Food Product Example’;
title2 h=1 ’Maximum D-Efficiency Found Over Time’;
plot e * n / vaxis=axis1;
symbol i=join;
axis1 order=(0 to 90 by 10);
run; quit;

title2;
The plot of maximum D-efficiency as a function of PROC OPTEX run number clearly shows that the gain in
efficiency that comes from a large number of iterations is very slight.

If you have a lot of time to search for a good design, you can specify some of the time and maximum number
of iteration parameters. Sometimes you will get lucky and find a better design. In this next example, max-
time=300 300 60 was specified. This give the macro up to 300 minutes for the algorithm search step, 300
minutes for the design search step, and 60 minutes for the refinement step. The option maxiter= increases
the number iterations to 10000 for each of the three steps (or the maximum time). With this specification, you
would expect the macro to run overnight. See the macro documentation (starting on page 327) for more iteration
options. Note that you must increase the number of iterations and the maximum amount of time if you want the
macro to run longer. With this specification, the macro performs 1800 OPTEX iterations initially (compared to
60 by default).
200 TS-677E Multinomial Logit, Discrete Choice Modeling

title ’Consumer Food Product Example’;

%macro bad;
bad = (x1 < 4) + (x2 < 4) + (x5 < 3) + (x6 < 3) + (x8 < 3);
bad = abs(bad - 3) * ((bad < 2) | (bad > 4));
%mend;

%mktex( 4 4 2 2 3 3 2 3, n=26, interact=x2*x3 x2*x4 x3*x4 x6*x7,

restrictions=bad, seed=151,
maxtime=300 300 60, maxiter=10000 )
The results from this step are not shown.

Examining the Design

We can use the %MktEval macro to start to evaluate the design.
%mkteval(data=[Link]);

Here are the results.

Consumer Food Product Example

Canonical Correlations Between the Factors
There are 4 Canonical Correlations Greater Than 0.316

x1 x2 x3 x4 x5 x6 x7 x8

x1 1 0.37 0.11 0.11 0.42 0.27 0.18 0.25

x2 0.37 1 0 0 0.38 0.50 0.25 0.08
x3 0.11 0 1 0.08 0.09 0.19 0.15 0.09
x4 0.11 0 0.08 1 0.25 0 0 0.09
x5 0.42 0.38 0.09 0.25 1 0.14 0.24 0.18
x6 0.27 0.50 0.19 0 0.14 1 0.10 0.29
x7 0.18 0.25 0.15 0 0.24 0.10 1 0.19
x8 0.25 0.08 0.09 0.09 0.18 0.29 0.19 1

Consumer Food Product Example

Canonical Correlations > 0.316 Between the Factors
There are 4 Canonical Correlations Greater Than 0.316

r r Square

x2 x6 0.50 0.25
x1 x5 0.42 0.17
x2 x5 0.38 0.14
x1 x2 0.37 0.13
Food Product Example with Asymmetry and Availability Cross Effects 201

Consumer Food Product Example

Summary of Frequencies
There are 4 Canonical Correlations Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

* x1 6 5 7 8
* x2 6 6 6 8
x3 13 13
x4 13 13
* x5 8 9 9
* x6 10 8 8
* x7 14 12
* x8 9 9 8
* x1 x2 2 2 0 2 1 0 2 2 1 2 2 2 2 2 2 2
* x1 x3 3 3 3 2 3 4 4 4
* x1 x4 3 3 3 2 3 4 4 4
* x1 x5 2 1 3 2 1 2 2 2 3 2 5 1
* x1 x6 2 2 2 1 2 2 3 1 3 4 3 1
* x1 x7 4 2 3 2 3 4 4 4
* x1 x8 2 3 1 2 1 2 2 2 3 3 3 2
* x2 x3 3 3 3 3 3 3 4 4
* x2 x4 3 3 3 3 3 3 4 4
* x2 x5 1 3 2 2 2 2 1 3 2 4 1 3
* x2 x6 0 3 3 4 1 1 2 1 3 4 3 1
* x2 x7 3 3 2 4 4 2 5 3
* x2 x8 2 2 2 2 2 2 2 2 2 3 3 2
* x3 x4 7 6 6 7
* x3 x5 4 5 4 4 4 5
* x3 x6 4 4 5 6 4 3
* x3 x7 6 7 8 5
* x3 x8 5 4 4 4 5 4
* x4 x5 5 3 5 3 6 4
* x4 x6 5 4 4 5 4 4
* x4 x7 7 6 7 6
* x4 x8 5 4 4 4 5 4
* x5 x6 3 3 2 3 3 3 4 2 3
* x5 x7 3 5 5 4 6 3
* x5 x8 3 2 3 3 3 3 3 4 2
* x6 x7 6 4 4 4 4 4
* x6 x8 2 5 3 3 2 3 4 2 2
* x7 x8 4 6 4 5 3 4
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1

Some of the canonical correlations are bigger than we would like. They all involve attributes in different alterna-
tives, so they should not pose huge problems. Still, they are large enough to make some researchers uncomfort-
able. The frequencies are pretty close to balanced. Perfect balance is not possible with 26 choice sets and this
design. If we were willing to consider blocking the design, we might do better with more choice sets.
202 TS-677E Multinomial Logit, Discrete Choice Modeling

Designing the Choice Experiment, More Choice Sets

Let’s run the %MktRuns macro to see what looks good. For now, we will ignore the interactions.
%mktruns( 4 4 2 2 3 3 2 3 )

Consumer Food Product Example

Design Summary

Number of
Levels Frequency

2 3
3 3
4 2

Consumer Food Product Example

Saturated = 16
Full Factorial = 3,456

Some Reasonable Cannot Be

Design Sizes Violations Divided By

144 0
72 1 16
48 3 9
96 3 9
192 3 9
24 4 9 16
120 4 9 16
168 4 9 16
36 7 8 16
108 7 8 16

The smallest suggestion larger than 26 is 36. With this mix of factor levels, we would have to have 144 runs
to get an orthogonal design, so we will definitely want to stick with a nonorthogonal design. Balance will be
possible in 36 runs, but 36 cannot be divided by 2  4 = 8 and 4  4 = 16. With 36 runs, a blocking factor will
be required (2 blocks of 18 or 3 blocks of 12). We would like the shelf-talker to appear in half of the choice sets
within block, so with two blocks, we will want the number of choice sets to be divisible by 2  2 = 4, and 36
can be divided by 4. The %MktRuns macro cannot provide us with much guidance with the interactions. We
“tricked” it in the past by substituting products of levels, like 9 = 3  3, but in this case, factors like x2, x3, and
x4 interact multiple times, so it would not be that simple. We will try making a design in 36 runs, and see how
it looks.
title ’Consumer Food Product Example’;

%macro bad;
bad = (x1 < 4) + (x2 < 4) + (x5 < 3) + (x6 < 3) + (x8 < 3);
bad = abs(bad - 3) * ((bad < 2) | (bad > 4));
%mend;

%mktex( 4 4 2 2 3 3 2 3, n=36, interact=x2*x3 x2*x4 x3*x4 x6*x7,

restrictions=bad, seed=377, outr=[Link] )

%mkteval;
Food Product Example with Asymmetry and Availability Cross Effects 203

Here is the last part of the output from the %MktEx macro.

Consumer Food Product Example

Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 94.6814 89.3664 94.3049 0.8333

D-efficiency at 94.68% looks good. Here is part of the %MktEval results.

Consumer Food Product Example

Canonical Correlations Between the Factors
There is 1 Canonical Correlation Greater Than 0.316

x1 x2 x3 x4 x5 x6 x7 x8

x1 1 0.20 0 0 0.18 0.17 0.16 0.18

x2 0.20 1 0 0 0.13 0.39 0.28 0.18
x3 0 0 1 0.11 0.07 0.07 0 0.07
x4 0 0 0.11 1 0.07 0.07 0 0.07
x5 0.18 0.13 0.07 0.07 1 0.05 0.07 0.15
x6 0.17 0.39 0.07 0.07 0.05 1 0.07 0.08
x7 0.16 0.28 0 0 0.07 0.07 1 0.07
x8 0.18 0.18 0.07 0.07 0.15 0.08 0.07 1

Consumer Food Product Example

Canonical Correlations > 0.316 Between the Factors
There is 1 Canonical Correlation Greater Than 0.316

r r Square

x2 x6 0.39 0.15

Consumer Food Product Example

Summary of Frequencies
There is 1 Canonical Correlation Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

* x1 8 10 8 10
* x2 10 8 8 10
x3 18 18
x4 18 18
* x5 9 11 16
* x6 13 11 12
x7 18 18
* x8 12 9 15
.
.
.
204 TS-677E Multinomial Logit, Discrete Choice Modeling

The correlations are better, although we still have one a bit larger than we would like. However, the biggest
problem is that the balance is much worse than we would like. We can run the macro again, this time specifying
balance=2, which forces better balance. The specification of 2 allows the maximum frequency for a level in a
factor to be no more than two greater than the minimum frequency.
%mktex( 4 4 2 2 3 3 2 3, n=36, interact=x2*x3 x2*x4 x3*x4 x6*x7,
restrictions=bad, seed=377, outr=[Link], balance=2 )
%mkteval;

Here is the last part of the output from the %MktEx macro.

Consumer Food Product Example

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 94.0824 88.3946 93.3172 0.8333

The D-efficiency looks good. It is a little lower than before, but not much. Here is the first part of the output
from the %MktEval macro.

Consumer Food Product Example

Canonical Correlations Between the Factors
There is 1 Canonical Correlation Greater Than 0.316

x1 x2 x3 x4 x5 x6 x7 x8

x1 1 0.24 0.08 0.08 0.18 0.21 0.13 0.11

x2 0.24 1 0 0 0.24 0.17 0.10 0.32
x3 0.08 0 1 0 0.12 0.12 0.17 0
x4 0.08 0 0 1 0.07 0.07 0.06 0
x5 0.18 0.24 0.12 0.07 1 0.17 0.10 0.14
x6 0.21 0.17 0.12 0.07 0.17 1 0.04 0.10
x7 0.13 0.10 0.17 0.06 0.10 0.04 1 0.08
x8 0.11 0.32 0 0 0.14 0.10 0.08 1

Consumer Food Product Example

Canonical Correlations > 0.316 Between the Factors
There is 1 Canonical Correlation Greater Than 0.316

r r Square

x2 x8 0.32 0.10
Food Product Example with Asymmetry and Availability Cross Effects 205

The canonical correlations look good. One is just large enough to be flagged (0:32 > 0:316), but r2 is only 0.1.
Here is the last part of the output from the %MktEval macro.

Consumer Food Product Example

Summary of Frequencies
There is 1 Canonical Correlation Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

* x1 8 10 9 9
* x2 8 8 10 10
x3 18 18
x4 18 18
* x5 11 13 12
* x6 11 12 13
* x7 17 19
x8 12 12 12
* x1 x2 2 2 2 2 2 2 2 4 2 3 2 2 2 1 4 2
* x1 x3 4 4 5 5 5 4 4 5
* x1 x4 4 4 5 5 5 4 4 5
* x1 x5 3 2 3 3 4 3 2 3 4 3 4 2
* x1 x6 2 3 3 2 4 4 3 2 4 4 3 2
* x1 x7 3 5 5 5 4 5 5 4
* x1 x8 3 2 3 3 4 3 3 3 3 3 3 3
* x2 x3 4 4 4 4 5 5 5 5
* x2 x4 4 4 4 4 5 5 5 5
* x2 x5 2 3 3 2 3 3 4 2 4 3 5 2
* x2 x6 2 3 3 2 2 4 4 3 3 3 4 3
* x2 x7 4 4 3 5 5 5 5 5
* x2 x8 2 2 4 2 2 4 4 4 2 4 4 2
x3 x4 9 9 9 9
* x3 x5 5 6 7 6 7 5
* x3 x6 5 7 6 6 5 7
* x3 x7 7 11 10 8
x3 x8 6 6 6 6 6 6
* x4 x5 6 6 6 5 7 6
* x4 x6 6 6 6 5 6 7
* x4 x7 8 10 9 9
x4 x8 6 6 6 6 6 6
* x5 x6 4 4 3 3 4 6 4 4 4
* x5 x7 6 5 6 7 5 7
* x5 x8 4 3 4 4 4 5 4 5 3
* x6 x7 5 6 6 6 6 7
* x6 x8 4 4 3 4 4 4 4 4 5
* x7 x8 6 6 5 6 6 7
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

This design looks much better. It is possible to get designs with better balance by specifying balance=1, but
for this problem, the price in efficiency is too high. We do want to ensure that x4, the shelf talker factor is
balanced, since we will be dividing the design into two parts, depending on whether the shelf talker is there or
not. It is balanced in this design. If it had not been, we could have switched it with another two-level factor
or tried again with a different seed. If nothing else worked, we could have added it after the fact by blocking
(running the %MktBlock macro as if we were adding a blocking factor).
206 TS-677E Multinomial Logit, Discrete Choice Modeling

The balance= option in the %MktEx macro works by adding restrictions to the design. The approach it uses
often works quite well, but sometimes it does not. Forcing balance gives the macro much less freedom in its
search, and makes it easy for the macro to get stuck in suboptimal designs. Most of our restrictions are imposed
within each row. Those kinds of restrictions do not pose a problem for the macro. Balance restrictions are
imposed across rows within a column. We know of better ways to impose balance, but they tend to be very slow.
This is an area where more research is needed, and the way this option works will quite likely be different in
future releases. If perfect balance is critical, try the %MktBal macro.

Examining the Subdesigns

As we mentioned previously, “it is sometimes necessary and good practice to check the ranks of the subdesigns
for more complex models (Lazari and Anderson 1994).” Here is a way to do that with PROC OPTEX. This is the
only usage of PROC OPTEX in this report that is too specialized to be run from one of the %Mkt macros (because
not all variables are designated as class variables). For convenience, we call PROC OPTEX from an ad hoc
macro, since it must be run five times, once per alternative, with only a change in the where statement. We
need to evaluate the design when the client’s alternative is available (x1 ne 4), when the client line extension
alternative is available (x2 ne 4), when the regional competitor is available (x5 ne 3), when the private
label competitor is available (x6 ne 3), and when the national competitor is available (x8 ne 3). We
need to use a model statement that lists all of the main effects and interactions. We do not designate all of the
variables on the class statement because we only have enough runs to consider linear price effects within each
availability group. The statement generate method=sequential initdesign=desv specifies that
we will be evaluating the initial design desv, using the sequential algorithm, which ensures no swaps between
the candidate set and the initial design. The other option of note here appears on the class statement, and that
is param=orthref. This specifies an orthogonal parameterization of the effects that gives us a nice 0 to 100
scale for the D-efficiencies.
%macro evaleff(where);
data desv / view=desv; set [Link](where=(&where)); run;

proc optex data=desv;

class x3 x4 x7 / param=orthref;
model x1-x8 x2*x3 x2*x4 x3*x4 x6*x7;
generate method=sequential initdesign=desv;
run; quit;

%mkteval(data=desv)
%mend;

%evaleff(x1 ne 4)
%evaleff(x2 ne 4)
%evaleff(x5 ne 3)
%evaleff(x6 ne 3)
%evaleff(x8 ne 3)
Each step took just over two seconds. We hope to not see any efficiencies of zero, and we hope to not get the mes-
sage WARNING: Can’t estimate model parameters in the final design. Here are some
of the results.

Consumer Food Product Example

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 66.9776 54.6662 82.8184 0.6939
Food Product Example with Asymmetry and Availability Cross Effects 207

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 73.7719 66.4421 87.3561 0.7071

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 69.9539 59.0519 83.0652 0.7360

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 70.8672 57.5401 80.8955 0.7518

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 65.7047 50.8521 84.5539 0.7360

Examining the Aliasing Structure

It is also good to look at the aliasing structure of the design. We use PROC GLM to do this, so we must create a
dependent variable. We will use a constant y=1. The first PROC GLM step just checks the model to make sure
none of the specified effects are aliased with each other. This step is not necessary since our D-efficiency value
greater than zero already guarantees this.
data temp;
set [Link];
y = 1;
run;
proc glm data=temp;
model y = x1-x8 x2*x3 x2*x4 x3*x4 x6*x7 / e aliasing;
run;
Here are the results, ignoring the ANOVA and regression tables, which are not of interest. Each of these lines is
a linear combination that is estimable. It is simply a list of the effects.

Intercept
x1
x2
x3
x4
x5
x6
x7
x8
208 TS-677E Multinomial Logit, Discrete Choice Modeling

x2*x3
x2*x4
x3*x4
x6*x7

Contrast this with a specification that includes all simple effects and two-way and three-way interactions. We
specify the model of interest first, so all of those terms will be listed first, then we specify all main effects
and two-way and three-way interactions using the notation x1|x2|x3|x4|x5|x6|x7|x8@3. This list will
generate all of the effects of interest, x1-x8 x2*x3 x2*x4 x3*x4 x6*x7, and all of the two-way and
three-way interactions. It is not a problem that some of the interactions were both explicitly specified and also
generated by the x1|x2|x3|x4|x5|x6|x7|x8@3 list since PROC GLM automatically eliminates duplicate
terms.
proc glm data=temp;
model y = x1-x8 x2*x3 x2*x4 x3*x4 x6*x7
x1|x2|x3|x4|x5|x6|x7|x8@3 / e aliasing;
run;

Intercept - 174.72*x4*x6 - 476.67*x1*x4*x6 - 490.37*x2*x4*x6 - 359.77*x3*x4*x6

+ 204.67*x5*x6 + 742.7*x1*x5*x6 + 698.48*x2*x5*x6 + 444.15*x3*x5*x6 -
30.82*x4*x5*x6 - 30.013*x1*x7 + 174.13*x2*x7 + 412.08*x1*x2*x7 + 149.33*x3*x7
+ 291.03*x1*x3*x7 + 748.79*x2*x3*x7 + 33.824*x4*x7 - 34.861*x1*x4*x7 +
453.21*x2*x4*x7 + 260.2*x3*x4*x7 + 197.6*x5*x7 + 430.1*x1*x5*x7 +
871.72*x2*x5*x7 + 623.82*x3*x5*x7 + 288.4*x4*x5*x7 + 59.967*x1*x6*x7 +
335.16*x2*x6*x7 + 313.37*x3*x6*x7 - 202.24*x4*x6*x7 + 701.86*x5*x6*x7 -
28.098*x1*x8 + 35.022*x2*x8 + 190.57*x1*x2*x8 - 49.406*x3*x8 - 134.53*x1*x3*x8
- 158.84*x2*x3*x8 - 138.5*x4*x8 - 505.81*x1*x4*x8 - 454.8*x2*x4*x8 -
347.43*x3*x4*x8 + 190.55*x5*x8 + 659.05*x1*x5*x8 + 470.76*x2*x5*x8 +
344.34*x3*x5*x8 - 19.226*x4*x5*x8 + 55.161*x6*x8 + 32.377*x1*x6*x8 +
260.89*x2*x6*x8 - 48.524*x3*x6*x8 - 547.3*x4*x6*x8 + 829.99*x5*x6*x8 -
120.95*x7*x8 - 338.09*x1*x7*x8 + 48.151*x2*x7*x8 + 112.34*x3*x7*x8 -
290.99*x4*x7*x8 + 417.15*x5*x7*x8 - 61.492*x6*x7*x8
x1 + 36.198*x4*x6 + 98.299*x1*x4*x6 + 102.07*x2*x4*x6 + 71.034*x3*x4*x6 -
28.544*x5*x6 - 111.22*x1*x5*x6 - 98.978*x2*x5*x6 - 68.98*x3*x5*x6 +
28.763*x4*x5*x6 + 1.2817*x1*x7 - 25.266*x2*x7 - 74.585*x1*x2*x7 - 29.591*x3*x7
- 69.14*x1*x3*x7 - 131.14*x2*x3*x7 - 8.8034*x4*x7 - 10.134*x1*x4*x7 -
78.619*x2*x4*x7 - 51.281*x3*x4*x7 - 46.501*x5*x7 - 118.36*x1*x5*x7 -
172.7*x2*x5*x7 - 132.56*x3*x5*x7 - 72.363*x4*x5*x7 - 23.906*x1*x6*x7 -
36.943*x2*x6*x7 - 59.657*x3*x6*x7 + 38.937*x4*x6*x7 - 132.19*x5*x6*x7 +
15.948*x1*x8 + 3.4064*x2*x8 + 29.796*x1*x2*x8 + 6.4151*x3*x8 + 27.738*x1*x3*x8
+ 40.475*x2*x3*x8 + 25.464*x4*x8 + 108.13*x1*x4*x8 + 99.485*x2*x4*x8 +
61.056*x3*x4*x8 - 37.732*x5*x8 - 111.6*x1*x5*x8 - 65.647*x2*x5*x8 -
70.116*x3*x5*x8 + 3.9099*x4*x5*x8 - 13.614*x6*x8 + 7.0483*x1*x6*x8 -
30.651*x2*x6*x8 + 1.3753*x3*x6*x8 + 106.47*x4*x6*x8 - 137.07*x5*x6*x8 +
18.312*x7*x8 + 59.267*x1*x7*x8 + 11.27*x2*x7*x8 - 34.681*x3*x7*x8 +
40.508*x4*x7*x8 - 106.81*x5*x7*x8 - 9.4494*x6*x7*x8
.
.
.

Again, we have a list of linear combinations that are estimable. This shows that the Intercept cannot be
estimated independently of the x4*x6 interaction and a bunch of others including four-way though eight-way
interactions which were not specified and hence not shown. Similarly, x1 is confounded with a bunch of in-
teractions, and so on. This is why we want to be estimable the two-way interactions between factors that are
combined to create an alternative. We did not want something like x2*x3, the client-line extension’s price and
microwave/stove top interaction to be confounded with say another brand’s price.
Food Product Example with Asymmetry and Availability Cross Effects 209

Blocking the Design

At 36 choice sets, this design is a bit large, so we will block it into two blocks of 18 choice sets. Within each
block we will want the shelf talker to be on half the time.
%mktblock(data=[Link], out=[Link], nblocks=2, seed=289)

The first attempt (not shown) produced a design where x4, the shelf talker did not occur equally often within
each block. Changing the seed took care of the problem. Here are the canonical correlations.

Consumer Food Product Example

Canonical Correlations Between the Factors
There is 1 Canonical Correlation Greater Than 0.316

Block x1 x2 x3 x4 x5 x6 x7 x8

Block 1 0.08 0 0 0 0.07 0.07 0.06 0.14

x1 0.08 1 0.24 0.08 0.08 0.18 0.21 0.13 0.11
x2 0 0.24 1 0 0 0.24 0.17 0.10 0.32
x3 0 0.08 0 1 0 0.12 0.12 0.17 0
x4 0 0.08 0 0 1 0.07 0.07 0.06 0
x5 0.07 0.18 0.24 0.12 0.07 1 0.17 0.10 0.14
x6 0.07 0.21 0.17 0.12 0.07 0.17 1 0.04 0.10
x7 0.06 0.13 0.10 0.17 0.06 0.10 0.04 1 0.08
x8 0.14 0.11 0.32 0 0 0.14 0.10 0.08 1

The blocking variable is not highly correlated with any of the factors. Here are some of the frequencies.

Consumer Food Product Example

Summary of Frequencies
There is 1 Canonical Correlation Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

Block 18 18
* x1 8 10 9 9
* x2 8 8 10 10
x3 18 18
x4 18 18
* x5 11 13 12
* x6 11 12 13
* x7 17 19
x8 12 12 12
* Block x1 4 5 4 5 4 5 5 4
* Block x2 4 4 5 5 4 4 5 5
Block x3 9 9 9 9
Block x4 9 9 9 9
* Block x5 6 6 6 5 7 6
* Block x6 6 6 6 5 6 7
* Block x7 9 9 8 10
* Block x8 5 6 7 7 6 5
210 TS-677E Multinomial Logit, Discrete Choice Modeling

The blocking variable is perfectly balanced, as it is guaranteed to be if the number of blocks divides the number
of runs. Balance within blocks, that is the cross-tabulations of the factors with the blocking variable, looks good.
The macro also prints canonical correlations within blocking variables. These can sometimes be quite high, even
1.0, but that is not a problem. Here is the design, as it is printed by the %MktBlock macro.

Consumer Food Product Example

Block Run x1 x2 x3 x4 x5 x6 x7 x8

1 1 3 2 1 1 1 1 2 3
2 4 4 1 2 2 2 1 3
3 3 4 2 2 1 3 1 2
4 4 3 1 2 1 1 1 1
5 4 4 2 1 3 1 2 1
6 3 1 2 1 2 2 2 3
7 1 1 1 2 1 2 2 3
8 2 1 1 1 3 3 2 3
9 2 2 1 2 3 3 1 3
10 1 3 2 2 3 2 1 2
11 2 2 2 1 2 3 1 2
12 2 4 1 2 2 1 2 1
13 1 1 2 1 1 3 1 1
14 3 3 1 2 3 1 2 2
15 4 3 2 2 2 3 1 3
16 2 4 2 1 2 2 1 2
17 4 3 2 1 1 1 2 2
18 1 2 1 1 3 2 2 1

Consumer Food Product Example

Block Run x1 x2 x3 x4 x5 x6 x7 x8

2 1 4 2 2 2 1 2 1 1
2 3 1 1 2 3 2 1 1
3 4 1 2 2 2 3 2 3
4 3 4 1 1 2 3 1 1
5 2 3 1 1 1 2 1 3
6 2 3 2 2 1 3 2 1
7 4 3 1 1 3 2 2 2
8 2 4 1 1 1 2 2 2
9 3 3 2 1 3 1 1 1
10 1 3 1 1 2 3 2 1
11 2 4 2 2 2 2 2 1
12 2 1 2 2 3 1 1 2
13 3 2 1 2 2 3 2 2
14 3 2 2 1 3 3 2 3
15 1 4 2 1 1 1 1 3
16 4 1 1 1 2 1 1 2
17 1 2 2 2 2 1 2 3
18 1 4 1 2 3 3 2 2

 Ideally, each subject would only make one choice, since the choice model is based on this assumption (which is almost always ignored).
As the number of blocks increases, the correlations will mostly go to one, and ultimately be undefined when there is only one choice set per
block.
Food Product Example with Asymmetry and Availability Cross Effects 211

The Final Design

The following code creates the final choice design, stored in [Link], sorted by the blocking and
shelf-talker variable. We will use the %MktLab macro to assign values, formats, and labels to the design. Previ-
ously, we have used the %MktLab macro to reassign factor names when we wanted something more descriptive
than the default, x1, x2, and so on, and when we wanted to reassign the names of two m-level factors to min-
imize the problems associated with correlated factors. This time, we will use the %MktLab macro primarily to
deal with the asymmetry in the price factors. Recall our factor levels.

Factors and Levels

Alternative Factor Levels Brand Description

1 X1 4 Client 1.29, 1.69, 2.09 + absent

2 X2 4 Client Line Extension 1.39, 1.89, 2.39, + absent

X3 2 microwave/stove-top
X4 2 shelf-talker yes/no

3 X5 3 Regional 1.99, 2.49 + absent

4 X6 3 Private Label 1.49, 2.29 absent

X7 2 microwave/stove-top

5 X8 3 National 1.99 + 2.39 + absent

The choice design will need a quantitative price factor, make from all five of the linear price factors, that contains
the prices of each of the alternatives. At this point, our factor x1 contains 1, 2, 3, 4, and not 1.29, 1.69, 2.09, and
absent, which is different from x2 and from all of the other factors. A 1 in x1 will need to become a price of
1.29 in the choice design, a 1 in x2 will need to become a price of 1.39 in the choice design, a 1 in x3 will need
to become a price of 1.99 in the choice design, and so on. Before we use the %MktRoll macro to turn the linear
design into a choice design, we need to use the %MktLab macro to assign the actual prices to the price factors.
The %MktLab macro is like the %MktRoll macro in the sense that it can use as input a key= data set that
contains the rules for customizing a design for our particular usage. In the %MktRoll macro, the key= data
set provides the rules for turning a linear design into a choice design. In contrast, in the %MktLab macro, the
key= data set contains the rules for turning a linear design into another linear design, changing one or more of
the following: factor names, factor levels, factor types (numeric to character), level formats, and factor labels.
We could use the %MktLab macro to change the names of the variables and their types, but we will not do that
for this example. Ultimately, we will use the %MktRoll macro to assign all of the price factors to a variable
called Price and similarly provide meaningful names for all of the factors in the choice design, just as we have
in previous examples. We could also change a variable like x3 with values of 1 and 2 to something like Stove
with values ’Stove’ and ’Micro’. We will not do that because we want to make a design with a simple list
of numeric factors, with simple names like x1-x8 that we can run through the %MktRoll macro to get the final
choice design. We will assign formats and labels, so we can print the design in a meaningful way, but ultimately,
our only goal at this step is to handle the price asymmetries by assigning the actual price values to the factors.
The key= data set contains the rules for customizing our design. The data set has as many rows as the maximum
number of levels, in this case four. Each variable is one of the factors in the design, and the values are the factor
levels that we want in the final design. The first factor, x1, is the price factor for the client brand. Its levels are
1.29, 1.69, and 2.09. In addition, one level is ’not available’, which is flagged by the SAS special missing value
.N. In order to read special missing values in an input data set, you must use the missing statement and name
the expected missing values. The factor x2 has the same structure as x1, but with different levels. The factor
x3 has two levels, hence the key= data set has missing values in the third and fourth row. Since the design
has only 1’s and 2’s for x3, this missing values will never be used. Notice that we are keeping x3 as a numeric
variable with values 1 and 2 using a format to supply the character levels ’micro’ and ’stove’. The other factors
212 TS-677E Multinomial Logit, Discrete Choice Modeling

are created in a similar fashion. You may not use ordinary missing ’.’ for levels. You may only use ordinary
missing values as place holders for factors that have fewer levels than the maximum. If you want missing values
in the levels, you must use one of the special missing values .A, .B, ..., .Z, and .- .
The %MktLab macro specification names the input SAS data set with the design. By default, it looks for an
input key= data set named KEY and creates an output SAS data set called FINAL. The data set is sorted by
block and shelf talker and printed.
proc format;
value yn 1 = ’ No’ 2 = ’Talker’;
value micro 1 = ’Micro’ 2 = ’Stove’;
run;

data key;
missing N;
input x1-x8;
format x1 x2 x5 x6 x8 dollar5.2
x4 yn. x3 x7 micro.;
label x1 = ’Client Brand’
x2 = ’Client Line Extension’
x3 = ’Client Micro/Stove’
x4 = ’Shelf Talker’
x5 = ’Regional Brand’
x6 = ’Private Label’
x7 = ’ Private Micro/Stove’
x8 = ’National Competitor’;
datalines;
1.29 1.39 1 1 1.99 1.49 1 1.99
1.69 1.89 2 2 2.49 2.29 2 2.39
2.09 2.39 . . N N . N
N N . . . . . .
;

%mktlab(data=[Link], key=key)

proc sort out=[Link](drop=run); by block x4; run;

proc print label; id block x4; by block x4; run;

The %MktLab macro prints the variable mapping that it uses, old names followed by new names. In this case,
none of the names change, but it is good to make sure that you have the expected correspondence.
Variable Mapping:
x1 : x1
x2 : x2
x3 : x3
x4 : x4
x5 : x5
x6 : x6
x7 : x7
x8 : x8
 Note that the ’.’ in ’.N’ is not typed in the data, nor is it typed in the missing statement. Furthermore, it does not appear in
the printed output. However, you need to type it if you ever refer to a special missing value in code: if x1 eq .N then ....
Food Product Example with Asymmetry and Availability Cross Effects 213

Here is the design.

Consumer Food Product Example

Client Client Private

Shelf Client Line Micro/ Regional Private Micro/ National
Block Talker Brand Extension Stove Brand Label Stove Competitor

1 No $2.09 $1.89 Micro $1.99 $1.49 Stove N

N N Stove N $1.49 Stove $1.99
$2.09 $1.39 Stove $2.49 $2.29 Stove N
$1.69 $1.39 Micro N N Stove N
$1.69 $1.89 Stove $2.49 N Micro $2.39
$1.29 $1.39 Stove $1.99 N Micro $1.99
$1.69 N Stove $2.49 $2.29 Micro $2.39
N $2.39 Stove $1.99 $1.49 Stove $2.39
$1.29 $1.89 Micro N $2.29 Stove $1.99

1 Talker N N Micro $2.49 $2.29 Micro N

$2.09 N Stove $1.99 N Micro $2.39
N $2.39 Micro $1.99 $1.49 Micro $1.99
$1.29 $1.39 Micro $1.99 $2.29 Stove N
$1.69 $1.89 Micro N N Micro N
$1.29 $2.39 Stove N $2.29 Micro $2.39
$1.69 N Micro $2.49 $1.49 Stove $1.99
$2.09 $2.39 Micro N $1.49 Stove $2.39
N $2.39 Stove $2.49 N Micro N

2 No $2.09 N Micro $2.49 N Micro $1.99

$1.69 $2.39 Micro $1.99 $2.29 Micro N
N $2.39 Micro N $2.29 Stove $2.39
$1.69 N Micro $1.99 $2.29 Stove $2.39
$2.09 $2.39 Stove N $1.49 Micro $1.99
$1.29 $2.39 Micro $2.49 N Stove $1.99
$2.09 $1.89 Stove N N Stove N
$1.29 N Stove $1.99 $1.49 Micro N
N $1.39 Micro $2.49 $1.49 Micro $2.39

2 Talker N $1.89 Stove $1.99 $2.29 Micro $1.99

$2.09 $1.39 Micro N $2.29 Micro $1.99
N $1.39 Stove $2.49 N Stove N
$1.69 $2.39 Stove $1.99 N Stove $1.99
$1.69 N Stove $2.49 $2.29 Stove $1.99
$1.69 $1.39 Stove N $1.49 Micro $2.39
$2.09 $1.89 Micro $2.49 N Stove $2.39
$1.29 $1.89 Stove $2.49 $1.49 Stove N
$1.29 N Micro N N Stove $2.39
214 TS-677E Multinomial Logit, Discrete Choice Modeling

In contrast, here are the actual values without formats and labels.
proc print data=[Link]; format _numeric_; run;

Consumer Food Product Example

Obs x1 x2 x3 x4 x5 x6 x7 x8 Block

1 2.09 1.89 1 1 1.99 1.49 2 N 1

2 N N 2 1 N 1.49 2 1.99 1
3 2.09 1.39 2 1 2.49 2.29 2 N 1
4 1.69 1.39 1 1 N N 2 N 1
5 1.69 1.89 2 1 2.49 N 1 2.39 1
6 1.29 1.39 2 1 1.99 N 1 1.99 1
7 1.69 N 2 1 2.49 2.29 1 2.39 1
8 N 2.39 2 1 1.99 1.49 2 2.39 1
9 1.29 1.89 1 1 N 2.29 2 1.99 1
10 N N 1 2 2.49 2.29 1 N 1
11 2.09 N 2 2 1.99 N 1 2.39 1
12 N 2.39 1 2 1.99 1.49 1 1.99 1
13 1.29 1.39 1 2 1.99 2.29 2 N 1
14 1.69 1.89 1 2 N N 1 N 1
15 1.29 2.39 2 2 N 2.29 1 2.39 1
16 1.69 N 1 2 2.49 1.49 2 1.99 1
17 2.09 2.39 1 2 N 1.49 2 2.39 1
18 N 2.39 2 2 2.49 N 1 N 1
19 2.09 N 1 1 2.49 N 1 1.99 2
20 1.69 2.39 1 1 1.99 2.29 1 N 2
21 N 2.39 1 1 N 2.29 2 2.39 2
22 1.69 N 1 1 1.99 2.29 2 2.39 2
23 2.09 2.39 2 1 N 1.49 1 1.99 2
24 1.29 2.39 1 1 2.49 N 2 1.99 2
25 2.09 1.89 2 1 N N 2 N 2
26 1.29 N 2 1 1.99 1.49 1 N 2
27 N 1.39 1 1 2.49 1.49 1 2.39 2
28 N 1.89 2 2 1.99 2.29 1 1.99 2
29 2.09 1.39 1 2 N 2.29 1 1.99 2
30 N 1.39 2 2 2.49 N 2 N 2
31 1.69 2.39 2 2 1.99 N 2 1.99 2
32 1.69 N 2 2 2.49 2.29 2 1.99 2
33 1.69 1.39 2 2 N 1.49 1 2.39 2
34 2.09 1.89 1 2 2.49 N 2 2.39 2
35 1.29 1.89 2 2 2.49 1.49 2 N 2
36 1.29 N 1 2 N N 2 2.39 2

One issue remains to be resolved regarding this design and that concerns the role of the shelf-talker when the
client line extension is not available. The second part of each block of the design consists of choice sets in which
the shelf-talker is present and calls attention to the client line extension. However, in five of those choice sets,
the client line extension is unavailable. This problem can be handled in several ways. Here are a few:
 Rerun the design creation and evaluation programs excluding all choice sets with shelf-talker present and
client line extension unavailable. However, this requires changing the model because the excluded cell
will make unestimable the interaction between client-line-extension price and shelf-talker. Furthermore,
the shelf-talker variable will almost certainly no longer be balanced.
 Move the choice sets with client line extension unavailable to the no-shelf-talker block and rerandomize.
The shelf-talker is then on for all of the last nine choice sets.
 Let the shelf-talker go on and off as needed.
Food Product Example with Asymmetry and Availability Cross Effects 215

 Let the shelf-talker call attention to a brand that happens to be out of stock. It is easy to imagine this
happening in a real store.

Other options are available as well. No one approach is obviously superior to the alternatives. For this example,
we will take the latter approach and allow the shelf-talker to be on even when the client line extension is not
available. Note that if the shelf-talker is turned off when the client line extension is not available then the design
must be manually modified to reflect this fact.

Testing the Design Before Data Collection

This is a complicated design that will be used to fit a complicated model with alternative specific and availability
cross effects. Collecting data is time consuming and expensive. It is good practice, particularly when there are
cross effects, to make sure that the design will work with the most complicated model that we anticipate fitting.
We saw, starting on page 166, an example of generating artificial data to test the design before collecting real
data. Here, we will explore an alternative approach. Before we collect any data, we will convert the linear design
to a choice design and use the %ChoicEff macro to evaluate its efficiency for a multinomial logit model, with
availability cross effects.
For analysis, the design will have four factors, Brand, Price, Micro, Shelf. We will use the %MktRoll
macro and a key= data set (although not the same one as before) to make the choice design. Brand is the
alternative name; its values are directly read from the key=KEY in-stream data. Price is an attribute whose
values will be constructed from the factors x1, x2, x5, x6, and x8 in [Link] data set. Micro,
the microwave factor, is constructed from x3 for the client line extension and x7 for the private label. Shelf,
the shelf talker factor, is created from x4 for the extension. The keep= option on the %MktRoll macro is used
to keep the original price factors in the design, since we will need them for the price effects. Normally, they
would be dropped.
data key;
input Brand $ 1-10 (Price Micro Shelf) ($);
datalines;
Client x1 . .
Extension x2 x3 x4
Regional x5 . .
Private x6 x7 .
National x8 . .
None . . .
;

%mktroll(design=[Link], key=key, alt=brand, out=rolled,

keep=x1 x2 x5 x6 x8)

proc print data=[Link](obs=2); run;

proc print data=rolled(obs=12);

format price dollar5.2 shelf yn. micro micro.;
id set; by set;
run;
Consider the first two choice sets in the linear design.

Consumer Food Product Example

Obs x1 x2 x3 x4 x5 x6 x7 x8 Block

1 $2.09 $1.89 Micro No $1.99 $1.49 Stove N 1

2 N N Stove No N $1.49 Stove $1.99 1
216 TS-677E Multinomial Logit, Discrete Choice Modeling

Here they are in the rolled out choice design.

Consumer Food Product Example

Set Brand Price Micro Shelf x1 x2 x5 x6 x8

1 Client $2.09 . . $2.09 $1.89 $1.99 $1.49 N

Extension $1.89 Micro No $2.09 $1.89 $1.99 $1.49 N
Regional $1.99 . . $2.09 $1.89 $1.99 $1.49 N
Private $1.49 Stove . $2.09 $1.89 $1.99 $1.49 N
National N . . $2.09 $1.89 $1.99 $1.49 N
None . . . $2.09 $1.89 $1.99 $1.49 N

2 Client N . . N N N $1.49 $1.99

Extension N Stove No N N N $1.49 $1.99
Regional N . . N N N $1.49 $1.99
Private $1.49 Stove . N N N $1.49 $1.99
National $1.99 . . N N N $1.49 $1.99
None . . . N N N $1.49 $1.99

Set 1, Alternative 1

Brand = ’Client’ the brand for this alternative

Price = x1 = $2.09 the price of this alternative
Micro = . does not apply to this brand
Shelf = . does not apply to this brand
x1 = $2.09 the price of the client brand in this choice set
x2 = $1.89 the price of the extension in this choice set
x5 = $1.99 the price of the regional competitor in this choice set
x6 = $1.49 the price of the private label in this choice set
x8 = N national competitor unavailable in this choice set

Set 1, Alternative 2

Brand = ’Extension’ the brand for this alternative

Price = x2 = $1.89 the price of this alternative
Micro = Micro Microwave version
Shelf = No Shelf Talker, No
x1 = $2.09 the price of the client brand in this choice set
x2 = $1.89 the price of the extension in this choice set
x5 = $1.99 the price of the regional competitor in this choice set
x6 = $1.49 the price of the private label in this choice set
x8 = N national competitor unavailable in this choice set

Notice that x1 through x8 are constant within each choice set. The variable x1 is the price of alternative one,
which is the same no matter which alternative it is stored with.
We need to do a few more things to this design before we are ready to use it. Since we will be treating all of the
price factors as a quantitative (not a class variable), we need to convert the missing prices to zero. We also
need to convert the missings for when Micro and Shelf do not apply to 2 for ’Stove’ and 1 for ’No’. We also
need to assign formats. Eventually, we will also need to output just the alternatives that are available (those with
a nonzero price and also the none alternative). For now, we will just make a variable w that flags the available
alternative (w = 1).
Food Product Example with Asymmetry and Availability Cross Effects 217

data [Link](drop=i);
set rolled;
array x[6] price x1 -- x8;
do i = 1 to 6; if nmiss(x[i]) then x[i] = 0; end;
if nmiss(micro) then micro = 2;
if nmiss(shelf) then shelf = 1;
w = brand eq ’None’ or price ne 0;
format price dollar5.2 shelf yn. micro micro.;
run;

proc print data=[Link](obs=12); by set; id set; run;

Here are the first two choice sets.

Consumer Food Product Example

Set Brand Price Micro Shelf x1 x2 x5 x6 x8 w

1 Client $2.09 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1

Extension $1.89 Micro No $2.09 $1.89 $1.99 $1.49 $0.00 1
Regional $1.99 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1
Private $1.49 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1
National $0.00 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 0
None $0.00 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1

2 Client $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0

Extension $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0
Regional $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0
Private $1.49 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1
National $1.99 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1
None $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1

Now our choice design is done except for the final coding for the analysis. We can now use the %ChoicEff
macro to evaluate it for a choice model. Normally, you would use this macro to search a candidate set for an
efficient choice design. You can also use it to evaluate a design created by other means. Here is some sample
code, omitting for now the details of the model (indicated by model= ...). The complicated part of this is the
model due to the alternative-specific price effects and cross effects. For now, let’s concentrate on everything else.
%choiceff(data=[Link],
model= ..., /* model specification skipped for now */
nsets=36, nalts=6, weight=w,
beta=zero, init=[Link](keep=set),
intiter=0);
The way you check the efficiency of a design like this is to first name it on the data= option. This will be the
candidate set that contains all of the choice sets that we will consider. In addition, the same design is named on
the init= option. The full specification is init=[Link](keep=set). Just the variable
Set is kept. It will be used to bring in just the indicated choice sets from the data= design, which in this
case is all of them. The option nsets=36 specifies the number of choice sets, and nalts=6 specifies the
number of alternatives. This macro requires a constant number of alternatives in each choice set for ease of data
management. However, not all of the alternatives have to be used. In this case, we have an availability study. We
need to keep the unavailable alternatives in the design for this step, but we do not want them to contribute to the
analysis, so we specify a weight variable with weight=w and flag the available alternatives with w=1 and the
unavailable alternatives with w=0. The option beta=zero specifies that we are assuming for design evaluation
purpose all zero betas. We can specify other values and get other results for the variances and standard errors.
Finally, we specify intiter=0 which specifies zero internal iterations. We use zero internal iterations when
we want to evaluate an initial design, but not attempt to improve it. Here is the actual specification we will use,
complete with the model specification.
218 TS-677E Multinomial Logit, Discrete Choice Modeling

%choiceff(data=[Link],
model=class(brand / zero=’None’)
class(brand / zero=’None’ separators=’’ ’ ’) *
identity(price)
class(shelf micro / lprefix=5 0 zero=’No’ ’Stove’)
identity(x1 x2 x5 x6 x8) *
class(brand / zero=’None’ separators=’ ’ ’ on ’) /
lprefix=0 order=data,
nsets=36, nalts=6, weight=w,
beta=zero, init=[Link](keep=set),
intiter=0);
The specification class(brand / zero=’None’) specifies the brand effects. This specification will
create dummy variables for brand with the constant alternative being the reference brand. The option
zero=’None’ ensures that the reference level will be ’None’ instead of the default last sorted level (’Re-
gional’). Dummy variables will be created for the brands Client, Extension, Regional, Private, and National, but
not None. The zero=’None’ option, like zero=’Home’ and other zero=’literal-string’ options we have
used in previous examples, names the actual formatted value of the class variable that should be excluded from
the coded variables because the coefficient will be zero. Do not confuse zero=none and zero=’None’. The
zero=none option specifies that you want all dummy variables to be created, even including one for the last
level. In contrast, the option zero=’None’ (or zero= any quoted string) names a specific formatted value, in
this case ’None’, for which dummy variables are not to be created.
The specification class(brand / ...) * identity(price) creates the alternative-specific price
effects. They are specified as an interaction between a categorical variable Brand and a quantitative factor
Price. The separators=” ’ ’ option in the class specification specifies the separators that are used
to construct the labels for the main effect and interaction terms. The main-effects separator, which is the first
separators= value, ”, is ignored since lprefix=0. Specifying ’ ’ as the second value creates labels of
the form brand-blank-price instead of the default brand-blank-asterisk-blank-price.
The specification class(shelf micro / ...) names the shelf talker and microwave variables as cat-
egorical variables and creates dummy variables for the ’Talker’ category, not the ’No’ category and the
’Micro’ category not the ’Stove’ category. In zero=’No’ ’Stove’, the ’No’ applies to the first vari-
able, Shelf and the second value, ’Stove’, applies to second variable, Micro.
The specification identity(x1 x2 x5 x6 x8) * class(brand / ...) creates the cross effects.
The separators= option is specified with a second value of ’ on ’ to create cross effect labels like
’Client on Extension’. More will be said on the cross effects when we look at the actual coded values
in the next few pages.
Note that PROC TRANSREG produces the following warning twice.
WARNING: This usage of * sets one group’s slope to zero. Specify |
to allow all slopes and intercepts to vary. Alternatively,
specify CLASS(vars) * identity(vars) identity(vars) for
separate within group functions and a common intercept.
This is a change from Version 6.
This is because on two occasions class was interacted with identity using the asterisk instead of the vertical
bar. In a linear model, this may be a sign of a coding error, so the procedure prints a warning. If you get this
warning while coding a choice model specifying zero=’constant-alternative-level’, you can safely ignore it.
Still, it is always good to print out one or more coded choice sets to check the coding as we will do later. Here is
the last part of the output from the %ChoicEff macro.
Food Product Example with Asymmetry and Availability Cross Effects 219

Consumer Food Product Example

Standard
n Variable Name Label Variance DF Error

1 BrandClient Client 9.5192 1 3.08532

2 BrandExtension Extension 9.5688 1 3.09335
3 BrandRegional Regional 32.4993 1 5.70081
4 BrandPrivate Private 11.9475 1 3.45651
5 BrandNational National 35.8218 1 5.98513

6 BrandClientPrice Client Price 2.3237 1 1.52437

7 BrandExtensionPrice Extension Price 1.5338 1 1.23845
8 BrandRegionalPrice Regional Price 4.5450 1 2.13191
9 BrandPrivatePrice Private Price 1.7607 1 1.32693
10 BrandNationalPrice National Price 6.4173 1 2.53323

11 ShelfTalker Shelf Talker 0.9420 1 0.97055

12 MicroMicro Micro 0.4983 1 0.70587
13 x1BrandClient Client Brand on Client . 0 .
14 x1BrandExtension Client Brand on Extension 0.5613 1 0.74923
15 x1BrandRegional Client Brand on Regional 0.5515 1 0.74265
16 x1BrandPrivate Client Brand on Private 0.5311 1 0.72879
17 x1BrandNational Client Brand on National 0.6016 1 0.77565
18 x2BrandClient Client Line Extension on Client 0.3960 1 0.62926
19 x2BrandExtension Client Line Extension on Extension . 0 .
20 x2BrandRegional Client Line Extension on Regional 0.4368 1 0.66095
21 x2BrandPrivate Client Line Extension on Private 0.4050 1 0.63643
22 x2BrandNational Client Line Extension on National 0.3484 1 0.59029
23 x5BrandClient Regional Brand on Client 0.2629 1 0.51270
24 x5BrandExtension Regional Brand on Extension 0.2971 1 0.54504
25 x5BrandRegional Regional Brand on Regional . 0 .
26 x5BrandPrivate Regional Brand on Private 0.4222 1 0.64978
27 x5BrandNational Regional Brand on National 0.3078 1 0.55477
28 x6BrandClient Private Label on Client 0.3090 1 0.55590
29 x6BrandExtension Private Label on Extension 0.3491 1 0.59086
30 x6BrandRegional Private Label on Regional 0.4039 1 0.63554
31 x6BrandPrivate Private Label on Private . 0 .
32 x6BrandNational Private Label on National 0.4439 1 0.66624
33 x8BrandClient National Competitor on Client 0.2766 1 0.52595
34 x8BrandExtension National Competitor on Extension 0.2918 1 0.54015
35 x8BrandRegional National Competitor on Regional 0.3185 1 0.56435
36 x8BrandPrivate National Competitor on Private 0.4331 1 0.65809
37 x8BrandNational National Competitor on National . 0 .
==
32

First we see estimable brand effects for each of the five brands, excluding the constant alternative ’None’. Next
we see quantitative alternative-specific price effects for each of the brands. The next two effects are single df
effects for the shelf talker and the microwave option. Next we see five sets of cross effects, each consisting of
four effects of a brand on another brand, plus one more zero df cross effect of a bran on itself. The zero df
and missing variances and standard errors are correct since the cross effect of an alternative on itself is perfectly
aliased with the alternative-specific price effects. These results look fine. Everything that should be estimable is,
and everything that should not is not.
220 TS-677E Multinomial Logit, Discrete Choice Modeling

Next, we will run some further checks by looking at the coded design. Before we look at the coded design, recall
that the design for the first five choice sets is as follows.

Consumer Food Product Example

Client Client Private

Shelf Client Line Micro/ Regional Private Micro/ National
Block Talker Brand Extension Stove Brand Label Stove Competitor

1 No $2.09 $1.89 Micro $1.99 $1.49 Stove N

N N Stove N $1.49 Stove $1.99
$2.09 $1.39 Stove $2.49 $2.29 Stove N
$1.69 $1.39 Micro N N Stove N
$1.69 $1.89 Stove $2.49 N Micro $2.39

The coded design that the %ChoicEff macro creates is called TMP– CAND. We will look at the coded data
set in several ways. First, here are the Brand, Price, microwave and shelf talker factors, for just the available
alternatives for the first five choice sets.
proc print data=tmp_cand(obs=21) label;
var Brand Price Shelf Micro;
where w;
run;

Consumer Food Product Example

Obs Brand Price Shelf Micro

1 Client $2.09 No Stove

2 Extension $1.89 No Micro
3 Regional $1.99 No Stove
4 Private $1.49 No Stove
6 None $0.00 No Stove
10 Private $1.49 No Stove
11 National $1.99 No Stove
12 None $0.00 No Stove
13 Client $2.09 No Stove
14 Extension $1.39 No Stove
15 Regional $2.49 No Stove
16 Private $2.29 No Stove
18 None $0.00 No Stove
19 Client $1.69 No Stove
20 Extension $1.39 No Micro
24 None $0.00 No Stove
25 Client $1.69 No Stove
26 Extension $1.89 No Stove
27 Regional $2.49 No Stove
29 National $2.39 No Stove
30 None $0.00 No Stove

Unlike all previous examples, the number of alternatives is not the same in all of the choice sets. The first choice
set consists of five alternatives including ’None’. The national competitor is not available in this choice set.
The second choice set consists of three alternatives including ’None’. The client brand, extension, and regional
competitors are not available in this choice set. The third choice set consists of five alternatives including ’None’,
and so on.
Food Product Example with Asymmetry and Availability Cross Effects 221

Here are the coded factors for the brand effects and alternative-specific price effects for the first choice set.
proc print data=tmp_cand(obs=5) label;
id Brand;
var BrandClient -- BrandNational;
where w;
run;
proc print data=tmp_cand(obs=5) label;
id Brand Price;
var BrandClientPrice -- BrandNationalPrice;
where w;
run;

Consumer Food Product Example

Brand Client Extension Regional Private National

Client 1 0 0 0 0
Extension 0 1 0 0 0
Regional 0 0 1 0 0
Private 0 0 0 1 0
None 0 0 0 0 0

Consumer Food Product Example

Client Extension Regional Private National

Brand Price Price Price Price Price Price

Client $2.09 2.09 0.00 0.00 0.00 0

Extension $1.89 0.00 1.89 0.00 0.00 0
Regional $1.99 0.00 0.00 1.99 0.00 0
Private $1.49 0.00 0.00 0.00 1.49 0
None $0.00 0.00 0.00 0.00 0.00 0

The brand effects and alternative-specific price effect codings are similar to those we have used previously. The
difference is the presence of all zero columns for unavailable alternatives, in this case the national competitor.
Note that Brand Price are just an ID variables and do not enter into the analysis.
Here are the shelf talker and microwave coded factors (along with the Brand, Price, Shelf, and Micro
factors.
proc print data=tmp_cand(obs=5) label;
id Brand Price Shelf Micro;
var shelftalker micromicro;
where w;
run;

Consumer Food Product Example

Shelf
Brand Price Shelf Micro Talker Micro

Client $2.09 No Stove 0 0

Extension $1.89 No Micro 0 1
Regional $1.99 No Stove 0 0
Private $1.49 No Stove 0 0
None $0.00 No Stove 0 0
222 TS-677E Multinomial Logit, Discrete Choice Modeling

The following code prints the cross effects along with Brand and Price for the first choice set.
proc print data=tmp_cand(obs=5) label;
id Brand Price;
var x1Brand:;
where w;
run;
proc print data=tmp_cand(obs=5) label;
id Brand Price;
var x2Brand:;
where w;
run;
proc print data=tmp_cand(obs=5) label;
id Brand Price;
var x5Brand:;
where w;
run;
proc print data=tmp_cand(obs=5) label;
id Brand Price;
var x6Brand:;
where w;
run;
proc print data=tmp_cand(obs=5) label;
id Brand Price;
var x8Brand:;
where w;
run;
The cross effects are printed in panels. This first panel shows the terms that capture the effect of the client brand
being available at $2.09 on the utility of the other brands. The last panel shows that the national competitor,
which is unavailable, has no effect on any other brand’s utility in this choice set.

Consumer Food Product Example

Client Client Client Client Client

Brand on Brand on Brand on Brand on Brand on
Brand Price Client Extension Regional Private National

Client $2.09 2.09 0.00 0.00 0.00 0

Extension $1.89 0.00 2.09 0.00 0.00 0
Regional $1.99 0.00 0.00 2.09 0.00 0
Private $1.49 0.00 0.00 0.00 2.09 0
None $0.00 0.00 0.00 0.00 0.00 0

Consumer Food Product Example

Client Line Client Line Client Line Client Line Client Line
Extension Extension on Extension on Extension Extension on
Brand Price on Client Extension Regional on Private National

Client $2.09 1.89 0.00 0.00 0.00 0

Extension $1.89 0.00 1.89 0.00 0.00 0
Regional $1.99 0.00 0.00 1.89 0.00 0
Private $1.49 0.00 0.00 0.00 1.89 0
None $0.00 0.00 0.00 0.00 0.00 0
Food Product Example with Asymmetry and Availability Cross Effects 223

Consumer Food Product Example

Regional Regional Regional Regional Regional

Brand on Brand on Brand on Brand on Brand on
Brand Price Client Extension Regional Private National

Client $2.09 1.99 0.00 0.00 0.00 0

Extension $1.89 0.00 1.99 0.00 0.00 0
Regional $1.99 0.00 0.00 1.99 0.00 0
Private $1.49 0.00 0.00 0.00 1.99 0
None $0.00 0.00 0.00 0.00 0.00 0

Consumer Food Product Example

Private Private Private Private Private

Label on Label on Label on Label on Label on
Brand Price Client Extension Regional Private National

Client $2.09 1.49 0.00 0.00 0.00 0

Extension $1.89 0.00 1.49 0.00 0.00 0
Regional $1.99 0.00 0.00 1.49 0.00 0
Private $1.49 0.00 0.00 0.00 1.49 0
None $0.00 0.00 0.00 0.00 0.00 0

Consumer Food Product Example

National National National National National

Competitor Competitor Competitor Competitor Competitor
Brand Price on Client on Extension on Regional on Private on National

Client $2.09 0 0 0 0 0
Extension $1.89 0 0 0 0 0
Regional $1.99 0 0 0 0 0
Private $1.49 0 0 0 0 0
None $0.00 0 0 0 0 0

A column like ’Private Label on Client’ in the second last panel, for example captured the effect of
the private label brand being available at $1.49 on the utility of the client brand. In the previous pane, ’Re-
gional Brand on Extension’ captures the effect of the regional brand being available at $1.99 on the
utility of the extension.
The design looks good, it has reasonably good balance and correlations, it can be used to estimate all of the
effects of interest, and we have shown we know how to specify the model to get all the right codings. We are
ready to collect data.

Generating Artificial Data

We will not illustrate questionnaire generation for this example since we have done it several times before.
Instead we will go straight to data processing and analysis. This DATA step generates some artificial data.
Creating artificial data and trying the analysis before collecting data is another way to test the design before
going to the expense of data collection.
%let m = 6;
%let mm1 = %eval(&m - 1);
%let n = 36;
224 TS-677E Multinomial Logit, Discrete Choice Modeling

proc format;
value yn 1 = ’ No’ 2 = ’Talker’;
value micro 1 = ’Micro’ 2 = ’Stove’;
run;

data _null_;
array brands[&m] _temporary_ (5 7 1 2 3 -2);
array u[&m];
array x[&mm1] x1 x2 x5 x6 x8;

do rep = 1 to 300;
if mod(rep, 2) then put;
put rep 3. +2 @@;
do j = 1 to &n;
set [Link] point=j;
do brand = 1 to &m; u[brand] = brands[brand] + 2 * normal(7); end;
do brand = 1 to &mm1;
if n(x[brand]) then u[brand] + -x[brand]; else u[brand] = .;
end;
if n(u2) and x4 = 2 then u2 + 1; /* shelf-talker */
if n(u2) and x3 = 1 then u2 + 1; /* microwave */
if n(u4) and x7 = 1 then u4 + 1; /* microwave */
* Choose the most preferred alternative.;
m = max(of u1-u&m);

do brand = 1 to &m;
if n(u[brand]) then if abs(u[brand] - m) < 1e-4 then c = brand;
end;
put +(-1) c @@;
end;
end;
stop;
run;
This DATA step reads the data.
data results;
input Subj (choose1-choose&n) (1.) @@;
datalines;
1 252212542412222122622122115222212221 2 252222521452221422122122212222212226
3 242221122412222422122122212222112221 4 242222122312222122122112212222252211
5 241222122452222122124152232522212221 6 251222122452222122122122212222212521
.
.
.
297 242222122412221122122522242422252221 298 251122141315222522122121212222112221
299 352222122412222112112511212222212221 300 242222122452222522512112212222112211
;

Processing the Data

The analysis proceeds in a fashion similar to before. We have already made the choice design, so we just have to
merge it with the data. The data and design are merged in the usual way using the %MktMerge macro. Notice
at this point that the unavailable alternatives are still in the design. The %MktMerge macro has an nalts=
alternative and expects a constant number of alternatives in each choice set.
%mktmerge(design=[Link], data=results, out=res2,
nsets=&n, nalts=&m, setvars=choose1-choose&n)

proc print data=res2(obs=12); id subj set; by subj set; run;

Food Product Example with Asymmetry and Availability Cross Effects 225

Here are the data and design for the first two choice sets for the first subject, including the unavailable alternatives.

Subj Set Brand Price Micro Shelf x1 x2 x5 x6 x8 w c

1 1 Client $2.09 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2

Extension $1.89 Micro No $2.09 $1.89 $1.99 $1.49 $0.00 1 1
Regional $1.99 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2
Private $1.49 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2
National $0.00 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 0 2
None $0.00 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2
1 2 Client $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0 2
Extension $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0 2
Regional $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0 2
Private $1.49 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1 2
National $1.99 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1 1
None $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1 2

These next steps aggregate the data. The data set is fairly large at 64,800 observations, and aggregating greatly
reduces its size, which makes both the TRANSREG and the PHREG steps run in just a few seconds. This step
also excludes the unavailable alternatives. When w is 1 (true) the alternative is available and counted, otherwise
when w is 0 (false) the alternative is unavailable and excluded by the where clause and not counted. There is
nothing in subsequent steps that assumes a fixed number of alternatives.
proc summary data=res2 nway;
class set brand price shelf micro x1 x2 x5 x6 x8 c;
output out=agg(drop=_type_);
where w; /* exclude unavailable, w = 0 */
run;

proc print; where set = 1; run;

All of the variables used in the analysis are named as class variables in PROC SUMMARY, which reduces the
data set from 64,800 observations to 292. Here are the aggregated data for the first choice set.

Consumer Food Product Example

Obs Set Brand Price Shelf Micro x1 x2 x5 x6 x8 c _FREQ_

1 1 Client $2.09 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 1 42

2 1 Client $2.09 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 2 258
3 1 Extension $1.89 No Micro $2.09 $1.89 $1.99 $1.49 $0.00 1 255
4 1 Extension $1.89 No Micro $2.09 $1.89 $1.99 $1.49 $0.00 2 45
5 1 None $0.00 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 2 300
6 1 Private $1.49 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 1 1
7 1 Private $1.49 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 2 299
8 1 Regional $1.99 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 1 2
9 1 Regional $1.99 No Stove $2.09 $1.89 $1.99 $1.49 $0.00 2 298

In the first choice set, the client brand was chosen (c = 1) a total of - freq- = 42 times and not chosen (c = 2)
a total of - freq- = 258 times. Each alternative was chosen and not chosen a total of 300 times, which is the
number of subjects. These next steps code and run the analysis.
226 TS-677E Multinomial Logit, Discrete Choice Modeling

Cross Effects
This next step codes the design for analysis. This coding was discussed on page 217. PROC TRANSREG is
run like before, except now the data set AGG is specified and the ID variable includes - freq- (the frequency
variable) but not Subj (the subject number variable).
proc transreg data=agg design=5000 nozeroconstant norestoremissing;
model class(brand / zero=’None’)
class(brand / zero=’None’ separators=’’ ’ ’) * identity(price)
class(shelf micro / lprefix=5 0 zero=’No’ ’Stove’)
identity(x1 x2 x5 x6 x8) *
class(brand / zero=’None’ separators=’ ’ ’ on ’) /
lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
id set c _freq_;
label x1 = ’CE, Client’
x2 = ’CE, Extension’
x5 = ’CE, Regional’
x6 = ’CE, Private’
x8 = ’CE, National’
shelf = ’Shelf Talker’
micro = ’Microwave’;
run;
Note that like we saw in the %ChoicEff macro, PROC TRANSREG produces the following warning twice.
WARNING: This usage of * sets one group’s slope to zero. Specify |
to allow all slopes and intercepts to vary. Alternatively,
specify CLASS(vars) * identity(vars) identity(vars) for
separate within group functions and a common intercept.
This is a change from Version 6.
This is because on two occasions class was interacted with identity using the asterisk instead of the vertical
bar. In a linear model, this may be a sign of a coding error, so the procedure prints a warning. If you get this
warning while coding a choice model specifying zero=’constant-alternative-level’, you can safely ignore it.
Analysis is the same as we have done previously with aggregate data. PROC PHREG is run to fit the mother
logit model, complete with availability cross effects.
proc phreg data=coded;
strata set;
model c*c(2) = &_trgind / ties=breslow;
freq _freq_;
run;

Multinomial Logit Model Results

These steps produced the following results. (Recall that we used %phchoice(on)on page 79 to customize the
output from PROC PHREG.)
Food Product Example with Asymmetry and Availability Cross Effects 227

Consumer Food Product Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Frequency Variable _FREQ_
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Set Alternatives Alternatives Chosen

1 1 1500 300 1200

2 2 900 300 600
3 3 1500 300 1200
4 4 900 300 600
5 5 1500 300 1200
6 6 1500 300 1200
7 7 1500 300 1200
8 8 1500 300 1200
9 9 1500 300 1200
10 10 900 300 600
11 11 1200 300 900
12 12 1500 300 1200
13 13 1500 300 1200
14 14 900 300 600
15 15 1500 300 1200
16 16 1500 300 1200
17 17 1500 300 1200
18 18 900 300 600
19 19 1200 300 900
20 20 1500 300 1200
21 21 1200 300 900
22 22 1500 300 1200
23 23 1500 300 1200
24 24 1500 300 1200
25 25 900 300 600
26 26 1200 300 900
27 27 1500 300 1200
28 28 1500 300 1200
29 29 1500 300 1200
30 30 900 300 600
31 31 1500 300 1200
32 32 1500 300 1200
33 33 1500 300 1200
34 34 1500 300 1200
35 35 1500 300 1200
36 36 900 300 600
---------------------------------------------------------------
Total 48000 10800 37200
228 TS-677E Multinomial Logit, Discrete Choice Modeling

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 154978.05 135268.63

AIC 154978.05 135332.63
SBC 154978.05 135565.83

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 19709.4180 32 <.0001

Score 22091.0481 32 <.0001
Wald 7080.2597 32 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Client 1 4.24512 0.51680 67.4740 <.0001

Extension 1 5.48777 0.54501 101.3878 <.0001
National 1 2.69212 0.74517 13.0519 0.0003
Private 1 2.93103 0.57137 26.3150 <.0001
Regional 1 2.58486 1.17748 4.8191 0.0281
Client Price 1 0.22857 0.32495 0.4948 0.4818
Extension Price 1 0.28959 0.27805 1.0848 0.2976
National Price 1 -0.73909 0.35790 4.2646 0.0389
Private Price 1 -0.50607 0.26142 3.7476 0.0529
Regional Price 1 -1.12619 0.46588 5.8435 0.0156
Shelf Talker 1 0.63195 0.06938 82.9781 <.0001
Micro 1 0.67436 0.06136 120.7997 <.0001
CE, Client on Client 0 0 . . .
CE, Client on Extension 1 0.97241 0.31232 9.6938 0.0018
CE, Client on National 1 0.78460 0.30882 6.4546 0.0111
CE, Client on Private 1 0.81216 0.30519 7.0820 0.0078
CE, Client on Regional 1 0.62573 0.34844 3.2250 0.0725
CE, Extension on Client 1 1.13135 0.26635 18.0422 <.0001
CE, Extension on Extension 0 0 . . .
CE, Extension on National 1 0.89205 0.26587 11.2574 0.0008
CE, Extension on Private 1 0.75865 0.26804 8.0110 0.0046
CE, Extension on Regional 1 0.81508 0.29154 7.8160 0.0052
CE, Regional on Client 1 -0.14977 0.21001 0.5086 0.4757
CE, Regional on Extension 1 -0.09684 0.21131 0.2100 0.6467
CE, Regional on National 1 -0.19080 0.21256 0.8057 0.3694
CE, Regional on Private 1 -0.21540 0.21625 0.9921 0.3192
CE, Regional on Regional 0 0 . . .
Food Product Example with Asymmetry and Availability Cross Effects 229

CE, Private on Client 1 0.39355 0.19335 4.1429 0.0418

CE, Private on Extension 1 0.38933 0.19579 3.9543 0.0468
CE, Private on National 1 0.35647 0.19772 3.2504 0.0714
CE, Private on Private 0 0 . . .
CE, Private on Regional 1 0.23482 0.22396 1.0994 0.2944
CE, National on Client 1 -0.37645 0.23758 2.5107 0.1131
CE, National on Extension 1 -0.33986 0.23888 2.0241 0.1548
CE, National on National 0 0 . . .
CE, National on Private 1 -0.32288 0.23670 1.8608 0.1725
CE, National on Regional 1 -0.27579 0.26135 1.1136 0.2913

Since the number of alternatives is not constant within each choice set, the summary table has nonconstant
numbers of alternatives and numbers not chosen. The number chosen, 300 (or one per subject per choice set),
is constant, since each subject always chooses one alternative from each choice set regardless of the number of
alternatives. The first choice set has 1500 alternatives, 5 available times 300 subjects; whereas the fifth choice
set has 900 alternatives, 3 available times 300 subjects.
The most to least preferred brands are: client line extension, client brand, private label, national brand, and
regional competitor, and finally the none alternative (with an implicit part-worth utility of zero). The price effects
are mostly negative, and the positive effects are nonsignificant. Both the shelf-talker and the microwaveable
option have positive utility. The cross effects are mostly nonsignificant. The most significant cross effect is the
effect of the extension on the original client brand.

Modeling Subject Attributes

This example uses the same design and data as we just saw, but this time we have some demographic information
about our respondents that we wish to model. The following DATA step reads a subject number, the choices, and
respondent age and income (in thousands of dollars). The data from two subjects appear on one line.
data results;
input Subj (choose1-choose&n) (1.) age income;
datalines;
1 252212542412222122622122115222212221 33 44
2 252222521452221422122122212222212226 52 82
3 242221122412222422122122212222112221 51 136
4 242222122312222122122112212222252211 60 108
.
.
.
299 352222122412222112112511212222212221 48 49
300 242222122452222522512112212222112211 38 51
;
Merging the data and design is no different from what we saw previously.
%mktmerge(design=[Link], data=results, out=res2,
nsets=&n, nalts=&m, setvars=choose1-choose&n)

proc print data=res2;

by subj set; id subj set;
where (subj = 1 and set = 1) or
(subj = 2 and set = 2) or
(subj = 3 and set = 3) or
(subj = 300 and set = 36);
run;
230 TS-677E Multinomial Logit, Discrete Choice Modeling

Here is a small sample of the data. Note that like before, the unavailable alternatives are required for the merge
step.

Consumer Food Product Example

Subj Set Age Income Brand Price Micro Shelf x1 x2 x5 x6 x8 w c

1 1 33 44 Client $2.09 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2

33 44 Extension $1.89 Micro No $2.09 $1.89 $1.99 $1.49 $0.00 1 1
33 44 Regional $1.99 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2
33 44 Private $1.49 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2
33 44 National $0.00 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 0 2
33 44 None $0.00 Stove No $2.09 $1.89 $1.99 $1.49 $0.00 1 2

2 2 52 82 Client $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0 2

52 82 Extension $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0 2
52 82 Regional $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 0 2
52 82 Private $1.49 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1 2
52 82 National $1.99 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1 1
52 82 None $0.00 Stove No $0.00 $0.00 $0.00 $1.49 $1.99 1 2

3 3 51 136 Client $2.09 Stove No $2.09 $1.39 $2.49 $2.29 $0.00 1 2

51 136 Extension $1.39 Stove No $2.09 $1.39 $2.49 $2.29 $0.00 1 1
51 136 Regional $2.49 Stove No $2.09 $1.39 $2.49 $2.29 $0.00 1 2
51 136 Private $2.29 Stove No $2.09 $1.39 $2.49 $2.29 $0.00 1 2
51 136 National $0.00 Stove No $2.09 $1.39 $2.49 $2.29 $0.00 0 2
51 136 None $0.00 Stove No $2.09 $1.39 $2.49 $2.29 $0.00 1 2

300 36 38 51 Client $1.29 Stove No $1.29 $0.00 $0.00 $0.00 $2.39 1 1

38 51 Extension $0.00 Micro Talker $1.29 $0.00 $0.00 $0.00 $2.39 0 2
38 51 Regional $0.00 Stove No $1.29 $0.00 $0.00 $0.00 $2.39 0 2
38 51 Private $0.00 Stove No $1.29 $0.00 $0.00 $0.00 $2.39 0 2
38 51 National $2.39 Stove No $1.29 $0.00 $0.00 $0.00 $2.39 1 2
38 51 None $0.00 Stove No $1.29 $0.00 $0.00 $0.00 $2.39 1 2

You can see that the demographic information matches the raw data and is constant within each subject. The
rest of the data processing is virtually the same as well. Since we have demographic information, we will not
aggregate. There would have to be ties in both the demographics and choice for aggregation to have any effect.
We use PROC TRANSREG to code, adding Age and Income to the analysis.
proc transreg data=res2 design=5000 nozeroconstant norestoremissing;
model class(brand / zero=’None’)
identity(age income) * class(brand / zero=’None’ separators=’’ ’, ’)
class(brand / zero=’None’ separators=’’ ’ ’) * identity(price)
class(shelf micro / lprefix=5 0 zero=’No’ ’Stove’)
identity(x1 x2 x5 x6 x8) *
class(brand / zero=’None’ separators=’ ’ ’ on ’) /
lprefix=0 order=data;
Food Product Example with Asymmetry and Availability Cross Effects 231

output out=coded(drop=_type_ _name_ intercept);

where w; /* exclude unavailable, w = 0 */
id subj set c;
label x1 = ’CE, Client’
x2 = ’CE, Extension’
x5 = ’CE, Regional’
x6 = ’CE, Private’
x8 = ’CE, National’
shelf = ’Shelf Talker’
micro = ’Microwave’;
run;
The Age and Income variables are incorporated into the analysis by interacting them with Brand. Demo-
graphic variables must be interacted with attributes to have any effect. If identity(age income) had been
specified instead of identity(age income) * class(brand / ...) the coefficients for age and
income would be zero. This is because age and income are constant within each choice set and subject combi-
nation, which means they are constant within each stratum. The second separator ’, ’ is used to create names
for the brand/demographic interaction terms like ’Age, Client’.
These next steps print the first coded choice set.
proc print data=coded(obs=5) label;
id brand price;
var BrandClient -- BrandPrivate Shelf Micro c;
run;
proc print data=coded(obs=5 drop=Age) label;
id brand price;
var Age:;
run;
proc print data=coded(obs=5 drop=Income) label;
id brand price;
var Income:;
run;
proc print data=coded(obs=5) label;
id brand price;
var BrandClientPrice -- BrandPrivatePrice;
format BrandClientPrice -- BrandPrivatePrice best4.;
run;
proc print data=coded(obs=5 drop=x1) label;
id brand price; var x1:; format x1: best4.;
run;
proc print data=coded(obs=5 drop=x2) label;
id brand price; var x2:; format x2: best4.;
run;
proc print data=coded(obs=5 drop=x5) label;
id brand price; var x5:; format x5: best4.;
run;
proc print data=coded(obs=5 drop=x6) label;
id brand price; var x6:; format x6: best4.;
run;
proc print data=coded(obs=5 drop=x8) label;
id brand price; var x8:; format x8: best4.;
run;
Here is the coded data set for the first subject and choice set. The part that is new is the second and third panel,
which will be used to capture the brand by age and brand by income effects.
232 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are the attributes and the brand effects.

Consumer Food Product Example

Shelf
Brand Price Client Extension Regional Private Talker Microwave c

Client $2.09 1 0 0 0 No Stove 2

Extension $1.89 0 1 0 0 No Micro 1
Regional $1.99 0 0 1 0 No Stove 2
Private $1.49 0 0 0 1 No Stove 2
None $0.00 0 0 0 0 No Stove 2

Here are the age by brand effects.

Consumer Food Product Example

Age, Age, Age, Age, Age,

Brand Price Client Extension Regional Private National

Client $2.09 33 0 0 0 0
Extension $1.89 0 33 0 0 0
Regional $1.99 0 0 33 0 0
Private $1.49 0 0 0 33 0
None $0.00 0 0 0 0 0

Here are the income by brand effects.

Consumer Food Product Example

Income, Income, Income, Income, Income,

Brand Price Client Extension Regional Private National

Client $2.09 44 0 0 0 0
Extension $1.89 0 44 0 0 0
Regional $1.99 0 0 44 0 0
Private $1.49 0 0 0 44 0
None $0.00 0 0 0 0 0

Here are the alternative-specific price effects.

Consumer Food Product Example

Client Extension Regional Private

Brand Price Price Price Price Price

Client $2.09 2.09 0 0 0

Extension $1.89 0 1.89 0 0
Regional $1.99 0 0 1.99 0
Private $1.49 0 0 0 1.49
None $0.00 0 0 0 0
Food Product Example with Asymmetry and Availability Cross Effects 233

Here are the client cross effects.

Consumer Food Product Example

CE, Client CE, Client CE, Client

CE, Client on on CE, Client on
Brand Price on Client Extension Regional on Private National

Client $2.09 2.09 0 0 0 0

Extension $1.89 0 2.09 0 0 0
Regional $1.99 0 0 2.09 0 0
Private $1.49 0 0 0 2.09 0
None $0.00 0 0 0 0 0

Here are the extension cross effects.

Consumer Food Product Example

CE, CE, CE,

CE, Extension Extension CE, Extension
Extension on on Extension on
Brand Price on Client Extension Regional on Private National

Client $2.09 1.89 0 0 0 0

Extension $1.89 0 1.89 0 0 0
Regional $1.99 0 0 1.89 0 0
Private $1.49 0 0 0 1.89 0
None $0.00 0 0 0 0 0

Here are the regional competitor cross effects.

Consumer Food Product Example

CE, CE, CE,

CE, Regional Regional CE, Regional
Regional on on Regional on
Brand Price on Client Extension Regional on Private National

Client $2.09 1.99 0 0 0 0

Extension $1.89 0 1.99 0 0 0
Regional $1.99 0 0 1.99 0 0
Private $1.49 0 0 0 1.99 0
None $0.00 0 0 0 0 0

Here are the private label cross effects.

Consumer Food Product Example

CE, CE, CE, CE, CE,

Private on Private on Private on Private on Private on
Brand Price Client Extension Regional Private National

Client $2.09 1.49 0 0 0 0

Extension $1.89 0 1.49 0 0 0
Regional $1.99 0 0 1.49 0 0
Private $1.49 0 0 0 1.49 0
None $0.00 0 0 0 0 0
234 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are the national competitor cross effects.

Consumer Food Product Example

CE, CE, CE,

CE, National National CE, National
National on on National on
Brand Price on Client Extension Regional on Private National

Client $2.09 0 0 0 0 0
Extension $1.89 0 0 0 0 0
Regional $1.99 0 0 0 0 0
Private $1.49 0 0 0 0 0
None $0.00 0 0 0 0 0

The PROC PHREG specification is the same as we have used before with nonaggregated data.
proc phreg data=coded brief;
model c*c(2) = &_trgind / ties=breslow;
strata subj set;
run;
This step took just about one minute and produced the following results.

Consumer Food Product Example

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Number of Chosen Not

Pattern Choices Alternatives Alternatives Chosen

1 2400 3 1 2
2 1200 4 1 3
3 7200 5 1 4

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 31776.351 12053.607

AIC 31776.351 12137.607
SBC 31776.351 12443.674
Food Product Example with Asymmetry and Availability Cross Effects 235

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 19722.7440 42 <.0001

Score 22093.7119 42 <.0001
Wald 7064.8884 42 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Client 1 5.15121 0.64435 63.9112 <.0001

Extension 1 6.34979 0.67179 89.3407 <.0001
Regional 1 3.68135 1.26766 8.4335 0.0037
Private 1 3.72866 0.69082 29.1321 <.0001
National 1 3.58319 0.84306 18.0643 <.0001
Age, Client 1 -0.03218 0.01306 6.0733 0.0137
Age, Extension 1 -0.03271 0.01338 5.9778 0.0145
Age, Regional 1 -0.04536 0.01657 7.4964 0.0062
Age, Private 1 -0.03551 0.01317 7.2717 0.0070
Age, National 1 -0.02887 0.01346 4.6048 0.0319
Income, Client 1 0.00773 0.00548 1.9941 0.1579
Income, Extension 1 0.00876 0.00560 2.4448 0.1179
Income, Regional 1 0.01336 0.00676 3.9103 0.0480
Income, Private 1 0.01154 0.00551 4.3830 0.0363
Income, National 1 0.00579 0.00564 1.0539 0.3046
Client Price 1 0.23736 0.32667 0.5280 0.4675
Extension Price 1 0.29648 0.27957 1.1247 0.2889
Regional Price 1 -1.12989 0.46647 5.8670 0.0154
Private Price 1 -0.50333 0.26187 3.6944 0.0546
National Price 1 -0.74511 0.35881 4.3124 0.0378
Shelf Talker 1 0.63199 0.06938 82.9766 <.0001
Micro 1 0.67453 0.06136 120.8327 <.0001
CE, Client on Client 0 0 . . .
CE, Client on Extension 1 0.98125 0.31408 9.7608 0.0018
CE, Client on Regional 1 0.63475 0.35006 3.2880 0.0698
CE, Client on Private 1 0.82032 0.30695 7.1425 0.0075
CE, Client on National 1 0.79361 0.31055 6.5306 0.0106
CE, Extension on Client 1 1.13830 0.26794 18.0478 <.0001
CE, Extension on Extension 0 0 . . .
CE, Extension on Regional 1 0.82281 0.29302 7.8852 0.0050
CE, Extension on Private 1 0.76600 0.26958 8.0737 0.0045
CE, Extension on National 1 0.89865 0.26742 11.2925 0.0008
CE, Regional on Client 1 -0.15260 0.21106 0.5228 0.4697
CE, Regional on Extension 1 -0.09966 0.21236 0.2202 0.6389
CE, Regional on Regional 0 0 . . .
CE, Regional on Private 1 -0.21835 0.21734 1.0093 0.3151
CE, Regional on National 1 -0.19370 0.21362 0.8221 0.3646
CE, Private on Client 1 0.39602 0.19388 4.1723 0.0411
CE, Private on Extension 1 0.39183 0.19632 3.9837 0.0459
CE, Private on Regional 1 0.23759 0.22445 1.1205 0.2898
CE, Private on Private 0 0 . . .
CE, Private on National 1 0.35882 0.19823 3.2766 0.0703
236 TS-677E Multinomial Logit, Discrete Choice Modeling

CE, National on Client 1 -0.38253 0.23881 2.5657 0.1092

CE, National on Extension 1 -0.34594 0.24010 2.0760 0.1496
CE, National on Regional 1 -0.28219 0.26249 1.1557 0.2824
CE, National on Private 1 -0.32953 0.23791 1.9185 0.1660
CE, National on National 0 0 . . .

In previous examples, when we used the brief option to produce a brief summary of the strata, the table
had only one line. In this case, since our choice sets have 3, 4, or 5 alternatives, we have three rows, one for
each choice set size. The coefficients for the age and income variables are generally not very significant in this
analysis.
Allocation of Prescription Drugs 237

Allocation of Prescription Drugs

This example discusses an allocation study, which is a technique often used in the area of prescription drug
marketing research. This example discusses designing the allocation experiment, processing the data, analyzing
frequencies, analyzing proportions, coding, analysis, and results. The principles of designing an allocation study
are the same as for designing a first-choice experiment, as is the coding and final analysis. However, processing
the data before analysis is different.
The previous examples have all modeled simple choice. However, sometimes the response of interest is not
simple first choice. For example, in prescription drug marketing, researchers often use allocation studies where
multiple, not single choices are made. Physicians are asked questions like “For the next ten prescriptions you
write for a particular condition, how many would you write for each of these drugs?” The response, for example,
could be “5 for drug 1, none for drug 2, 3 for drug 3, and 2 for drug 4.”

Designing the Allocation Experiment

In this study, physicians were asked to specify which of ten drugs they would prescribe to their next ten patients.
In this study, ten drugs, Drug 1 Drug 10, were available each at three different prices, $50, $75, and $100.
In real studies, real brand names would be used and there would probably be more attributes. Since design has
been covered in some detail in other examples, we chose a simple design for this experiment so that we could
concentrate on data processing. First, we use the %MktRuns autocall macro to suggest a design size. (All of
the autocall macros used in this report are documented starting on page 287.) We specify 3 ** 10 for the 10
three-level factors.
title ’Allocation of Prescription Drugs’;

%mktruns( 3 ** 10 )

Allocation of Prescription Drugs

Design Summary

Number of
Levels Frequency

3 10

Allocation of Prescription Drugs

Saturated = 21
Full Factorial = 59,049

Some Reasonable Cannot Be

Design Sizes Violations Divided By

27 * 0
36 * 0
45 0
54 * 0
21 45 9
24 45 9
30 45 9
33 45 9
39 45 9
42 45 9

* - 100% Efficient Design can be made with the MktEx Macro.

238 TS-677E Multinomial Logit, Discrete Choice Modeling

Allocation of Prescription Drugs

n Design Reference

27 3 ** 13 Fractional-factorial
36 2 ** 11 3 ** 12 Taguchi, 1987
36 2 ** 4 3 ** 13 Taguchi, 1987
36 2 ** 2 3 ** 12 6 ** 1 Wang and Wu, 1991
36 3 ** 13 4 ** 1 Dey, 1985
36 3 ** 12 12 ** 1 Wang and Wu, 1991
54 2 ** 1 3 ** 25 Taguchi, 1987
54 3 ** 24 6 ** 1 Hedayat, Sloane, and Stufken, 1999
54 3 ** 18 18 ** 1 Hedayat, Sloane, and Stufken, 1999

We need at least 21 choice sets and we see the optimal sizes are all divisible by nine. We will use 27 choice sets,
which can give us up to 13 three-level factors.
Next, we use the %MktEx macro to create the design. In addition, one more factor is added to the design. This
factor will be used to block the design into three blocks of size 9.
%let nalts = 10;

%mktex(3 ** &nalts 3, n=27, seed=7654321)

The macro finds a 100% D-efficient design.

Allocation of Prescription Drugs

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 100.0000 100.0000 Tab
1 End 100.0000

Allocation of Prescription Drugs

The OPTEX Procedure

Class Level Information

Class Levels -Values-

x1 3 1 2 3
x2 3 1 2 3
x3 3 1 2 3
x4 3 1 2 3
x5 3 1 2 3
x6 3 1 2 3
x7 3 1 2 3
x8 3 1 2 3
x9 3 1 2 3
x10 3 1 2 3
x11 3 1 2 3
Allocation of Prescription Drugs 239

Allocation of Prescription Drugs

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 100.0000 100.0000 100.0000 0.9230

The %MktEx macro always creates factor names of x1, x2, and so on with values of 1, 2, .... You can create
a data set with the names and values you want and use it to rename the factors and reset the levels. This first
step creates a data set with 11 variables, Block and Brand1 - Brand10. Block has values 1, 2, and 3,
and the brand variables have values of 50, 75, and 100 with a dollar format. The %MktLab macro takes the
data=Randomized design data set and uses the names, values, and formats in the key=Key data set to make
the out=Final data set. This data set is sorted by block and printed. The %MktEval macro is called to check
the results.
data key(drop=i);
input Block Brand1;
array Brand[10];
do i = 2 to 10; brand[i] = brand1; end;
format brand: dollar4.;
datalines;
1 50
2 75
3 100
;

proc print; run;

%mktlab(key=key);

proc sort out=[Link]; by block; run;

proc print; id block; by block; run;

%mkteval(blocks=block)
Here is the key= data set.

Allocation of Prescription Drugs

Obs Block Brand1 Brand2 Brand3 Brand4 Brand5 Brand6 Brand7 Brand8 Brand9 Brand10

1 1 $50 $50 $50 $50 $50 $50 $50 $50 $50 $50
2 2 $75 $75 $75 $75 $75 $75 $75 $75 $75 $75
3 3 $100 $100 $100 $100 $100 $100 $100 $100 $100 $100
240 TS-677E Multinomial Logit, Discrete Choice Modeling

The %MktLab macro prints the following mapping information.

Variable Mapping:
x1 : Block
x2 : Brand1
x3 : Brand2
x4 : Brand3
x5 : Brand4
x6 : Brand5
x7 : Brand6
x8 : Brand7
x9 : Brand8
x10 : Brand9
x11 : Brand10

Here is the design.

Allocation of Prescription Drugs

Block Brand1 Brand2 Brand3 Brand4 Brand5 Brand6 Brand7 Brand8 Brand9 Brand10

1 $100 $75 $100 $100 $75 $100 $50 $50 $75 $100
$50 $100 $75 $100 $100 $100 $100 $75 $75 $75
$75 $50 $50 $100 $50 $100 $75 $100 $75 $50
$75 $75 $100 $50 $75 $75 $100 $75 $50 $75
$75 $100 $75 $75 $100 $50 $50 $50 $100 $100
$50 $50 $50 $50 $50 $75 $50 $50 $50 $100
$100 $50 $50 $75 $50 $50 $100 $75 $100 $75
$100 $100 $75 $50 $100 $75 $75 $100 $50 $50
$50 $75 $100 $75 $75 $50 $75 $100 $100 $50

2 $75 $100 $50 $75 $75 $75 $75 $75 $75 $100
$100 $50 $100 $75 $100 $75 $50 $100 $75 $75
$100 $100 $50 $50 $75 $100 $100 $50 $100 $50
$100 $75 $75 $100 $50 $50 $75 $75 $50 $100
$50 $100 $50 $100 $75 $50 $50 $100 $50 $75
$75 $75 $75 $50 $50 $100 $50 $100 $100 $75
$50 $75 $75 $75 $50 $75 $100 $50 $75 $50
$50 $50 $100 $50 $100 $100 $75 $75 $100 $100
$75 $50 $100 $100 $100 $50 $100 $50 $50 $50

3 $100 $75 $50 $100 $100 $75 $100 $100 $100 $100
$75 $100 $100 $75 $50 $100 $100 $100 $50 $100
$100 $50 $75 $75 $75 $100 $75 $50 $50 $75
$100 $100 $100 $50 $50 $50 $50 $75 $75 $50
$75 $75 $50 $50 $100 $50 $75 $50 $75 $75
$75 $50 $75 $100 $75 $75 $50 $75 $100 $50
$50 $100 $100 $100 $50 $75 $75 $50 $100 $75
$50 $75 $50 $75 $100 $100 $50 $75 $50 $50
$50 $50 $75 $50 $75 $50 $100 $100 $75 $100
Allocation of Prescription Drugs 241

Here are some of the evaluation results.

Allocation of Prescription Drugs

Canonical Correlations Between the Factors
There are 0 Canonical Correlations Greater Than 0.316

Block Brand1 Brand2 Brand3 Brand4 Brand5 Brand6 Brand7 Brand8 Brand9 Brand10

Block 1 0 0 0 0 0 0 0 0 0 0
Brand1 0 1 0 0 0 0 0 0 0 0 0
Brand2 0 0 1 0 0 0 0 0 0 0 0
Brand3 0 0 0 1 0 0 0 0 0 0 0
Brand4 0 0 0 0 1 0 0 0 0 0 0
Brand5 0 0 0 0 0 1 0 0 0 0 0
Brand6 0 0 0 0 0 0 1 0 0 0 0
Brand7 0 0 0 0 0 0 0 1 0 0 0
Brand8 0 0 0 0 0 0 0 0 1 0 0
Brand9 0 0 0 0 0 0 0 0 0 1 0
Brand10 0 0 0 0 0 0 0 0 0 0 1

Allocation of Prescription Drugs

Summary of Frequencies
There are 0 Canonical Correlations Greater Than 0.316

Frequencies

Block 9 9 9
Brand1 9 9 9
Brand2 9 9 9
Brand3 9 9 9
Brand4 9 9 9
Brand5 9 9 9
Brand6 9 9 9
Brand7 9 9 9
Brand8 9 9 9
Brand9 9 9 9
Brand10 9 9 9

Block Brand1 3 3 3 3 3 3 3 3 3
Block Brand2 3 3 3 3 3 3 3 3 3
Block Brand3 3 3 3 3 3 3 3 3 3
Block Brand4 3 3 3 3 3 3 3 3 3
Block Brand5 3 3 3 3 3 3 3 3 3
Block Brand6 3 3 3 3 3 3 3 3 3
Block Brand7 3 3 3 3 3 3 3 3 3
Block Brand8 3 3 3 3 3 3 3 3 3
Block Brand9 3 3 3 3 3 3 3 3 3
Block Brand10 3 3 3 3 3 3 3 3 3
.
.
.
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1
242 TS-677E Multinomial Logit, Discrete Choice Modeling

Processing the Data

Questionnaires are generated and data collected using a minor modification of the methods discussed in earlier
examples. The difference is instead of asking for first choice data, allocation data are collected instead. Each
row of the input data set contains a block, subject, and set number, followed by the number of times each of the
ten alternatives was chosen. If all of the choice frequencies are zero, then the constant alternative was chosen.
The if statement is used to check data entry. For convenience, choice set number is recoded to run from 1 to 27
instead of consisting of three blocks of nine sets. This gives us one fewer variable on which to stratify.
data results;
input Block Subject Set @9 (freq1-freq&nalts) (2.);
if not (sum(of freq:) in (0, &nalts)) then put _all_;
set = (block - 1) * 9 + set;
datalines;
1 1 1 0 0 8 0 2 0 0 0 0 0
1 1 2 0 0 8 0 0 0 2 0 0 0
1 1 3 0 0 0 0 0 0 0 010 0
1 1 4 1 0 0 1 3 3 0 0 2 0
1 1 5 2 0 8 0 0 0 0 0 0 0
1 1 6 0 1 3 1 0 0 0 0 1 4
1 1 7 0 1 3 1 1 2 0 0 2 0
1 1 8 0 0 3 0 0 2 1 0 0 4
1 1 9 0 2 5 0 0 0 0 0 3 0
2 2 1 1 1 0 2 0 3 0 1 1 1
2 2 2 1 0 3 1 0 1 1 0 2 1
.
.
.
;
In the first step, in creating an analysis data set for an allocation study, we reformat the data from one row per
choice set per block per subject (9  3  100 = 2700 observations) to one per alternative (including the constant)
per choice set per block per subject ((10 + 1)  9  3  100 = 29700 observations). For each choice set, 11
observations are written storing the choice frequency in the variable Count and the brand in the variable Brand.
If no alternative is chosen, then the constant alternative is chosen ten times, otherwise it is chosen zero times.
data allocs(keep=block set brand count);
set results;

array freq[&nalts];

* Handle the &nalts alternatives;

do b = 1 to &nalts;
Brand = ’Brand ’ || put(b, 2.);
Count = freq[b];
output;
end;
* Constant alt choice is implied if nothing else is chosen.
brand = ’ ’ is used to flag the constant alternative.;

brand = ’ ’;
count = 10 * (sum(of freq:) = 0);
output;
run;
proc print data=results(obs=3) label noobs; run;
proc print data=allocs(obs=33); id block set; by block set; run;
The PROC PRINT steps show how the first three observations of the RESULTS data set are transposed into the
first 33 observations of the ALLOCS data set.
Allocation of Prescription Drugs 243

Allocation of Prescription Drugs

Block Subject Set Freq1 Freq2 Freq3 Freq4 Freq5 Freq6 Freq7 Freq8 Freq9 Freq10

1 1 1 0 0 8 0 2 0 0 0 0 0
1 1 2 0 0 8 0 0 0 2 0 0 0
1 1 3 0 0 0 0 0 0 0 0 10 0

Allocation of Prescription Drugs

Block Set Brand Count

1 1 Brand 1 0
Brand 2 0
Brand 3 8
Brand 4 0
Brand 5 2
Brand 6 0
Brand 7 0
Brand 8 0
Brand 9 0
Brand 10 0
0

1 2 Brand 1 0
Brand 2 0
Brand 3 8
Brand 4 0
Brand 5 0
Brand 6 0
Brand 7 2
Brand 8 0
Brand 9 0
Brand 10 0
0

1 3 Brand 1 0
Brand 2 0
Brand 3 0
Brand 4 0
Brand 5 0
Brand 6 0
Brand 7 0
Brand 8 0
Brand 9 10
Brand 10 0
0

The next step aggregates the data. It stores in the variable Count the number of times each alternative of each
choice set was chosen. This creates a data set with 297 observations (3 blocks  9 sets  11 alternatives = 297).
* Aggregate, store the results back in count.;

proc summary data=allocs nway missing;

class set brand;
output sum(count)=Count out=allocs(drop=_type_ _freq_);
run;
244 TS-677E Multinomial Logit, Discrete Choice Modeling

These next steps prepare the design for analysis. We need to create a data set KEY that describes how the factors
in our design will be used for analysis. It will contain all of the factor names, Brand1, Brand2, ... Brand10.
We can run the %MktKey macro to get these names in the SAS log for cutting and pasting into the program
without typing them.
%mktkey(Brand1-Brand10)

The %MktKey macro produced the following line.

Brand1 Brand2 Brand3 Brand4 Brand5 Brand6 Brand7 Brand8 Brand9 Brand10
The next step rolls out the experimental design data set to match the choice allocations data set. The data set
is transposed from one row per choice set to one row per alternative per choice set. This data set also has 297
observations. As we saw in many previous examples, the %MktRoll macro can be used to process the design.
data key(keep=Brand Price);
input Brand $ 1-8 Price $;
datalines;
Brand 1 Brand1
Brand 2 Brand2
Brand 3 Brand3
Brand 4 Brand4
Brand 5 Brand5
Brand 6 Brand6
Brand 7 Brand7
Brand 8 Brand8
Brand 9 Brand9
Brand 10 Brand10
. .
;
%mktroll(design=[Link], key=key, alt=brand, out=rolled)

proc print data=rolled(obs=11); format price dollar4.; run;

Allocation of Prescription Drugs

Obs Set Brand Price

1 1 Brand 1 $100
2 1 Brand 2 $75
3 1 Brand 3 $100
4 1 Brand 4 $100
5 1 Brand 5 $75
6 1 Brand 6 $100
7 1 Brand 7 $50
8 1 Brand 8 $50
9 1 Brand 9 $75
10 1 Brand 10 $100
11 1 .

Both data sets must be sorted the same way before they can be merged. The constant alternative, indicated by a
missing brand, is last in the design choice set and hence is out of order. Missing must come before nonmissing
for the merge. The order is correct in the ALLOCS data set since it was created by PROC SUMMARY with
Brand as a class variable.
proc sort data=rolled; by set brand; run;
Allocation of Prescription Drugs 245

The data are merged along with error checking to ensure that the merge proceeded properly. Both data sets should
have the same observations and Set and Brand variables, so the merge should be one to one.
data allocs2;
merge allocs(in=flag1) rolled(in=flag2);
by set brand;
if flag1 ne flag2 then put ’ERROR: Merge is not 1 to 1.’;
format price dollar4.;
run;
proc print data=allocs2(obs=22);
var brand price count;
sum count;
by notsorted set;
id set;
run;
In the aggregate and combined data set, we see how often each alternative was chosen for each choice set. For
example, in the first choice set, the constant alternative was chosen zero times, Brand 1 at $100 was chosen 103
times, and so on. The 11 alternatives were chosen a total of 1000 times, 100 subjects times 10 choices each.

Allocation of Prescription Drugs

Set Brand Price Count

1 . 0
Brand 1 $100 103
Brand 2 $75 58
Brand 3 $100 318
Brand 4 $100 99
Brand 5 $75 54
Brand 6 $100 83
Brand 7 $50 71
Brand 8 $50 58
Brand 9 $75 100
Brand 10 $100 56
--- -----
1 1000

2 . 10
Brand 1 $50 73
Brand 2 $100 76
Brand 3 $75 342
Brand 4 $100 55
Brand 5 $100 50
Brand 6 $100 77
Brand 7 $100 95
Brand 8 $75 71
Brand 9 $75 72
Brand 10 $75 79
--- -----
2 1000

At this point, the data set contains 297 observations (27 choice sets times 11 alternatives) showing the number
of times each alternative was chosen. This data set must be augmented to also include the number of times each
alternative was not chosen. For example, in the first choice set, brand 1 was chosen 103 times, which means it
was not chosen 0 + 58 + 318 + 99 + 54 + 83 + 71 + 58 + 100 + 56 = 897 times. We use a macro, %MktAllo
for “marketing allocation study” to process the data. We specify the input data=allocs2 data set, the output
out=allocs3 data set, the number of alternatives including the constant (nalts=%eval(&nalts + 1)),
246 TS-677E Multinomial Logit, Discrete Choice Modeling

the variables in the data set except the frequency variable (vars=set brand price), and the frequency
variable (freq=Count). The macro counts how many times each alternative was chosen and not chosen and
writes the results to the out= data set along with the usual c = 1 for chosen and c = 2 for unchosen.
%mktallo(data=allocs2, out=allocs3, nalts=%eval(&nalts + 1),
vars=set brand price, freq=Count)

proc print data=allocs3(obs=22);

var set brand price count c;
run;
The first 22 records of the allocation data set are shown next.

Allocation of Prescription Drugs

Obs Set Brand Price Count c

1 1 . 0 1
2 1 . 1000 2
3 1 Brand 1 $100 103 1
4 1 Brand 1 $100 897 2
5 1 Brand 2 $75 58 1
6 1 Brand 2 $75 942 2
7 1 Brand 3 $100 318 1
8 1 Brand 3 $100 682 2
9 1 Brand 4 $100 99 1
10 1 Brand 4 $100 901 2
11 1 Brand 5 $75 54 1
12 1 Brand 5 $75 946 2
13 1 Brand 6 $100 83 1
14 1 Brand 6 $100 917 2
15 1 Brand 7 $50 71 1
16 1 Brand 7 $50 929 2
17 1 Brand 8 $50 58 1
18 1 Brand 8 $50 942 2
19 1 Brand 9 $75 100 1
20 1 Brand 9 $75 900 2
21 1 Brand 10 $100 56 1
22 1 Brand 10 $50 944 2

In the first choice set, the constant alternative is chosen zero times and not chosen 1000 times, Brand 1 is chosen
103 times and not chosen 1000 103 = 897 times, Brand 2 is chosen 58 times and not chosen 1000 58 =
942 times, and so on. Note that allocation studies do not always have fixed sums, so it is important to use
the %MktAllo macro or some other approach that actually counts the number of times each alternative was
unchosen. It is not always sufficient to simply subtract from a fixed constant (in this case 1000).

Coding and Analysis

The next step codes the design for analysis. Dummy variables are created for Brand and Price. All of the
PROC TRANSREG options have been discussed in other examples.
proc transreg design data=allocs3 nozeroconstant norestoremissing;
model class(brand price / zero=none) / lprefix=0;
output out=coded(drop=_type_ _name_ intercept);
id set c count;
run;
Allocation of Prescription Drugs 247

Analysis proceeds like it has in all other examples. We stratify by choice set number. We do not need to stratify
by Block since choice set number does not repeat within block.
proc phreg data=coded;
where count > 0;
model c*c(2) = &_trgind / ties=breslow;
freq count;
strata set;
run;
We used the where statement to exclude observations with zero frequency; otherwise PROC PHREG complains
about them.

Multinomial Logit Model Results

Here are the results. Recall that we used %phchoice(on)on page 79 to customize the output from PROC
PHREG.

Allocation of Prescription Drugs

The PHREG Procedure

Model Information

Data Set [Link]

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Frequency Variable Count
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Set Alternatives Alternatives Chosen

1 1 11000 1000 10000

2 2 11000 1000 10000
3 3 11000 1000 10000
4 4 11000 1000 10000
5 5 11000 1000 10000
6 6 11000 1000 10000
7 7 11000 1000 10000
8 8 11000 1000 10000
9 9 11000 1000 10000
10 10 11000 1000 10000
11 11 11000 1000 10000
12 12 11000 1000 10000
13 13 11000 1000 10000
14 14 11000 1000 10000
15 15 11000 1000 10000
16 16 11000 1000 10000
17 17 11000 1000 10000
18 18 11000 1000 10000
248 TS-677E Multinomial Logit, Discrete Choice Modeling

19 19 11000 1000 10000

20 20 11000 1000 10000
21 21 11000 1000 10000
22 22 11000 1000 10000
23 23 11000 1000 10000
24 24 11000 1000 10000
25 25 11000 1000 10000
26 26 11000 1000 10000
27 27 11000 1000 10000
---------------------------------------------------------------
Total 297000 27000 270000

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Allocation of Prescription Drugs

The PHREG Procedure

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 502505.13 489061.18

AIC 502505.13 489085.18
SBC 502505.13 489183.62

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 13443.9511 12 <.0001

Score 18342.3475 12 <.0001
Wald 14088.6926 12 <.0001

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 2.09201 0.06766 955.9229 <.0001

Brand 2 1 2.08414 0.06769 947.9356 <.0001
Brand 3 1 3.53501 0.06484 2972.3245 <.0001
Brand 4 1 2.09005 0.06767 953.9288 <.0001
Brand 5 1 2.07819 0.06771 941.9267 <.0001
Brand 6 1 2.02826 0.06790 892.2654 <.0001
Brand 7 1 2.06215 0.06777 925.8259 <.0001
Brand 8 1 2.07868 0.06771 942.4280 <.0001
Brand 9 1 2.11000 0.06760 974.2866 <.0001
Brand 10 1 2.05658 0.06779 920.2659 <.0001
$50 1 0.02024 0.01627 1.5466 0.2136
$75 1 0.00665 0.01632 0.1660 0.6837
$100 0 0 . . .
Allocation of Prescription Drugs 249

The output shows that there are 27 strata, one per choice set, each consisting of 1000 chosen alternatives (10
choices by 100 subjects) and 10,000 unchosen alternatives. All of the brand coefficients are “significant,” with
the Brand 3 effect being by far the strongest. (We will soon see that statistical significance should be ignored
with allocation studies.) There is no price effect.

Analyzing Proportions
Recall that we collected data by asking physicians to report which brands they would prescribe the next ten times
they write prescriptions. Alternatively, we could ask them to report the proportion of time they would prescribe
each brand. We can simulate having proportion data by dividing our count data by 10. This means our frequency
variable will no longer contain integers, so we need to specify the notruncate option on PROC PHREG freq
statement to allow noninteger “frequencies.”
data coded2;
set coded;
count = count / 10;
run;

proc phreg data=coded2;

where count > 0;
model c*c(2) = &_trgind / ties=breslow;
freq count / notruncate;
strata set;
run;
When we do this, we see the number of alternatives and the number chosen and not chosen decrease by a factor of
10 as do all of the Chi-Square tests. The coefficients are unchanged. This implies that market share calculations
are invariant to the different scalings of the frequencies. However, the p-values are not invariant. The sample
size is artificially inflated when counts are used so p-values are not interpretable in an allocation study. When
proportions are used, each subject is contributing 1 to the number chosen instead of 10, just like a normal choice
study, so p-values have meaning.

Allocation of Prescription Drugs

The PHREG Procedure

Model Information

Data Set WORK.CODED2

Dependent Variable c
Censoring Variable c
Censoring Value(s) 2
Frequency Variable Count
Ties Handling BRESLOW

Summary of Subjects, Sets, and Chosen and Unchosen Alternatives

Number of Chosen Not

Stratum Set Alternatives Alternatives Chosen

1 1 1100.0 100.0 1000.0

2 2 1100.0 100.0 1000.0
3 3 1100.0 100.0 1000.0
4 4 1100.0 100.0 1000.0
5 5 1100.0 100.0 1000.0
250 TS-677E Multinomial Logit, Discrete Choice Modeling

6 6 1100.0 100.0 1000.0

7 7 1100.0 100.0 1000.0
8 8 1100.0 100.0 1000.0
9 9 1100.0 100.0 1000.0
10 10 1100.0 100.0 1000.0
11 11 1100.0 100.0 1000.0
12 12 1100.0 100.0 1000.0
13 13 1100.0 100.0 1000.0
14 14 1100.0 100.0 1000.0
15 15 1100.0 100.0 1000.0
16 16 1100.0 100.0 1000.0
17 17 1100.0 100.0 1000.0
18 18 1100.0 100.0 1000.0
19 19 1100.0 100.0 1000.0
20 20 1100.0 100.0 1000.0
21 21 1100.0 100.0 1000.0
22 22 1100.0 100.0 1000.0
23 23 1100.0 100.0 1000.0
24 24 1100.0 100.0 1000.0
25 25 1100.0 100.0 1000.0
26 26 1100.0 100.0 1000.0
27 27 1100.0 100.0 1000.0
---------------------------------------------------------------
Total 29700.0 2700.0 27000.0

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Allocation of Prescription Drugs

The PHREG Procedure

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 37816.553 36472.158

AIC 37816.553 36496.158
SBC 37816.553 36566.970

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 1344.3951 12 <.0001

Score 1834.2348 12 <.0001
Wald 1408.8693 12 <.0001
Allocation of Prescription Drugs 251

Multinomial Logit Parameter Estimates

Parameter Standard
DF Estimate Error Chi-Square Pr > ChiSq

Brand 1 1 2.09201 0.21397 95.5923 <.0001

Brand 2 1 2.08414 0.21406 94.7936 <.0001
Brand 3 1 3.53501 0.20504 297.2324 <.0001
Brand 4 1 2.09005 0.21399 95.3929 <.0001
Brand 5 1 2.07819 0.21413 94.1927 <.0001
Brand 6 1 2.02826 0.21472 89.2265 <.0001
Brand 7 1 2.06215 0.21432 92.5826 <.0001
Brand 8 1 2.07868 0.21412 94.2428 <.0001
Brand 9 1 2.11000 0.21377 97.4287 <.0001
Brand 10 1 2.05658 0.21438 92.0266 <.0001
$50 1 0.02024 0.05146 0.1547 0.6941
$75 1 0.00665 0.05160 0.0166 0.8975
$100 0 0 . . .
252 TS-677E Multinomial Logit, Discrete Choice Modeling

Chair Design with Generic Attributes

This study illustrates creating an experimental design for a purely generic choice model. This example discusses
generic attributes, alternative swapping, choice set swapping, and constant alternatives. In a purely generic study,
there are no brands, just bundles of attributes. Say a manufacturer is interested in designing one or more new
chairs. The manufacturer can vary the attributes of the chairs, present subjects with competing chair designs, and
model the effects of the attributes on choice. Here are the attributes of interest.
Factor Attribute Levels
X1 Color 3 Colors
X2 Back 3 Styles
X3 Seat 3 Styles
X4 Arm Rest 3 Styles
X5 Material 3 Materials
Since seeing descriptions of chairs is not the same as seeing and sitting in the actual chairs, the manufacturer is
going to actually make sample chairs for people to try and choose from. Subjects will be shown groups of three
chairs at a time. If we were to make our design using the approach discussed in previous examples, we would use
the %MktEx autocall macro to create a design with 15 factors, five for the first chair, five for the second chair,
and five for the third chair. This design would have to have at least 15  (3 1) + 1 = 31 runs and 93 sample
chairs. Here is how we could have made the design.
title ’Generic Chair Attributes’;

* This design will not be used;

%mktex(3 ** 15, n=36, seed=238)

data key;
input (x1-x5) ($) @@;
datalines;
x1 x2 x3 x4 x5
x6 x7 x8 x9 x10
x11 x12 x13 x14 x15
;

%mktroll(design=randomized, key=key, out=cand);

The %MktEx approach to designing an experiment like this allows you to fit very general models including
models with alternative-specific effects and even mother logit models. However, at analysis time for this purely
generic model, we will fit a model with 10 parameters, two for each of the five factors, class(x1-x5).
Creating a design with over 31  3 = 93 chairs is way too expensive. In ordinary linear designs, we need at least
as many runs as parameters. In choice designs, we need to count the total number of alternatives across all choice
sets, subtract the number the number of choice sets, and this number must be at least as large as the number of
parameters. Equivalently, each choice set allows us to estimate m 1 parameters, where m is the number of
alternatives in that choice set. In this case, we could fit our purely generic model with as few as 10=(3 1) = 5
choice sets.
Since we only need a simple generic model model for this example, and since our chair manufacturing for our
research will be expensive, we will not use the %MktEx approach for designing our choice experiment. Instead,
we will use a different approach that will allow us to get a smaller design that is adequate for our model and
budget. Recall the discussion of linear design efficiency, choice model design efficiency, and using linear design
efficiency as a surrogate for choice design goodness from the ’Preliminaries’ section starting on page 76. Instead
of using linear design efficiency as a surrogate for choice design goodness, we can directly optimize choice
design efficiency given an assumed model and parameter vector . This approach uses the %ChoicEff macro.
Chair Design with Generic Attributes 253

Generic Attributes, Alternative Swapping, Large Candidate Set

This part of the example illustrates using the %ChoicEff macro for efficient choice designs, using its algorithm
that builds a design from candidate alternatives (as opposed to candidates consisting of entire choice sets). First,
we will use the %MktRuns macro to suggest a candidate-set size.
%mktruns(3 ** 5)

Here are some of the results.

Generic Chair Attributes

Design Summary

Number of
Levels Frequency

3 5

Generic Chair Attributes

Saturated = 11
Full Factorial = 243

Some Reasonable Cannot Be

Design Sizes Violations Divided By

18 * 0
27 * 0
36 * 0
12 10 9
15 10 9
21 10 9
24 10 9
30 10 9
33 10 9
11 15 3 9

* - 100% Efficient Design can be made with the MktEx Macro.

Generic Chair Attributes

n Design Reference

18 2 ** 1 3 ** 7 Taguchi, 1987
18 3 ** 6 6 ** 1 Taguchi, 1987
27 3 ** 13 Fractional-factorial
27 3 ** 9 9 ** 1 Fractional-factorial
36 2 ** 11 3 ** 12 Taguchi, 1987
36 2 ** 4 3 ** 13 Taguchi, 1987
36 2 ** 2 3 ** 12 6 ** 1 Wang and Wu, 1991
36 2 ** 1 3 ** 8 6 ** 2 Zhang, Lu, and Pang, 1999
36 3 ** 13 4 ** 1 Dey, 1985
36 3 ** 12 12 ** 1 Wang and Wu, 1991
36 3 ** 7 6 ** 3 Finney, 1982
254 TS-677E Multinomial Logit, Discrete Choice Modeling

We could use candidate sets of size: 18, 27 or 36. Additionally, since this problem is small, we could try an
81-run fractional-factorial design or the 243-run full-factorial design. We will choose the 243-run full-factorial
design, since it is reasonably small and it should give us a good design. 
We will use the %MktEx macro to create a candidate set. The candidate set will consist of 5 three-level factors,
one for each of the five generic attributes. We will add three flag variables to the candidate set, f1-f3, one
for each alternative. Since there are three alternatives, the candidate set must contain those observations that
may be used for alternative 1, those observations that may be used for alternative 2, and those observations that
may be used for alternative 3. The flag variable for each alternative consists of ones for those candidates that
may be included for that alternative and zeros or missings for those candidates that may not be included for
that alternative. The candidates for the different alternatives may be all different, all the same, or something in
between depending on the problem. For example, the candidate set may contain one observation that is only used
for the last, constant alternative. In this purely generic case, each flag variable consists entirely of ones indicating
that any candidate can appear in any alternative. The %MktEx macro will not allow you to create constant or
one-level factors. We can instead use the %MktLab macro to add the flag variables, essentially by specifying
that we have multiple intercepts. The option int=f1-f3 creates three variables with values all one. The default
output data set is called FINAL. The following code creates the candidates.
%mktex(3 ** 5, n=243)
%mktlab(data=design, int=f1-f3)

proc print data=final(obs=27); run;

The columns f1-f3 are the flags, and x1-x5 are the generic attributes. Here is part of the candidate set.

Generic Chair Attributes

Obs f1 f2 f3 x1 x2 x3 x4 x5

1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 2
3 1 1 1 1 1 1 1 3
4 1 1 1 1 1 1 2 1
5 1 1 1 1 1 1 2 2
6 1 1 1 1 1 1 2 3
7 1 1 1 1 1 1 3 1
8 1 1 1 1 1 1 3 2
9 1 1 1 1 1 1 3 3
10 1 1 1 1 1 2 1 1
11 1 1 1 1 1 2 1 2
12 1 1 1 1 1 2 1 3
13 1 1 1 1 1 2 2 1
14 1 1 1 1 1 2 2 2
15 1 1 1 1 1 2 2 3
16 1 1 1 1 1 2 3 1
17 1 1 1 1 1 2 3 2
18 1 1 1 1 1 2 3 3
19 1 1 1 1 1 3 1 1
20 1 1 1 1 1 3 1 2
21 1 1 1 1 1 3 1 3
22 1 1 1 1 1 3 2 1
23 1 1 1 1 1 3 2 2
24 1 1 1 1 1 3 2 3
25 1 1 1 1 1 3 3 1
26 1 1 1 1 1 3 3 2
27 1 1 1 1 1 3 3 3

 Later, we will see we could have chosen 18.

Chair Design with Generic Attributes 255

Next, we will search that candidate set for an efficient design for the model specification class(x1-x5) and
the assumption = 0. We will use the %ChoicEff autocall macro to do this. (All of the autocall macros used
in this report are documented starting on page 287.) This approach is based on the work of Huber and Zwerina
(1996) who proposed constructing efficient experimental designs for choice experiments under an assumed model
and . The %ChoicEff macro uses a modified Federov algorithm (Federov, 1972; Cook and Nachtsheim, 1980)
to optimize the choice model variance matrix. We will be using the largest possible candidate set for this problem,
the full-factorial design, and we will ask for more than the default number of iterations, so run time will be slower
than it could be. However, we will be requesting a very small number of choice sets. Building the chairs will be
expensive, so we want to get a really good but small design. This specification requests a generic design with six
choice sets each consisting of three alternatives.
%choiceff(data=final, model=class(x1-x5), nsets=6, maxiter=100,
seed=121, flags=f1-f3, beta=zero);
The data=final option names the input data set of candidates. The model=class(x1-x5) option speci-
fies the most general model that will be considered at analysis time. The nsets=6 option specifies the number
of choice sets. Note that this is considerably smaller than the minimum of 31 that would be required if we
were just using the %MktEx linear-design approach (6  3 = 18 chairs instead of 31  3 = 93 chairs). The
maxiter=100 option requests 100 designs based on 100 random initial designs (by default, maxiter=2).
The seed=121 option specifies the random number seed. The flags=f1-f3 specifies the flag variables for
alternatives 1 to 3. Implicitly, this option also specifies the fact that there are three alternatives since three flag
variables were specified. The beta=zero option specifies the assumption = 0. A vector of numbers like
beta=-1 0 -1 0 -1 0 -1 0 -1 0 -1 0 could be specified. When you wish to assume all parameters
are zero, you can specify beta=zero instead of typing a vector of the zeros. You can also omit the beta=
option if you just want the macro to list the parameters. You can use this list to ensure that you specify the
parameters in the right order.
The first part of the output from the macro is a list of all of the effects generated and the assumed values of . It
is very important to check this list and make sure it is correct. In particular, when you are explicitly specifying
the vector, you need to make sure you specified all of the values in the right order.

Generic Chair Attributes

n Name Beta Label

1 x11 0 x1 1
2 x12 0 x1 2
3 x21 0 x2 1
4 x22 0 x2 2
5 x31 0 x3 1
6 x32 0 x3 2
7 x41 0 x4 1
8 x42 0 x4 2
9 x51 0 x5 1
10 x52 0 x5 2
256 TS-677E Multinomial Logit, Discrete Choice Modeling

Next, the macro produces the iteration history, which is different from the iteration histories we are used to
seeing in the %MktEx macro. The %ChoicEff macro uses PROC IML and a modified Federov algorithm to
iteratively improve the efficiency of the choice design given the specified candidates, model, and . Note that
these efficiencies are not on a 0 to 100 scale. This step took about 12 minutes. Here are some of the results.

Generic Chair Attributes

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0.352304 2.838455
1 0.946001 1.057081
2 1.001164 0.998838
3 1.041130 0.960494

Design Iteration D-Efficiency D-Error

----------------------------------------------
2 0 0.792568 1.261721
1 1.513406 0.660761
2 1.732051 0.577350
3 1.732051 0.577350
.
.
.

Design Iteration D-Efficiency D-Error

----------------------------------------------
34 0 0.469771 2.128698
1 0.919074 1.088051
2 1.058235 0.944970
3 1.154701 0.866025
4 1.154701 0.866025
.
.
.

Design Iteration D-Efficiency D-Error

----------------------------------------------
100 0 0.456308 2.191501
1 1.006320 0.993719
2 1.042702 0.959046
3 1.042702 0.959046

Next, the macro shows which design it chose and the final efficiency and D-Error (D-Efficiency = 1 / D-Error).

Final Results: Design = 34

Efficiency = 1.1547005384
D-Error = 0.8660254038
Chair Design with Generic Attributes 257

Next, it shows the variance, standard error, and df for each effect. It is important to ensure that each effect is
estimable: (df = 1). Usually, when all of the variances are constant, like we see in this table, it means that the
macro has found the optimal design.

Generic Chair Attributes

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 1 1 1
2 x12 x1 2 1 1 1
3 x21 x2 1 1 1 1
4 x22 x2 2 1 1 1
5 x31 x3 1 1 1 1
6 x32 x3 2 1 1 1
7 x41 x4 1 1 1 1
8 x42 x4 2 1 1 1
9 x51 x5 1 1 1 1
10 x52 x5 2 1 1 1
==
10

The data set BEST contains the final, best design found.
proc print; by set; id set; run;

The data set contains: Design - the number of the design with the maximum efficiency, Efficiency - the
efficiency of this design, Index - the candidate set observation number, Set - the choice set number, Prob -
the probability that this alternative will be chosen given , n - the observation number, x1-x5 - the design, and
f1-f3 - the flags.

Generic Chair Attributes

Set Design Efficiency Index Prob n f1 f2 f3 x1 x2 x3 x4 x5

1 34 1.15470 183 0.33333 595 1 1 1 3 1 3 1 3

34 1.15470 62 0.33333 596 1 1 1 1 3 1 3 2
34 1.15470 121 0.33333 597 1 1 1 2 2 2 2 1
2 34 1.15470 217 0.33333 598 1 1 1 3 3 1 1 1
34 1.15470 45 0.33333 599 1 1 1 1 2 2 3 3
34 1.15470 104 0.33333 600 1 1 1 2 1 3 2 2
3 34 1.15470 215 0.33333 601 1 1 1 3 2 3 3 2
34 1.15470 147 0.33333 602 1 1 1 2 3 2 1 3
34 1.15470 4 0.33333 603 1 1 1 1 1 1 2 1
4 34 1.15470 78 0.33333 604 1 1 1 1 3 3 2 3
34 1.15470 178 0.33333 605 1 1 1 3 1 2 3 1
34 1.15470 110 0.33333 606 1 1 1 2 2 1 1 2
5 34 1.15470 90 0.33333 607 1 1 1 2 1 1 3 3
34 1.15470 46 0.33333 608 1 1 1 1 2 3 1 1
34 1.15470 230 0.33333 609 1 1 1 3 3 2 2 2
6 34 1.15470 195 0.33333 610 1 1 1 3 2 1 2 3
34 1.15470 11 0.33333 611 1 1 1 1 1 2 1 2
34 1.15470 160 0.33333 612 1 1 1 2 3 3 3 1

This design has 18 runs (6 choice sets  3 alternatives). Notice that in this design, each level occurs exactly
once in each factor and each choice set. To use this design for analysis, you would only need the variables Set
and x1-x5. Since it is already in choice design format, it would not need to be processed using the %MktRoll
258 TS-677E Multinomial Logit, Discrete Choice Modeling

macro. Since data collection, processing, and analysis have already been covered in detail in other examples, this
example will concentrate solely on experimental design.

Generic Attributes, Alternative Swapping, Small Candidate Set

In this part of this example, we will try to make an equivalent design to the one we just made, only this time
using a smaller candidate set. Here is the code.
%mktex(3 ** 5, n=18)

%mktlab(data=design, int=f1-f3)

%choiceff(data=final, model=class(x1-x5), nsets=6, maxiter=20,

seed=121, flags=f1-f3, beta=zero);

proc print; run;

This time, instead of creating a full-factorial candidate set, we asked for 5 three-level factors from the L18 , an
orthogonal tabled design in 18 runs. We also asked for fewer iterations in the %ChoicEff macro. Since the
candidate set is much smaller, the macro should be able to find the best design available in this candidate set
fairly easily. Here are some of the results.

Generic Chair Attributes

n Name Beta Label

1 x11 0 x1 1
2 x12 0 x1 2
3 x21 0 x2 1
4 x22 0 x2 2
5 x31 0 x3 1
6 x32 0 x3 2
7 x41 0 x4 1
8 x42 0 x4 2
9 x51 0 x5 1
10 x52 0 x5 2

Generic Chair Attributes

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0 .
1 0.913290 1.094943
2 1.008888 0.991191
3 1.042878 0.958885
4 1.154701 0.866025
5 1.154701 0.866025
.
.
.
Chair Design with Generic Attributes 259

Design Iteration D-Efficiency D-Error

----------------------------------------------
20 0 0.364703 2.741954
1 0.851038 1.175036
2 1.008888 0.991191
3 1.042878 0.958885
4 1.154701 0.866025
5 1.154701 0.866025

Generic Chair Attributes

Final Results: Design = 1

Efficiency = 1.1547005384
D-Error = 0.8660254038

Generic Chair Attributes

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 1 1 1
2 x12 x1 2 1 1 1
3 x21 x2 1 1 1 1
4 x22 x2 2 1 1 1
5 x31 x3 1 1 1 1
6 x32 x3 2 1 1 1
7 x41 x4 1 1 1 1
8 x42 x4 2 1 1 1
9 x51 x5 1 1 1 1
10 x52 x5 2 1 1 1
==
10

Generic Chair Attributes

Obs Design Efficiency Index Set Prob n f1 f2 f3 x1 x2 x3 x4 x5

1 1 1.15470 11 1 0.33333 1 1 1 1 2 3 1 3 1
2 1 1.15470 13 1 0.33333 2 1 1 1 3 1 2 1 2
3 1 1.15470 4 1 0.33333 3 1 1 1 1 2 3 2 3
4 1 1.15470 3 2 0.33333 4 1 1 1 1 2 1 3 2
5 1 1.15470 12 2 0.33333 5 1 1 1 2 3 2 1 3
6 1 1.15470 14 2 0.33333 6 1 1 1 3 1 3 2 1
7 1 1.15470 5 3 0.33333 7 1 1 1 1 3 2 2 1
8 1 1.15470 8 3 0.33333 8 1 1 1 2 1 3 3 2
9 1 1.15470 15 3 0.33333 9 1 1 1 3 2 1 1 3
10 1 1.15470 9 4 0.33333 10 1 1 1 2 2 2 2 2
11 1 1.15470 1 4 0.33333 11 1 1 1 1 1 1 1 1
12 1 1.15470 18 4 0.33333 12 1 1 1 3 3 3 3 3
13 1 1.15470 10 5 0.33333 13 1 1 1 2 2 3 1 1
14 1 1.15470 17 5 0.33333 14 1 1 1 3 3 1 2 2
15 1 1.15470 2 5 0.33333 15 1 1 1 1 1 2 3 3
16 1 1.15470 6 6 0.33333 16 1 1 1 1 3 3 1 2
17 1 1.15470 7 6 0.33333 17 1 1 1 2 1 1 2 3
18 1 1.15470 16 6 0.33333 18 1 1 1 3 2 2 3 1
260 TS-677E Multinomial Logit, Discrete Choice Modeling

Notice we got the same D-efficiency and variances as before (D-efficiency = 1.1547005384 and all variances 1).
Also notice the Index variable in the design (which is the candidate set row number). Each candidate appears
in the design exactly once. We have frequently found for problems like this (all generic attributes, no brands, no
constant alternative, total number of alternatives equal to the number of runs in an orthogonal design, all factors
available in that orthogonal design, and an assumed vector of zero) that the optimal design can be created by
optimally sorting the rows of an orthogonal design into choice sets, and the %ChoicEff macro can do this quite
well.
Six choice sets is a bit small. If you can afford a larger number, it would be good to try a larger design. In this
case, nine choice sets are requested using a fractional-factorial candidate set in 27 runs. Notice that like before,
the number of runs in the candidate set was chosen to be the product of the number of choice sets and the number
of alternatives in each choice set.
%mktex(3 ** 5, n=27)

%mktlab(data=design, int=f1-f3)

%choiceff(data=final, model=class(x1-x5), nsets=9, maxiter=20,

seed=121, flags=f1-f3, beta=zero);

proc print; id set; by set; var index prob x:; run;

Here are the variances and the design.

Generic Chair Attributes

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 0.66667 1 0.81650

2 x12 x1 2 0.66667 1 0.81650
3 x21 x2 1 0.66667 1 0.81650
4 x22 x2 2 0.66667 1 0.81650
5 x31 x3 1 0.66667 1 0.81650
6 x32 x3 2 0.66667 1 0.81650
7 x41 x4 1 0.66667 1 0.81650
8 x42 x4 2 0.66667 1 0.81650
9 x51 x5 1 0.66667 1 0.81650
10 x52 x5 2 0.66667 1 0.81650
==
10

Generic Chair Attributes

Set Index Prob x1 x2 x3 x4 x5

1 25 0.33333 3 3 1 1 2
2 0.33333 1 1 2 3 3
15 0.33333 2 2 3 2 1

2 10 0.33333 2 1 1 2 1
23 0.33333 3 2 2 1 2
9 0.33333 1 3 3 3 3

3 24 0.33333 3 2 3 3 1
11 0.33333 2 1 2 1 3
7 0.33333 1 3 1 2 2
Chair Design with Generic Attributes 261

4 13 0.33333 2 2 1 1 3
3 0.33333 1 1 3 2 2
26 0.33333 3 3 2 3 1

5 20 0.33333 3 1 2 2 3
6 0.33333 1 2 3 1 1
16 0.33333 2 3 1 3 2

6 8 0.33333 1 3 2 1 1
22 0.33333 3 2 1 2 3
12 0.33333 2 1 3 3 2

7 5 0.33333 1 2 2 2 2
18 0.33333 2 3 3 1 3
19 0.33333 3 1 1 3 1

8 1 0.33333 1 1 1 1 1
14 0.33333 2 2 2 3 2
27 0.33333 3 3 3 2 3

9 17 0.33333 2 3 2 2 1
4 0.33333 1 2 1 3 3
21 0.33333 3 1 3 1 2

Notice that like before, the variances are constant, 2/3, and each candidate appears once. This is an optimal
design in 9 choice sets.

Generic Attributes, a Constant Alternative, and Alternative Swapping

Now let’s make a design for the same problem but this time with a constant alternative. We will first use the
%MktEx macro just like before to make a design for the nonconstant alternatives. We will then use a DATA step
to add the flags and a constant alternative.
title ’Generic Chair Attributes’;

%mktex(3 ** 5, n=243)

data final(drop=i);
set design end=eof;
retain f1-f3 1 f4 0;
output;
if eof then do;
array x[9] x1-x5 f1-f4;
do i = 1 to 9; x[i] = i le 5 or i eq 9; end;
output;
end;
run;

proc print data=final(where=(x1 eq x3 and x2 eq x4 and x3 eq x5 or f4)); run;

262 TS-677E Multinomial Logit, Discrete Choice Modeling

Here is a sample of the observations in the candidate set.

Obs x1 x2 x3 x4 x5 f1 f2 f3 f4

1 1 1 1 1 1 1 1 1 0
31 1 2 1 2 1 1 1 1 0
61 1 3 1 3 1 1 1 1 0
92 2 1 2 1 2 1 1 1 0
122 2 2 2 2 2 1 1 1 0
152 2 3 2 3 2 1 1 1 0
183 3 1 3 1 3 1 1 1 0
213 3 2 3 2 3 1 1 1 0
243 3 3 3 3 3 1 1 1 0
244 1 1 1 1 1 0 0 0 1

The first 243 observations may be used for any of the first three alternatives and the 244th observation may only
be used for fourth or constant alternative. In this example, the constant alternative is composed solely from the
first level of each factor. Of course this could be changed depending on the situation. The %ChoicEff macro
invocation is the same as before, except now we have four flags.
%choiceff(data=final, model=class(x1-x5), nsets=6, maxiter=100,
seed=121, flags=f1-f4, beta=zero);

proc print; by set; id set; run;

You can see in the final design that there are now four alternatives and the last alternative in each choice set is
constant and is always flagged by f4=1. In the interest of space, most of the iteration histories are omitted.

Generic Chair Attributes

n Name Beta Label

1 x11 0 x1 1
2 x12 0 x1 2
3 x21 0 x2 1
4 x22 0 x2 2
5 x31 0 x3 1
6 x32 0 x3 2
7 x41 0 x4 1
8 x42 0 x4 2
9 x51 0 x5 1
10 x52 0 x5 2

Generic Chair Attributes

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0.424723 2.354476
1 0.900662 1.110294
2 0.939090 1.064861
3 0.943548 1.059830
.
.
.
Chair Design with Generic Attributes 263

Design Iteration D-Efficiency D-Error

----------------------------------------------
13 0 0.494007 2.024263
1 0.873818 1.144404
2 0.915135 1.092735
3 0.960392 1.041241
4 0.999769 1.000231
5 1.003398 0.996614

.
.
.

Design Iteration D-Efficiency D-Error

----------------------------------------------
100 0 0.528399 1.892509
1 0.883854 1.131408
2 0.924346 1.081846
3 0.939811 1.064044
4 0.942047 1.061518

Generic Chair Attributes

Final Results: Design = 13

Efficiency = 1.0033975924
D-Error = 0.9966139121

Generic Chair Attributes

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 1.14695 1 1.07096

2 x12 x1 2 1.33333 1 1.15470
3 x21 x2 1 1.14695 1 1.07096
4 x22 x2 2 1.33333 1 1.15470
5 x31 x3 1 1.19793 1 1.09450
6 x32 x3 2 1.27439 1 1.12889
7 x41 x4 1 1.13102 1 1.06350
8 x42 x4 2 1.27439 1 1.12889
9 x51 x5 1 1.13102 1 1.06350
10 x52 x5 2 1.27439 1 1.12889
==
10

Generic Chair Attributes

Set Design Efficiency Index Prob n x1 x2 x3 x4 x5 f1 f2 f3 f4

1 13 1.00340 152 0.25 289 2 3 2 3 2 1 1 1 0

13 1.00340 213 0.25 290 3 2 3 2 3 1 1 1 0
13 1.00340 18 0.25 291 1 1 2 3 3 1 1 1 0
13 1.00340 244 0.25 292 1 1 1 1 1 0 0 0 1

2 13 1.00340 154 0.25 293 2 3 3 1 1 1 1 1 0

13 1.00340 15 0.25 294 1 1 2 2 3 1 1 1 0
13 1.00340 197 0.25 295 3 2 1 3 2 1 1 1 0
13 1.00340 244 0.25 296 1 1 1 1 1 0 0 0 1
264 TS-677E Multinomial Logit, Discrete Choice Modeling

3 13 1.00340 108 0.25 297 2 1 3 3 3 1 1 1 0

13 1.00340 220 0.25 298 3 3 1 2 1 1 1 1 0
13 1.00340 38 0.25 299 1 2 2 1 2 1 1 1 0
13 1.00340 244 0.25 300 1 1 1 1 1 0 0 0 1

4 13 1.00340 121 0.25 301 2 2 2 2 1 1 1 1 0

13 1.00340 182 0.25 302 3 1 3 1 2 1 1 1 0
13 1.00340 63 0.25 303 1 3 1 3 3 1 1 1 0
13 1.00340 244 0.25 304 1 1 1 1 1 0 0 0 1

5 13 1.00340 111 0.25 305 2 2 1 1 3 1 1 1 0

13 1.00340 77 0.25 306 1 3 3 2 2 1 1 1 0
13 1.00340 178 0.25 307 3 1 2 3 1 1 1 1 0
13 1.00340 244 0.25 308 1 1 1 1 1 0 0 0 1

6 13 1.00340 228 0.25 309 3 3 2 1 3 1 1 1 0

13 1.00340 52 0.25 310 1 2 3 3 1 1 1 1 0
13 1.00340 86 0.25 311 2 1 1 2 2 1 1 1 0
13 1.00340 244 0.25 312 1 1 1 1 1 0 0 0 1

When there were three alternatives, each alternative had a probability of choice of 1/3, and now with four al-
ternatives, the probability is 1/4. They are all equal because of the assumption = 0. With other assumptions
about , typically the probabilities will not all be equal. To use this design for analysis, you would only need
the variables Set and x1-x5. Since it is already in choice design format (one row per alternative), it would
not need to be processed using the %MktRoll macro. Note that when you make designs with the %ChoicEff
macro, the model statement in PROC TRANSREG should match or be no more complicated than the model
specification that generated the design:
model class(x1-x5);

A model with fewer degrees of freedom is safe, although the design will be suboptimal. For example, if x1-x5
are numeric, this would be safe:
model identity(x1-x5);

However, specifying interactions, or using this design in a branded study and specifying alternative-specific
effects like this could lead to quite a few unestimable parameters.
* Bad idea for this design!!;
model class(x1-x5 x1*x2 x4*x5);

* Another bad idea for this design!!;

model class(brand)
class(brand * x1 brand * x2 brand * x3 brand * x4 brand * x5);

Generic Attributes, a Constant Alternative, and Choice Set Swapping

The %ChoicEff macro can be used in a very different way. Instead of providing a candidate set of alternatives
to swap in and out of the design, you can provide a candidate set of entire choice sets. For this particular example,
swapping alternatives will almost certainly be better (see page 267). However, sometimes, if you need to impose
restrictions on which alternative can appear with which other alternative, then you must use the set-swapping
options. We will start by using the %MktEx macro to make a candidate design, with one run per choice set
and one factor for each attribute of each alternative (just like we did in the vacation, fabric softener, and food
examples). We will then process the candidates from one row per choice set to one row per alternative per choice
set using the %MktRoll macro.
Chair Design with Generic Attributes 265

%mktex(3 ** 15, n=81 * 81, seed=522)

%mktkey(x1-x15)

data key;
input (x1-x5) ($);
datalines;
x1 x2 x3 x4 x5
x6 x7 x8 x9 x10
x11 x12 x13 x14 x15
. . . . .
;

%mktroll(design=randomized, key=key, out=rolled)

* Code the constant alternative;

data final;
set rolled;
if _alt_ = ’4’ then do; x1 = 1; x2 = 1; x3 = 1; x4 = 1; x5 = 1; end;
run;

proc print; by set; id set; where set in (1, 100, 1000, 5000, 6561); run;
The %MktKey macro produced the following line, which we copied, pasted, and edited to make the KEY data
set.
x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15
Here are a few of the candidate choice sets.

Generic Chair Attributes

Set _Alt_ x1 x2 x3 x4 x5

1 1 3 2 2 1 2
2 3 2 1 1 2
3 2 1 3 3 1
4 1 1 1 1 1

100 1 3 3 2 2 3
2 3 1 3 3 2
3 1 3 3 3 1
4 1 1 1 1 1

1000 1 3 2 2 2 2
2 3 3 3 2 1
3 1 2 3 2 1
4 1 1 1 1 1

5000 1 1 2 2 3 3
2 3 3 3 3 2
3 2 2 1 3 3
4 1 1 1 1 1

6561 1 3 3 1 3 2
2 3 2 1 2 1
3 1 1 1 1 1
4 1 1 1 1 1
266 TS-677E Multinomial Logit, Discrete Choice Modeling

Next, we will then run the %ChoicEff macro, only this time we will specify nalts=4 instead of flags=f1-
f4. Since there are no alternative flag variables to count, we have to tell the macro how many alternatives are in
each choice set. We will also ask for fewer iterations since the candidate set is large.
%choiceff(data=final, model=class(x1-x5), nsets=6, nalts=4, maxiter=10,
beta=zero, seed=109);

Generic Chair Attributes

n Name Beta Label

1 x11 0 x1 1
2 x12 0 x1 2
3 x21 0 x2 1
4 x22 0 x2 2
5 x31 0 x3 1
6 x32 0 x3 2
7 x41 0 x4 1
8 x42 0 x4 2
9 x51 0 x5 1
10 x52 0 x5 2

Generic Chair Attributes

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0.518092 1.930159
1 0.800630 1.249017
2 0.861910 1.160214
3 0.861910 1.160214
.
.
.

Design Iteration D-Efficiency D-Error

----------------------------------------------
8 0 0.523312 1.910907
1 0.850993 1.175097
2 0.878594 1.138182
3 0.878594 1.138182
.
.
.

Design Iteration D-Efficiency D-Error

----------------------------------------------
10 0 0.552471 1.810051
1 0.833732 1.199427
2 0.844183 1.184577
3 0.844183 1.184577

Generic Chair Attributes

Final Results: Design = 8

Efficiency = 0.8785943904
D-Error = 1.1381816352
Chair Design with Generic Attributes 267

Generic Chair Attributes

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 1.23879 1 1.11301

2 x12 x1 2 1.99174 1 1.41129
3 x21 x2 1 1.11908 1 1.05787
4 x22 x2 2 1.84621 1 1.35875
5 x31 x3 1 1.34469 1 1.15961
6 x32 x3 2 1.87653 1 1.36987
7 x41 x4 1 1.40455 1 1.18514
8 x42 x4 2 1.47021 1 1.21252
9 x51 x5 1 1.51281 1 1.22996
10 x52 x5 2 1.29878 1 1.13964
==
10

This design is less efficient than we found using the alternative-swapping algorithm, so we will not use it.

Design Algorithm Comparisons

It is instructive to compare the three approaches outlined in this report in the context of this problem. There are
335 = 14; 348; 907 choice sets for this problem (three-level factors and 3 alternatives times 5 factors per alter-
native). If we were to use the pure linear design approach using the %MktEx macro, we could never begin to con-
sider all possible candidate choice sets. Similarly, with the choice-set-swapping algorithm of the %ChoicEff
macro, we could never begin to consider all possible candidate choice sets. Furthermore, with the linear de-
sign approach, we could not create a design with six choice sets since the minimum size is 2  15 + 1 = 31.
Now consider the alternative-swapping algorithm. It uses at most a candidate set with only 244 observations
(35 + 1). From it, every possible choice set can potentially be constructed, although the macro will only consider
a tiny fraction of the possibilities. Hence, the alternative swapping will usually find a better design, because the
candidate set does not limit it.
Both uses of the %ChoicEff macro have the advantage that they are explicitly minimizing the variances of the
parameter estimates given a model and a vector. They can be used to produce smaller, more specialized, and
better designs. However, if the vector or model is badly misspecified, the designs could be horrible. How
badly do things have to be misspecified before you will have problems? Who knows. More research is needed.
In contrast, the linear model %MktEx approach is very conservative and safe in that it should let you specify
a very general model and still produce estimable parameters. The cost is you may be using many more choice
sets than you need, particularly for nonbranded generic attributes. If you really have some information about
your parameters, you should use them to produce a smaller and better design. However, if you have little or no
information about parameters and if you anticipate specifying very general models like mother logit, then you
probably want to use the linear design approach.
268 TS-677E Multinomial Logit, Discrete Choice Modeling

Initial Designs
This section illustrates some design strategies that involve improving on or augmenting initial designs. We will
not actually use any designs from this section.

Improving an Existing Design

Sometimes, it is useful to try to improve an existing design. In this example, we use the %MktEx macro to create
a design in 80 runs for 25 four-level factors. In the next step, we specify init=, and the macro goes straight into
the design refinement history seeking to refine the input design. You might want to do this for example whenever
you have a good, but not 100% efficient design, and you are willing to wait a few minutes to see if the macro can
make it any better.
title ’Try to Improve an Existing Design’;

%mktex(4 ** 25, n=80, seed=368)

%mktex(4 ** 25, n=80, seed=306, init=design)

Here is the D-efficiency of the final design from the first step.

Try to Improve an Existing Design

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 91.2636 83.9694 97.8111 0.9747

Here are the results from the second step.

Try to Improve an Existing Design

Design Refinement History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 91.2636 91.2636 Ini

1 Start 81.8822 Pre,Mut,Ann

1 End 89.5862

2 Start 82.3429 Pre,Mut,Ann

2 End 89.2088

NOTE: Quitting the refinement step after 7.33 minutes and 2 designs.

The macro skips the normal first steps, algorithm search and design search, and goes straight into the design
refinement search. No improvements were found, which is usually the case.
Initial Designs 269

When Some Choice Sets are Fixed in Advance

Sometimes certain runs or choice sets are fixed in advance and must be included in the design. Stated differently,
the %MktEx macro can be used to efficiently augment a starting design with other choice sets. Suppose that you
can make a choice design from the L36 (211 312 ). In addition, you want to optimally add four more choice sets to
use as holdouts. First we will look at how to do this using the fixed= option. This option can be used for fairly
general design augmentation and refinement problems. Later we will see an easier way to handle this particular
problem using the holdouts= option.
You can create the design in 36 runs as before. Next, a DATA step is used to add a flag variable f that has values
of 1 for the original 36 runs. In addition, four more runs are added (just copies of the last run) but with a flag
value of missing. When this variable is specified on the fixed=f option, it indicates that the first 36 runs of
the init=init design may not change. The remaining 4 runs are to be randomly initialized and optimally
refined to maximize the D-efficiency of the overall 40-run design. We specified options=nosort so that the
additional runs would stay at the end of the design.
title ’Augment a Design’;

%mktex(n=36, seed=292)
data init;
set randomized end = eof;
f = 1;
output;
if eof then do;
f = .;
do i = 1 to 4; output; end;
drop i;
end;
run;

proc print; run;

%mktex(2 ** 11 3 ** 12, n=40, init=init, fixed=f, seed=513, options=nosort)

proc print; run;

Here is the initial design.

Augment a Design

O x x x x x x x x x x x x x x
b x x x x x x x x x 1 1 1 1 1 1 1 1 1 1 2 2 2 2
s 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 f

1 1 1 1 2 1 1 1 2 1 2 1 3 2 1 2 3 3 1 1 1 3 1 3 1
2 1 1 2 1 2 1 2 1 1 1 1 3 1 1 3 2 3 3 2 1 1 3 2 1
3 2 1 1 2 2 2 2 2 2 1 1 2 2 1 1 2 1 2 3 1 2 3 3 1
4 1 2 1 2 1 2 2 1 1 1 2 1 3 3 3 2 3 1 1 2 2 3 3 1
5 2 2 1 1 1 1 1 1 2 1 1 2 3 2 3 3 1 1 2 1 2 1 2 1
6 2 2 2 2 2 2 1 1 1 2 1 3 3 3 1 1 3 3 3 1 2 1 1 1
7 1 2 2 2 2 1 1 2 2 1 2 1 1 3 2 3 1 1 3 1 1 3 1 1
8 2 1 2 1 1 2 1 2 1 1 2 2 1 3 3 1 1 3 1 1 3 2 3 1
9 1 1 1 1 2 2 1 1 2 2 2 2 2 2 3 3 3 3 3 2 3 3 1 1
10 1 1 1 1 2 2 1 1 2 2 2 1 3 1 1 1 2 1 2 1 1 2 3 1
270 TS-677E Multinomial Logit, Discrete Choice Modeling

11 1 1 1 2 1 1 1 2 1 2 1 2 3 3 3 1 2 2 3 3 1 3 2 1
12 2 2 1 1 1 1 1 1 2 1 1 1 1 1 1 1 3 2 1 3 3 3 1 1
13 1 1 1 2 1 1 1 2 1 2 1 1 1 2 1 2 1 3 2 2 2 2 1 1
14 1 1 2 1 2 1 2 1 1 1 1 2 2 3 1 3 2 1 1 3 2 2 1 1
15 2 1 2 1 1 2 1 2 1 1 2 3 3 1 2 3 2 2 2 2 2 3 1 1
16 1 2 1 2 1 2 2 1 1 1 2 2 2 1 2 1 1 3 2 3 1 1 1 1
17 1 2 2 2 2 1 1 2 2 1 2 3 2 2 3 1 3 2 2 3 2 2 3 1
18 2 1 1 2 2 2 2 2 2 1 1 1 3 3 2 3 3 3 2 3 3 2 2 1
19 1 1 2 1 2 1 2 1 1 1 1 1 3 2 2 1 1 2 3 2 3 1 3 1
20 2 1 2 1 1 2 1 2 1 1 2 1 2 2 1 2 3 1 3 3 1 1 2 1
21 2 2 2 2 2 2 1 1 1 2 1 1 2 1 3 3 1 2 1 2 1 2 2 1
22 2 1 2 2 1 1 2 1 2 2 2 2 3 2 2 2 3 2 1 1 1 2 1 1
23 2 2 1 1 2 1 2 2 1 2 2 3 3 1 3 2 1 1 3 3 3 2 1 1
24 2 1 2 2 1 1 2 1 2 2 2 3 2 3 1 1 1 1 2 2 3 3 2 1
25 1 2 1 2 1 2 2 1 1 1 2 3 1 2 1 3 2 2 3 1 3 2 2 1
26 2 2 2 2 2 2 1 1 1 2 1 2 1 2 2 2 2 1 2 3 3 3 3 1
27 1 2 2 1 1 2 2 2 2 2 1 3 3 2 1 3 1 3 1 3 1 3 3 1
28 1 2 2 1 1 2 2 2 2 2 1 1 2 3 3 2 2 2 2 1 3 1 1 1
29 2 2 1 1 2 1 2 2 1 2 2 2 1 3 1 3 3 2 2 2 1 1 3 1
30 1 1 1 1 2 2 1 1 2 2 2 3 1 3 2 2 1 2 1 3 2 1 2 1
31 2 2 1 1 2 1 2 2 1 2 2 1 2 2 2 1 2 3 1 1 2 3 2 1
32 2 1 1 2 2 2 2 2 2 1 1 3 1 2 3 1 2 1 1 2 1 1 1 1
33 1 2 2 2 2 1 1 2 2 1 2 2 3 1 1 2 2 3 1 2 3 1 2 1
34 2 1 2 2 1 1 2 1 2 2 2 1 1 1 3 3 2 3 3 3 2 1 3 1
35 2 2 1 1 1 1 1 1 2 1 1 3 2 3 2 2 2 3 3 2 1 2 3 1
36 1 2 2 1 1 2 2 2 2 2 1 2 1 1 2 1 3 1 3 2 2 2 2 1
37 1 2 2 1 1 2 2 2 2 2 1 2 1 1 2 1 3 1 3 2 2 2 2 .
38 1 2 2 1 1 2 2 2 2 2 1 2 1 1 2 1 3 1 3 2 2 2 2 .
39 1 2 2 1 1 2 2 2 2 2 1 2 1 1 2 1 3 1 3 2 2 2 2 .
40 1 2 2 1 1 2 2 2 2 2 1 2 1 1 2 1 3 1 3 2 2 2 2 .

Here is the iteration history for the augmentation.

Augment a Design

Design Refinement History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 97.0381 97.0381 Ini

1 Start 97.0328 Pre,Mut,Ann

1 37 3 97.0612 97.0612
1 37 7 97.0618 97.0618
1 37 8 97.0810 97.0810
1 37 9 97.0994 97.0994
1 37 12 97.1028 97.1028
1 37 16 97.1074 97.1074
1 38 5 97.1116 97.1116
1 38 9 97.1220 97.1220
1 38 10 97.1224 97.1224
1 38 12 97.1224 97.1224
Initial Designs 271

1 38 14 97.1225 97.1225
1 39 2 97.1460 97.1460
1 39 3 97.1569 97.1569
1 39 5 97.1629 97.1629
1 39 6 97.1815 97.1815
1 39 7 97.1878 97.1878
1 39 8 97.1899 97.1899
1 39 9 97.1965 97.1965
1 39 11 97.1986 97.1986
1 39 12 97.1986 97.1986
1 39 13 97.2002 97.2002
1 39 17 97.2002 97.2002
1 40 10 97.2002 97.2002
1 40 13 97.2002 97.2002
1 37 1 97.2023 97.2023
1 37 6 97.2043 97.2043
1 End 97.2033

2 Start 97.1737 Pre,Mut,Ann

2 End 97.2038

3 Start 97.1949 Pre,Mut,Ann

3 39 5 97.2043 97.2043
3 39 14 97.2043 97.2043
3 39 23 97.2043 97.2043
3 40 10 97.2043 97.2043
3 40 16 97.2043 97.2043
3 37 3 97.2043 97.2043
3 End 97.2033

4 Start 97.1918 Pre,Mut,Ann

4 End 97.2033

5 Start 97.1825 Pre,Mut,Ann

5 End 97.2033

6 Start 97.1940 Pre,Mut,Ann

6 38 9 97.2043 97.2043
6 38 13 97.2043 97.2043
6 38 18 97.2043 97.2043
6 39 13 97.2049 97.2049
6 39 23 97.2049 97.2049
6 37 14 97.2049 97.2049
6 37 19 97.2049 97.2049
6 37 12 97.2049 97.2049
6 37 21 97.2049 97.2049
6 39 15 97.2049 97.2049
6 End 97.2033

7 Start 97.1923 Pre,Mut,Ann

7 39 6 97.2049 97.2049
7 39 11 97.2049 97.2049
7 39 17 97.2049 97.2049
7 39 19 97.2049 97.2049
7 39 22 97.2049 97.2049
7 40 19 97.2054 97.2054
7 End 97.2023

8 Start 97.1839 Pre,Mut,Ann

8 End 97.2023
272 TS-677E Multinomial Logit, Discrete Choice Modeling

9 Start 97.1915 Pre,Mut,Ann

9 End 97.2038

10 Start 97.1862 Pre,Mut,Ann

10 End 97.2038

Notice that the macro goes straight into the design refinement stage. Also notice that in the iteration history, only
rows 37 through 40 are changed. Here is the design. The last four rows are the holdouts.

Augment a Design

O x x x x x x x x x x x x x x
b x x x x x x x x x 1 1 1 1 1 1 1 1 1 1 2 2 2 2
s 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 f

1 1 1 1 2 1 1 1 2 1 2 1 3 2 1 2 3 3 1 1 1 3 1 3 1
2 1 1 2 1 2 1 2 1 1 1 1 3 1 1 3 2 3 3 2 1 1 3 2 1
3 2 1 1 2 2 2 2 2 2 1 1 2 2 1 1 2 1 2 3 1 2 3 3 1
4 1 2 1 2 1 2 2 1 1 1 2 1 3 3 3 2 3 1 1 2 2 3 3 1
5 2 2 1 1 1 1 1 1 2 1 1 2 3 2 3 3 1 1 2 1 2 1 2 1
6 2 2 2 2 2 2 1 1 1 2 1 3 3 3 1 1 3 3 3 1 2 1 1 1
7 1 2 2 2 2 1 1 2 2 1 2 1 1 3 2 3 1 1 3 1 1 3 1 1
8 2 1 2 1 1 2 1 2 1 1 2 2 1 3 3 1 1 3 1 1 3 2 3 1
9 1 1 1 1 2 2 1 1 2 2 2 2 2 2 3 3 3 3 3 2 3 3 1 1
10 1 1 1 1 2 2 1 1 2 2 2 1 3 1 1 1 2 1 2 1 1 2 3 1
11 1 1 1 2 1 1 1 2 1 2 1 2 3 3 3 1 2 2 3 3 1 3 2 1
12 2 2 1 1 1 1 1 1 2 1 1 1 1 1 1 1 3 2 1 3 3 3 1 1
13 1 1 1 2 1 1 1 2 1 2 1 1 1 2 1 2 1 3 2 2 2 2 1 1
14 1 1 2 1 2 1 2 1 1 1 1 2 2 3 1 3 2 1 1 3 2 2 1 1
15 2 1 2 1 1 2 1 2 1 1 2 3 3 1 2 3 2 2 2 2 2 3 1 1
16 1 2 1 2 1 2 2 1 1 1 2 2 2 1 2 1 1 3 2 3 1 1 1 1
17 1 2 2 2 2 1 1 2 2 1 2 3 2 2 3 1 3 2 2 3 2 2 3 1
18 2 1 1 2 2 2 2 2 2 1 1 1 3 3 2 3 3 3 2 3 3 2 2 1
19 1 1 2 1 2 1 2 1 1 1 1 1 3 2 2 1 1 2 3 2 3 1 3 1
20 2 1 2 1 1 2 1 2 1 1 2 1 2 2 1 2 3 1 3 3 1 1 2 1
21 2 2 2 2 2 2 1 1 1 2 1 1 2 1 3 3 1 2 1 2 1 2 2 1
22 2 1 2 2 1 1 2 1 2 2 2 2 3 2 2 2 3 2 1 1 1 2 1 1
23 2 2 1 1 2 1 2 2 1 2 2 3 3 1 3 2 1 1 3 3 3 2 1 1
24 2 1 2 2 1 1 2 1 2 2 2 3 2 3 1 1 1 1 2 2 3 3 2 1
25 1 2 1 2 1 2 2 1 1 1 2 3 1 2 1 3 2 2 3 1 3 2 2 1
26 2 2 2 2 2 2 1 1 1 2 1 2 1 2 2 2 2 1 2 3 3 3 3 1
27 1 2 2 1 1 2 2 2 2 2 1 3 3 2 1 3 1 3 1 3 1 3 3 1
28 1 2 2 1 1 2 2 2 2 2 1 1 2 3 3 2 2 2 2 1 3 1 1 1
29 2 2 1 1 2 1 2 2 1 2 2 2 1 3 1 3 3 2 2 2 1 1 3 1
30 1 1 1 1 2 2 1 1 2 2 2 3 1 3 2 2 1 2 1 3 2 1 2 1
31 2 2 1 1 2 1 2 2 1 2 2 1 2 2 2 1 2 3 1 1 2 3 2 1
32 2 1 1 2 2 2 2 2 2 1 1 3 1 2 3 1 2 1 1 2 1 1 1 1
33 1 2 2 2 2 1 1 2 2 1 2 2 3 1 1 2 2 3 1 2 3 1 2 1
34 2 1 2 2 1 1 2 1 2 2 2 1 1 1 3 3 2 3 3 3 2 1 3 1
35 2 2 1 1 1 1 1 1 2 1 1 3 2 3 2 2 2 3 3 2 1 2 3 1
36 1 2 2 1 1 2 2 2 2 2 1 2 1 1 2 1 3 1 3 2 2 2 2 1
37 2 2 2 1 2 1 2 2 2 2 1 2 2 3 2 1 3 3 3 1 1 1 3 .
38 2 1 1 1 1 2 2 1 2 1 1 3 3 2 1 2 1 2 2 1 1 1 3 .
39 2 1 2 2 1 2 1 2 2 2 2 3 3 3 2 3 2 3 2 3 3 3 3 .
40 1 2 1 1 1 2 2 2 1 2 2 1 3 1 3 2 3 3 1 2 1 3 3 .
Initial Designs 273

This code does the same thing only using the holdouts=4 option instead.
title ’Augment a Design’;

%mktex(n=36, seed=292)
%mktex(2 ** 11 3 ** 12, n=40, init=randomized,
holdouts=4, seed=513, options=nosort)

proc print data=design(firstobs=37); run;

Here are the holdout observations, which are the same as we saw previously.

Augment a Design

O x x x x x x x x x x x x x x
b x x x x x x x x x 1 1 1 1 1 1 1 1 1 1 2 2 2 2
s 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 w

37 2 2 2 1 2 1 2 2 2 2 1 2 2 3 2 1 3 3 3 1 1 1 3 .
38 2 1 1 1 1 2 2 1 2 1 1 3 3 2 1 2 1 2 2 1 1 1 3 .
39 2 1 2 2 1 2 1 2 2 2 2 3 3 3 2 3 2 3 2 3 3 3 3 .
40 1 2 1 1 1 2 2 2 1 2 2 1 3 1 3 2 3 3 1 2 1 3 3 .
274 TS-677E Multinomial Logit, Discrete Choice Modeling

Partial Profiles and Restrictions

Partial-profile designs (Chrzan and Elrod, 1995) are used when there are many attributes but no more than a
few of them are allowed to vary at a time. Chrzan and Elrod show an example where respondents must choose
between vacuum cleaners that vary along 20 different attributes: Brand, Price, Warranty, Horsepower, and so
on. It is difficult for respondents to simultaneously evaluate that many attributes, so it is better if they are only
exposed to a few at a time.

Pair-wise Partial Profile Choice Design

Here for example is a partial profile design for 20 two-level factors, with 5 varying at a time, with the factors that
are not shown printed with a dash.

- - - - - - - - - 1 2 2 2 2 - - - - - -
- - 1 1 1 2 - - - - - - - - - - - - - 1
- - - - - - - - 1 1 1 1 1 - - - - - - -
- - - - - - - - - - 2 1 1 - 1 2 - - - -
- 1 1 2 2 2 - - - - - - - - - - - - - -
- - - - - - - - - - 2 1 2 - - 1 2 - - -
- - - - - - - - - - 1 2 1 1 - - 2 - - -
- - - - - - - - - - - - - 1 1 - 2 2 1 -
- - - - 2 1 - - 1 2 2 - - - - - - - - -
- - - - 1 2 - 2 - 2 - 1 - - - - - - - -
- - - - - - - - - - 1 1 - 2 - - 2 1 - -
- - - - - - - - - - - - - 2 1 2 - - 2 2
- - - - - - - - - - 1 1 2 2 2 - - - - -
- - - - - - - - - - - 2 2 - 1 2 - 1 - -
- - - - - - - - - - - - - 1 - 2 1 2 1 -
2 1 - - - - - - - - - - - - - - - 1 1 1
- 1 - - - - 1 - 2 - 2 1 - - - - - - - -
- - - - - - - - - - - - - - 2 2 2 2 2 -
1 2 - 1 2 - - - - - - - - - - - 1 - - -
- - - - 2 2 - 2 2 - 1 - - - - - - - - -
2 2 1 - - - - - - - - - - - - - - 2 - 2
- - - - - - - - - - - - - 1 2 - - 1 2 2
- - - - - 1 2 2 1 - 2 - - - - - - - - -
- 2 - - - - - - - - - - - - - 1 - 2 2 1
2 2 1 - - 1 - - - - - - - - - - - - 2 -
- 2 2 2 - - - - - - - - - - - - 1 - 1 -
- - - - 2 1 - - - - - - - - - - - 2 1 2
- - - 1 1 - 1 - - - - - - - - - - - 1 2
- - - - - 1 - - - 1 - - - 1 2 1 - - - -
- - - - 1 2 2 1 1 - - - - - - - - - - -
2 - - 2 2 - 1 - - - - - - - - - - - - 1
- - - - - - - - - - - - 1 2 - 1 1 - 2 -
1 - 1 - - - 2 1 - - - - - - - - - - - 1
- - - - - 2 1 1 1 2 - - - - - - - - - -
2 - 2 1 - - 2 - - - - - - - - - - - - 2
1 - 2 2 1 2 - - - - - - - - - - - - - -
- - - - - - - - - - 1 2 - - 1 1 1 - - -
- - - 2 1 - - 1 2 2 - - - - - - - - - -
- - - - - - - - - - - 2 - 2 2 1 - - 1 -
- - - - - 2 2 - 2 2 1 - - - - - - - - -
Partial Profiles and Restrictions 275

A design like this could be used to make a binary choice experiment. For example, the first run has factors 10
through 14 varying.

- - - - - - - - - 1 2 2 2 2 - - - - - -

Assume they are all yes-no factors (1 yes, 2 no). Subjects could be offered a choice between these two profiles:

x10 = yes, x11 = no, x12 = no, x13 = no, x14 = no

x10 = no, x11 = yes, x12 = yes, x13 = yes, x14 = yes

The first profile came directly from the design and the second came from shifting the design: yes ! no, and no
! yes. Partial-profile designs have become very popular among some researchers.
Here is the code that generated and printed the partial profile design above.
%mktex(3 ** 20, n=41, partial=5, seed=292)

%mktlab(values=0 1 2)

proc format; value part 0 = ’ -’; run;

data _null_; set final(firstobs=2); put (x1-x20) (part2.); run;

A 320 design is requested in 41 runs. The three levels are yes, no, and not shown. Forty-one runs will give us 40
partial profiles and one more run with just all attributes not shown (all ones). When we ask for partial profiles, in
this case partial=5, we are imposing a constraint that the number of 2’s and 3’s in each run equal 5 and the
number of 1’s equal 15. This makes the sum of the coded variables constant in each run and hence introduces
a linear dependency (the sum of the coded variables is proportional to the intercept). The way we avoid having
the linear dependency is by adding this additional row where all attributes are set to the not shown level. The
sum of the coded variables for this row will be different than the constant sum for the other rows and hence will
eliminate the linear dependency we would otherwise have.
The %MktLab macro reassigns the levels 1, 2, 3 to 0, 1, 2 where 0 will mean not shown, then a format is written
to print zeros as dashes. A DATA step prints the design using the format excluding the constant (all not shown)
first row.
This next section of code takes this design and turns it into a partial-profile choice design. It reads each profile in
the design, and outputs it. If the level is not zero, the code changes 1 to 2 and 2 to 1 and outputs the new profile.
The next step uses the %ChoicEff macro to evaluate the design. We specified zero=none for now to see
exactly which parameters we can estimate and which ones we cannot. This usage of the %ChoicEff macro is
similar to what we saw in the food product example on page 217. Our choice design is specified on data= and
the same data set, with just the Set variable kept, is specified on the init= option. The number of choice sets,
20 (we drop the constant choice set), number of alternatives, 2, and assumed betas, a vector of zeros, are also
specified. Zero internal iterations are requested since we want a design evaluation, not an attempt to improve the
design.
276 TS-677E Multinomial Logit, Discrete Choice Modeling

data des(drop=i);
set = _n_;
set final(firstobs=2);
array x[20];
output;
do i = 1 to 20;
if x[i] then do; if x[i] = 1 then x[i] = 2; else x[i] = 1; end;
end;
output;
run;

%choiceff(data=des,
model=class(x1-x20 / zero=none),
nsets=20, nalts=2,
beta=zero, init=des(keep=set),
intiter=0)
Here is the last part of the output.

Partial Profiles

Variable Standard
n Name Label Variance DF Error

1 x10 x1 0 . 0 .
2 x11 x1 1 2.4989 1 1.58080
3 x12 x1 2 . 0 .
4 x20 x2 0 . 0 .
5 x21 x2 1 4.3585 1 2.08770
6 x22 x2 2 . 0 .
7 x30 x3 0 . 0 .
8 x31 x3 1 5.7179 1 2.39121
9 x32 x3 2 . 0 .
10 x40 x4 0 . 0 .
11 x41 x4 1 19.0020 1 4.35913
12 x42 x4 2 . 0 .
13 x50 x5 0 . 0 .
14 x51 x5 1 1.4018 1 1.18396
15 x52 x5 2 . 0 .
16 x60 x6 0 . 0 .
17 x61 x6 1 2.9092 1 1.70564
18 x62 x6 2 . 0 .
19 x70 x7 0 . 0 .
20 x71 x7 1 3.6474 1 1.90982
21 x72 x7 2 . 0 .
22 x80 x8 0 . 0 .
23 x81 x8 1 5.5731 1 2.36075
24 x82 x8 2 . 0 .
25 x90 x9 0 . 0 .
26 x91 x9 1 5.6681 1 2.38077
27 x92 x9 2 . 0 .
28 x100 x10 0 . 0 .
29 x101 x10 1 1.2731 1 1.12831
30 x102 x10 2 . 0 .
31 x110 x11 0 . 0 .
32 x111 x11 1 1.0522 1 1.02577
33 x112 x11 2 . 0 .
34 x120 x12 0 . 0 .
35 x121 x12 1 1.5623 1 1.24993
36 x122 x12 2 . 0 .
Partial Profiles and Restrictions 277

37 x130 x13 0 . 0 .
38 x131 x13 1 7.9449 1 2.81868
39 x132 x13 2 . 0 .
40 x140 x14 0 . 0 .
41 x141 x14 1 3.7175 1 1.92809
42 x142 x14 2 . 0 .
43 x150 x15 0 . 0 .
44 x151 x15 1 3.5820 1 1.89261
45 x152 x15 2 . 0 .
46 x160 x16 0 . 0 .
47 x161 x16 1 4.7737 1 2.18488
48 x162 x16 2 . 0 .
49 x170 x17 0 . 0 .
50 x171 x17 1 5.7060 1 2.38872
51 x172 x17 2 . 0 .
52 x180 x18 0 . 0 .
53 x181 x18 1 10.0597 1 3.17170
54 x182 x18 2 . 0 .
55 x190 x19 0 . 0 .
56 x191 x19 1 6.5273 1 2.55486
57 x192 x19 2 . 0 .
58 x200 x20 0 . 0 .
59 x201 x20 1 6.5214 1 2.55370
60 x202 x20 2 . 0 .
==
20

We see that one parameter is estimable for each factor and that is the parameter for the 1 or yes level. In effect,
we have two reference levels, one for the not shown level and the expected one for the no level. The %ChoicEff
macro prints a list of all redundant variables.

Redundant Variables:
x10 x12 x20 x22 x30 x32 x40 x42 x50 x52 x60 x62 x70 x72 x80 x82 x90 x92 x100
x102 x110 x112 x120 x122 x130 x132 x140 x142 x150 x152 x160 x162 x170 x172
x180 x182 x190 x192 x200 x202

We can cut this list into our program and drop those terms.
%choiceff(data=des,
model=class(x1-x20 / zero=none),
nsets=20, nalts=2,
beta=zero, init=des(keep=set),
intiter=0, drop=
x10 x12 x20 x22 x30 x32 x40 x42 x50 x52 x60 x62 x70 x72 x80 x82 x90 x92 x100
x102 x110 x112 x120 x122 x130 x132 x140 x142 x150 x152 x160 x162 x170 x172
x180 x182 x190 x192 x200 x202);
278 TS-677E Multinomial Logit, Discrete Choice Modeling

Here is the last part of the output.

Partial Profiles

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 2.4989 1 1.58080

2 x21 x2 1 4.3585 1 2.08770
3 x31 x3 1 5.7179 1 2.39121
4 x41 x4 1 19.0020 1 4.35913
5 x51 x5 1 1.4018 1 1.18396
6 x61 x6 1 2.9092 1 1.70564
7 x71 x7 1 3.6474 1 1.90982
8 x81 x8 1 5.5731 1 2.36075
9 x91 x9 1 5.6681 1 2.38077
10 x101 x10 1 1.2731 1 1.12831
11 x111 x11 1 1.0522 1 1.02577
12 x121 x12 1 1.5623 1 1.24993
13 x131 x13 1 7.9449 1 2.81868
14 x141 x14 1 3.7175 1 1.92809
15 x151 x15 1 3.5820 1 1.89261
16 x161 x16 1 4.7737 1 2.18488
17 x171 x17 1 5.7060 1 2.38872
18 x181 x18 1 10.0597 1 3.17170
19 x191 x19 1 6.5273 1 2.55486
20 x201 x20 1 6.5214 1 2.55370
==
20

Linear Partial Profile Design

Here is another example. Say you would like to make a design in 36 runs with 12 three-level factors, but you
want only four of them to be considered at a time. You would need to create four-level factors with one of the
levels meaning not shown. You also need to ask for a design in 37 runs, because with partial profiles, one run
must be all-constant. Here is a partial profile request with the %MktEx macro using the partial= option.
title ’Partial Profiles’;

%mktex(4 ** 12, n=37, partial=4, seed=201)

proc print; run;

The iteration history will proceed like before, so we will not discuss it. Here is the final D-efficiency.

Partial Profiles

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 41.2020 17.9653 100.0000 1.0000
Partial Profiles and Restrictions 279

With partial-profile designs, D-efficiency will typically be much less than we are accustomed to seeing with other
types of designs. Here is the design.

Partial Profiles

Obs x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12

1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 1 4 2 3 4
3 1 1 1 1 1 1 1 1 4 3 2 3
4 1 1 1 1 1 1 1 2 3 3 4 1
5 1 1 1 1 1 1 2 4 4 2 1 1
6 1 1 1 1 1 1 4 3 4 4 1 1
7 1 1 1 1 1 2 4 1 2 1 4 1
8 1 1 1 1 1 4 3 3 2 1 1 1
9 1 1 1 1 2 4 2 1 3 1 1 1
10 1 1 1 1 3 3 3 1 3 1 1 1
11 1 1 1 1 3 4 4 2 1 1 1 1
12 1 1 1 1 4 1 4 4 3 1 1 1
13 1 1 1 2 2 1 2 3 1 1 1 1
14 1 1 1 2 4 4 1 1 1 1 4 1
15 1 1 1 3 2 2 1 4 1 1 1 1
16 1 1 2 1 4 2 1 3 1 1 1 1
17 1 1 2 4 2 3 1 1 1 1 1 1
18 1 2 1 1 1 1 1 2 2 4 1 1
19 1 2 1 4 1 2 3 1 1 1 1 1
20 1 2 2 3 1 1 2 1 1 1 1 1
21 1 3 3 1 2 1 3 1 1 1 1 1
22 1 3 4 3 1 1 1 1 1 1 1 3
23 1 4 3 1 1 3 1 1 1 1 3 1
24 1 4 3 4 3 1 1 1 1 1 1 1
25 1 4 4 2 1 1 1 1 1 1 1 2
26 2 1 4 1 1 1 1 1 1 1 2 4
27 2 2 3 1 1 1 1 1 1 1 1 2
28 2 3 2 2 1 1 1 1 1 1 1 1
29 2 4 1 3 4 1 1 1 1 1 1 1
30 3 1 1 1 1 1 1 1 3 2 2 1
31 3 1 4 1 1 1 1 1 4 1 3 1
32 3 3 1 1 1 1 1 1 1 3 1 4
33 3 4 1 1 1 1 1 1 1 1 4 3
34 4 1 1 1 1 1 1 1 1 4 3 3
35 4 1 3 1 1 1 1 2 1 1 1 4
36 4 2 4 1 2 1 1 1 1 1 1 1
37 4 3 1 1 1 1 1 1 1 1 2 2

Notice that the first run is constant. For all other runs, exactly four factors vary and have levels not equal to 1.
280 TS-677E Multinomial Logit, Discrete Choice Modeling

Choice from Triples; Partial Profiles Constructed Using Restrictions

The approach we just saw, constructing partial profiles using the partial= option, would be fine for a full-
profile conjoint study or a pair-wise choice study with level shifts. However, it would not be good for a more
general choice experiment with more alternatives. For a choice experiment, you would have to have full profile
restrictions on each alternative, and you must have the same attributes varying in each choice set. There is
currently no automatic way to request this in the %MktEx macro, so you have to program the restrictions yourself.
To specify restrictions for choice designs, you need to take into consideration the number of attributes that may
vary within each alternative, which ones, and which attributes go with which alternatives. Fortunately, that is not
too difficult. See page 195 for other examples of restrictions.
In this section, we will construct a partial profile design for a purely generic study (unbranded), with ten attributes
and three alternatives. Each attribute will have three levels, and each alternative will be a bundle of attributes.
Partial-profile designs have the advantage that subjects do not have to consider all attributes at once. However,
this is also a bit of a disadvantage as well in the sense that the subjects must constantly shift from considering
one set of attributes to considering a different set. For this reason, it can be helpful to get more information out
of each choice, and having more than two alternatives per choice set accomplishes this.
This example will have several parts. As we mentioned in the chair study, we will usually not directly use the
%MktEx macro to generate designs for generic studies. Instead, we will use the %MktEx macro to generate
a candidate set of partial profile choice sets. Next, the design will be checked and turned into a candidate set
of generic choice sets. Next, the %MktDups macro will be called to ensure there are no duplicate choice sets.
Finally, the %ChoicEff macro will be used to create an efficient generic partial-profile choice design.
Before we go into any more detail on making this design, let’s skip ahead and look at a couple of potential choice
sets so it will be clear what we are trying to accomplish and why. Here are two potential choice, sets still in linear
design format.

2 2 1 3 1 2 1 2 2 2 2 1 3 2 1 1 2 1 2 2 2 3 2 1 1 3 3 3 2 2
2 2 1 3 2 2 1 3 2 3 3 2 2 3 1 2 1 1 1 1 1 2 3 3 3 2 1 2 3 2

Here are the same two potential choice sets, but now arrayed in choice design format.

Partial Profiles

Set x1 x2 x3 x4 x5 x6 x7 x8 x9 x10

1 2 2 1 3 1 2 1 2 2 2
2 1 3 2 1 1 2 1 2 2
2 3 2 1 1 3 3 3 2 2

2 2 2 1 3 2 2 1 3 2 3
3 2 2 3 1 2 1 1 1 1
1 2 3 3 3 2 1 2 3 2

Each choice set has 10 three-level factors and three alternatives. Four attributes are constant in each choice set:
x1, x5, x9, and x10 in the first choice set, and x2, x4, x6, and x7 in the second choice set. We do not need
an all-constant choice set like we saw in our earlier partial-profile designs, nor do we need an extra level for not
varying. In this approach, we will simply construct choice sets for four constant attributes (they may be constant
at 1, 2, or 3) and six varying attributes (with levels: 1, 2, and 3). Respondents will be given a choice task along
the lines of “Given a set of products that differ on these attributes but are identical in all other respects, which
one would you choose?”. They would then be shown a list of differences.
Partial Profiles and Restrictions 281

Here is the code for making the candidate set.

title ’Partial Profiles’;

%macro partprof;
sum = 0;
do k = 1 to 10;
sum = sum + (x[k] = x[k+10] & x[k] = x[k+20] & x[k+10] = x[k+20]);
end;
bad = abs(sum - 4);
%mend;

%mktex(3 ** 30, n=198,

optiter=0, tabiter=0, maxdesigns=1, mutate=, anneal=, options=accept,
out=[Link], restrictions=partprof, seed=201)
We requested a design in 198 runs with 30 three-level factors. The 198 was chosen arbitrarily as a number
divisible by 3  3 = 9 that would give us approximately 200 candidate sets. The first ten factors, x1-x10, will
make the first alternative, the next ten, x11-x20, will make the second alternative, and the last ten, x21-x30,
will make the third alternative. We will want six attributes to be nonconstant at a time. The PartProf macro
will count the number of constant alternatives: x1 = x11 = x21, x2 = x12 = x22, ..., and x10 = x20
= x30. If the number of constant alternatives is four, our choice set conforms. If it is more or less than four, our
choice set is in violation of the restrictions. The badness is the absolute difference between four and the number
of constant attributes.
Our goal in this step is to make a candidate set of potential partial-profile choice sets, not to make a final experi-
mental design. All we really need is a design with most of the runs defining valid partial-profile candidate choice
sets. Ideally, it would be nice if we had more than random candidates it would be nice if our candidate gen-
eration code at least made some attempt to ensure that our attributes are approximately orthogonal and balanced
across attributes both between and within alternatives. It is not critical that we allow the macro to spend a great
deal of time optimizing linear model D-efficiency. We will specify options so that the macro will just create one
candidate design using the coordinate-exchange algorithm with a random initialization.
This is a big problem (30 factors and 198 runs) with restrictions, so %MktEx macro will run slowly by default.
For this reason, we use some of the more esoteric number-of-iterations options. We specify optiter=0, which
specifies no OPTEX iterations, since with large partial profile studies, we will never have a good candidate set
for PROC OPTEX to search. We also specify tabiter=0 since a tabled initial design will be horrible for this
problem. We specify maxdesigns=1 which will just give us one design using the first method, which will now
be coordinate exchange with a random initial design. We also specify mutate= and anneal= (both with null
arguments) so there will be no random mutations or simulated annealing. When we have very strong restrictions
like this, the simulated annealing will regularly destroy the structure of the design and make the iterations take
longer, so we turn if off. We also specify options=accept so that the macro will still output a design even
if every choice set does not meet the restrictions. We have used the other options in several previous examples.
Here is a small part of the output.

Partial Profiles

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 85.1169 Ran
1 170 1 95.1279 95.1279 Conforms
1 170 3 95.1305 95.1305
1 171 16 95.1314 95.1314
1 171 6 95.1346 95.1346
282 TS-677E Multinomial Logit, Discrete Choice Modeling

.
.
.
1 60 7 96.5153 96.5153
1 66 3 96.5166 96.5166
1 End 96.5166
.
.
.

Partial Profiles

The OPTEX Procedure

Average
Prediction
Design Standard
Number D-Efficiency A-Efficiency G-Efficiency Error
------------------------------------------------------------------------
1 96.5165 93.4268 95.9714 0.5551

The macro finds a design that conforms to the restrictions (shown by the Conforms note) that is 95.1% D-
efficient, and the final design is just over 96.5% D-efficient. This step took almost 28 minutes.
Here is the rest of the code for making the partial-profile choice design.
data des(drop=k sum);
set [Link];
array x[30];
sum = 0;
do k = 1 to 10;
sum = sum + (x[k] = x[k+10] & x[k] = x[k+20] & x[k+10] = x[k+20]);
end;
if sum eq 4;
run;

%mktkey(x1-x30)

data key;
input (x1-x10) ($);
datalines;
x1 x2 x3 x4 x5 x6 x7 x8 x9 x10
x11 x12 x13 x14 x15 x16 x17 x18 x19 x20
x21 x22 x23 x24 x25 x26 x27 x28 x29 x30
;

%mktroll(design=des, key=key, out=rolled)

%mktdups(generic, data=rolled, out=nodups, factors=x1-x10, nalts=3);

proc print data=nodups(obs=9); id set; by set; run;

%choiceff(data=nodups, model=class(x1-x10), seed=155,

iter=10, nsets=27, nalts=3, options=nodups, beta=zero)

proc print data=best; id set; by notsorted set; var x1-x10; run;

Partial Profiles and Restrictions 283

We use a DATA step to check each profile to see if it conforms using essentially the same code we saw in the
restrictions macro. The subsetting if statement outputs just those choice sets with four constant attributes. This
step is not necessary for this particular problem because we saw in the output that the design conformed to the
restrictions, however, it could be necessary for other problems. The %MktKey macro is run to generate the
full list of names in the range x1 - x30 for pasting into the next step. A KEY data set is created and the
%MktRoll macro is run to create a generic choice design from the linear candidate design.
The next step runs the %MktDups macro, which we have not used in previous examples. The %MktDups macro
can check a design to see if there are any duplicate runs and output just the unique sets. For a generic study like
this, it can also check to make sure there are no duplicate choice sets taking into account the fact that two choice
sets can be duplicates even if the alternatives are not in the same order. The %MktDups step names in a positional
parameter the type of design as a generic choice design. It names the input data set and the output data set
that will contain the design with any duplicates removed. It names the factors in the choice design x1-x10 and
the number of alternatives. The result is a data set called NODUPS. Here are the first 3 candidate choice sets.

Partial Profiles

Set _Alt_ x1 x2 x3 x4 x5 x6 x7 x8 x9 x10

1 1 1 1 1 1 1 3 1 3 3 3
2 1 2 2 3 1 2 1 1 2 3
3 1 3 3 2 1 2 1 2 1 3

2 1 1 1 1 1 3 2 2 1 3 1
2 3 1 1 1 2 2 1 3 3 3
3 2 1 3 1 1 2 3 2 3 2

3 1 1 1 1 2 1 3 3 2 3 3
2 3 1 1 2 3 1 1 1 1 3
3 2 1 1 2 2 2 2 3 2 3

The %ChoicEff macro is called to search for an efficient choice design. The specification
model=class(x1-x10) specifies a generic model with 10 attributes. The option iter=10 specifies
more than the default number of iterations (the default is 2 designs). We ask for a design with 27 sets and 3
alternatives. Furthermore, we ask for no duplicate choice sets and specify an assumed beta vector of zero. Here
are some of the results from the %ChoicEff macro.

Partial Profiles

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 2.410695 0.414818
1 2.846091 0.351359
2 2.875865 0.347721
3 2.881470 0.347045
.
.
.

Design Iteration D-Efficiency D-Error

----------------------------------------------
8 0 2.317748 0.431453
1 2.857337 0.349976
2 2.902798 0.344495
3 2.902798 0.344495
284 TS-677E Multinomial Logit, Discrete Choice Modeling

.
.
.

Partial Profiles

Final Results: Design = 8

Efficiency = 2.9027977631
D-Error = 0.3444952358

Partial Profiles

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 0.43314 1 0.65814

2 x12 x1 2 0.45528 1 0.67474
3 x21 x2 1 0.40651 1 0.63758
4 x22 x2 2 0.38230 1 0.61831
5 x31 x3 1 0.39510 1 0.62857
6 x32 x3 2 0.44795 1 0.66929
7 x41 x4 1 0.45472 1 0.67433
8 x42 x4 2 0.43250 1 0.65765
9 x51 x5 1 0.43399 1 0.65878
10 x52 x5 2 0.43367 1 0.65854
11 x61 x6 1 0.40985 1 0.64020
12 x62 x6 2 0.41572 1 0.64477
13 x71 x7 1 0.45571 1 0.67506
14 x72 x7 2 0.45254 1 0.67271
15 x81 x8 1 0.40567 1 0.63693
16 x82 x8 2 0.36000 1 0.60000
17 x91 x9 1 0.39250 1 0.62649
18 x92 x9 2 0.41535 1 0.64447
19 x101 x10 1 0.43561 1 0.66001
20 x102 x10 2 0.40682 1 0.63783
==
20

Here is part of the design.

Partial Profiles

Set x1 x2 x3 x4 x5 x6 x7 x8 x9 x10

16 1 1 2 3 2 1 3 3 3 3
3 1 2 2 1 2 3 1 3 2
2 1 2 1 3 3 3 2 3 1

196 3 3 3 3 1 2 3 2 1 1
1 2 3 2 2 3 3 2 1 2
2 1 3 1 3 1 3 2 1 3
.
.
.

77 2 1 2 2 3 1 1 1 2 2
1 2 2 1 3 2 1 3 1 2
3 3 2 3 3 3 1 2 3 2
Partial Profiles and Restrictions 285

The choice set number corresponds to the original set numbers in the candidate design.

Advanced Restrictions
There is one more aspect to restrictions that must be understood for the most sophisticated usages of restrictions.
The macro that imposes the restrictions is defined and called in four distinct places in the %MktEx macro. First,
the restrictions macro is called in a separate, preliminary IML step, just to catch some syntax errors you might
have made. Next, it is called in between calling PROC PLAN or FACTEX and calling PROC OPTEX. Here, the
restrictions macro is used to impose restrictions on the candidate set. Next it is used in the obvious way during
design creation and the coordinate-exchange algorithm. Finally, when options=accept is specified, which
means that restriction violations are acceptable, the macro is called after all of the iterations have completed to
report on restriction violations in the final design. For some advanced restrictions, we will not want exactly the
same code running in all four places. When the restrictions are purely written in terms of restrictions on x, which
is the ith row of the design matrix, there is no problem. The same macro will work fine for all uses. However,
when xmat (the full x matrix) or i or j1 (the row or column number) are used, the same code typically cannot
be used for all applications, although sometimes it does not matter. Next are some notes on each of the four
phases.
Syntax Check In this phase, the macro is defined and called just to check for syntax errors. This step allows the
macro to end more gracefully than it would otherwise if there are errors. Your restrictions macro can recognize
when it is in this phase because the macro variable &main is set to 0 and the macro variable &pass is set to
null. The pass variable is null before the iterations begin, 1 for the algorithm search phase, 2 for the design search
phase, 3 for the design refinement stage, and 4 after the iterations end. You can conditionally execute code in this
step or not using the following macro statements.
%if &main eq 0 and &pass eq %then %do; /* execute in syntax check */
%if not (&main eq 0 and &pass eq) %then %do; /* not execute in syntax check */
You will usually not need to worry about this step. It just calls the macro once and ignores the results to check for
syntax errors. For this step, xmat (and hence x) is a vector of ones (since the design does not exist yet) and j1
= j2 = j3 = i = 1. If you have complicated restrictions involving the row or column exchange indices
(i, j1, j2, j3) you may need to worry about this step. You may need to either not execute your restrictions
in this step or conditionally execute some assignment statements (just for this step) that set up j1, j2, and j3
more appropriately. If you have syntax errors in your restrictions macro and you cannot figure out what they are,
sometimes the best thing to do is directly submit the statements in your restrictions macro to IML so you can see
the normal IML syntax errors. First submit the following statements.
%let n = 27; /* substitute number of runs */
%let m = 10; /* substitute number of factors */
proc iml;
xmat = j(&n, &m, 1);
i = 1; j1 = 1; j2 = 1; j3 = 1; bad = 0; x = xmat[i,];
Candidate Check In this phase, the macro is used to impose restrictions on the candidate set created by PROC
PLAN or PROC FACTEX before it is searched by PROC OPTEX. For some problems, such as most partial profile
problems, the restrictions are so severe that virtually none of the candidates will conform. Also, restrictions that
are based on row number and column number do not make sense in the context of a candidate design. Your
restrictions macro can recognize when it is in this phase because the macro variable &main is set to 0 and the
macro variable &pass is set to 1 or 2. You can conditionally execute code in this step or not using the following
macro statements.
%if &main eq 0 and &pass ge 1 and &pass le 2)
%then %do; /* execute on candidates */
%if not (&main eq 0 and &pass ge 1 and &pass le 2)
%then %do; /* not execute on candidates */
286 TS-677E Multinomial Logit, Discrete Choice Modeling

For simple restrictions, not involving the column exchange indices (j1, j2, j3) you probably do not need to
worry about this step. If you use j1, j2, or j3, you will need to either not execute your restrictions in this step
or conditionally execute some assignment statements that set up j1, j2, and j3 appropriately. Ordinarily for
this step, xmat contains the candidate design, x contains the ith row, j1 = 0; j2 = 0; j3 = 0; and i
is set to the candidate row number.
Main Coordinate-Exchange Algorithm In this phase, the macro is used to impose restrictions on the design as
it is being built in the coordinate-exchange algorithm. Your restrictions macro can recognize when it is in this
phase because the macro variable &main is set to 1 and the macro variable &pass is set to 1, 2, or 3. You can
conditionally execute code in this step or not using the following macro statements.
%if &main eq 1 and &pass ge 1 and &pass le 3)
%then %do; /* execute on coordinate exchange */
%if not (&main eq 1 and &pass ge 1 and &pass le 3)
%then %do; /* not execute on coordinate exchange */
For this step, xmat contains the candidate design, x contains the ith row, j1 contains the column index, j2 and
j3 are zero (unless you are using exchange=, in which case j1 and j2 are indexes of other columns being
exchanged), and i is the row number.
Restrictions Violations Check In this phase, the macro is used to check the design when there are restrictions and
options=accept. Your restrictions macro can recognize when it is in this phase because the macro variable
&main is set to 1 and the macro variable &pass is greater than 3. You can conditionally execute code in this
step or not using the following macro statements.
%if &main eq 1 and &pass gt 3) %then %do; /* execute on final check */
%if not (&main eq 1 and &pass gt 3) %then %do; /* not execute on final check */
For this step, xmat contains the candidate design, x contains the ith row, j1 = 0; j2 = 0; j3 = 0; and
i is the row number.
Here is an example of a partial profile macro that does what the partial=4 option does.
%macro partprof;
nvary = sum(x ^= 1);
%if &main %then %do;
if i = 1 then bad = nvary;
else bad = abs(nvary - 4);
%end;
%else %do;
bad = ^ (nvary = 0 | nvary = 4);
%end;
%mend;
In the main algorithm, when imposing restrictions on the design, we restrict the first run to be constant and all
other runs to have four attributes varying. For the candidate-set restrictions, when MAIN is zero, any observation
with zero or four varying factors is acceptable. For the candidate-set restrictions, there is no reason to count the
number of violations. A candidate run is either acceptable or not. We do not worry about the syntax error or final
check steps; both versions will work fine in either.
The Macros 287

The Macros
The autocall macros that are used in this report are documented in this section on the indicated pages.

Macro Page Release Purpose

%ChoicEff 288 8.0 efficient choice design
%MktAllo 303 8.1 processing allocation data
%MktBal 305 9.0 balanced main-effects designs
%MktBlock 307 9.0 block a linear or choice design
%MktDes 314 8.0 efficient linear experimental design
%MktDups 319 9.0 identify duplicate choice sets or runs
%MktEval 325 8.1 evaluate an experimental design
%MktEx 327 9.0 efficient linear experimental design
%MktKey 344 9.0 aid creation of key= data set
%MktLab 345 9.0 relabel and rename design factors
%MktMerge 353 8.1 merging a choice design with choice data
%MktOrth 354 9.0 list orthogonal designs that %MktEx can make
%MktRoll 356 8.1 rolling a linear design into a choice design
%MktRuns 360 8.0 experimental design size
%PhChoice 364 8.0 customizing the printed output from a choice model

The “Release” column indicates the first SAS release in which each macro was distributed. If your site has
installed the autocall libraries supplied by SAS and uses the standard configuration of SAS supplied soft-
ware, you need to ensure that the SAS system option mautosource is in effect to begin using the auto-
call macros. Note however, that Version 9.0 was finished before the macros were finalized and this book
finished. Hence there are a few differences between the macros used in this book and those shipped with
Version 9.0 of SAS. If you are running version 8.2 or version 9.0 of SAS, get the latest macros from the
web at [Link] [Link] or by writing [Link]@[Link]. These
macros will not work with Version 6.12. You should install all of these macros, not just one. Some of the macros
call other macros and will not work if the other macros are not there or if only older versions of the other macros
are there. For example, the %MktEx macro calls the %MktRuns and %MktDes macros.
The macros do not have to be included (for example, with a %include statement). They can be called directly
once they are properly installed. For more information about autocall libraries, refer to SAS Macro Language:
Reference. On a PC for example, the autocall library may be installed in the statnsasmacro directory off of
your SAS root directory. The name or your SAS root directory could be anything, but it is probably something
like SAS or SASnV8. One way to find the right directory is to use Start ! Find to find one of the existing
autocall macros such as [Link] or [Link].
Unix should have a similar directory structure to the PC. The autocall library in Unix may be installed in the
stat/sasmacro directory off of your SAS root directory. On MVS, each macro will be a different member
of a PDS. For details on installing autocall macros, consult your host documentation.
Usually, an autocall library is a directory containing individual files, each of which contains one macro definition.
An autocall library can also be a SAS catalog. To use a directory as a SAS autocall library, store the source
code for each macro in a separate file in the directory. The name of the file must be the same as the macro
name, typically followed by .sas. For example, the macro %MktEx must typically be stored in a file named
[Link]. On most hosts, the reserved fileref sasautos is assigned at invocation time to the autocall
library supplied by SAS or another one designated by your site. If you are specifying your own autocall libraries,
remember to concatenate the autocall library supplied by SAS with your autocall libraries so that these macros
will also be available. For details, refer to your host documentation and SAS macro language documentation.
288 TS-677E Multinomial Logit, Discrete Choice Modeling

Macro Errors
Usually, if you make a mistake in specifying macro options, the macro will print an informative message and
quit. These macros go to great lengths to check their input and issue informative errors. However, complete error
checking is impossible in macros, and sometimes you will get a cascade of less than helpful error messages.
In that case, you will have to check the input and hunt for errors. One of the more common errors is a missing
comma between options. Sometimes for harder errors, specifying options mprint; will help you locate the
problem. Once you think you know which option is involved, be sure to also check the option before and after in
your macro invocation, because that might be where the problem really is. If you have problems with a %MktEx
macro restrictions macro, see page 285 for a suggestion on how to diagnose the problem.
The %MktRuns and %PhChoice macros use PROC TEMPLATE and ODS to create customized output tables.
Typically, the instructions for this customization, created by PROC TEMPLATE, are stored in a file under the
sasuser directory with a host dependent name. On some hosts, this name is templat.sas7bitm. On
other hosts, the name is some variation of the name templat. Sometimes this file can be corrupted. When this
happens, these macros will not run correctly, and you will see error messages including errors about invalid pages.
The solution is to find the corrupt file under sasuser and delete it (using your ordinary operating system file
deletion method). After that, these macros should run fine again. If you have run any other PROC TEMPLATE
customizations, you will need to rerun them after deleting the file. For more information, see “Template Store”
or “Item Store” in the SAS ODS documentation.
Sometimes, you will run the %MktEx macro, and everything will seem to run fine in the entire job, but at the end
of your SAS log, you will see the message:
ERROR: Errors printed on page ....
Typically, this is caused by one or more PROC FACTEX steps failing to find the requested design. When this
happens, the macro recovers and continues searching. The macro does not always know in advance if PROC
FACTEX will succeed. The only way for it to find out is for it to try. The macro suppresses the PROC FACTEX
error messages along with most other notes and warnings that would ordinarily come out. However, SAS still
knows that a procedure tried to print an error message, and prints an error at the end of the log. This error can be
ignored.

%ChoicEff Macro
The %ChoicEff autocall macro is used to find efficient experimental designs for choice experiments. You
supply sets of candidate alternatives. The macro searches the candidates for an efficient experimental design a
design in which the variances of the parameter estimates are minimized, given an assumed parameter vector .
There are two ways you can use the macro:

 You can create a candidate set of alternatives, and the macro will create a design consisting of choice
sets built from the alternatives you supplied. You must designate for each candidate alternative the design
alternative(s) for which it is a candidate. For a branded study with m brands, you must create m lists of
candidate alternatives, one for each brand.
 You can create a candidate set of choice sets, and the macro will build a design from the choice sets that
you supplied. Typically, you would only use this approach when there are restrictions on across alternative
restrictions (certain alternatives may not appear with certain other alternatives).

 If this happens, please write [Link]@[Link], and I will see if I can make the macros better handle that problem in the next
release. Send all the code necessary to reproduce what you have done.
The Macros 289

The %ChoicEff macro uses a modified Federov algorithm, just like PROC OPTEX and the %MktEx macro.
First, the %ChoicEff macro either constructs a random initial design from the candidates or it uses an initial
design that you specified. The macro considers swapping out every design alternative/set and replacing it with
each candidate alternative/set. Typically, you use as a candidate set a full-factorial, fractional-factorial, or a tabled
design created with the %MktEx macro. Swaps that increase efficiency are performed. The process of evaluating
and swapping continues until efficiency stabilizes. This process is repeated with different initial designs, and the
best design is output for use. The key differences between the %ChoicEff macro and the %MktEx macro are as
follows. The %ChoicEff macro requires you to specify the true (or assumed true) parameters and it optimizes
the variance matrix for a multinomial logit model, whereas PROC OPTEX and the %MktEx macro optimize the
variance matrix for a linear model, which does not depend on the parameters.
Here is an example. This example creates a design for a generic model with 3 three-level factors. First, the
%MktEx macro is used to create a set of candidate alternatives, where x1-x3 are the factors. Note that the n=
specification allows expressions. Our candidate set must also contain flag variables, one for each alternative, that
flag which candidates can be used for which alternative(s). Since this is a generic model, each candidate can
appear in any alternative, so we need to add flags that are constant: f1=1 f2=1 f3=1. The %MktEx macro
does not allow you to create constant factors. Instead, we can use the %MktLab macro to add the flag variables,
essentially by specifying that we have multiple intercepts. The option int=f1-f3 creates three variables with
values all one. The default output data set is called FINAL. Next, the %ChoicEff macro is run to find an
efficient design for the unbranded, purely generic model assuming = 0. Here is the code.
%mktex(3 ** 3, n=3**3, seed=238)
%mktlab(int=f1-f3)

%choiceff(data=final, model=class(x1-x3), nsets=9,

flags=f1-f3, beta=zero, seed=145)

proc print; var x1-x3; id set; by set; run;

The option data=final names the input data set, model=class(x1-x3) specifies the PROC TRANSREG
model statement for coding the design, nsets=9 specifies nine choice sets, flags=f1-f3 specifies the
three alternative flag variables, beta=zero specifies all zero parameters, and seed=145 specifies the random
number seed. Here is the output.

n Name Beta Label

1 x11 0 x1 1
2 x12 0 x1 2
3 x21 0 x2 1
4 x22 0 x2 2
5 x31 0 x3 1
6 x32 0 x3 2

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0.817815 1.222770
1 1.677761 0.596032
2 1.702866 0.587245
3 1.702866 0.587245

Design Iteration D-Efficiency D-Error

----------------------------------------------
2 0 0.993579 1.006463
1 1.662712 0.601427
2 1.710398 0.584659

Final Results: Design = 1

Efficiency = 1.7028655003
D-Error = 0.58724544
290 TS-677E Multinomial Logit, Discrete Choice Modeling

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 0.68354 1 0.82677

2 x12 x1 2 0.71089 1 0.84314
3 x21 x2 1 0.70177 1 0.83772
4 x22 x2 2 0.67544 1 0.82185
5 x31 x3 1 0.67544 1 0.82185
6 x32 x3 2 0.67544 1 0.82185
==
6

Set x1 x2 x3

1 2 1 1
1 2 2
3 3 3

2 2 2 3
3 3 2
1 1 1

3 2 1 1
3 2 3
1 3 2

4 3 2 2
2 3 1
1 1 3

5 1 2 1
3 1 3
2 3 2

6 1 2 2
3 3 1
2 1 3

7 1 3 3
3 2 1
2 1 2

8 1 1 2
2 2 3
3 3 1

9 3 1 2
1 3 3
2 2 1

The output from the %ChoicEff macro consists of a list of the parameter names, values and labels, followed by
two iteration histories (each based on a different random initial design), then a brief report on the most efficient
design found, and finally a table with the parameter names, variances, df, and standard errors. The design is
printed using PROC PRINT.
The Macros 291

Here is another example. These next steps directly create an optimal design for this generic model and evaluate
its efficiency using the %ChoicEff macro and the initial design options. The DATA step creates a cyclic design.
In a cyclic design, the factor levels increase cyclically from one alternative to the next. The levels for a factor for
the three alternatives will always be one of the following: (1, 2, 3) or (2, 3, 1) or (3, 1, 2).
* Cyclic (Optimal) Design;
data x(keep=f1-f3 x1-x3);
retain f1-f3 1;
d1 = ceil(_n_ / 3); d2 = mod(_n_ - 1, 3) + 1; input d3 @@;
do i = -1 to 1;
x1 = mod(d1 + i, 3) + 1;
x2 = mod(d2 + i, 3) + 1;
x3 = mod(d3 + i, 3) + 1;
output;
end;
datalines;
1 2 3 3 1 2 2 3 1
;

proc print data=x; var x: f:; run;

Here is part of the cyclic design. Notice the cyclical pattern. Each level in the second or third alternative is one
greater than the level in the previous alternative, where 3+1 is defined to be 1. The flag variables f1-f3 contain
all ones showing that each candidate can be used in any alternative.

Obs x1 x2 x3 f1 f2 f3

1 1 1 1 1 1 1
2 2 2 2 1 1 1
3 3 3 3 1 1 1
4 1 2 2 1 1 1
5 2 3 3 1 1 1
6 3 1 1 1 1 1
.
.
.
25 3 3 1 1 1 1
26 1 1 2 1 1 1
27 2 2 3 1 1 1

This is the code that evaluates the design.

%choiceff(data=x, model=class(x1-x3), nsets=9, flags=f1-f3,
beta=zero, init=x, initvars=x1-x3, intiter=0);
The option init=x specifies the initial design, initvars=x1-x3 specifies the factors in the initial design,
and intiter=0 specifies the number of internal iterations. Specify intiter=0 when you just want to eval-
uate the efficiency of a given design. Here is the output from the %ChoicEff macro.

n Name Beta Label

1 x11 0 x1 1
2 x12 0 x1 2
3 x21 0 x2 1
4 x22 0 x2 2
5 x31 0 x3 1
6 x32 0 x3 2
292 TS-677E Multinomial Logit, Discrete Choice Modeling

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 1.732051 0.577350

Final Results: Design = 1

Efficiency = 1.7320508076
D-Error = 0.5773502692

Variable Standard
n Name Label Variance DF Error

1 x11 x1 1 0.66667 1 0.81650

2 x12 x1 2 0.66667 1 0.81650
3 x21 x2 1 0.66667 1 0.81650
4 x22 x2 2 0.66667 1 0.81650
5 x31 x3 1 0.66667 1 0.81650
6 x32 x3 2 0.66667 1 0.81650
==
6

These next steps use the %MktEx and %MktRoll macros to create a candidate set of choice sets and the
%ChoicEff macro to search for an efficient design using the candidate-set-swapping algorithm.
%mktex(3 ** 9, n=2187)

data key;
input (x1-x3) ($);
datalines;
x1 x2 x3
x4 x5 x6
x7 x8 x9
;

%mktroll(design=design, key=key, out=rolled)

%choiceff(data=rolled, model=class(x1-x3), nsets=9, nalts=3,

beta=zero, seed=446);
The first steps create a candidate set of choice sets. The %MktEx macro creates a design with nine factors, three
for each of the three alternatives. The KEY data set specifies that the first alternative is made from the linear
design factors x1-x3, the second alternative is made from x4-x6, and the third alternative is made from x7-
x9. The %MktRoll macro turns a linear design into a choice design using the rules specified in the KEY data
set.
In the %ChoicEff macro, the nalts=3 option specifies that there are three alternatives. There must always
be a constant number of alternatives in each choice set, even if all of the alternatives will not be used. When a
nonconstant number of alternatives is desired, you must use a weight variable to flag those alternatives that the
subject will not see. When you swap choice sets, you need to specify nalts=. The output from these steps is
not appreciably different from what we saw previously, so it is not shown.
This next example has brand effects and uses the alternative-swapping algorithm.
%mktex(3 ** 4, n = 3**4)
%mktlab(data=design, vars=x1-x3 Brand)
The Macros 293

data full(drop=i);
set final;
array f[3];
do i = 1 to 3; f[i] = (brand eq i); end;
run;

proc print data=full(obs=9); run;

The %MktEx macro makes the linear candidate design. The %MktLab macro changes the name of the variable
x4 to Brand while retaining the original names for x1-x3 and original values for all factors of 1, 2, and 3. The
DATA step creates the flags. The flag f1 flags brand 1 candidates as available for the first alternative, f2 flags
brand 2 candidates as available for the second alternative, and so on. The Boolean expression (brand eq i)
evaluates to 1 if true and 0 if false. Here is the first part of the candidate set.

Obs x1 x2 x3 Brand f1 f2 f3

1 1 1 1 1 1 0 0
2 1 1 1 2 0 1 0
3 1 1 1 3 0 0 1
4 1 1 2 1 1 0 0
5 1 1 2 2 0 1 0
6 1 1 2 3 0 0 1
7 1 1 3 1 1 0 0
8 1 1 3 2 0 1 0
9 1 1 3 3 0 0 1

Here is the %ChoicEff macro call for making the design.

%choiceff(data=full, seed=538,
model=class(brand brand*x1 brand*x2 brand*x3 / zero=’ ’),
nsets=15, flags=f1-f3, beta=zero, converge=1e-12);
The model= specification states that Brand and x1-x3 are classification or categorical variables and brand
effects and brand by attribute interactions (alternative-specific effects) are desired. The zero=’ ’ specification
is like zero=none except zero=none applies to all factors in the specification whereas zero=’ ’ applies
to just the first. This zero=’ ’ specification specifies that there is no reference level for the first factor (brand),
and last level will by default be the reference category for the other factors (x1-x3). Hence, binary variables are
created for all three brands, but only two binary variables are created for the 3 three-level factors. We need to do
this because we need the alternative-specific effects for all brands, including brand 3.
The option converge=1e-12 specifies a convergence criterion smaller than the default. Notice that the can-
didate set consists of branded alternatives with flags such that only brand n is considered for the nth alternative
of each choice set. In the interest of space, not all of the output is shown. Here is the output.

n Name Beta Label

1 Brand1 0 Brand 1
2 Brand2 0 Brand 2
3 Brand3 0 Brand 3
4 Brand1x11 0 Brand 1 * x1 1
5 Brand1x12 0 Brand 1 * x1 2
6 Brand2x11 0 Brand 2 * x1 1
7 Brand2x12 0 Brand 2 * x1 2
8 Brand3x11 0 Brand 3 * x1 1
9 Brand3x12 0 Brand 3 * x1 2
.
.
.
294 TS-677E Multinomial Logit, Discrete Choice Modeling

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0 .
1 0 .
0.300889 (Ridged)
2 0 .
0.304093 (Ridged)
3 0 .
0.304853 (Ridged)
4 0 .
0.304967 (Ridged)
5 0 .
0.304967 (Ridged)

Design Iteration D-Efficiency D-Error

----------------------------------------------
2 0 0 .
1 0 .
0.300306 (Ridged)
.
.
.
6 0 .
0.303801 (Ridged)

Final Results: Design = 1

Efficiency = 0
D-Error = .

Variable Standard
n Name Label Variance DF Error

1 Brand1 Brand 1 3.79498 1 1.94807

2 Brand2 Brand 2 5.58130 1 2.36248
3 Brand3 Brand 3 . 0 .
4 Brand1x11 Brand 1 * x1 1 2.10386 1 1.45047
5 Brand1x12 Brand 1 * x1 2 2.32036 1 1.52327
6 Brand2x11 Brand 2 * x1 1 2.51756 1 1.58668
7 Brand2x12 Brand 2 * x1 2 1.87117 1 1.36791
8 Brand3x11 Brand 3 * x1 1 2.27177 1 1.50724
9 Brand3x12 Brand 3 * x1 2 2.08819 1 1.44506
.
.
.
21 Brand3x32 Brand 3 * x3 2 2.19338 1 1.48100
==
20

The following is printed to the log.

Redundant Variables:

Brand3

Notice that at each step, the efficiency is zero, but a nonzero ridged value is printed. This model contains a
structural-zero coefficient in Brand3. While we need alternative-specific effects for Brand 3 (like Brand3x11
and Brand3x12), we do not need the Brand 3 effect (Brand3), which is a structural-zero effect. This can
be seen from both the ’Redundant Variables’ list and from looking at the variance and df table. The inclusion
The Macros 295

of the Brand3 term in the model makes the efficiency of the design zero. However, the %ChoicEff macro
can still optimize the goodness of the design by optimizing a ridged efficiency criterion. That is what is shown
in the iteration history. The option converge=1e-12 was specified because for this example, iteration stops
prematurely with the default convergence criterion. These next steps switch to a full-rank coding, dropping the
redundant variable Brand3, and using the output from the last step as the initial design.
%choiceff(data=full, init=best(keep=index), drop=brand3,
model=class(brand brand*x1 brand*x2 brand*x3 / zero=’ ’),
nsets=15, flags=f1-f3, beta=zero, converge=1e-12);
The option drop=brand3 is used to drop the parameter with the zero coefficient. We could have moved
the brand specification into its own class specification (separate from the alternative-specific effects) and not
specified zero=’ ’ with it (see for example page 296). However, sometimes it is easier to specify a model
with more terms than you really need, and then list the terms to drop, so that is what we illustrate here.
In this usage of init= with alternative swapping, the only part of the initial design that is required is the Index
variable. It contains indices into the candidate set of the alternatives that are used to make the initial design. This
usage is for the situation where the initial design was output from the macro. (In contrast, in the example usage
on page 291, the option initvars=x1-x3 was specified because the initial design was not created by the
%ChoicEff macro.) Here is some of the output. Notice that now there are no zero parameters so D-efficiency
can be directly computed.

Design Iteration D-Efficiency D-Error

----------------------------------------------
1 0 0.683297 1.463493
1 0.683297 1.463493

Final Results: Design = 1

Efficiency = 0.683296784
D-Error = 1.4634929117

Variable Standard
n Name Label Variance DF Error

1 Brand1 Brand 1 3.79498 1 1.94807

2 Brand2 Brand 2 5.58130 1 2.36248
3 Brand1x11 Brand 1 * x1 1 2.10386 1 1.45047
4 Brand1x12 Brand 1 * x1 2 2.32036 1 1.52327
5 Brand2x11 Brand 2 * x1 1 2.51756 1 1.58668
6 Brand2x12 Brand 2 * x1 2 1.87117 1 1.36791
7 Brand3x11 Brand 3 * x1 1 2.27177 1 1.50724
8 Brand3x12 Brand 3 * x1 2 2.08819 1 1.44506
9 Brand1x21 Brand 1 * x2 1 2.35839 1 1.53570
10 Brand1x22 Brand 1 * x2 2 2.19604 1 1.48190
11 Brand2x21 Brand 2 * x2 1 2.42179 1 1.55621
12 Brand2x22 Brand 2 * x2 2 2.14472 1 1.46449
13 Brand3x21 Brand 3 * x2 1 2.67577 1 1.63578
14 Brand3x22 Brand 3 * x2 2 2.29487 1 1.51488
15 Brand1x31 Brand 1 * x3 1 2.24390 1 1.49797
16 Brand1x32 Brand 1 * x3 2 2.24520 1 1.49840
17 Brand2x31 Brand 2 * x3 1 2.05763 1 1.43444
18 Brand2x32 Brand 2 * x3 2 2.15244 1 1.46712
19 Brand3x31 Brand 3 * x3 1 2.17488 1 1.47475
20 Brand3x32 Brand 3 * x3 2 2.19338 1 1.48100
==
20
296 TS-677E Multinomial Logit, Discrete Choice Modeling

These next steps handle the same problem, only this time, we use the set-swapping algorithm, and we will specify
a parameter vector that is not zero. At first, we will omit the beta= option, just to see the coding. We specified
the effects option in the PROC TRANSREG class specification to get -1, 0, 1 coding.
%mktex(3 ** 9, n=2187)

data key;
input (Brand x1-x3) ($);
datalines;
1 x1 x2 x3
2 x4 x5 x6
3 x7 x8 x9
;

%mktroll(design=design, key=key, alt=brand, out=rolled)

%choiceff(data=rolled, nsets=15, nalts=3,

model=class(brand)
class(brand*x1 brand*x2 brand*x3 / effects zero=’ ’))
The output tells us the parameter names and the order in which we need to specify parameters.

n Name Beta Label

1 Brand1 . Brand 1
2 Brand2 . Brand 2
3 Brand1x11 . Brand 1 * x1 1
4 Brand1x12 . Brand 1 * x1 2
5 Brand2x11 . Brand 2 * x1 1
6 Brand2x12 . Brand 2 * x1 2
7 Brand3x11 . Brand 3 * x1 1
8 Brand3x12 . Brand 3 * x1 2
9 Brand1x21 . Brand 1 * x2 1
10 Brand1x22 . Brand 1 * x2 2
11 Brand2x21 . Brand 2 * x2 1
12 Brand2x22 . Brand 2 * x2 2
13 Brand3x21 . Brand 3 * x2 1
14 Brand3x22 . Brand 3 * x2 2
15 Brand1x31 . Brand 1 * x3 1
16 Brand1x32 . Brand 1 * x3 2
17 Brand2x31 . Brand 2 * x3 1
18 Brand2x32 . Brand 2 * x3 2
19 Brand3x31 . Brand 3 * x3 1
20 Brand3x32 . Brand 3 * x3 2

Now that we are sure we know the order of the parameters, we can specify the assumed betas on the beta=
option. These numbers are based on prior research or our expectations of approximately what we expect the
parameter estimates will be. We also specified n=100 on this run, which is a sample size we are considering.
%choiceff(data=rolled, nsets=15, nalts=3, n=100, seed=543,
beta=1 2 -0.5 0.5 -0.75 0.75 -1 1
-0.5 0.5 -0.75 0.75 -1 1 -0.5 0.5 -0.75 0.75 -1 1,
model=class(brand)
class(brand*x1 brand*x2 brand*x3 / effects zero=’ ’))
The Macros 297

Here is some of the output. Notice that parameters and test statistics are incorporated into the output. The n=
value is incorporated into the variance matrix and hence the efficiency statistics, variances and tests.

Prob >
Variable Assumed Standard Squared
n Name Label Variance Beta DF Error Wald Wald

1 Brand1 Brand 1 0.012495 1.00 1 0.11178 8.9462 0.0001

2 Brand2 Brand 2 0.029691 2.00 1 0.17231 11.6068 0.0001
3 Brand1x11 Brand 1 * x1 1 0.012907 -0.50 1 0.11361 -4.4011 0.0001
4 Brand1x12 Brand 1 * x1 2 0.009894 0.50 1 0.09947 5.0267 0.0001
5 Brand2x11 Brand 2 * x1 1 0.012299 -0.75 1 0.11090 -6.7628 0.0001
6 Brand2x12 Brand 2 * x1 2 0.013671 0.75 1 0.11692 6.4144 0.0001
7 Brand3x11 Brand 3 * x1 1 0.022617 -1.00 1 0.15039 -6.6494 0.0001
8 Brand3x12 Brand 3 * x1 2 0.020324 1.00 1 0.14256 7.0145 0.0001
9 Brand1x21 Brand 1 * x2 1 0.010342 -0.50 1 0.10170 -4.9166 0.0001
10 Brand1x22 Brand 1 * x2 2 0.011604 0.50 1 0.10772 4.6416 0.0001
11 Brand2x21 Brand 2 * x2 1 0.010035 -0.75 1 0.10018 -7.4868 0.0001
12 Brand2x22 Brand 2 * x2 2 0.012963 0.75 1 0.11386 6.5873 0.0001
13 Brand3x21 Brand 3 * x2 1 0.020136 -1.00 1 0.14190 -7.0471 0.0001
14 Brand3x22 Brand 3 * x2 2 0.015137 1.00 1 0.12303 8.1279 0.0001
15 Brand1x31 Brand 1 * x3 1 0.011747 -0.50 1 0.10838 -4.6133 0.0001
16 Brand1x32 Brand 1 * x3 2 0.009208 0.50 1 0.09596 5.2105 0.0001
17 Brand2x31 Brand 2 * x3 1 0.012416 -0.75 1 0.11143 -6.7308 0.0001
18 Brand2x32 Brand 2 * x3 2 0.012723 0.75 1 0.11280 6.6491 0.0001
19 Brand3x31 Brand 3 * x3 1 0.016991 -1.00 1 0.13035 -7.6717 0.0001
20 Brand3x32 Brand 3 * x3 2 0.013543 1.00 1 0.11637 8.5931 0.0001
==
20

These next steps create a design for a cross-effects model with five brands at three prices and a constant alterna-
tive. Note the choice-set-swapping algorithm can handle cross effects but not the alternative-swapping algorithm.
%mktex(3 ** 5, n=3**5)
data key;
input (Brand Price) ($);
datalines;
1 x1
2 x2
3 x3
4 x4
5 x5
. .
;
%mktroll(design=design, key=key, alt=brand, out=rolled, keep=x1-x5)

proc print; by set; id set; where set in (1, 48, 101, 243); run;
The keep= option on the %MktRoll macro is used to keep the price variables that are needed to make the cross
effects. Here are a few of the candidate choice sets.
298 TS-677E Multinomial Logit, Discrete Choice Modeling

Set Brand Price x1 x2 x3 x4 x5

1 1 1 1 1 1 1 1
2 1 1 1 1 1 1
3 1 1 1 1 1 1
4 1 1 1 1 1 1
5 1 1 1 1 1 1
. 1 1 1 1 1

48 1 1 1 2 3 1 3
2 2 1 2 3 1 3
3 3 1 2 3 1 3
4 1 1 2 3 1 3
5 3 1 2 3 1 3
. 1 2 3 1 3

101 1 2 2 1 3 1 2
2 1 2 1 3 1 2
3 3 2 1 3 1 2
4 1 2 1 3 1 2
5 2 2 1 3 1 2
. 2 1 3 1 2

243 1 3 3 3 3 3 3
2 3 3 3 3 3 3
3 3 3 3 3 3 3
4 3 3 3 3 3 3
5 3 3 3 3 3 3
. 3 3 3 3 3

Notice that x1 contains the price for Brand 1, x2 contains the price for Brand 2, and so on, and the price of brand
i in a choice set is the same, no matter which alternative it is stored with.
Here is the %ChoicEff macro call for creating the choice design with cross effects.
%choiceff(data=rolled, seed=17,
model=class(brand brand*price / zero=none)
identity(x1-x5) * class(brand / zero=none),
nsets=20, nalts=6, beta=zero);
Cross effects are created by interacting the price factors with brand. See pages 179 and 217 for more information
about cross effects.
Here is the redundant variable list from the log.

Redundant Variables:

Brand1Price3 Brand2Price3 Brand3Price3 Brand4Price3 Brand5Price3 x1Brand1

x2Brand2 x3Brand3 x4Brand4 x5Brand5

Next, we will run the macro again, this time requesting a full-rank model. The list of dropped names was created
by copying from the redundant variable list. Also, zero=none was changed to zero=’ ’ so no level would
be zeroed for Brand but the last level of Price would be zeroed.
%choiceff(data=rolled, seed=17,
model=class(brand brand*price / zero=’ ’)
identity(x1-x5) * class(brand / zero=none),
drop=x1Brand1 x2Brand2 x3Brand3 x4Brand4 x5Brand5,
nsets=20, nalts=6, beta=zero);
The Macros 299

Here is the last part of the output. Notice that we have five brand parameters, two price parameters for each of
the five brands, and four cross effect parameters for each of the five brands.

Variable Standard
n Name Label Variance DF Error

1 Brand1 Brand 1 13.8149 1 3.71683

2 Brand2 Brand 2 13.5263 1 3.67782
3 Brand3 Brand 3 13.2895 1 3.64547
4 Brand4 Brand 4 13.5224 1 3.67728
5 Brand5 Brand 5 16.3216 1 4.04000
6 Brand1Price1 Brand 1 * Price 1 2.8825 1 1.69779
7 Brand1Price2 Brand 1 * Price 2 3.5118 1 1.87399
8 Brand2Price1 Brand 2 * Price 1 2.8710 1 1.69441
9 Brand2Price2 Brand 2 * Price 2 3.5999 1 1.89733
10 Brand3Price1 Brand 3 * Price 1 2.8713 1 1.69448
11 Brand3Price2 Brand 3 * Price 2 3.5972 1 1.89662
12 Brand4Price1 Brand 4 * Price 1 2.8710 1 1.69441
13 Brand4Price2 Brand 4 * Price 2 3.5560 1 1.88574
14 Brand5Price1 Brand 5 * Price 1 2.8443 1 1.68649
15 Brand5Price2 Brand 5 * Price 2 3.8397 1 1.95953
16 x1Brand2 x1 * Brand 2 0.7204 1 0.84878
17 x1Brand3 x1 * Brand 3 0.7209 1 0.84908
18 x1Brand4 x1 * Brand 4 0.7204 1 0.84878
19 x1Brand5 x1 * Brand 5 0.7204 1 0.84877
20 x2Brand1 x2 * Brand 1 0.7178 1 0.84722
21 x2Brand3 x2 * Brand 3 0.7178 1 0.84724
22 x2Brand4 x2 * Brand 4 0.7178 1 0.84720
23 x2Brand5 x2 * Brand 5 0.7248 1 0.85133
24 x3Brand1 x3 * Brand 1 0.7178 1 0.84722
25 x3Brand2 x3 * Brand 2 0.7178 1 0.84721
26 x3Brand4 x3 * Brand 4 0.7178 1 0.84720
27 x3Brand5 x3 * Brand 5 0.7248 1 0.85133
28 x4Brand1 x4 * Brand 1 0.7178 1 0.84722
29 x4Brand2 x4 * Brand 2 0.7178 1 0.84721
30 x4Brand3 x4 * Brand 3 0.7178 1 0.84724
31 x4Brand5 x4 * Brand 5 0.7293 1 0.85402
32 x5Brand1 x5 * Brand 1 0.7111 1 0.84325
33 x5Brand2 x5 * Brand 2 0.7180 1 0.84737
34 x5Brand3 x5 * Brand 3 0.7248 1 0.85135
35 x5Brand4 x5 * Brand 4 0.7179 1 0.84731
==
35

%ChoicEff Macro Options

The following options can be used with the %ChoicEff macro. You must specify both the model= and
nsets= options and either the flags= or nalts= options. You can omit beta= if you just want a listing of
effects, however you must specify beta= to create a design. The rest of the options are optional.

model= model-specification
specifies a PROC TRANSREG model statement list of effects. There are many potential forms for the model
specification and a number of options. See the SAS/STAT PROC TRANSREG documentation.
300 TS-677E Multinomial Logit, Discrete Choice Modeling

Generic effects example:

model=class(x1-x3),

Brand and alternative-specific effects example:

model=class(b)
class(b*x1 b*x2 b*x3 / effects zero=’ ’),
Brand, alternative-specific, and cross effects:
model=class(b b*p / zero=’ ’)
identity(x1-x5) * class(b / zero=none),

nsets= n
specifies the number of choice sets desired.

You must specify exactly one of these next two options. When the candidate set consists of individual alternatives
to be swapped, specify the alternative flags with flags=. When the candidate set consists of entire sets of
alternatives to be swapped, specify the number of alternatives in each set with nalts=.

flags= variable-list
specifies variables that flag the alternative(s) for which each candidate may be used. There must be one flag
variable per alternative. If every candidate can be used in all alternatives, then the flags are constant. For example,
with three alternatives, create these constant flags: f1=1 f2=1 f3=1. Otherwise, with three alternatives,
specify flags=f1-f3 and create a candidate set where: alternative 1 candidates are indicated by f1=1 f2=0
f3=0, alternative 2 candidates are indicated by f1=0 f2=1 f3=0, and alternative 3 candidates are indicated
by f1=0 f2=0 f3=1.

nalts= n
specifies the number of alternatives in each choice set.

The rest of the parameters are optional. You may specify zero or more of them.

bestcov= SAS-data-set
specifies a name for the data set containing the covariance matrix for the best design. By default, this data set is
called BESTCOV.

bestout= SAS-data-set
specifies a name for the data set containing the best design. By default, this data set is called BEST.

beta= list
specifies the true parameters. By default, when beta= is not specified, the macro just reports on coding. You
can specify beta=zero to assume all zeros. Otherwise specify a number list: beta=1 -1 2 -2 1 -1.
The Macros 301

converge= n
specifies the D-efficiency convergence criterion. By default, converge=0.005.

cov= SAS-data-set
specifies a name for the data set containing all of the covariance matrices for all of the designs. By default, this
data set is called COV.

data= SAS-data-set
specifies the input choice candidate set. By default, the macro uses the last data set created.

drop= variable-list
specifies a list of variables to drop from the model. If you specified a less-than-full-rank model= specification,
you can use drop= to produce a full rank coding. When there are redundant variables, the macro prints a list
that you can use in the drop= option on a subsequent run.

fixed= variable-list
specifies the variable that flags the fixed alternatives. When fixed=variable is specified, the init= data
set must contain the named variable, which indicates which alternatives are fixed (cannot be swapped out) and
which ones may be changed. Example: fixed=fixed, init=init, initvars=x1-x3

 1 - means this alternative can never be swapped out.

 0 - means this alternative is used in the initial design, but it may be swapped out.
 . - means this alternative should be randomly initialized, and it may be swapped out.

fixed= may be specified only when both init= and initvars= are specified.

init= SAS-data-set
specifies an input initial design data set. Null means a random start. One usage is to specify the bestout=
data set for an initial start. When flags= is specified, init= must contain the index variable. Example:
init=best(keep=index). When nalts= is specified, init= must contain the choice set variable. Ex-
ample: init=best(keep=set).
Alternatively, the init= data set can contain an arbitrary design, potentially created outside this macro. In that
case, you must also specify initvars=factors, where factors are the factors in the design, for example
initvars=x1-x3. When alternatives are swapped, this data set must also contain the flags= variables.
When init= is specified with initvars=, the data set may also contain a variable specified on the fixed=
option, which indicates which alternatives are fixed, and which ones can be swapped in and out.

intiter= n
specifies the maximum number of internal iterations. Specify intiter=0 to just evaluate efficiency of an
existing design. By default, intiter=10.

initvars= variable-list
specifies the factor variables in the init= data set that must match up with the variables in the data= data set.
See init=. All of these variables must be of the same type.
302 TS-677E Multinomial Logit, Discrete Choice Modeling

maxiter= n
iter= n
specifies the maximum iterations (designs to create). By default, maxiter=10.

morevars= variable-list
specifies more variables to add to the model. This option gives you the ability to specify a list of variables to
copy along as is, through the TRANSREG coding, then add them to the model.

n= n
specifies the number of observations to use in the variance matrix formula. By default, n=1.

options= options-list
specifies binary options. By default, none of these options are specified. Specify one or more of the following
values after options=.

coded
prints the coded candidate set.

detail
prints the details of the swaps.

nocode
skips the PROC TRANSREG coding stage, assuming that [Link]– CAND was created by a previ-
ous step. This is most useful with set swapping when the candidate set can be big. It is important with
options=nocode to note that the effect of morevars= and drop= in previous runs has already been
taken care of, so do not specify them (unless for instance you want to drop still more variables).

nodups
prevents the same choice set from coming out more than once. This option does not affect the initialization,
so the random initial design may have duplicates. This options forces duplicates out during the iterations,
so do not set intiter= to a small value. It may take several iterations to eliminate all duplicates. It is
possible that efficiency will decrease as duplicates are forced out. With set swapping, this macro checks
the candidate choice set numbers to avoid duplicates. With alternative swapping, this macro checks the
candidate alternative index to avoid duplicates. The macro does not look at the actual factors. This makes
the checks faster, but if the candidate set contains duplicate choice sets or alternatives, the macro may not
succeed in eliminating all duplicates. Run the %MktDups macro (which looks at the actual factors) on the
design to check and make sure all duplicates are eliminated. If you are using set swapping to make a generic
design make sure you run the %MktDups macro on the candidate set to eliminate duplicate choice sets in
advance.

notes
stops the macro from submitting the statement options nonotes.

notests
suppresses printing the diagonal of the covariance matrix, and hypothesis tests for this n and . When
is not zero, the results include a Wald test statistic ( divided by the standard error), which is normally
distributed, and the probability of a larger squared Wald statistic.

orthcan
orthogonalizes the candidate set.
The Macros 303

out= SAS-data-set
specifies a name for the output SAS data set with the final designs. The default is out=results.

seed= n
specifies the random number seed. By default, seed=0, and clock time is used as the random number seed. By
specifying a random number seed, results should be reproducible within a SAS release for a particular operating
system. However, due to machine differences, some results may not be exactly reproducible on other machines,
although you would expect the efficiency differences to be slight.

submat= number-list
specifies a submatrix for which efficiency calculations are desired. Specify an index vector. For example, with
3 three-level factors, a, b, and c, and the model class(a b c a*b), specify submat=1:6, to see the
efficiency of just the 6  6 matrix of main effects. Specify submat=3:6, to see the efficiency of just the 4  4
matrix of b and c main effects.

types= integer-list
specifies the number of sets of each type to put into the design. This option is used when you have multiple
types of choice sets and you want the design to consist of only certain numbers of each type. This option can
be specified with the set-swapping algorithm. The argument is an integer list. When you specify types=, you
must also specify typevar=. Say you are creating a design with 30 choice sets, and you want the first 10 sets
to consist of sets whose typevar= variable in the candidate set is type 1, and you want the rest to be type 2.
You would specify types=10 20.

typevar= variable
specifies a variable in the candidate data set that contains choice set types. The types must be integers starting
with 1. This option can only be specified with the set-swapping algorithm. When you specify typevar=, you
must also specify types=.

weight= weight-variable
specifies an optional weight variable. Typical usage is with an availability design. Give unavailable alternatives
a weight of zero and available alternatives a weight of one. The number of alternatives must always be constant,
so varying numbers of alternatives are handled by giving unavailable or unseen alternatives a weight of zero.

%MktAllo Macro
The %MktAllo autocall macro is used for manipulating data for an allocation choice experiment. It takes as
input a data set with one row for each alternative of each choice set. For example, in a study with 10 brands plus
a constant alternative and 27 choice sets, there are 27  11 = 297 observations in the input data set. Here is
an example of an input data set. It contains a choice set variable, product attributes (Brand and Price) and a
frequency variable (Count) that contains the total number of times that each alternative was chosen.

Obs Set Brand Price Count

1 1 0
2 1 Brand 1 $50 103
3 1 Brand 2 $75 58
4 1 Brand 3 $50 318
5 1 Brand 4 $100 99
6 1 Brand 5 $100 54
304 TS-677E Multinomial Logit, Discrete Choice Modeling

7 1 Brand 6 $100 83
8 1 Brand 7 $75 71
9 1 Brand 8 $75 58
10 1 Brand 9 $75 100
11 1 Brand 10 $50 56
.
.
.
296 27 Brand 9 $100 94
297 27 Brand 10 $50 65

The end result is a data set with twice as many observations that contains the number of times each alternative
was chosen and the number of times it was not chosen. This data set also contains a variable c with values 1 for
first choice and 2 for second or subsequent choice.

Obs Set Brand Price Count c

1 1 0 1
2 1 1000 2
3 1 Brand 1 $50 103 1
4 1 Brand 1 $50 897 2
5 1 Brand 2 $75 58 1
6 1 Brand 2 $75 942 2
7 1 Brand 3 $50 318 1
8 1 Brand 3 $50 682 2
.
.
.
593 27 Brand 10 $50 65 1
594 27 Brand 10 $50 935 2

Here is an example of usage:

%mktallo(data=allocs2, out=allocs3, nalts=11,
vars=set brand price, freq=Count)
The option data= names the input data set, out= names the output data set, nalts= specifies the number of
alternatives, vars= names the variables in the data set that will be used in the analysis excluding the freq=
variable, and freq= names the frequency variable.

%MktAllo Macro Options

The following options can be used with the %MktAllo macro. You must specify the nalts=, freq=, and
vars= options.

data= SAS-data-set
specifies the input SAS data set. By default, the macro uses the last data set created.

freq= variable
specifies the frequency variable, which contains the number of times this alternative was chosen. This option
must be specified.
The Macros 305

nalts= n
specifies the number of alternatives (including if appropriate the constant alternative). This option must be
specified.

out= SAS-data-set
specifies the output SAS data set. The default is out=allocs.

vars= variable-list
specifies the variables in the data set that will be used in the analysis but not the freq= variable. This option
must be specified.

%MktBal Macro
The %MktBal macro creates linear experimental designs using an algorithm that ensures that the design is
perfectly balanced, or in the case when the number of levels of a factor does not divide the number of runs,
as close to perfectly balanced as possible. Do not use the %MktBal macro until you have tried the %MktEx
macro and determined that it does not make a design that is balanced enough for your needs. The %MktEx
macro can directly create hundreds of orthogonal and balanced designs that the %MktBal algorithm will never
be able to find. Even when the %MktEx macro cannot create an orthogonal and balanced design, it will usually
find a nearly balanced design. Designs created with the %MktBal macro, while perfectly balanced, may be less
efficient than designs found with the %MktEx macro, and for large problems, the %MktBal macro can be slow.
The %MktBal macro has several options that can make it run faster for large problems, but at a price of an
even further decrease in D-efficiency. It is likely that the current algorithm used by the %MktBal macro will be
changed in the future to use some now unknown algorithm that is both faster and better.
The %MktBal macro is not a full-featured experimental design generator. For example, you cannot specify
interactions that you want to estimate or specify restrictions such as which levels may or may not appear together.
You must use the %MktEx macro for that. The %MktBal macro builds a design by creating a balanced first
factor, optimally blocking it to create the second factor, then optimally blocking the first two factors to create the
third, and so on. Once it creates all factors, it refines each factor. Each factor is in turn removed from the design,
and the rest of the design is reblocked, replacing the initial factor if the new design is more D-efficient.
Here is a simple example of creating a design with 2 two-level factors and 3 three-level factors in 18 runs. The
%MktEval macro evaluates the results. This design is in fact optimal.
%mktbal(2 2 3 3 3, n=18, seed=151)
%mkteval;
In all cases, the factors are named x1, x2, x3, ... and so on.
This next example, at 120 runs and with factor levels greater than 5, is starting to get big and hence, by default,
will run slowly. You can use the maxstarts= and maxiter= options to make the macro run more quickly.
For example, the second example below runs much faster than the first.
%mktbal(2 3 4 5 6 7 8 9 10, n=120, options=progress, seed=17)

%mktbal(2 3 4 5 6 7 8 9 10, n=120, options=progress, seed=17,

maxstarts=1, maxiter=1)
306 TS-677E Multinomial Logit, Discrete Choice Modeling

%MktBal Macro Options

The following options can be used with the %MktBal macro.

list
specifies a list of the numbers of levels of all the factors. For example, for 3 two-level factors specify ei-
ther 2 2 2 or 2 ** 3. Lists of numbers, like 2 2 3 3 4 4 or a levels**number of factors syntax like:
2**2 3**2 4**2 can be used, or both can be combined: 2 2 3**4 5 6. The specification 3**4 means
four three-level factors. You must specify a list. Note that the factor list is a positional parameter. This means it
must come first, and unlike all other parameters, it is not specified after a name and an equal sign.

n= n
specifies the number of runs in the design. You must specify n=. You can use the %MktRuns macro to get
suggestions for values of n=.

out= SAS-data set

specifies the output experimental design. The default is out=design.

These next options, control some of the details of the %MktBal macro.

maxiter= n
iter= n
specifies the maximum iterations (designs to create). By default, maxiter=5.

maxstarts= n
specifies the maximum number of random starts for each factor. With larger values, the macro tends to find
slightly better designs at a cost of slower run times. The default is maxstarts=10.

maxtries= n
specifies the maximum number of times to try refining each factor after the initialization stage. The default is
maxtries=10.

options= options-list
specifies binary options. By default, none of these options are specified. Specify one or more of the following
values after options=.

noprint
specifies that the final D-efficiency should not be printed.

progress
reports on the macro’s progress. For large numbers of factors, a large number or runs, or when the number
of levels is large, this macro is slow. The options=progress specification gives you information about
which step is being executed.
The Macros 307

seed= n
specifies the random number seed. By default, seed=0, and clock time is used as the random number seed. By
specifying a random number seed, results should be reproducible within a SAS release for a particular operating
system. However, due to machine differences, some results may not be exactly reproducible on other machines,
although you would expect the efficiency differences to be slight.

%MktBlock Macro
The %MktBlock autocall macro is used to block a choice design or an ordinary linear experimental design.
When a choice design is too large to show all choice sets to each subject, the design is blocked and a block of
choice sets is shown to each subject. For example, if there are 36 choice sets, instead of showing each subject
36 sets, you could instead create 2 blocks and show 2 groups of subjects 18 sets each. You could also create 3
blocks of 12 choice sets or 4 blocks of 9 choice sets. You can also request just one block if you want to see the
correlations and frequencies among all of the attributes of all of the alternatives of a choice design.
The design can be in one of two formats. Typically, a choice design has one row for each alternative of each
choice set and one column for each of the attributes. Typically, this kind of design is produced by either the
%ChoicEff or %MktRoll macro. Alternatively, a “linear” design is an intermediate step in preparing some
choice designs. The linear design has one row for each choice set and one column for each attribute of each
alternative. Typically, the linear design is produced by the %MktEx macro. The output from the %MktBlock
macro is a data set containing the design, with the blocking variable added and hence not in the original order,
with runs or choice sets nested within blocks.
The macro tries to create a blocking factor that is uncorrelated with every attribute of every alternative. In other
words, the macro is trying to optimally add one additional factor, a blocking factor, to the linear design. It is
trying to make a factor that is orthogonal to all of the attributes of all of the alternatives. For linear designs, you
can usually just ask for a blocking factor directly as just another factor in the design, and then use the %MktLab
macro to provide a name like Block, or you can use the %MktBlock macro.
Here is an example of creating the blocking variable directly.
%mktex(3 ** 7, n=27)

%mktlab(vars=x1-x6 Block)
Here is an example of creating a design then blocking it.
%mktex(3 ** 6, n=27, seed=350)

%mktblock(data=randomized, nblocks=3, seed=377)

The output shows that the blocking factor is uncorrelated with all of the factors in the design. This output comes
from the %MktEval macro, which is called by the %MktBlock macro.

Canonical Correlations Between the Factors

There are 0 Canonical Correlations Greater Than 0.316

Block x1 x2 x3 x4 x5 x6

Block 1 0 0 0 0 0 0
x1 0 1 0 0 0 0 0
x2 0 0 1 0 0 0 0
x3 0 0 0 1 0 0 0
x4 0 0 0 0 1 0 0
x5 0 0 0 0 0 1 0
x6 0 0 0 0 0 0 1
308 TS-677E Multinomial Logit, Discrete Choice Modeling

Summary of Frequencies
There are 0 Canonical Correlations Greater Than 0.316

Frequencies

Block 9 9 9
x1 9 9 9
x2 9 9 9
x3 9 9 9
x4 9 9 9
x5 9 9 9
x6 9 9 9
Block x1 3 3 3 3 3 3 3 3 3
Block x2 3 3 3 3 3 3 3 3 3
Block x3 3 3 3 3 3 3 3 3 3
Block x4 3 3 3 3 3 3 3 3 3
Block x5 3 3 3 3 3 3 3 3 3
Block x6 3 3 3 3 3 3 3 3 3
x1 x2 3 3 3 3 3 3 3 3 3
x1 x3 3 3 3 3 3 3 3 3 3
x1 x4 3 3 3 3 3 3 3 3 3
x1 x5 3 3 3 3 3 3 3 3 3
x1 x6 3 3 3 3 3 3 3 3 3
x2 x3 3 3 3 3 3 3 3 3 3
x2 x4 3 3 3 3 3 3 3 3 3
x2 x5 3 3 3 3 3 3 3 3 3
x2 x6 3 3 3 3 3 3 3 3 3
x3 x4 3 3 3 3 3 3 3 3 3
x3 x5 3 3 3 3 3 3 3 3 3
x3 x6 3 3 3 3 3 3 3 3 3
x4 x5 3 3 3 3 3 3 3 3 3
x4 x6 3 3 3 3 3 3 3 3 3
x5 x6 3 3 3 3 3 3 3 3 3
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1

Canonical Correlations Between the Factors by Block

Block x1 x2 x3 x4 x5 x6

1 x1 1 0 0.58 0.58 0.58 0.58

x2 0 1 0 0 0 0
x3 0.58 0 1 0 0.58 0.58
x4 0.58 0 0 1 0.58 0.58
x5 0.58 0 0.58 0.58 1 0
x6 0.58 0 0.58 0.58 0 1

2 x1 1 0 0.58 0.58 0.58 0.58

x2 0 1 0 0 0 0
x3 0.58 0 1 0 0.58 0.58
x4 0.58 0 0 1 0.58 0.58
x5 0.58 0 0.58 0.58 1 0
x6 0.58 0 0.58 0.58 0 1
The Macros 309

3 x1 1 0 0 0 0 0
x2 0 1 0 0 0 0
x3 0 0 1 0 1.00 0
x4 0 0 0 1 0 1.00
x5 0 0 1.00 0 1 0
x6 0 0 0 1.00 0 1

Notice that even with a perfect blocking variable like we have in this example, canonical correlations within each
block will not be all zero.
Here is the blocked linear design (3 blocks of nine choice sets). Note that in the linear version of the design,
there is one row for each choice set and all of the attributes of all of the alternatives are in the same row.

Block Run x1 x2 x3 x4 x5 x6

1 1 3 2 2 2 3 3
2 3 3 1 3 2 1
3 3 1 2 3 2 3
4 1 2 1 1 2 2
5 1 3 3 2 1 3
6 2 1 3 1 3 1
7 2 2 3 3 1 1
8 2 3 2 1 3 2
9 1 1 1 2 1 2

Block Run x1 x2 x3 x4 x5 x6

2 1 3 2 1 1 1 1
2 3 3 3 2 3 2
3 3 1 3 1 1 2
4 2 3 1 3 1 3
5 1 1 2 3 3 1
6 1 3 2 1 2 1
7 1 2 3 3 3 3
8 2 1 1 2 2 3
9 1 2 3 2 1 3

Block Run x1 x2 x3 x4 x5 x6

3 1 1 3 1 3 3 2
2 2 2 1 1 3 3
3 2 1 2 3 1 2
4 2 3 3 2 2 1
5 1 1 3 1 2 3
6 3 2 3 3 2 2
7 3 1 1 2 3 1
8 1 2 2 2 1 1
9 3 3 2 1 1 3
310 TS-677E Multinomial Logit, Discrete Choice Modeling

Next, we’ll create and block a choice design with two blocks of nine sets instead of blocking the linear version
of a choice design.
%mktex(3 ** 6, n=3**6)

* Create an efficient choice design;

data key;
input (x1-x3) ($);
datalines;
x1 x2 x3
x4 x5 x6
;

%mktroll(design=design, key=key, out=out)

%choiceff(data=out, model=class(x1-x3), nsets=18, nalts=2,

beta=zero, options=nodups, seed=151)

* Block the choice design. Ask for 2 blocks;

%mktblock(data=best, nalts=2, nblocks=2, factors=x1-x3, seed=472)
(Note that if this had been a branded example, and if you were running SAS version 8.2 or an earlier release,
specify id=brand; do not add your brand variable to the factor list. For version 9 and later SAS versions, it is
fine to add your brand variable to the factor list.)
Both the design and the blocking are not as good this time. The variable names in the output are composed of
Alt, the alternative number, and the factor name. Since there are two alternatives each composed of three factors
plus one blocking variable (2  3 + 1 = 7), a 7  7 correlation matrix is reported. Here is some of the output.

Canonical Correlations Between the Factors

There are 11 Canonical Correlations Greater Than 0.316

Block Alt1_x1 Alt1_x2 Alt1_x3 Alt2_x1 Alt2_x2 Alt2_x3

Block 1 0.13 0 0 0.15 0.13 0

Alt1_x1 0.13 1 0.36 0.33 0.63 0.29 0.26
Alt1_x2 0 0.36 1 0.47 0.34 0.59 0.47
Alt1_x3 0 0.33 0.47 1 0.37 0.30 0.60
Alt2_x1 0.15 0.63 0.34 0.37 1 0.23 0.36
Alt2_x2 0.13 0.29 0.59 0.30 0.23 1 0.35
Alt2_x3 0 0.26 0.47 0.60 0.36 0.35 1

Summary of Frequencies
There are 11 Canonical Correlations Greater Than 0.316
* - Indicates Unequal Frequencies

Frequencies

Block 9 9
* Alt1_x1 7 7 4
* Alt1_x2 8 2 8
* Alt1_x3 8 4 6
* Alt2_x1 5 5 8
* Alt2_x2 5 9 4
* Alt2_x3 4 8 6
The Macros 311

* Block Alt1_x1 4 3 2 3 4 2
* Block Alt1_x2 4 1 4 4 1 4
* Block Alt1_x3 4 2 3 4 2 3
* Block Alt2_x1 2 3 4 3 2 4
* Block Alt2_x2 3 4 2 2 5 2
* Block Alt2_x3 2 4 3 2 4 3
* Alt1_x1 Alt1_x2 3 1 3 4 1 2 1 0 3
* Alt1_x1 Alt1_x3 4 2 1 3 1 3 1 1 2
* Alt1_x1 Alt2_x1 0 4 3 2 0 5 3 1 0
* Alt1_x1 Alt2_x2 3 3 1 1 4 2 1 2 1
* Alt1_x1 Alt2_x3 1 4 2 2 2 3 1 2 1
* Alt1_x2 Alt1_x3 4 1 3 2 0 0 2 3 3
* Alt1_x2 Alt2_x1 1 3 4 1 0 1 3 2 3
* Alt1_x2 Alt2_x2 0 5 3 1 0 1 4 4 0
* Alt1_x2 Alt2_x3 2 2 4 0 2 0 2 4 2
* Alt1_x3 Alt2_x1 3 3 2 1 1 2 1 1 4
* Alt1_x3 Alt2_x2 2 5 1 2 1 1 1 3 2
* Alt1_x3 Alt2_x3 0 5 3 1 0 3 3 3 0
* Alt2_x1 Alt2_x2 1 3 1 1 3 1 3 3 2
* Alt2_x1 Alt2_x3 0 3 2 2 2 1 2 3 3
* Alt2_x2 Alt2_x3 0 3 2 3 3 3 1 2 1
N-Way 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

Note that in this example, the input is a choice design (as opposed to the linear version of a choice design) so the
results are in choice design format. There is one row for each alternative of each choice set.

Block Set Alt x1 x2 x3

1 1 1 3 1 3
2 2 3 1

1 2 1 2 1 1
2 1 2 3

.
.
.

Block Set Alt x1 x2 x3

2 1 1 2 1 1
2 3 2 3

2 2 1 2 2 1
2 1 3 2

.
.
.
312 TS-677E Multinomial Logit, Discrete Choice Modeling

%MktBlock Macro Options

The following options can be used with the %MktBlock macro.

alt= variable
specifies the variable to contain the alternative number. If this variable is in the input data set, it is excluded from
the factor list. The default is alt=Alt.

block= variable
specifies the variable to contain the block number. If this variable is in the input data set, it is excluded from the
factor list. The default is block=Block.

data= SAS-data-set
specifies either the choice or linear design. The choice design has one row for each alternative of each choice
set and one column for each of the attributes. Typically this design is produced by either the %ChoicEff
or %MktRoll macro. For choice designs, you must also specify the nalts= option. The default is
data=- last- . The linear design has one row for each choice set and one column for each attribute of each
alternative. Typically this design is produced by the %MktEx macro. This is the design that is input into the
%MktRoll macro.

factors= variable-list
vars= variable-list
specifies the factors in the design. By default, all numeric variables are used, except variables with names
matching those in the block=, set=, and alt= options. (By default, the variables Block, Set, Run, and
Alt are excluded from the factor list.) If you are using version 8.2 or an earlier SAS release with a branded
choice design (assuming the brand factor is called Brand), specify id=Brand. Do not add the brand factor to
the factor list unless you are using version 9.0 or a later SAS release.

id= variable-list
specifies variables in the data= data set to copy to the output data set. If you are using version 8.2 or an earlier
SAS release with a branded choice design (assuming the brand factor is called Brand), specify id=Brand. Do
not add the brand factor to the factor list unless you are using version 9.0 or a later SAS release.

initblock= variable
specifies the name of a variable in the data set that is to be used as the initial blocking variable for the first
iteration.

maxiter= n
iter= n
specifies the number of times to try to block the design starting with a different random blocking. By default,
the macro tries five random starts, and iteratively refines each until D-efficiency quits improving, then in the end
selects the blocking with the best D-efficiency.

nalts= n
specifies the number of alternatives in each choice set. If you are inputting a choice design, you must specify
nalts=, otherwise the macro assumes you are inputting a linear design.
The Macros 313

nblocks= n
specifies the number of blocks to create. The option nblocks=1 just reports information about the design. The
nblocks= option must be specified.

next= n
specifies how far into the design to go to look for the next swap. The specification next=1 specifies that
the macro should try swapping the level for each run with the level for the next run and all other runs. The
specification next=2 considers swaps with half of the other runs, which makes the algorithm run more quickly.
The macro considers swapping the level for run i with run i + 1 then uses the next= value to find the next
potential swaps. Other values, including nonintegers can be specified as well. For example next=1.5 considers
swapping observation 1 with observations 2, 4, 5, 7, 8, 10, 11, and so on. With smaller values, the macro tends
to find a slightly better blocking variable at a cost of much slower run time.

out= SAS-data-set
specifies the output data set with the block numbers. The default is out=blocked.

print= print-options
specifies both the %MktBlock and the %MktEval macro printing options, which control the printing of the
results. The default is print=normal. Values include:
corr canonical correlations
list list of big canonical correlations
freqs long frequencies list
summ frequency summaries
block canonical correlations within blocks
design blocked design
note blocking note
all all of the above
noprint no printed output
normal corr list summ block design note
short corr summ note

ridge= n
specifies the value to add to the diagonal of (X0 X) 1 to make it nonsingular. Usually, you will not need to
change this value. If you do, you probably will not notice any effect. Specify ridge=0 to use a generalized
inverse instead of ridging. The default is ridge=0.01.

seed= n
specifies the random number seed. By default, seed=0, and clock time is used as the random number seed. By
specifying a random number seed, results should be reproducible within a SAS release for a particular operating
system. However, due to machine differences, some results may not be exactly reproducible on other machines,
although you would expect the efficiency differences to be slight.

set= variable
specifies the variable to contain the choice set number. When nalts= is specified, the default is Set, otherwise
the default is Run. If this variable is in the input data set, it is excluded from the factor list.
314 TS-677E Multinomial Logit, Discrete Choice Modeling

%MktDes Macro
The %MktDes autocall macro creates efficient experimental designs. Throughout this report, we used the
%MktEx autocall macro, which calls the %MktDes macro, to design our experiments, Usually, we will not
need to call the %MktDes macro directly. At the heart of the %MktDes macro are PROC PLAN, PROC FAC-
TEX, and PROC OPTEX. We use macros instead of calling these procedures directly because the macros have a
simpler syntax. In extreme cases, a single-line macro call can generate hundreds of lines of otherwise tedious to
write procedure code.
The %MktDes macro creates efficient experimental designs. You specify the names of the factors and the number
of levels for each factor. You also specify the number of runs you want in your final design. Here for example is
how you can create a design in 18 runs with 2 two-level factors (x1 and x2) and 3 three-level factors (x3, x4,
and x5).
%mktdes(factors=x1-x2=2 x3-x5=3, n=18)

You can also optionally specify interactions that you want to be estimable. The macro creates a candidate design
in which every effect you want to be estimable is estimable, but the candidate design is bigger than you want.
By default, the candidate set is stored in a SAS data set called CAND1. The macro then uses PROC OPTEX to
search the candidate design for an efficient final design. By default, the final experimental design is stored in a
SAS data set called DESIGN.
When the full-factorial design is small (by default less than 2189 runs, although sizes up to 5000 or 6000 runs
are reasonably small) the experimental design problem is straightforward. First, the macro uses PROC PLAN to
create a full-factorial candidate set. Next, PROC OPTEX searches the full-factorial candidate set. For very small
problems (a few hundred candidates) PROC OPTEX will often find the optimal design, and for larger problems,
it may not find the optimal design, but given sufficient iteration (for example, specify iter=100 or more) it will
find very good designs. Run time will typically be a few seconds or a few minutes, but it could run longer. Here is
a typical example of using the %MktDes macro to find an optimal nonorthogonal design when the full-factorial
design is small (108 runs):
*---Two two-level factors and 3 three-level factors in 18 runs---;
%mktdes(factors=x1-x2=2 x3-x5=3, n=18, maxiter=500)
When the full-factorial design is larger, the macro uses PROC FACTEX to create a fractional-factorial candidate
set. In those cases, the other methods found in the %MktEx macro usually make better designs than those found
with the %MktDes macro.

%MktDes Macro Options

The following options can be used with the %MktDes macro.

big= n
specifies the size at which the candidate set is considered to be big. By default, big=2188. If the size of the
full-factorial design is less than or equal to this size, and if PROC PLAN is in the run= list, the macro uses
PROC PLAN instead of PROC FACTEX to create the candidate set. The default of 2188 is max(211 ; 37 ) + 1).
Specifying values as large as big=6000 or even slightly more is often reasonable. However, run time is slower
as the size of the candidate set increases. The %MktEx macro coordinate-exchange algorithm will usually work
better than a candidate-set search when the full-factorial design has more than several thousand runs.
The Macros 315

cand= SAS-data-set
specifies the output data set with the candidate design (from PROC FACTEX or PROC PLAN). The default name
is Cand followed by the step number, for example: Cand1 for step 1, Cand2 for step 2, and so on. You should
only use this option when you are reading an external candidate set. When you specify step= values greater
than 1, the macro assumes the default candidate set names, CAND1, CAND2, and so on, were used in previous
steps. Specify just a data set name, no data set options.

classopts= options
specifies PROC OPTEX class statement options. The default, is classopts=param=orthref. You
probably never want to change this option.

coding= name
specifies the PROC OPTEX coding= option. This option is usually not needed.

examine= I | V
specifies the matrices that you want to examine. The option examine=I prints the information matrix, X0 X;
examine=V prints the variance matrix, (X0 X) 1 ; and examine=I V prints both. By default, these matrices
are not printed.

facopts= options
specifies PROC FACTEX statement options.

factors= factor-list
specifies the factors and the number of levels for each factor. The factors= option must be specified. All other
options are optional. Optionally, the number of pseudo-factors can also be specified. Here is a simple example
of creating a design with 10 two-level factors.
%mktdes(factors=x1-x10=2)

First, a factor list, which is a valid SAS variable list, is specified. The factor list must be followed by an equal
sign and an integer, which gives the number of levels. Multiple lists can be specified. For example, to create 5
two-level factors, 5 three-level factors, and 5 five-level factors, specify:
%mktdes(factors=x1-x5=2 x6-x10=3 x11-x15=5)

By default, this macro creates each factor from a minimum number of pseudo-factors. Pseudo-factors are not
output. They are used to create the factors of interest and then discarded. For example, with nlev=2, a three-
level factor x1 is created from 2 two-level pseudo-factors (- 1 and - 2) and their interaction by coding down:
(_1=1, _2=1) -> x1=1
(_1=1, _2=2) -> x1=2
(_1=2, _2=1) -> x1=3
(_1=2, _2=2) -> x1=1
This creates imbalance the 1 level appears twice as often as 2 and 3. Somewhat better balance can be obtained
by instead using three pseudo-factors. The number of pseudo-factors is specified in parentheses after the number
of levels. Example:
%mktdes(factors=x1-x5=2 x6-x10=3(3))
316 TS-677E Multinomial Logit, Discrete Choice Modeling

The levels 1 to 8 are coded down to 1 2 3 1 2 3 1 3, which is a little better balanced. The cost is candidate-set
size may increase and efficiency may actually decrease. Some researchers are willing to sacrifice a little bit of
efficiency in order to achieve better balance.

generate= options
specifies the PROC OPTEX generate statement options. By default, additional options are not added to the
generate statement.

interact= interaction-list
specifies interactions that must be estimable. By default, no interactions are guaranteed to be estimable. Exam-
ples:
interact=x1*x2
interact=x1*x2 x3*x4*x5
interact=x1|x2|x3|x4|x5@2
The interaction syntax is like PROC GLM’s and many of the other modeling procedures. It uses “*” for simple
interactions (x1*x2 is the interaction between x1 and x2), “|” for main effects and interactions (x1|x2|x3
is the same as x1 x2 x1*x2 x3 x1*x3 x2*x3 x1*x2*x3) and “@” to eliminate higher-order interac-
tions (x1|x2|x3@2 eliminates x1*x2*x3 and is the same as x1 x2 x1*x2 x3 x1*x3 x2*x3). The
specification “@2” allows only main effects and two-way interactions. Only “@” values of 2 or 3 are allowed.

iter= n
maxiter= n
specifies the PROC OPTEX iter= option which creates n designs. By default, iter=10.

keep= n
specifies the PROC OPTEX keep= option which keeps n designs. By default, keep=5.

nlev= n
specifies the number of levels from which factors are constructed through pseudo-factors and coding down. The
value must be a prime or a power of a prime: 2, 3, 4, 5, 7, 8, 9, 11 .... This option is used with PROC FACTEX:
factors factors / nlev=&nlev;

By default, the macro uses the minimum prime or power of a prime from the factors= list or 2 if no suitable
value is found.

method= name
specifies the PROC OPTEX method= search method option. The default is method=m- federov (modified
Federov).

n= n|SATURATED
specifies the PROC OPTEX n= option, which is the number of runs in the final design. The default is the PROC
OPTEX default and depends on the problem. Typically, you will not want to use the default. Instead, you should
pick a value using the information produced by the %MktRuns macro as guidance. The n=saturated option
creates a design with the minimum number of runs.
The Macros 317

options= options-list
specifies binary options. By default, none of these options are specified. Specify one or more of the following
values after options=.

check
checks the efficiency of a given design, specified in cand=.

nocode
suppresses printing the PROC PLAN, PROC FACTEX, and PROC OPTEX code.

allcode
shows all code, even code that will not be run.

otherfac= variable-list
specifies other terms to mention in the factors statement of PROC FACTEX. These terms are not guaranteed
to be estimable. By default, there are no other factors.

otherint= terms
specifies interaction terms that will only be specified with PROC OPTEX for multi-step macro invocations.
By default, no interactions are guaranteed to be estimable. Normally, interactions that are specified via the
interact= option affect both the PROC FACTEX and the PROC OPTEX model statements. In multi-step
problems, part of an interaction may not be in a particular PROC FACTEX step. In that case, the interaction
term must only appear in the PROC OPTEX step. For example, if x1 is created in one step and x4 is created in
another, and if the x1*x4 interaction must be estimable, specify otherint=x1*x4 on the final step, the one
that runs PROC OPTEX.
%mktdes(step=1, factors=x1-x3=2, n=30, run=factex)

%mktdes(step=2, factors=x4-x6=3, n=30, run=factex)

%mktdes(step=3, factors=x7-x9=5, n=30, run=factex optex,

otherint=x1*x4)

out= SAS-data-set
specifies the output experimental design (from PROC OPTEX). By default, out=design.

procopts= options
specifies PROC OPTEX statement options. By default, no options are added to the PROC OPTEX statement.

run= procedure-list
specifies the list of procedures that the macro may run. Normally, the macro runs either PROC FACTEX or PROC
PLAN and then PROC OPTEX. By default, run=plan factex optex. You can skip steps by omitting
procedure names from this list. When both PLAN and FACTEX are in the list, the macro chooses between them
based on the size of the full-factorial design and the value of big=. When PLAN is not in the list, the macro
generates code for PROC FACTEX.
318 TS-677E Multinomial Logit, Discrete Choice Modeling

seed= n
specifies the random number seed. By default, seed=0, and clock time is used as the random number seed. By
specifying a random number seed, results should be reproducible within a SAS release for a particular operating
system. However, due to machine differences, some results may not be exactly reproducible on other machines,
although you would expect the efficiency differences to be slight.

size= n|MIN
specifies the candidate-set size. Start with the default size=min and see how big that design is. If you want,
subsequently you can specify larger values that are nlev=n multiples of the minimum size. This option is used
with PROC FACTEX:
size design=&size;

Say you specified nlev=2 or the macro defaulted to nlev=2. Increase the size= value by a factor of two
each time. For example, if size=min implies size=128, then 256, 512, 1024, and 2048 are reasonable sizes
to try. Integer expressions like size=128*4 are allowed.

step= n
specifies the step number. By default, there is only one step. However, sometimes, a better design can be found
using a multi-step approach. Do not specify the cand= option on any step of a multistep run. Consider the
problem of making a design with 3 two-level factors, 3 three-level factors, and 3 five-level factors. The simplest
approach is to do something like this create a design from two-level factors using pseudo-factors and coding
down.
%mktdes(factors=x1-x3=2 x4-x6=3 x7-x9=5, n=30)

However, for small problems like this, the following three-step approach will usually be better.
%mktdes(step=1, factors=x1-x3=2, n=30, run=factex)
%mktdes(step=2, factors=x4-x6=3, n=30, run=factex)
%mktdes(step=3, factors=x7-x9=5, n=30, run=factex optex)
Note however, that the following %MktEx macro call will usually be better still.
%mktex(2 2 2 3 3 3 5 5 5, n=30)

Returning to the %MktDes macro, the first step uses PROC FACTEX to create a fractional-factorial design for the
two-level factors. The second step uses PROC FACTEX to create a fractional-factorial design for the three-level
factors and cross it with the two-level factors. The third step uses PROC FACTEX to create a fractional-factorial
design for the five-level factors and cross it with the design for the two and three-level factors and then run PROC
OPTEX.
Each step globally stores two macro variables (&class1 and &inter1 for the first step, &class2 and &in-
ter2 for the second step, ...) that are used to construct the PROC OPTEX class and model statements. When
step > 1, variables from the previous steps are used in the class and model statements. In this example, the
following PROC OPTEX code is created by step 3:
proc optex data=Cand3;
class
x1-x3
x4-x6
x7-x9
/ param=orthref;
model
x1-x3
x4-x6
x7-x9
;
The Macros 319

generate n=30 iter=10 keep=5 method=m_federov;

output out=Design;
run; quit;
This step uses the previously stored macro variables &class1=x1-x3 and &class2=x4-x6.

where= where-clause
specifies a SAS where clause for the candidate design, which is used to restrict the candidates. By default, the
candidate design is not restricted.

%MktDups Macro
The %MktDups autocall macro detects duplicate choice sets and duplicate alternatives within generic choice
sets. For example, consider a simple experiment with these two choice sets. These choice sets are completely
different and are not duplicates.

a b c a b c
1 2 1 1 1 1
2 1 2 2 2 2
1 1 2 2 2 1
2 1 1 1 2 2

Now consider these two choice sets:

a b c a b c
1 2 1 2 1 2
2 1 2 1 1 2
1 1 2 2 1 1
2 1 1 1 2 1

They are the same for a generic study because all of the same alternatives are there, they are just in a different
order. However, for a branded study they are different. For a branded study, there would be a different brand
for each alternative, so the choice sets would be the same only if all the same alternatives appeared in the same
order. For both a branded and generic study, these choice sets are duplicates:

a b c a b c
1 2 1 1 2 1
2 1 2 2 1 2
1 1 2 1 1 2
2 1 1 2 1 1

Now consider these choice sets for a generic study.

a b c a b c
1 2 1 1 2 1
2 1 1 1 2 1
1 1 2 1 1 2
2 1 1 2 1 1

First, each of these choice sets has duplicate alternatives (2 1 1 in the first and 1 2 1 in the second). Second,
these two choice sets are flagged as duplicates, even though they are not exactly the same. They are flagged as
duplicates because every alternative in choice set one is also in choice set two, and every alternative in choice set
two is also in choice set one. In generic studies, two choice sets are considered duplicates unless one has one or
more alternatives that are not in the other choice set.
320 TS-677E Multinomial Logit, Discrete Choice Modeling

Here is an example. A design is created with the %ChoicEff macro choice-set-swapping algorithm for a
branded study, then the %MktDups macro is run to check for and eliminate duplicate choice sets.
%mktex(3 ** 9, n=27, seed=424)

data key;
input (Brand x1-x3) ($);
datalines;
Acme x1 x2 x3
Ajax x4 x5 x6
Widgit x7 x8 x9
;

%mktroll(design=randomized, key=key, alt=brand, out=cand);

%choiceff(data=cand, model=class(brand x1-x3), seed=420,

nsets=18, nalts=3, beta=zero);

proc freq; tables set; run;

%mktdups(branded, data=best, factors=brand x1-x3, nalts=3, out=out)

proc freq; tables set; run;

The first PROC FREQ output shows us that several candidate choice sets occur twice in the design.

The FREQ Procedure

Cumulative Cumulative
Set Frequency Percent Frequency Percent
--------------------------------------------------------
1 3 5.56 3 5.56
3 3 5.56 6 11.11
4 6 11.11 12 22.22
8 3 5.56 15 27.78
10 3 5.56 18 33.33
11 3 5.56 21 38.89
16 3 5.56 24 44.44
19 3 5.56 27 50.00
21 6 11.11 33 61.11
22 6 11.11 39 72.22
24 3 5.56 42 77.78
25 3 5.56 45 83.33
26 3 5.56 48 88.89
27 6 11.11 54 100.00

The output from the %MktDups macro contains the following tables:

Design: Branded
Factors: brand x1-x3
Brand
x1 x2 x3
Duplicate Sets: 4
The Macros 321

Duplicate
Choice Choice Sets
Set To Delete

1 5

2 12

4 18

11 14

The first line of the first table tells us that this is a branded design as opposed to generic. The second line tells
us the factors as specified on the factors= option. These are followed by the actual variable names for the
factors. The last line reports the number of duplicates. The second table tells us that choice set 1 is the same as
choice set 5. Similarly, 2 and 12 are the same as are 4 and 18, and also 11 and 14. The out= data set will contain
the design with the duplicate choice set eliminated.
Now consider an example with purely generic alternatives.
%mktex(2 ** 5, n=2**5, seed=93)
%mktlab(int=f1-f4)

%choiceff(data=final, model=class(x1-x5), seed=109,

nsets=42, flags=f1-f4, beta=zero);

%mktdups(generic, data=best, factors=x1-x5, nalts=4, out=out)

The macro produces the following tables:

Design: Generic
Factors: x1-x5
x1 x2 x3 x4 x5
Sets w Dup Alts: 2
Duplicate Sets: 2

Duplicate
Choice Choice Sets
Set To Delete

3 36

5 Alternatives

24 Alternatives

35 42

For each choice set listed in the choice set column, either the other choice sets it duplicates are listed or the word
’Alternatives’ is printed if the problem is with duplicate alternatives.
322 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are just the choice sets with duplication problems:

Set x1 x2 x3 x4 x5

3 2 1 1 1 2
1 1 1 1 1
1 2 2 2 2
2 2 2 2 1

5 1 1 2 1 2
1 1 2 1 2
2 2 1 2 1
2 2 1 2 1

24 1 2 2 1 2
1 2 2 1 2
2 1 1 2 1
2 1 1 2 1

35 2 1 2 1 1
2 2 1 2 2
1 1 2 1 2
1 2 1 2 1

36 2 1 1 1 2
1 1 1 1 1
1 2 2 2 2
2 2 2 2 1

42 1 1 2 1 2
2 1 2 1 1
1 2 1 2 1
2 2 1 2 2

You can see that the macro detects duplicates even though the alternatives do not always appear in the same order
in the different choice sets.
Now consider another example.
%mktex(2 ** 6, n=2**6)

data key;
input (x1-x2) ($) @@;
datalines;
x1 x2 x3 x4 x5 x6
;

%mktroll(design=design, key=key, out=cand);

%mktdups(generic, data=cand, factors=x1-x2, nalts=3, out=out)

proc print; by set; id set; run;

The Macros 323

Here is some of the output. The output lists, for each set of duplicates, the choice set that will be kept (in the first
column) and all the matching choice sets that will be deleted (in the second column).

Design: Generic
Factors: x1-x2
x1 x2
Sets w Dup Alts: 40
Duplicate Sets: 50

Duplicate
Choice Choice Sets
Set To Delete

1 Alternatives

2 Alternatives
5
6
17
18
21

.
.
.

Here are the unique choice sets.

Set _Alt_ x1 x2

7 1 1 1
2 1 2
3 2 1

8 1 1 1
2 1 2
3 2 2

12 1 1 1
2 2 1
3 2 2

28 1 1 2
2 2 1
3 2 2

This next example creates a conjoint design and tests it for duplicates.
%mktex( 3 ** 3 2 ** 2, n=19, seed=513)

%mktdups(linear, factors=x1-x5);
324 TS-677E Multinomial Logit, Discrete Choice Modeling

Design: Linear
Factors: x1-x5
x1 x2 x3 x4 x5
Duplicate Runs: 1

Duplicate
Runs
Run To Delete

9 10

%MktDups Macro Options

The following options can be used with the %MktDups macro.

options
For the first option, specify one or more of the following. You may specify noprint and one of the following:
generic, branded, or linear.

branded
specifies that since one of the factors is brand, the macro only needs to compare corresponding alternatives
in each choice set.

generic
specifies a generic design and is the default. This means that there are no brands, so options are interchange-
able, so the macro needs to compare each alternative with every other alternative in every choice set.

linear
specifies a linear not a choice design. Specify linear for a full-profile conjoint design, for an ANOVA design,
or for the linear version of a branded choice design.

noprint
specifies no printed output. This option will be used when you are only interested in the output data set or
macro variable.

Example:
%mktdups(branded noprint, nalts=3)

This next option is mandatory with choice designs.

nalts= n
specifies the number of alternatives. This option must be specified with generic or branded designs. It is ignored
with linear designs. For generic or branded designs, the data= data set must contain nalts= observations for
the first choice set, nalts= observations for the second choice set, and so on.
The Macros 325

Here are the other options.

data= SAS-data-set
specifies the input choice design. By default, the macro uses the last data set created.

out= SAS-data-set
specifies an output data set that contains the design with duplicate choice sets excluded. By default, no data set
is created, and the macro just reports on duplicates.

outlist= SAS-data-set
specifies the output data set with the list of duplicates. By default, outlist=outdups.

vars= variable-list
factors= variable-list
specifies the factors in the design. By default, all numeric variables are used.

%MktEval Macro
The %MktEval autocall macro helps you evaluate an experimental design. This macro reports on balance and
orthogonality. Typically, you will call it immediately after running the %MktEx macro. The output from this
macro contains two default tables. The first table shows the canonical correlations between pairs of coded factors.
A canonical correlation is the maximum correlation between linear combinations of the coded factors. All zeros
off the diagonal show that the design is orthogonal for main effects. Off-diagonal canonical correlations greater
than 0.316 (r2 > 0:1) are listed in a separate table.
For nonorthogonal designs and designs with interactions, the canonical-correlation matrix is not a substitute for
looking at the variance matrix with the %MktEx macro. It just provides a quick and more-compact picture of the
correlations between the factors. The variance matrix is sensitive to the actual model specified and the coding.
The canonical-correlation matrix just tells you if there is some correlation between the main effects. When is
a canonical correlation too big? You will have to decide that for yourself. In part, the answer depends on the
factors and how the design will be used. A high correlation between the client’s and the main competitor’s price
factor is a serious problem meaning you will need to use a different design. In contrast, a moderate correlation in
a choice design between one brand’s minor attribute and another brand’s minor attribute may be perfectly fine.
The macro also prints one-way, two-way and n-way frequencies. Equal one-way frequencies occur when the
design is balanced. Equal two-way frequencies occur when the design is orthogonal. Equal n-way frequencies,
all equal to one, occur when there are no duplicate runs or choice sets.

%MktEval Macro Options

The following options can be used with the %MktEval macro.

blocks= variable
specifies a blocking variable. This option prints separate canonical correlations within each bloc. By default,
there is one block.

data= SAS-data-set
specifies the input SAS data set with the experimental design. By default, the macro uses the last data set created.
326 TS-677E Multinomial Logit, Discrete Choice Modeling

factors= variable-list
vars= variable-list
specifies a list of the factors in the experimental design. The default is all of the numeric variables in the data set.

freqs= frequency-list
specifies the frequencies to print. By default, freqs=1 2 n, and 1-way, 2-way, and n-way frequencies are
printed. Do not specify the exact number of ways instead of n. For ways other than n, the macro checks for
and prints zero cell frequencies. For n-ways, the macro does not output or print zero frequencies. Only the
full-factorial design will have nonzero cells, so specifying something like freqs=1 2 20 will make the macro
take a long time and create huge data sets, whereas freqs=1 2 n runs very reasonably.

format= format
specifies the format for printing canonical correlations. The default format is 4.2.

list= n
specifies the minimum canonical correlation to list. The default is 0.316, the square root of r2 = 0:1.

outcorr= SAS-data-set
specifies the output SAS data set for the canonical correlation matrix. The default data set name is CORR.

outcb= SAS-data-set
specifies the output SAS data set for the with-block canonical correlation matrices. The default data set name is
CB.

outlist= SAS-data-set
specifies the output data set for the list of largest canonical correlations. The default data set name is LIST.

outfreq= SAS-data-set
specifies the output data set for the frequencies. The default data set name is FREQ.

outfsum= SAS-data-set
specifies the output data set for the frequency summaries. The default data set name is FSUM.

print= short|corr|list|freqs|summ|all
controls the printing of the results. Specify one or more values from the following list.
corr prints the canonical correlations matrix.
block prints the canonical correlations within block.
list prints the list of canonical correlations greater than the list= value.
freqs prints the frequencies, specified by the freqs= option.
summ prints the frequency summaries.
all prints all of the above.
short is the default and is equivalent to: corr list summ block.
noprint specifies no printed output.

By default, the frequency list, which contains the factor names, levels, and frequencies is not printed, but the
more compact frequency summary list, which contains the factors and frequencies but not the levels is printed.
The Macros 327

%MktEx Macro
The %MktEx autocall macro is designed for marketing researchers and any one else who wants to make good,
efficient experimental designs. This macro is designed to be very simple to use, and to run in seconds for trivial
problems, minutes for small problems, and in less than an hour for larger and difficult problems. This macro is
a full-featured linear designer that can handle simple problems like main-effects designs and more complicated
problems including designs with interactions and restrictions on which levels can appear together. The macro
is particularly designed to easily create the kinds of linear designs that marketing researches need for conjoint
and choice experiments. For any linear design problem, you can simply run the macro once, specifying only the
number of runs and the numbers of levels of all the factors. You will no longer have to try different algorithms
and different approaches to see which one works best. The macro does all of that for you.
Here is an example of using the %MktEx macro to create a design with 5 two-level factors, 4 three-level factors,
3 five-level factors, 2 six-level factors, all in 60 runs (rows or conjoint profiles or choice sets).
%mktex( 2 ** 5 3 ** 4 5 5 5 6 6, n=60 )

The notation m ** n means mn or n m-level factors. For example 2 ** 5 means 2  2  2  2  2 or 5

two-level factors.
The %MktEx macro creates efficient linear experimental designs using several approaches. The macro will try
to create a tabled design, it will search a set of candidate runs (rows of the design), and it will use a coordinate-
exchange algorithm using both random initial designs and also a partial tabled design initialization. The macro
stops if at any time it finds a perfect, 100% efficient, orthogonal and balanced design. This first phase is the
algorithm search phase. In it, the macro determines which approach is working best for this problem. At the
end of this phase, the macro chooses the method that has produced the best design and performs another set of
iterations using exclusively the chosen approach. Finally, the macro performs a third set of iterations where it
takes the best design it found so far and tries to improve it.
In all phases, the macro attempts to optimize D-efficiency (also known as D-optimality), which is a standard
measure of the goodness of the experimental design. As D-efficiency increases, the standard errors of the param-
eter estimates in the linear model decrease. A perfect design is orthogonal and balanced and has 100% efficiency.
A design is orthogonal when all of the parameter estimates are uncorrelated. A design is balanced when all of the
levels within each of the factors occur equally often. A design is orthogonal and balanced when the variance ma-
trix, which is proportional to (X0 X) 1 is diagonal, where X is a suitable orthogonal coding of the design matrix
(see page 80). See Kuhfeld, Tobias, and Garratt (1994) on page 25for more information on efficient experimental
designs.
For most problems, you only need to specify the levels of all the factors and the number of runs. For more
complicated problems, you may need to also specify the interactions that you want to be estimable or restrictions
on which levels may not appear together. Other than that, you should not need any other options. This macro is
not like other design tools that you have to tell what to do. With this macro, you just tell it what you want, and
it figures out a good way to do it. For some problems, the sophisticated user, with a lot of work, may be able
to adjust the options to come up with a better design. However, this macro should always produce a very good
design with minimal effort for even the most unsophisticated users.
For certain designs, you can just specify n=, which specifies the number of runs or rows in the design, and the
macro will give you a 100% efficient design. Certain tabled designs, fractional and full-factorial designs, and
Hadamard designs can be requested just by specifying n=. The tables show some of the designs up through
n=100 that can be obtained by just specifying n=. The macro will also output certain full-factorial designs with
just a specification of n=. For example n=162 implies 2 3 ** 4. The n= value is factored into a list of prime
factors, and the full-factorial design for that list is created. The tables show some of the 100% efficient designs
that the macro can directly create.
328 TS-677E Multinomial Logit, Discrete Choice Modeling

Implied Designs When Only n= is Specified

N Design Reference N Design Reference N Design Reference
4 23 Hadamard 32 231 Hadamard 64 263 Hadamard
8 27 Hadamard 36 211 312 Taguchi 1987 68 267 Hadamard
9 34 Fractional 40 239 Hadamard 72 223324 Dey 1985
12 211 Hadamard 44 243 Hadamard 76 275 Hadamard
16 215 Hadamard 48 211 412 Suen 1989 80 279 Hadamard
18 21 37 Taguchi 1987 49 78 Fractional 81 340 Fractional
20 219 Hadamard 50 21511 Taguchi 1987 84 283 Hadamard
24 223 Hadamard 52 251 Hadamard 88 287 Hadamard
25 56 Fractional 54 2325 Taguchi 1987 92 291 Hadamard
27 313 Fractional 56 255 Hadamard 96 295 Hadamard
28 227 Hadamard 60 259 Hadamard 100 299 Hadamard

Some of the Orthogonal Designs Available from the %MktEx Macro

N Design Reference N Design Reference

4 23 Hadamard 28 227 Hadamard
6 21 31 Full-factorial 28 212 71 Suen 1989
8 27 Hadamard 30 21 31 51 Full-factorial
9 34 Fractional-factorial 32 231 Hadamard
10 21 51 Full-factorial 32 228 41 Fractional-factorial
12 211 Hadamard 32 225 42 Fractional-factorial
12 24 31 Hedayat, Sloane & Stufken 32 224 81 Fractional-factorial
12 22 61 Hedayat, Sloane & Stufken 32 222 43 Fractional-factorial
12 31 41 Full-factorial 32 221 41 81 Fractional-factorial
14 21 71 Full-factorial 32 219 44 Fractional-factorial
15 31 51 Full-factorial 32 218 42 81 Fractional-factorial
16 215 Hadamard 32 216 161 Fractional-factorial
16 212 41 Fractional-factorial 32 216 45 Fractional-factorial
16 29 42 Fractional-factorial 32 215 43 81 Fractional-factorial
16 26 43 Fractional-factorial 32 213 46 Fractional-factorial
16 23 44 Fractional-factorial 32 212 44 81 Fractional-factorial
16 45 Fractional-factorial 32 210 47 Fractional-factorial
18 21 37 Taguchi 1987 32 29 45 81 Fractional-factorial
18 36 61 Taguchi 1987 32 27 48 Fractional-factorial
20 219 Hadamard 32 26 46 81 Fractional-factorial
20 28 51 Wang & Wu 1992 32 24 49 Fractional-factorial
20 22 101 Hedayat, Sloane & Stufken 32 23 47 81 Fractional-factorial
20 41 51 Full-factorial 32 48 81 Fractional-factorial
21 31 71 Full-factorial 33 31 111 Full-factorial
22 21 111 Full-factorial 34 21 171 Full-factorial
24 223 Hadamard 35 51 71 Full-factorial
24 220 41 Hadamard 36 235 Hadamard
24 216 31 Wang & Wu 1991 36 227 31 Hedayat, Sloane & Stufken
24 214 61 Wang & Wu 1991 36 220 32 Hedayat, Sloane & Stufken
24 213 31 41 Wang & Wu 1991 36 218 31 61 Hedayat, Sloane & Stufken
24 212 121 Hedayat, Sloane & Stufken 36 213 34 Suen 1989
24 211 41 61 Wang & Wu 1991 36 213 91 Suen 1989
24 31 81 Full-factorial 36 211 312 Taguchi 1987
25 56 Fractional-factorial 36 211 32 61 Hedayat, Sloane & Stufken
26 21 131 Full-factorial 36 29 31 62 Hedayat, Sloane & Stufken
27 313 Fractional-factorial 36 24 313 Taguchi 1987
27 39 91 Fractional-factorial 36 24 33 61 Hedayat, Sloane & Stufken
The Macros 329

Some of the Orthogonal Designs Available from the %MktEx Macro

N Design Reference N Design Reference

36 22 312 61 Wang & Wu 1991 48 220 49 Suen 1989
36 22 32 62 Hedayat, Sloane & Stufken 48 220 46 61 Suen 1989
36 22 181 Hedayat, Sloane & Stufken 48 219 31 47 Suen 1989
36 21 38 62 Zhang, Lu & Pang 1999 48 218 46 121 Suen 1989
36 21 63 SAS Procedure OPTEX 48 217 410 Suen 1989
36 313 41 Dey 1985 48 217 47 61 Suen 1989
36 312 121 Wang & Wu 1991 48 216 31 48 Suen 1989
36 37 63 Finney 1982 48 215 47 121 Suen 1989
38 21 191 Full-factorial 48 214 411 Suen 1989
39 31 131 Full-factorial 48 214 48 61 Suen 1989
40 239 Hadamard 48 213 31 49 Suen 1989
40 236 41 Hadamard 48 212 48 121 Suen 1989
40 228 51 Hedayat, Sloane & Stufken 48 211 412 Suen 1989
40 222 101 Hedayat, Sloane & Stufken 48 211 49 61 Suen 1989
40 220 201 Dey 1985 48 210 31 410 Suen 1989
40 220 41 51 Dey 1985 48 29 49 121 Suen 1989
42 21 31 71 Full-factorial 48 28 410 61 Suen 1989
44 243 Hadamard 48 27 31 411 Suen 1989
45 32 51 Full-factorial 48 26 410 121 Suen 1989
46 21 231 Full-factorial 48 25 411 61 Suen 1989
48 247 Hadamard 48 24 31 412 Suen 1989
48 244 41 Suen 1989 48 23 411 121 Suen 1989
48 241 42 Suen 1989 48 22 412 61 Suen 1989
48 240 31 Suen 1989 48 31 413 Suen 1989
48 238 61 Suen 1989 48 412 121 Suen 1989
48 238 43 Suen 1989 49 78 Fractional-factorial
48 237 31 41 Suen 1989 50 21 511 Taguchi 1987
48 236 121 Suen 1989 50 510 101 Hedayat, Sloane & Stufken
48 235 44 Suen 1989 52 251 Hadamard
48 235 41 61 Suen 1989 54 21 325 Taguchi 1987
48 234 31 42 Suen 1989 54 324 61 Hedayat, Sloane & Stufken
48 233 41 121 Suen 1989 54 318 181 Hedayat, Sloane & Stufken
48 232 45 Suen 1989 55 51 111 Full-factorial
48 232 42 61 Suen 1989 56 255 Hadamard
48 231 31 43 Suen 1989 56 252 41 Hedayat, Sloane & Stufken
48 230 42 121 Suen 1989 58 21 291 Full-factorial
48 229 46 Suen 1989 60 259 Hadamard
48 229 43 61 Suen 1989 62 21 311 Full-factorial
48 228 31 44 Suen 1989 63 31 211 Full-factorial
48 227 43 121 Suen 1989 64 263 Hadamard
48 226 47 Suen 1989 64 260 41 Fractional-factorial
48 226 44 61 Suen 1989 64 257 42 Fractional-factorial
48 225 31 45 Suen 1989 64 256 81 Fractional-factorial
48 224 31 81 Hedayat, Sloane & Stufken 64 254 43 Fractional-factorial
48 224 241 Hedayat, Sloane & Stufken 64 253 41 81 Fractional-factorial
48 224 44 121 Suen 1989 64 251 44 Fractional-factorial
48 223 48 Suen 1989 64 250 42 81 Fractional-factorial
48 223 45 61 Suen 1989 64 249 82 Fractional-factorial
48 222 31 46 Suen 1989 64 248 45 Fractional-factorial
48 221 45 121 Suen 1989 64 247 43 81 Fractional-factorial
330 TS-677E Multinomial Logit, Discrete Choice Modeling

Some of the Orthogonal Designs Available from the %MktEx Macro

N Design Reference N Design Reference

64 246 41 82 Fractional-factorial 64 217 413 81 Fractional-factorial
64 245 46 Fractional-factorial 64 217 46 84 Fractional-factorial
64 244 44 81 Fractional-factorial 64 216 411 82 Fractional-factorial
64 243 42 82 Fractional-factorial 64 216 44 85 Fractional-factorial
64 242 47 Fractional-factorial 64 215 416 Fractional-factorial
64 242 83 Fractional-factorial 64 215 49 83 Fractional-factorial
64 241 45 81 Fractional-factorial 64 215 42 86 Fractional-factorial
64 240 43 82 Fractional-factorial 64 214 414 81 Fractional-factorial
64 239 48 Fractional-factorial 64 214 47 84 Fractional-factorial
64 239 41 83 Fractional-factorial 64 214 87 Fractional-factorial
64 238 46 81 Fractional-factorial 64 213 412 82 Fractional-factorial
64 237 44 82 Fractional-factorial 64 213 45 85 Fractional-factorial
64 236 49 Fractional-factorial 64 212 417 Fractional-factorial
64 236 42 83 Fractional-factorial 64 212 410 83 Fractional-factorial
64 235 47 81 Fractional-factorial 64 212 43 86 Fractional-factorial
64 235 84 Fractional-factorial 64 211 415 81 Fractional-factorial
64 234 45 82 Fractional-factorial 64 211 48 84 Fractional-factorial
64 233 410 Fractional-factorial 64 211 41 87 Fractional-factorial
64 233 43 83 Fractional-factorial 64 210 413 82 Fractional-factorial
64 232 48 81 Fractional-factorial 64 210 46 85 Fractional-factorial
64 232 41 84 Fractional-factorial 64 29 418 Fractional-factorial
64 231 46 82 Fractional-factorial 64 29 411 83 Fractional-factorial
64 230 411 Fractional-factorial 64 29 44 86 Fractional-factorial
64 230 44 83 Fractional-factorial 64 28 416 81 Fractional-factorial
64 229 49 81 Fractional-factorial 64 28 49 84 Fractional-factorial
64 229 42 84 Fractional-factorial 64 28 42 87 Fractional-factorial
64 228 47 82 Fractional-factorial 64 27 414 82 Fractional-factorial
64 228 85 Fractional-factorial 64 27 47 85 Fractional-factorial
64 227 412 Fractional-factorial 64 27 88 Fractional-factorial
64 227 45 83 Fractional-factorial 64 26 419 Fractional-factorial
64 226 410 81 Fractional-factorial 64 26 412 83 Fractional-factorial
64 226 43 84 Fractional-factorial 64 26 45 86 Fractional-factorial
64 225 48 82 Fractional-factorial 64 25 417 81 Fractional-factorial
64 225 41 85 Fractional-factorial 64 25 410 84 Fractional-factorial
64 224 413 Fractional-factorial 64 24 415 82 Fractional-factorial
64 224 46 83 Fractional-factorial 64 24 48 85 Fractional-factorial
64 223 411 81 Fractional-factorial 64 24 41 88 Fractional-factorial
64 223 44 84 Fractional-factorial 64 23 420 Fractional-factorial
64 222 49 82 Fractional-factorial 64 23 413 83 Fractional-factorial
64 222 42 85 Fractional-factorial 64 23 46 86 Fractional-factorial
64 221 414 Fractional-factorial 64 421 Fractional-factorial
64 221 47 83 Fractional-factorial 64 47 86 Fractional-factorial
64 221 86 Fractional-factorial 65 51 131 Full-factorial
64 220 412 81 Fractional-factorial 66 21 31 111 Full-factorial
64 220 45 84 Fractional-factorial 66 21 331 Full-factorial
64 219 410 82 Fractional-factorial 68 267 Hadamard
64 219 43 85 Fractional-factorial 69 31 231 Full-factorial
64 218 415 Fractional-factorial 70 21 51 71 Full-factorial
64 218 48 83 Fractional-factorial 72 271 Hadamard
64 218 41 86 Fractional-factorial 72 268 41 Hadamard
The Macros 331

Some of the Orthogonal Designs Available from the %MktEx Macro

N Design Reference N Design Reference

72 263 31 Hedayat, Sloane & Stufken 80 270 43 Wang 1996
72 256 32 Hedayat, Sloane & Stufken 80 249 51 Hedayat, Sloane & Stufken
72 254 31 61 Hedayat, Sloane & Stufken 80 243 101 Hedayat, Sloane & Stufken
72 249 34 Hedayat, Sloane & Stufken 80 241 201 Hedayat, Sloane & Stufken
72 249 91 Hedayat, Sloane & Stufken 80 241 41 51 Hedayat, Sloane & Stufken
72 247 312 Wang 1996 81 340 Fractional-factorial
72 247 32 61 Hedayat, Sloane & Stufken 81 336 91 Fractional-factorial
72 245 31 62 Hedayat, Sloane & Stufken 81 332 92 Fractional-factorial
72 240 313 Wang & Wu 1991 81 328 93 Fractional-factorial
72 240 33 61 Hedayat, Sloane & Stufken 81 324 94 Fractional-factorial
72 238 312 61 Hedayat, Sloane & Stufken 81 320 95 Fractional-factorial
72 238 32 62 Hedayat, Sloane & Stufken 81 316 96 Fractional-factorial
72 238 181 Hedayat, Sloane & Stufken 81 312 97 Fractional-factorial
72 237 38 62 Hedayat, Sloane & Stufken 81 38 98 Fractional-factorial
72 237 63 Hedayat, Sloane & Stufken 81 34 99 Fractional-factorial
72 236 313 41 Hedayat, Sloane & Stufken 81 910 Fractional-factorial
72 236 312 121 Hedayat, Sloane & Stufken 82 21 411 Full-factorial
72 236 37 63 Hedayat, Sloane & Stufken 84 283 Hadamard
72 236 361 Hedayat, Sloane & Stufken 85 51 171 Full-factorial
72 223 324 Dey 1985 86 21 431 Full-factorial
72 220 324 41 Wang 1996 88 287 Hadamard
72 216 325 Wang 1996 90 21 32 51 Full-factorial
72 214 324 6 1 Wang 1996 92 291 Hadamard
72 213 325 41 Wang 1996 93 31 311 Full-factorial
72 212 324 121 Hedayat, Sloane & Stufken 94 21 471 Full-factorial
72 211 324 41 61 Wang 1996 96 295 Hadamard
72 325 81 Hedayat, Sloane & Stufken 96 292 41 Hedayat, Sloane & Stufken
72 324 241 Hedayat, Sloane & Stufken 96 289 42 Hedayat, Sloane & Stufken
74 21 371 Full-factorial 96 286 43 Hedayat, Sloane & Stufken
75 31 52 Full-factorial 96 283 44 Hedayat, Sloane & Stufken
76 275 Hadamard 96 280 45 Hedayat, Sloane & Stufken
77 71 111 Full-factorial 96 277 46 Hedayat, Sloane & Stufken
78 21 391 Full-factorial 96 274 47 Hedayat, Sloane & Stufken
80 279 Hadamard 98 21 491 Full-factorial
80 276 41 Wang 1996 99 32 111 Full-factorial
80 273 42 Wang 1996 100 299 Hadamard
Other relevant references include Rao (1947), Addelman (1962); Bose (1947); and Hadamard (1893). Most of
these designs were created using information found in Hedayat, Sloane, and Stufken (1999); Dey (1985); and
Neil Sloane’s very useful web site: [Link] njas/oadir/. Many of the designs above could
have multiple references. Most of the above reference list was derived from Hedayat et. al.
When n=n is a multiple of 4, the %MktEx macro can construct orthogonal designs with up to n 1 two-
level factors. The two-level designs are constructed from Hadamard matrices (Hadamard, 1893; Paley, 1933;
Williamson, 1944; Hedayat, Sloane, and Stufken, 1999). The next table shows the available sizes up through
n=1000:
332 TS-677E Multinomial Logit, Discrete Choice Modeling

Hadamard Matrix Sizes Up to n=1000

4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64
68 72 76 80 84 88 92 96 100 104 108 112 116 120 124 128
132 136 140 144 148 152 156 160 164 168 172 176 180 184 188 192
196 200 204 208 212 216 220 224 228 232 236 240 244 248 252 256
264 272 276 280 284 288 296 300 304 308 312 316 320 328 332 336
344 348 352 360 364 368 376 380 384 388 392 396 400 408 416 420
424 432 440 444 448 456 460 464 468 472 480 484 488 492 496 500
504 512 516 524 528 540 544 548 552 556 560 564 568 572 576 588
592 600 608 616 620 624 628 632 636 640 644 656 660 664 672 676
684 688 692 696 700 704 708 720 728 736 740 748 752 760 768 776
780 784 788 792 796 800 804 812 816 820 828 832 840 844 848 860
864 868 880 884 888 896 900 908 912 916 920 924 928 936 944 948
960 968 972 976 984 992 1000

Larger sizes are available as well. The %MktEx macro can construct these designs when n is a multiple of 4 and
one or more of the following hold:

 n  256
 n 1 is prime
 n=2 1 is prime and mod(n=2; 4) = 2
 n is a power of 2 (2, 4, 8, 16, ...) times the size of a smaller Hadamard matrix that is available.

For some of these sizes, the macro can create orthogonal designs with a small number (say m) four-level factors
in place of 3  m of the two-level factors (for example, 270 43 in 80 runs and 274 47 in 96 runs).
Here is a simple example of using the %MktEx macro to request the L36 design, which has 11 two-level factors
and 12 three-level factors.
%mktex( n=36 )

No iterations are needed, and the macro immediately creates the L36 , which is 100% efficient. This example
runs in a few seconds. The factors are always named x1, x2, ... and the levels are always consecutive integers
starting with 1. You can use the %MktLab macro to get different names and levels.
By default, the macro creates two output data sets with the design.
 out=Design - the experimental design, sorted by the factor levels.
 outr=Randomized - the randomized experimental design.

The designs are equivalent and have the same D-efficiency. The out=Design data set is sorted and hence is
usually easier to look at, however the outr=Randomized design is the better one to use. The randomized
design has the rows sorted into a random order, and all of the factor levels are randomly reassigned. For example
with two-level factors, approximately half of the original (1, 2) mappings will be reassigned (2, 1). Similarly,
with three level factors, the mapping (1, 2, 3) will be changed to one of the following: (1, 2, 3), (1, 3, 2), (2, 1,
3), (2, 3, 1), (3, 1, 2), or (3, 2, 1). The reassignment of levels is usually not critical for the iteratively derived
designs, but it can be very important for some of the tabled designs, which have all ones in the first row.
The Macros 333

%MktEx Macro Notes

The %MktEx macro prints notes to the SAS log to show you what it is doing while it is running. Most of the
notes that would normally come out of the macro’s procedure and DATA steps are suppressed by default by an
options nonotes statement. The macro will usually start by printing one of the following notes (filling in a
value after n=).
NOTE: Generating the Hadamard design, n=.
NOTE: Generating the full-factorial design, n=.
NOTE: Generating the fractional-factorial design, n=.
NOTE: Generating the tabled design, n=.
These messages tell you which type of tabled design the macro is constructing. The design may be the final
design, or it may provide an initialization for the coordinate exchange algorithm. In some cases, it may not have
the same number of runs, n, as the final design. Usually this step is fast, but constructing some fractional-factorial
designs may be time consuming.
If the macro is going to use PROC OPTEX to search a candidate set, it will print this note.
NOTE: Generating the candidate set.
This step will usually be fast. Next, when a candidate set is searched, the macro will print this next note,
substituting in values for the ellipses.
NOTE: Performing ... searches of ... candidates.
This step may be take a while depending on the size of the candidate set and the size of the design. When there
are a lot of restrictions and a fractional-factorial candidate set is being used, the candidate set may be so restricted
that it does not contain enough information to make the design. In that case, you will get this message.
NOTE: The candidate-set initialization failed,
but the MKTEX macro is continuing.
Even though part of the macro’s algorithm failed, it is not a problem. The macro just goes on to the coordinate-
exchange algorithm, which will almost certainly work better than searching any severely-restricted candidate
set.
Sometimes you will get this note.
NOTE: Stopping since it appears that no improvement is possible.
When the macro keeps finding the same maximum D-efficiency over and over again in different designs, it may
stop early. This may mean that the macro has found the optimal design, or it may mean that the macro keeps
finding a very attractive local optimum. Either way, it is unlikely that the macro will do any better. You can
control this using the stopearly= option.
The macro has options that control the amount of time it spends trying different techniques. When time expires,
the macro may switch to other techniques before it completes the usual maximum number of iterations. When
this happens, the macro tells you.
NOTE: Switching to a random initialization after ... minutes and
... designs.
NOTE: Quitting the algorithm search after ... minutes and ... designs.
NOTE: Quitting the design search after ... minutes and ... designs.
NOTE: Quitting the refinement step after ... minutes and ... designs.
When there are restrictions or you specify that you do not want duplicate runs, but you also specify
options=accept, which means you are willing to accept designs that violate the restrictions, the macro
will tell you if the restrictions are not met.
NOTE: The restrictions were not met.
NOTE: The design has duplicate runs.
334 TS-677E Multinomial Logit, Discrete Choice Modeling

The macro ends with one of the following two messages.

NOTE: The MKTEX macro used ... seconds.
NOTE: The MKTEX macro used ... minutes.

%MktEx Macro Iteration History

This section provides information on interpreting the iteration history table produced by the %MktEx macro.
Here is part of a table.

Vacation Example

Algorithm Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
1 Start 82.2172 82.2172 Can
1 End 82.2172

2 Start 78.5039 Tab,Ran

2 5 14 83.2098 83.2098
2 6 14 83.3917 83.3917
2 6 15 83.5655 83.5655
2 7 14 83.7278 83.7278
2 7 15 84.0318 84.0318
2 7 15 84.3370 84.3370
2 8 14 85.1449 85.1449
.
.
.
2 End 98.0624
.
.
.

12 Start 51.8915 Ran,Mut,Ann

12 End 93.0214
.
.
.

Vacation Example

Design Search History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 98.8933 98.8933 Ini

1 Start 80.4296 Tab,Ran

1 End 98.8567
.
.
.
The Macros 335

Vacation Example

Design Refinement History

Current Best
Design Row,Col D-Efficiency D-Efficiency Notes
----------------------------------------------------------
0 Initial 98.9438 98.9438 Ini

1 Start 94.7490 Pre,Mut,Ann

1 End 92.1336
.
.
.

The first column, Design, is a design number. Each design corresponds to a complete iteration using a differ-
ent initialization. Initial designs are numbered zero. The second column is Row,Col, which shows the design
row and column that is changing in the coordinate-exchange algorithm. This column also contains Start for
displaying the initial efficiency, End for displaying the final efficiency, and Initial for displaying the effi-
ciency of a previously created (perhaps externally, perhaps in a previous step) initial design. The Current
D-Efficiency column contains the D-efficiency for the design including starting, intermediate and final val-
ues. The next column is Best D-efficiency. Values are put in this column for initial designs and when
a design is found that is as good as or better than the previous best design. The last column, Notes, contains
assorted algorithm and explanatory details. Values are added to the table at the beginning of an iteration, at the
end of an iteration, when a better design is found, and when a design first conforms to restrictions. Details of
the candidate search iterations are not shown. Only the D-efficiency for the best design found through candidate
search is shown.
Here are the notes.

Can - the results of a candidate-set search

Tab - tabled initialization (full or in part)
Ran - random initialization (full or in part)
Unb - unbalanced initialization (usually in part)
Ini - initial design
Mut - random mutations of the initial design were performed
Ann - simulated annealing was used in this iteration
Pre - using previous best design as a starting point
Conforms - design conforms to restrictions

Sometimes, more than one note appears. For example, the triples Ran,Mut,Ann and Pre,Mut,Ann fre-
quently appear together.
The iteration history consists of three tables.

Algorithm Search History - searches for a design and the best algorithm for this problem
Design Search History - uses the best algorithm to search further
Design Refinement History - tries to refine the best design
336 TS-677E Multinomial Logit, Discrete Choice Modeling

%MktEx Macro Options

The following options can be used with the %MktEx macro.

list
specifies a list of the numbers of levels of all the factors. For example, for 3 two-level factors specify ei-
ther 2 2 2 or 2 ** 3. Lists of numbers, like 2 2 3 3 4 4 or a levels**number of factors syntax like:
2**2 3**2 4**2 can be used, or both can be combined: 2 2 3**4 5 6. The specification 3**4 means
four three-level factors. Note that the factor list is a positional parameter. This means that if it is specified, it must
come first, and unlike all other parameters, it is not specified after a name and an equal sign. Usually, you have
to specify a list. However, in some cases, you can just specify n= and omit the list and a default list is implied
(see page 328). For example, n=18 implies a list of 2 3 ** 7. When the list is omitted, and if there are no
interactions, restrictions, or duplicate exclusions, then by default there are no OPTEX iterations (optiter=0).

n= n
specifies the number of runs in the design. You must specify n=. You can use the %MktRuns macro to get
suggestions for values of n=.
Example:
%mktruns( 4 2 ** 5 3 ** 5 )

In this case, this macro suggests several sizes including an orthogonal design with n=72 runs and some smaller
nonorthogonal designs including n=36, 24, 48, 60.

Basic Options
This next group of options contains some of the more commonly used options.

balance= n
specifies the maximum allowed level-frequency range. The balance= option allows you to tell the macro that
it should make an extra effort to ensure that the design is nearly balanced. By default, the macro does not try to
ensure balance beyond the fact that lack of balance decreases D-efficiency. Specify a positive integer, usually 1
or 2, that specifies the degree of imbalance that is acceptable. You may need to also specify options=accept
with balance=. The macro usually does a good job of producing nearly balanced design, but if balance is
critically important, and your designs are not balanced enough, you can sometimes achieve better balance by
specifying balance=1, but usually at the price of worse efficiency, sometimes much worse. The balance=
option specifies additional restrictions (see restrictions=) that help achieve better balance. By default, no
additional restrictions are added. The balance=n option specifies that for each factor, the difference between
the frequencies, for the most and least frequently occuring levels, should be no larger than n. You may specify
balance=0, however this usually is not a good idea. The macro needs the flexibility to have imbalance as it
refines the design. Another option is to instead use the %MktBal macro, which produces perfectly balanced
main effects plans. It is likely that the algorithms used by both the balance= option and the %MktBal macro
will be changed in the future to use some now unknown algorithms that are both faster and better.

examine= I | V
specifies the matrices that you want to examine. The option examine=I prints the information matrix, X0 X;
examine=V prints the variance matrix, (X0 X) 1 ; and examine=I V prints both. By default, these matrices
are not printed.
The Macros 337

interact= interaction-list
specifies interactions that must be estimable. By default, no interactions are guaranteed to be estimable. Exam-
ples:
interact=x1*x2
interact=x1*x2 x3*x4*x5
interact=x1|x2|x3|x4|x5@2
the interaction syntax is like PROC GLM’s and many of the other modeling procedures. It uses “*” for simple
interactions (x1*x2 is the interaction between x1 and x2), “|” for main effects and interactions (x1|x2|x3 is
the same as x1 x2 x1*x2 x3 x1*x3 x2*x3 x1*x2*x3) and “@” to eliminate higher-order interactions
(x1|x2|x3@2 eliminates x1*x2*x3 and is the same as x1 x2 x1*x2 x3 x1*x3 x2*x3). The specifi-
cation “@2” allows only main effects and two-way interactions. Only “@” values of 2 or 3 are allowed. For the
factor names, you must specify either the actual variable names (for example, x1 x2 ...) or you can just specify
the number without the “x” (for example, x1*x2 is equivalent to 1*2).

options= options-list
specifies binary options. By default, none of these options are specified. Specify one or more of the following
values after options=.

accept
allows the macro to output designs that violate restrictions imposed by restrictions=, balance=, or
partial=, or have duplicates with options=nodups. Normally the macro will not output such designs.
With options=accept, a design becomes eligible for output when the macro can no longer improve on
the restrictions or eliminate duplicates. Without options=accept, a design is only eligible when all
restrictions are met and all duplicates are eliminated.

check
checks the efficiency of a given design, specified in init=, and disables the out=, outr=, and outall=
options. If init= is not specified, options=check is ignored.

nodups
eliminates duplicate runs.

nofinal
skips calling PROC OPTEX to print the efficiency of the final experimental design.

nohistory
does not print the iteration history.

nosort
does not sort the design. One use of this option is with Hadamard matrices. Hadamard matrices are generated
with a banded structure that is lost when the design is sorted. If you want to see the original Hadamard
matrix, and not just a design constructed from the Hadamard matrix, specify options=nosort.

partial= n
specifies a partial profile design. The default is an ordinary linear design. Specify for example partial=4
if you only want 4 attributes to vary in each row of the design (except the first run, in which none vary). This
option works by adding restrictions to the design (see restrictions=). Specifying options=accept or
balance= with partial= is not a good idea.
338 TS-677E Multinomial Logit, Discrete Choice Modeling

restrictions= macro-name
specifies the name of a macro that places restrictions on the design. By default, there are no restrictions. If
you have restrictions on the design, what combinations can appear with what other combinations, then you must
create a macro that creates a variable called bad that contains a numerical summary of how bad the row of the
design is. When everything is fine, set bad to zero. Otherwise set bad to 1 or a larger value. Ideally, set bad
to the number of violations so that the macro knows if changes to factor levels are moving in the right direction.
The macro must consist of PROC IML statements and possibly some macro statements.
Be sure to check the log when you specify restrictions=. The macro cannot always ensure that your
statements are syntax-error free and stop if they are not.
Your macro can look at several things in quantifying badness, and must store its results in bad.
i - is a scalar that contains the number of the row currently being changed.
x - is a row vector of factor levels, always containing integer values beginning with 1 and continuing on to
the number of levels for each factor.
x1 is the same as x[1], x2 is the same as x[2], and so on.
j1 - is a scalar that contains the number of the column currently being changed.
j2 - is a scalar that contains the number of the other column currently being changed (along with j1) with
exchange=2 and larger exchange= values.
j3 - is a scalar that contains the number of the third column currently being changed (along with j1 and
j2) with exchange=3 and larger exchange= values.
xmat - is the entire x matrix. Note that the ith row of xmat may not be x since x may contain information
on the swaps being considered.
bad - results: 0 - fine, or the number of violations of restrictions.
Do not use these names (other than bad) for intermediate values!
Other than that, you can create intermediate variables without worrying about conflicts with the names in the
macro. The levels of the factors for one row of the experimental design are stored in a vector x, and the first level
is always 1, the second always 2, and so on. All restrictions must be defined in terms of x[j] (or alternatively,
x1, x2, ..., and perhaps the other matrices). For example, if there are five three-level factors and if it is bad
if the level for a factor equals the level for the following factor, create a macro restrict as follows and specify
restrictions=restrict.
%macro restrict;
bad = (x1 = x2) +
(x2 = x3) +
(x3 = x4) +
(x4 = x5);
%mend;
Note that you specify just the macro name and no percents on the restrictions= option. Also note that IML
does not have the full set of Boolean operators that the DATA step and other parts of SAS have. For example,
these are not available: OR AND NOT GT LT GE LE EQ NE. Here are the operators you can use along with
their meaning.

= equals not: EQ
^ = or : = not equals not: NE
< less than not: LT
<= less than or equal to not: LE
> greater than not: GT
>= greater than or equal to not: GE
& and not: AND
j or not: OR
^ or : not not: NOT
The Macros 339

Restrictions seriously slow down the algorithm.

With restrictions, the ’Current D-Efficiency’ column of the iteration history table may contain values larger
than the ’Best D-Efficiency’ column. This is because the design corresponding to the current D-efficiency may
have restriction violations. Values are only reported in the best D-efficiency column after all of the restriction
violations have been removed. You can specify options=accept with restrictions= when it is okay if
the restrictions are not met. See pages 195, 280, and 285 for more information on restrictions.

seed= n
specifies the random number seed. By default, seed=0, and clock time is used as the random number seed. By
specifying a random number seed, results should be reproducible within a SAS release for a particular operating
system. However, due to machine differences, some results may not be exactly reproducible on other machines.
For most orthogonal and balanced designs, the results should be reproducible. When computerized searches
are done, it is likely that you will not get the same design across different operating systems and different SAS
releases, although you would expect the efficiency differences to be slight.

Data Set Options

These next options specify the names of the input and output data sets.

init= SAS-data-set
specifies the initial (input) experimental design. By default, there is no initial design. Use init= when you
want to evaluate the efficiency of a design (along with options=check) or when you want to try to improve
a design.

out= SAS-data-set
specifies the output experimental design. The default is out=Design. By default, this design is sorted unless
you specify options=nosort. This is the output data set to look at in evaluating the design. See the outr=
option (next) for a randomized version of the same design, which is generally more suitable for actual use.
Specify a null value for out= if you do not want this data set created.

outall= SAS-data-set
specifies the output data set containing all designs found. By default, this data set is not created.

outr= SAS-data-set
specifies the randomized output experimental design. The default is out=Randomized. Random levels are
assigned within factors, and the runs are sorted into a random order. When restrictions= or partial= is
specified, only the random sort is performed. Specify a null value for outr= if you do not want a randomized
design created.

Iteration Options
These next options control some of the details of the iterations. The macro can perform three sets of iterations.
The ’Algorithm Search’ set of iterations looks for efficient designs using three different approaches. It then
determines which approach appears to be working best and uses that approach exclusively in the second set of
’Design Search’ iterations. The third set or ’Design Refinement’ iterations tries to refine the best design found
so far by using level swaps combined with random mutations and simulated annealing.
340 TS-677E Multinomial Logit, Discrete Choice Modeling

The first set of iterations can have up to three parts. The first part uses either PROC PLAN or PROC FACTEX
followed by PROC OPTEX, called through the %MktDes macro, to create and search a candidate set for an
optimal initial design. The second part may use a tabled or fractional-factorial design as an initial design. The
next part consists of level exchanges starting with random initial designs.
In the first part, if the full-factorial design is manageable (arbitrarily defined as < 5186 runs), it is used as a
candidate set, otherwise a fractional-factorial candidate set is used. The macro tries optiter= iterations to
make an optimal design using the %MktDes macro and PROC OPTEX.
In the second part, the macro will try to generate and improve a standard tabled or fractional-factorial design.
Sometimes, this can lead immediately to an optimal design, for example with 211 312 and n = 36. In other cases,
when only part of the desired design matches some standard design, only part of the design is initialized with the
standard design and multiple iterations are run using the standard design as a partial initialization with the rest of
the design randomly initialized.
In the third part, the macro uses the coordinate-exchange algorithm with random initial designs.

anneal= n1 < n2 < n3 >>

specifies the starting probability for simulated annealing in the coordinate-exchange algorithm. The default is
anneal=.05 .05 .05. Specify a zero or null for no annealing. You can specify more than one value if
you would like to use a different value for the algorithm search, design search, and design refinement iterations.
Specifying a value (greater than zero and less than one, for example 0.1) allows the design to get worse with
decreasing probability as the number of iterations increases. This often helps the algorithm overcome local
efficiency maxima. Allowing efficiency to decrease can help get past the bumps on the efficiency function.
Examples: anneal= or anneal=0 specifies no annealing, anneal=0.1 specifies an annealing probability of
0.1 during all three sets of iterations, mutate=0 0.1 0.05 specifies no annealing during the initial iterations,
an annealing probability of 0.1 during the search iterations, and an annealing probability of 0.05 during the
refinement iterations.

anniter= n1 < n2 < n3 >>

specifies the first iteration to consider using annealing on the design. The default is anniter=. . ., which
means that the macro chooses values to use. The default is the first iteration that uses a fully random initial design
in each of the three sets of iterations. Hence by default, there is no random annealing in any part of the initial
design when part of the initial design comes from a tabled design.

canditer= n1 < n2 >

specifies the number of coordinate-exchange iterations that will be used to try to improve a candidate-set based,
OPTEX-generated initial design. The default is canditer=1 1. Note that optiter= controls the number
of OPTEX iterations. Unless you are using annealing or mutation, in the canditer= iterations (by default
you are not) or unless you are using options=nodups, do not change theses values. The default value of
canditer=1 1, along with the default mutiter= and anniter= values of missing, mean that the results
of the OPTEX iterations are presented once in the algorithm iteration history, and if appropriate, once in the
design search iteration history. Furthermore, by default, OPTEX generated designs are not improved with level
exchanges except in the design refinement phase.
The Macros 341

maxdesigns= n
specifies that the macro should stop after maxdesign= designs have been created. This option may be useful
for big, slow problems with restrictions. You could specify for example maxdesigns=3 and maxtime=0 and
the macro would perform one candidate-set-based iteration, one tabled design initialization iteration, and one
random initialization iteration and then stop. By default, this option is ignored and stopping is based on the other
iteration options.

maxiter= n1 < n2 < n3 >>

iter= n1 < n2 < n3 >>
specifies the maximum number of iterations or designs to generate. The default is maxiter=21 25 10.
With larger values, the macro tends to find better designs at a cost of much slower run times. You can specify
more than one value if you would like to use a different value for the algorithm search, design search, and
design refinement iterations. The second value is only used if the second set of iterations consists of coordinate-
exchange iterations. Otherwise, the number of iterations for the second set is specified with the tabiter=, or
canditer= and optiter= options. If you want more iterations, be sure to set the maxtime= option as well,
because iteration stops when the maximum number of iterations is reached or the maximum amount of time,
whichever comes first. Examples: maxiter=10 specifies 10 iterations for the initial, search, and refinement
iterations, and maxiter=10 10 5 specifies 10 initial iterations, followed by 10 search iterations, followed by
5 refinement iterations.

maxstages= n
specifies that the macro should stop after maxstages= algorithm stages have been completed. This option may
be useful for big, slow problems with restrictions. You could specify maxstages=1 and the macro will stop
after the algorithm search stage, or maxstages=2 and the macro will stop after the design search stage. The
default is maxstages=3, which means the macro will stop after the design refinement stage.

maxtime= n1 < n2 < n3 >>

specifies the approximate maximum amount of time in minutes to run each phase. The default is maxtime=10
20 5. When an iteration completes (a design is completed), if more than the specified amount of time has
elapsed, the macro quits iterating in that phase. Usually, run time will be no more than 10% or 20% larger
than the specified values. However, for large problems, with restrictions, and with exchange= values other
than 1, run time may be quite a bit larger than the specified value, since the macro only checks time after a
design finishes. You can specify more than one value if you would like to use a different value for the algorithm
search, design search, and design refinement iterations. By default, the macro spends up to 10 minutes on the
algorithm search iterations, 20 minutes on the design search iterations, and 5 minutes in the refinement stage.
Most problems run in much less time than this. Note that the second value is ignored for OPTEX iterations since
OPTEX does not have any timing options. This option also affects, in the algorithm search iterations, when the
macro switches between using a tabled initial design to using a random initial design. If the macro is not done
using tabled initializations, and one half of the first time value has passed, it switches. Examples: maxtime=60
specifies up to one hour for each phase. maxtime=20 30 10 specifies 20 minutes for the first phase and 30
minutes for the second, and 10 for the third. A null value, maxtime=, removes all time restrictions. The option
maxtime=0 provides a way to get a quick run, with no more than one iteration in each phase. However, even
with maxtime=0, run time can be several minutes or more for large problems. See the maxdesigns= and
maxstages= options (next) for other ways to drastically cut run time for large problems.
342 TS-677E Multinomial Logit, Discrete Choice Modeling

mutate= n1 < n2 < n3 >>

specifies the probability at which each value in an initial design may mutate or be assigned a different random
value before the coordinate-exchange iterations begin. The default is mutate=.05 .05 .05. Specify a zero
or null for no mutation. You can specify more than one value if you would like to use a different value for the
algorithm search, design search, and design refinement iterations. Examples: mutate= or mutate=0 specifies
no random mutations. The mutate=0.1 option specifies a mutation probability of 0.1 during all three sets
of iterations. The mutate=0 0.1 0.05 option specifies no mutations during the first iterations, a mutation
probability of 0.1 during the search iterations, and a mutation probability of 0.05 during the refinement iterations.

mutiter= n1 < n2 < n3 >>

specifies the first iteration to consider mutating the design. The default is mutiter=. . ., which means
that the macro chooses values to use. The default is the first iteration that uses a fully random initial design in
each of the three sets of iterations. Hence by default, there are no random mutations of any part of the initial
design when part of the initial design comes from a tabled design.

optiter= n1 < n2 >

specifies the number of iterations to use in the OPTEX candidate-set based searches in the algorithm and design
search iterations. The default is optiter=. ., which means that the macro chooses values to use. When
the first value is “.” (missing), the macro will choose a value usually no smaller than 20 for larger problems and
usually no larger than 200 for smaller problems. However, maxtime= values other than the defaults can make
the macro choose values outside this range. When the second value is missing, the macro will choose a value
based on how long the first OPTEX run took and the value of maxtime=, but no larger than 5000. When a
missing value is specified for the first optiter= value, the default, the macro may choose to not perform any
OPTEX iterations to save time if it thinks it can find a perfect design without them.

tabiter= n1 < n2 >

specifies the number of times to try to improve a tabled or fractional-factorial initial design. The default is
tabiter=10 200, which means 10 tries in the algorithm search iterations and 200 tries in the design search
iterations.

unbalanced= n1 < n2 >

specifies the proportion of the tabiter= iterations to consider using unbalanced factors in the initial design.
The default is unbalanced=.2 .1. One way that unbalanced factors occur is through coding down. Coding
down for example creates a three-level factor from a four-level factor: (1 2 3 4) ) (1 2 3 3) or a two-level
factor from a three-level factor: (1 2 3) ) (1 2 2). For any particular problem, this strategy is probably either
going to work really well or not well at all, without much variability in the results, so it is not tried very often
by default. This option will try to create two-level through five-level factors from three-level through six-level
factors. It will not attempt for example to code down a twenty-level factor into a nineteen-level factor (although
the macro is often capable of in effect doing this through level swaps).
The Macros 343

Miscellaneous Options
This section contains some miscellaneous options that some users may occasionally find useful.

big= n < choose >

specifies the full-factorial-design size that is considered to be big. The default is big=5186 choose. The
default value was chosen, because 5186 is approximately 5000 and greater than 26 34 = 5184, 212 = 4096, and
2  37 = 4374. When the full-factorial design is smaller than the big= value, the %MktEx macro searches a
full-factorial candidate set. Otherwise, it searches a fractional-factorial candidate set. When choose is specified
as well (the default), the macro is allowed to choose to use a fractional-factorial even if the full-factorial design
is not too big, if it appears that the final design can be created from the fractional-factorial design. This may be
useful for example when you are requesting a fractional-factorial design with interactions. Using FACTEX to
create the fractional-factorial design may be a better strategy than searching a full-factorial design with PROC
OPTEX.

exchange= n
specifies the number of factors to consider at a time when exchanging levels. The default is exchange=1,
which means that the macro works with one factor at a time. You can specify exchange=2 to do pair-wise
exchanges. Pair-wise exchanges are much slower, but may produce better designs.

fixed= variable
specifies an init= data set variable that indicates which runs are fixed (cannot be changed) and which ones may
be changed. By default, no runs are fixed.
1 - (or any nonmissing) means this run may never change.
0 - means this run is used in the initial design, but it may be swapped out.
. - means this run should be randomly initialized, and it may be swapped out.
This option can be used to add holdout runs to a conjoint design, but see holdouts= for an easier way.

holdouts= n
adds holdout observations to the init= data set. This option augments an initial design. Specifying
holdouts=n optimally adds n runs to the init= design. The option holdouts=n works by adding a
fixed= variable and extra runs to the init= data set. Do not specify both fixed= and holdouts=. The
number of rows in the init= design, plus the value specified in holdouts= must equal the n= value.

stopearly= n
specifies that the macro may stop early when it keeps finding the same maximum D-efficiency over and over
again in different designs. The default is stopearly=5. By default, during the design search iterations and
refinement iterations, the macro will stop early if 5 times, the macro finds a D-efficiency essentially equal to the
maximum but not greater than the maximum. This may mean that the macro has found the optimal design, or it
may mean that the macro keeps finding a very attractive local optimum. Either way, it is unlikely it will do any
better. When the macro stops for this reason, the macro will print
NOTE: Stopping since it appears that no improvement is possible.
Specify either 0 or a very large value to turn off the stop-early checking.
344 TS-677E Multinomial Logit, Discrete Choice Modeling

tabsize= n
specifies which tabled (or FACTEX or Hadamard) design is used for the partial initialization when an exact match
to a tabled design is not found. Specify the number of runs in the tabled design. By default, the macro chooses a
tabled design that bests matches the specified design.

target= n
specifies the target efficiency criterion. The default is target=100. The macro stops when it finds an efficiency
value greater than or equal to this number. If you know what the maximum efficiency criterion is, or you know
how big is big enough, you can sometimes make the macro run faster by allowing it to stop when it reaches the
specified efficiency.

Esoteric Options
This last set of options contains all of the other miscellaneous options. Most of the time, most users should not
specify options from this list.

annealfun= function
specifies the function that controls how the simulated annealing probability changes with each pass through
the design. The default is annealfun=anneal # 0.85. Note that the IML operator # performs ordinary
(scalar) multiplication. Most users will never need this option.

detfuzz= n
specifies the value used to determine if determinants are changing. The default is detfuzz=1e-8. If newde-
ter > olddeter * (1 + detfuzz) then the new determinant is larger. If newdeter > olddeter
* (1 - detfuzz) then the new determinant is the same. Otherwise the new determinant is smaller. Most
users will never need this option.

imlopts= options
specifies IML PROC statement options. For example, for very large problems, you can use this option to specify
the IML symsize= or worksize= options: imlopts=symsize=n worksize=m, substituting numeric
values for n and m. The defaults for these options are host dependent. Most users will never need this option.

ridge= n
specifies the value to add to the diagonal of X0 X to make it nonsingular. The default is ridge=1e-7. Usually,
for normal problems, you will not need to change this value. If you want the macro to create designs with more
parameters than runs, you must specify some other value, usually something like 0.01. By default, the macro
will quit when there are more parameters than runs. Specifying a ridge= value other than the default (even if
you just change the ’e’ in 1e-7 to ’E’) allows the macro to create a design with more parameters than runs. Most
users will never need this option.

%MktKey Macro
The %MktKey macro creates expanded lists of variable names.
%mktkey(x1-x15)

The %MktKey macro produced the following line.

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 x15
You can cut and paste this list to make it easier to construct the key= data set for the %MktRoll macro.
The Macros 345

data key;
input (x1-x5) ($);
datalines;
x1 x2 x3 x4 x5
x6 x7 x8 x9 x10
x11 x12 x13 x14 x15
. . . . .
;

%MktKey Macro Options

The only argument to the %MktKey macro is a variable list.

list
specifies a variable list. Note that the variable list is a positional parameter and it is not specified after a name
and an equal sign.

%MktLab Macro
The macro %MktLab is used to process an experimental design, usually created by the %MktEx macro, and
assign the final variable names and levels.
For example, say you used the %MktEx macro to create a design with 11 two-level factors (with default levels
of 1 and 2).
%mktex(n=12, options=nosort)

proc print noobs; run;

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11

1 2 1 2 2 2 1 1 1 2 1
1 1 2 1 2 2 2 1 1 1 2
2 1 1 2 1 2 2 2 1 1 1
1 2 1 1 2 1 2 2 2 1 1
1 1 2 1 1 2 1 2 2 2 1
1 1 1 2 1 1 2 1 2 2 2
2 1 1 1 2 1 1 2 1 2 2
2 2 1 1 1 2 1 1 2 1 2
2 2 2 1 1 1 2 1 1 2 1
1 2 2 2 1 1 1 2 1 1 2
2 1 2 2 2 1 1 1 2 1 1
2 2 2 2 2 2 2 2 2 2 2

The %MktLab macro can be used to assign levels of -1 and 1, add an intercept, and change the variable name
prefixes from x to Had. This creates a Hadamard matrix (although, of course, the Hadamard matrix can have
any set of variable names).
%mktlab(data=design, values=-1 1, int=Had0, prefix=Had);

proc print; run;

346 TS-677E Multinomial Logit, Discrete Choice Modeling

Here is the resulting Hadamard matrix:

Had0 Had1 Had2 Had3 Had4 Had5 Had6 Had7 Had8 Had9 Had10 Had11

1 -1 1 -1 1 1 1 -1 -1 -1 1 -1
1 -1 -1 1 -1 1 1 1 -1 -1 -1 1
1 1 -1 -1 1 -1 1 1 1 -1 -1 -1
1 -1 1 -1 -1 1 -1 1 1 1 -1 -1
1 -1 -1 1 -1 -1 1 -1 1 1 1 -1
1 -1 -1 -1 1 -1 -1 1 -1 1 1 1
1 1 -1 -1 -1 1 -1 -1 1 -1 1 1
1 1 1 -1 -1 -1 1 -1 -1 1 -1 1
1 1 1 1 -1 -1 -1 1 -1 -1 1 -1
1 -1 1 1 1 -1 -1 -1 1 -1 -1 1
1 1 -1 1 1 1 -1 -1 -1 1 -1 -1
1 1 1 1 1 1 1 1 1 1 1 1

Here is an alternative way of doing the same thing using a key= data set.
data key;
array Had[11];
input Had1 @@;
do i = 2 to 11; Had[i] = Had1; end;
drop i;
datalines;
-1 1
;
proc print data=key; run;

Here is the key= data set.

Obs Had1 Had2 Had3 Had4 Had5 Had6 Had7 Had8 Had9 Had10 Had11

1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
2 1 1 1 1 1 1 1 1 1 1 1

%mktlab(data=design, key=key, int=Had0);

The Hadamard matrix from this step (not shown) is exactly the same as above.
The key= data set contains all of the variables that you want in the design and all of their levels. This information
will be applied to the design, by default the one stored in a data set called RANDOMIZED, which is the default
outr= data set name from the %MktEx macro. The results are stored in a new data set, FINAL, with the desired
factor names and levels.
Consider the consumer food product example from Kuhfeld, Tobias, and Garratt (1994). Here is one possible
design.
data randomized;
input x1-x8 @@;
datalines;
4 2 1 1 1 2 2 2 2 1 1 2 1 3 1 3 3 4 2 2 1 3 2 3 4 3 2 1 3 2 2 3 4 1 2 1
1 1 1 1 2 4 1 2 1 2 1 1 1 2 1 2 3 3 2 1 2 2 2 2 2 2 2 3 1 4 2 1 1 2 2 2
3 2 2 1 3 1 2 1 1 4 1 2 2 3 1 2 1 3 2 2 1 3 1 1 3 2 1 2 2 1 2 3 3 4 1 1
3 1 1 3 4 1 2 2 2 1 2 1 2 3 2 1 2 3 2 2 2 1 2 1 3 3 1 3 4 2 2 2 1 3 1 2
2 4 2 2 3 1 1 2 3 1 2 2 3 2 1 2 3 3 1 1 2 3 1 1 4 4 2 1 2 2 1 3 1 1 1 1
3 2 1 2 4 3 1 2 3 3 2 2 1 2 2 1 2 1 1 3 1 3 1 1 1 1 2 3
;
The Macros 347

Designs created by the %MktEx macro always have factor names x1, x2, ..., and so on, and the levels are
consecutive integers beginning with 1 (1, 2 for two-level factors; 1, 2, 3 for three-level factors; and so on). The
%MktLab macro provides you with a convenient way to change the names and levels to more meaningful values.
The data set KEY contains the variable names and levels that you ultimately want.
data key;
missing N;
input Client ClientLineExtension ClientMicro $ ShelfTalker $
Regional Private PrivateMicro $ NationalLabel;
format _numeric_ dollar5.2;
datalines;
1.29 1.39 micro Yes 1.99 1.49 micro 1.99
1.69 1.89 stove No 2.49 2.29 stove 2.39
2.09 2.39 . . N N . N
N N . . . . . .
;
%mktlab(key=key);

proc sort; by shelftalker; run;

proc print; by shelftalker; run;

The variable Client with 4 levels will be made from x1, ClientLineExtension with 4 levels will be
made from x2, ClientMicro with 2 levels will be made from x3 The N (for not available) is treated as a
special missing value. The KEY data set has four rows because the maximum number of levels is four. Factors
with fewer than four levels are filled in with ordinary missing values. The macro %MktLab takes by default
the data=randomized data set from %MktEx and the key=key data set, and combines the information to
create the out=final data set, which is shown below, sorted by the shelf talker variable.
Here is some of the design:

----------------------------- ShelfTalker=No ------------------------------

Client
Line Client Private National
Obs Client Extension Micro Regional Private Micro Label

1 $1.69 $1.39 micro $1.99 N micro N

2 $2.09 N stove $1.99 N stove N
3 $1.69 N micro $1.99 $2.29 micro $1.99
.
.
.

----------------------------- ShelfTalker=Yes -----------------------------

Client
Line Client Private National
Obs Client Extension Micro Regional Private Micro Label

14 N $1.89 micro $1.99 $2.29 stove $2.39

15 N $2.39 stove N $2.29 stove N
16 N $1.39 stove $1.99 $1.49 micro $1.99
.
.
.
348 TS-677E Multinomial Logit, Discrete Choice Modeling

This macro creates the out= data set by repeatedly reading and rereading the key= data set, one datum at a
time, using the information in the data= data set to determine which levels to read from the key= data set. In
this example, for the first observation, x1=4 so the fourth value of the first key= variable is read, then x2=2 so
the second value of the second key= variable is read, then x3=1 so the first value of the third key= variable is
read, ..., then x8=2 so the second value of the eighth key= variable is read, then the first observation is output.
This continues for all observations. This is why the data= data set must have integer values beginning with 1.
This example creates the L36 , renames the two-level factors two1-two11 and assigns them values -1, 1, and
renames the three-level factors thr1-thr12 and assigns them values -1, 0, 1.
%mktex(n=36)

data key;
array x[23] two1-two11 thr1-thr12;
input two1 thr1;
do i = 2 to 11; x[i] = two1; end;
do i = 13 to 23; x[i] = thr1; end;
drop i;
datalines;
-1 -1
1 0
. 1
;
%mktlab(key=key);

proc print data=key noobs; var two:; run;

proc print data=key noobs; var thr:; run;

proc print data=final(obs=5) noobs; var two:; run;

proc print data=final(obs=5) noobs; var thr:; run;
Here is the KEY data set.

two1 two2 two3 two4 two5 two6 two7 two8 two9 two10 two11

-1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
1 1 1 1 1 1 1 1 1 1 1
. . . . . . . . . . .

thr1 thr2 thr3 thr4 thr5 thr6 thr7 thr8 thr9 thr10 thr11 thr12

-1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1
0 0 0 0 0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1 1 1 1 1

Here are the first five rows of the design.

two1 two2 two3 two4 two5 two6 two7 two8 two9 two10 two11

1 -1 1 1 1 1 -1 1 1 1 -1
-1 1 1 -1 -1 1 1 1 1 1 1
-1 -1 -1 -1 1 1 1 1 -1 -1 -1
1 1 1 -1 1 -1 -1 1 -1 -1 1
1 -1 1 1 -1 1 1 -1 -1 -1 1
The Macros 349

thr1 thr2 thr3 thr4 thr5 thr6 thr7 thr8 thr9 thr10 thr11 thr12

1 0 1 0 0 1 0 0 1 -1 1 1
1 1 1 1 -1 -1 0 -1 -1 -1 0 -1
-1 0 1 -1 0 0 -1 -1 0 -1 0 0
0 0 0 1 -1 -1 -1 1 1 -1 1 0
0 -1 1 -1 1 -1 1 -1 1 -1 -1 1

This next step creates a design and blocks it. This example shows that it is OK if not all of the variables in the
input design are used. The variables Block, Run, and x4 are just copied from the input to the output.
%mktex(n=18, seed=396)

%mktblock(nblocks=2, factors=x1-x4, seed=292)

data key;
input Brand $ Price Size;
format price dollar5.2;
datalines;
Acme 1.49 6
Apex 1.79 8
. 1.99 12
;

%mktlab(data=blocked, key=key)

proc print; id block run; by block; run;

Here are the results:

Block Run Brand Price Size x4

1 1 Acme $1.79 6 1
2 Acme $1.79 8 3
3 Acme $1.99 8 2
4 Acme $1.99 12 1
5 Apex $1.49 8 3
6 Apex $1.49 12 2
7 Apex $1.79 6 3
8 Apex $1.79 12 1
9 Apex $1.99 6 2

2 1 Acme $1.49 6 2
2 Acme $1.49 8 1
3 Acme $1.49 12 3
4 Acme $1.79 12 2
5 Acme $1.99 6 3
6 Apex $1.49 6 1
7 Apex $1.79 8 2
8 Apex $1.99 8 1
9 Apex $1.99 12 3
350 TS-677E Multinomial Logit, Discrete Choice Modeling

This next example illustrates using the labels= option. This option is more typically used with values=
input, rather than when you construct the key= data set yourself, but it can be used either way. This example is
from the Vacation Example.
%mktex(3 ** 15, n=36, seed=17, maxtime=0)

%mktblock(data=randomized, nblocks=2, factors=x1-x15, seed=448)

%macro lab;
label X1 = ’Hawaii, Accommodations’
X2 = ’Alaska, Accommodations’
X3 = ’Mexico, Accommodations’
X4 = ’California, Accommodations’
X5 = ’Maine, Accommodations’
X6 = ’Hawaii, Scenery’
X7 = ’Alaska, Scenery’
X8 = ’Mexico, Scenery’
X9 = ’California, Scenery’
X10 = ’Maine, Scenery’
X11 = ’Hawaii, Price’
X12 = ’Alaska, Price’
X13 = ’Mexico, Price’
X14 = ’California, Price’
X15 = ’Maine, Price’;
format x11-x15 dollar5.;
%mend;

data key;
length x1-x5 $ 16 x6-x10 $ 8 x11-x15 8;
input x1 & $ x6 $ x11;
x2 = x1; x3 = x1; x4 = x1; x5 = x1;
x7 = x6; x8 = x6; x9 = x6; x10 = x6;
x12 = x11; x13 = x11; x14 = x11; x15 = x11;
datalines;
Cabin Mountains 999
Bed & Breakfast Lake 1249
Hotel Beach 1499
;

%mktlab(data=blocked, key=key, labels=lab)

proc contents p; ods select position; run;

Here is the variable name, label, and format information.
The Macros 351

The CONTENTS Procedure

Variables in Creation Order

# Variable Type Len Format Label

1 x1 Char 16 Hawaii, Accommodations

2 x2 Char 16 Alaska, Accommodations
3 x3 Char 16 Mexico, Accommodations
4 x4 Char 16 California, Accommodations
5 x5 Char 16 Maine, Accommodations
6 x6 Char 8 Hawaii, Scenery
7 x7 Char 8 Alaska, Scenery
8 x8 Char 8 Mexico, Scenery
9 x9 Char 8 California, Scenery
10 x10 Char 8 Maine, Scenery
11 x11 Num 8 DOLLAR5. Hawaii, Price
12 x12 Num 8 DOLLAR5. Alaska, Price
13 x13 Num 8 DOLLAR5. Mexico, Price
14 x14 Num 8 DOLLAR5. California, Price
15 x15 Num 8 DOLLAR5. Maine, Price
16 Block Num 8
17 Run Num 8

%MktLab Macro Options

The following options can be used with the %MktLab macro.

data= SAS-data-set
specifies the input data set with the experimental design, usually created by the %MktEx macro. The default is
data=Randomized. The factor levels in the data= data set must be consecutive integers beginning with 1.

dolist= do-list
specifies the new values, using a do-list syntax (n TO m <BY p>), for example: 1 to 10 or 0 to 9. With
asymmetric designs (not all factors have the same levels), specify the levels for the largest number of levels. For
example, with two-level and three-level factors and dolist=0 to 2, the two-level factors will be assigned
levels 0 and 1, and the three-level factors will be assigned levels 0, 1, and 2. Do not specify both values= and
dolist=. By default, when key=, values=, and dolist= are all not specified, the default value list comes
from dolist=1 to 100.

int= variable-list
specifies the name of an intercept variable (column of ones), if you want an intercept added to the out= data set.
You can also specify a variable list instead of a variable name if you would like to make a list of variables with
values all one. This can be useful for example, for generic choice models, for creating flag variables when the
design is going to be used as a candidate set for the %ChoicEff macro.

key= SAS-data-set
specifies the input data set with the key to recoding the design. When values= or dolist= is specified, this
data set is made for you. By default, when key=, values=, and dolist= are all not specified, the default
value list comes from dolist=1 to 100.
352 TS-677E Multinomial Logit, Discrete Choice Modeling

labels= macro-name
specifies the name of a macro that provides labels, formats, or other additional information to the key= data
set. For a simple format specification, it is easier to use statements=. For more involved specifications, use
labels=. Note that you specify just the macro name, no percents on the labels= option. Example:
%mktex(3 ** 4, n=18, seed=205)
%macro labs;
label x1 = ’Sploosh’ x2 = ’Plumbob’
x3 = ’Platter’ x4 = ’Moosey’;
format x1-x4 dollar5.2;
%mend;
%mktlab(values=1.49 1.99 2.49, labels=labs)
proc print label; run;
proc print label; run;

Obs Sploosh Plumbob Platter Moosey

1 $1.49 $1.49 $1.99 $2.49

2 $2.49 $1.99 $1.49 $2.49
3 $1.49 $1.99 $1.49 $1.99
4 $1.49 $1.99 $1.99 $1.49
.
.
.

out= SAS-data-set
specifies the output data set with the final, recoded design. The default is out=final.

prefix= variable-prefix
specifies a prefix for naming variables when values= is specified. For example prefix=Var creates variables
Var1, Var2, and so on. By default, the variables are x1, x2, .... This option is ignored when vars= is specified.

statements= SAS-code
is an alternative to labels= that you can use to add extra statements to the key= data set. For a simple format
specification, it is easier to use statements=. For more involved specifications, use labels=. Example:
%mktex(3 ** 4, n=18, seed=205)

%mktlab(values=1.49 1.99 2.49,

vars=Sploosh Plumbob Platter Moosey,
statements=format Sploosh Plumbob Platter Moosey dollar5.2)

proc print; run;

Obs Sploosh Plumbob Platter Moosey

1 $1.49 $1.49 $1.99 $2.49

2 $2.49 $1.99 $1.49 $2.49
3 $1.49 $1.99 $1.49 $1.99
4 $1.49 $1.99 $1.99 $1.49
.
.
.
The Macros 353

values= value-list
specifies the new values for all of the variables. If all variables will have the same value, it is easier to specify
values= or dolist= than key=. When you specify values=, the key= data set is created for you. Specify
a list of levels separated by blanks. If your levels contain blanks, separate them with two blanks. With asymmetric
designs (not all factors have the same levels) specify the levels for the largest number of levels. For example,
with two-level and three-level factors and values=a b c, the two-level factors will be assigned levels ’a’
and ’b’, and the three-level factors will be assigned levels ’a’, ’b’, and ’c’. Do not specify both values=
and dolist=. By default, when key=, values=, and dolist= are all not specified, the default value list
comes from dolist=1 to 100.

vars= variable-list
specifies a list of variable names when values= or dolist= is specified. If vars= is not specified with
values=, then prefix= is used.

%MktMerge Macro
The %MktMerge autocall macro merges a data set containing an experimental design for a choice model with
the data for the choice model. Here is a typical usage of the macro.
%mktmerge(design=rolled, data=results, out=res2,
nsets=18, nalts=5, setvars=choose1-choose18)
The design= data set comes from the %MktRoll macro. The data= data set contains the data, and the
setvars= variables in the data= data set contain the numbers of the chosen alternatives for each of the 18
choice sets. The nsets= option specifies the number of choice sets, and the nalts= option specifies the
number of alternatives. The out= option names the output SAS data set that contains the experimental design
and a variable c that contains 1 for the chosen alternatives (first choice) and 2 for unchosen alternatives (second
or subsequent choice).
When the data= data set contains a blocking variable, name it on the blocks= option. When there is blocking,
it is assumed that the design= data set contains blocks of nalts  nsets observations. The blocks= variable
must contain values 1, 2, ..., n for n blocks. Here is an example of using the %MktMerge macro with blocking.
%mktmerge(design=rolled, data=results, out=res2, blocks=form,
nsets=18, nalts=5, setvars=choose1-choose18)

%MktMerge Macro Options

The following options can be used with the %MktMerge macro. You must specify the design=, nalts=,
nsets=, and setvars= options.

blocks= 1|variable
specifies either a 1 (the default) if there is no blocking or the name of a variable in the data= data set that
contains the block number. When there is blocking, it is assumed that the design= data set contains blocks
of nalts  nsets observations, one set per block. The blocks= variable must contain values 1, 2, ..., n for n
blocks.

data= SAS-data-set
specifies an input SAS data set with data for the choice model. By default, the data= data set is the last data set
created.
354 TS-677E Multinomial Logit, Discrete Choice Modeling

design= SAS-data-set
specifies an input SAS data set with the choice design. This data set could have been created for example with
the %MktRoll macro. This option must be specified.

nalts= n
specifies the number of alternatives. This option must be specified.

nsets= n
specifies the number of choice sets. This option must be specified.

out= SAS-data-set
specifies the output SAS data set. If out= is not specified, the DATAn convention is used. This data set contains
the experimental design and a variable c that contains 1 for the chosen alternatives (first choice) and 2 for
unchosen alternatives (second or subsequent choice).

setvars= variable-list
specifies a list of variables, one per choice set, in the data= data set that contain the numbers of the chosen
alternatives. It is assumed that the values of these variables range from 1 to nalts. This option must be specified.

stmts= SAS-statements
specifies additional statements like format and label statements. Example:
%mktmerge(design=rolled, data=results, out=res2, blocks=form,
nsets=&n, nalts=&m, setvars=choose1-choose&n,
stmts=%str(price = input(put(price, price.), 5.);
format scene scene. lodge lodge.;))

%MktOrth Macro
The %MktOrth macro lists some of the 100% orthogonal main-effects plans that the %MktEx macro can gen-
erate, up through 100 runs. Here is a typical usage.
%mktorth;

The macro creates data sets and no printed output.

NOTE: The data set [Link] has 345 observations and 53 variables.
NOTE: The data set [Link] has 345 observations and 3 variables.
Here are the first few and last few designs in the marketing design catalogue (data set MKTDESCAT).
proc print data=mktdescat(where=(n le 12 or n ge 98)); run;
The Macros 355

Obs n Design Reference

1 4 2 ** 3 Hadamard
2 6 2 ** 1 3 ** 1 Full-factorial
3 8 2 ** 7 Hadamard
4 9 3 ** 4 Fractional-factorial
5 10 2 ** 1 5 ** 1 Full-factorial
6 12 2 ** 11 Hadamard
7 12 2 ** 4 3 ** 1 Hedayat, Sloane, and Stufken, 1999
8 12 2 ** 2 6 ** 1 Hedayat, Sloane, and Stufken, 1999
9 12 3 ** 1 4 ** 1 Full-factorial
343 98 2 ** 1 49 ** 1 Full-factorial
344 99 3 ** 2 11 ** 1 Full-factorial
345 100 2 ** 99 Hadamard

If you just want to display a list of designs, possibly selecting on n, the number of runs, you can use the MKT-
DESCAT data set. However, if you would like to do more advanced processing, based on the numbers of levels of
some of the factors, you can use the outlev=mktdeslev data set to select potential designs. You can look at
the level information in MKTDESLEV and see the number of two-level factors in x2, the number of three-level
factors in x3, ..., and the number of fifty-level factors is in x50. The number of one level factors, x1, is always
zero, but x1 is available so you can make arrays (for example, array x[50]) and have x[2] refer to x2, the
number of two-level factors.
Say you are interested in the design 25 35 41 . Here are the ways in which it is available.
proc print data=mktdeslev(where=(x2 ge 5 and x3 ge 5 and x4 ge 1));
var n design reference;
run;

Obs n Design Reference

289 72 2 ** 36 3 ** 13 4 ** 1 Hedayat, Sloane, and Stufken, 1999

294 72 2 ** 20 3 ** 24 4 ** 1 Wang, 1996
297 72 2 ** 13 3 ** 25 4 ** 1 Wang, 1996
299 72 2 ** 11 3 ** 24 4 ** 1 6 ** 1 Wang, 1996

Here is how you can see all the designs in a certain range of sizes.
proc print; where 12 le n le 20; run;

Obs n Design Reference

6 12 2 ** 11 Hadamard
7 12 2 ** 4 3 ** 1 Hedayat, Sloane, and Stufken, 1999
8 12 2 ** 2 6 ** 1 Hedayat, Sloane, and Stufken, 1999
9 12 3 ** 1 4 ** 1 Full-factorial
10 14 2 ** 1 7 ** 1 Full-factorial
11 15 3 ** 1 5 ** 1 Full-factorial
12 16 2 ** 15 Hadamard
13 16 2 ** 12 4 ** 1 Fractional-factorial
14 16 2 ** 9 4 ** 2 Fractional-factorial
15 16 2 ** 6 4 ** 3 Fractional-factorial
16 16 2 ** 3 4 ** 4 Fractional-factorial
17 16 4 ** 5 Fractional-factorial
18 18 2 ** 1 3 ** 7 Taguchi, 1987
19 18 3 ** 6 6 ** 1 Taguchi, 1987
20 20 2 ** 19 Hadamard
356 TS-677E Multinomial Logit, Discrete Choice Modeling

21 20 2 ** 8 5 ** 1 Wang and Wu, 1992

22 20 2 ** 2 10 ** 1 Hedayat, Sloane, and Stufken, 1999
23 20 4 ** 1 5 ** 1 Full-factorial

%MktOrth Macro Options

The following options can be used with the %MktOrth macro.

outall= SAS-data-set
specifies the output data set with all designs. This is like the outlev= data set, except larger. The outall=
data set includes all of the %MktEx design catalogue, including all of the smaller designs that can be trivially
made from larger designs by dropping factors. For example, when the outlev= data set has x2=2 x3=2, then
the outall= data set has that design and also x1=2 x3=1, x1=1 x3=2, and x1=1 x2=1. This data set is not
created by default.

outcat= SAS-data-set
specifies the output data set with the catalogue of designs that the %MktEx macro can create. The default is
outcat=MktDesCat.

outlev= SAS-data-set
specifies the output data set with the list of designs and 50 more variables: x2 - number of two-level factors, x3
- number of three-level factors and so on. The default is outlev=MktDesLev.

%MktRoll Macro
The %MktRoll autocall macro is used for manipulating the experimental design for choice experiments. It
takes as input a SAS data set containing an experimental design with one row per choice set, for example a
design created by the %MktEx macro. This data set is specified in the design= option. This data set has one
variable for each attribute of each alternative in the choice experiment.
The output from this macro is an out= SAS data set containing the experimental design with one row per
alternative per choice set. There is one column for each different attribute. For example, in a simple branded
study, design= could contain the variables x1-x5 which contain the prices of each of five alternative brands.
The output data set would have one factor, Price, that contains the price of each of the five alternatives. In
addition, it would have the number (or optionally the name) of each alternative.
The rules for determining the mapping between factors in the design= data set and the out= data set are
contained in the key= data set. For example, assume that the design= data set contains the variables x1-x5
which contain the prices of each of five alternative brands: Brand A, B, C, D, and E. Here is how you would
create the key= data set. The choice design has two factors, Brand and Price. Brand A price is made from
x1, Brand B price is made from x2, ..., and Brand E price is made from x5.
A convenient way to get all the names in a variable list like x1-x5 is with the %MktKey macro.
%mktkey(x1-x5)

The %MktKey macro produced the following line.

x1 x2 x3 x4 x5
The Macros 357

Here is the KEY data set.

data key;
input (Brand Price) ($);
datalines;
A x1
B x2
C x3
D x4
E x5
;
This data set has two variables. Brand contains the brand names, and Price contains the names of the factors
that are used to make the price effects for each of the alternatives. The out= data set will contain the variables
with the same names as the variables in the key= data set.
Here is how you can create the design with one row per choice set:
%mktex(3 ** 5, n=12)

Here is how you can create the design with one row per alternative per choice set:
%mktroll(design=randomized, key=key, out=[Link], alt=brand)

For example, if the data set RANDOMIZED contains the row:

Obs x1 x2 x3 x4 x5

9 3 1 1 2 1

then the data set [Link] contains the rows:

41 9 A 3
42 9 B 1
43 9 C 1
44 9 D 2
45 9 E 1

The price for Brand A is made from x1=3, ..., and the price for Brand E is made from x5=1.
Now assume that there are three alternatives, each a different brand, and each composed of four factors: Price,
Size, Color, and Shape. In addition, there is a constant alternative. First, the %MktEx macro is used to
create a design with 12 factors, one for each attribute of each alternative.
%mktex(2 ** 12, n=16)

Next, the key= data set is created. It shows that there are three brands, A, B, and C, and also None.
data key;
input (Brand Price Size Color Shape) ($);
datalines;
A x1 x2 x3 x4
B x5 x6 x7 x8
C x9 x10 x11 x12
None . . . .
;
Brand A is created from Brand = ’A’, Price = x1, Size = x2, Color = x3, Shape = x4.
Brand B is created from Brand = ’B’, Price = x5, Size = x6, Color = x7, Shape = x8.
Brand C is created from Brand = ’C’, Price = x9, Size = x10, Color = x11, Shape = x12.
358 TS-677E Multinomial Logit, Discrete Choice Modeling

The constant alternative is created from Brand = ’None’ and none of the attributes. The “.” notation is used
to indicate missing values in input data sets. The actual values in the KEY data set will be blank (character
missing).
Here is how you create the design with one row per alternative per choice set:
%mktroll(key=key, design=randomized, out=[Link], alt=brand)

For example, if the data set RANDOMIZED contains the row:

Obs x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12

8 2 1 2 1 1 2 2 2 1 2 1 1

then the data set [Link] contains the rows:

29 8 A 2 1 2 1
30 8 B 1 2 2 2
31 8 C 1 2 1 1
32 8 None . . . .

Now assume like before that there are three branded alternatives, each composed of four factors: Price, Size,
Color, and Shape. In addition, there is a constant alternative. Also, there is an alternative-specific factor,
Pattern, that only applies to Brand A and Brand C. First, the %MktEx macro is used to create a design with
14 factors, one for each attribute of each alternative.
%mktex(2 ** 14, n=16)

Next, the key= data set is created. It shows that there are three brands, A, B, and C, plus None.
data key;
input (Brand Price Size Color Shape Pattern) ($);
datalines;
A x1 x2 x3 x4 x13
B x5 x6 x7 x8 .
C x9 x10 x11 x12 x14
None . . . . .
;
Brand A is created from Brand = ’A’, Price = x1, Size = x2, Color = x3, Shape = x4, Pattern =
x13.
Brand B is created from Brand = ’B’, Price = x5, Size = x6, Color = x7, Shape = x8.
Brand C is created from Brand = ’C’, Price = x9, Size = x10, Color = x11, Shape = x12, Pattern =
x14.
The constant alternative is Brand = ’None’ and none of the attributes.
Here is how you can create the design with one row per alternative per choice set:
%mktroll(key=key, design=randomized, out=[Link], alt=brand)
The Macros 359

For example, if the data set RANDOMIZED contains the row:

Obs x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14

8 1 2 2 1 1 2 1 2 2 1 2 1 2 2

then the data set [Link] contains the rows:

Obs Set Brand Price Size Color Shape Pattern

29 8 A 1 2 2 1 2
30 8 B 1 2 1 2 .
31 8 C 2 1 2 1 2
32 8 None . . . . .

Now assume we are going to fit a model with price cross effects so we need x1, x5, and x9 (the three price
effects) available in the out= data set.
%mktroll(key=key, design=randomized, out=[Link], alt=brand,
keep=x1 x5 x9)
Now the data set also contains the three original price variables.

Obs Set Brand Price Size Color Shape Pattern x1 x5 x9

29 8 A 1 2 2 1 2 1 1 2
30 8 B 1 2 1 2 . 1 1 2
31 8 C 2 1 2 1 2 1 1 2
32 8 None . . . . . 1 1 2

Every value in the key= data set must appear as a variable in the design= data set. The macro prints a warning
if it encounters a variable name in the design= data set that does not appear as a value in the key= data set.

%MktRoll Macro Options

The following options can be used with the %MktRoll macro. You must specify the design= and key=
options.

alt= variable
specifies the variable in the key= data set that contains the name of each alternative. Often this will be something
like alt=Brand. When alt= is not specified, the macro creates a variable - Alt- that contains the alternative
number.

design= SAS-data-set
specifies an input SAS data set with one row per choice set. The design= option must be specified.

keep= variable-list
specifies factors from the design= data set that should also be kept in the out= data set. This option is useful
to keep terms that will be used to create cross effects.

key= SAS-data-set
specifies an input SAS data set containing the rules for mapping the design= data set to the out= data set. The
key= option must be specified.
360 TS-677E Multinomial Logit, Discrete Choice Modeling

options= options-list
specifies binary options. By default, none of these options are specified. Specify one or more of the following
values after options=.

notes
do not specify options nonotes during most of the macro.

nowarn
do not print a warning when the design= data set contains variables not mentioned in the KEY= data set.
Sometimes this is perfectly fine.

out= SAS-data-set
specifies the output SAS data set. If out= is not specified, the DATAn convention is used.

set= variable
specifies the variable in the out= data set that will contain the choice set number. By default, this variable is
named Set.

%MktRuns Macro
The %MktRuns autocall macro suggests reasonable sizes for main-effects experimental designs. It tries to find
sizes in which perfect balance and orthogonality can occur, or at least sizes in which violations of orthogonality
and balance are minimized. Typically, the macro takes one argument, a list of the number of levels of each factor.
For example, with 3 two-level and 4 three-level factors, specify either of the following.
%mktruns( 2 2 2 3 3 3 3 )

%mktruns( 2 ** 3 3 ** 4 )
The output from the macro in this example is:

Design Summary

Number of
Levels Frequency

2 3
3 4
The Macros 361

Saturated = 12
Full Factorial = 648

Some Reasonable Cannot Be

Design Sizes Violations Divided By

36 * 0
72 * 0
18 3 4
54 3 4
12 6 9
24 6 9
48 6 9
60 6 9
30 9 4 9
42 9 4 9

* - 100% Efficient Design can be made with the MktEx Macro.

n Design Reference

36 2 ** 13 3 ** 4 Suen, 1989
36 2 ** 11 3 ** 12 Taguchi, 1987
36 2 ** 4 3 ** 13 Taguchi, 1987
72 2 ** 49 3 ** 4 Hedayat, Sloane, and Stufken, 1999
72 2 ** 47 3 ** 12 Wang, 1996
72 2 ** 40 3 ** 13 Wang and Wu, 1991
72 2 ** 38 3 ** 12 6 ** 1 Hedayat, Sloane, and Stufken, 1999
72 2 ** 37 3 ** 8 6 ** 2 Hedayat, Sloane, and Stufken, 1999
72 2 ** 36 3 ** 13 4 ** 1 Hedayat, Sloane, and Stufken, 1999
72 2 ** 36 3 ** 12 12 ** 1 Hedayat, Sloane, and Stufken, 1999
72 2 ** 36 3 ** 7 6 ** 3 Hedayat, Sloane, and Stufken, 1999
72 2 ** 23 3 ** 24 Dey, 1985
72 2 ** 20 3 ** 24 4 ** 1 Wang, 1996
72 2 ** 16 3 ** 25 Wang, 1996
72 2 ** 14 3 ** 24 6 ** 1 Wang, 1996
72 2 ** 13 3 ** 25 4 ** 1 Wang, 1996
72 2 ** 12 3 ** 24 12 ** 1 Hedayat, Sloane, and Stufken, 1999
72 2 ** 11 3 ** 24 4 ** 1 6 ** 1 Wang, 1996

The macro reports that the saturated design has 12 runs and that 36 is an optimal design size. The macro picks
36, because it is the smallest integer >= 12 that can be divided by 2, 3, 2  2, 2  3, and 3  3. The macro also
reports 18 as a reasonable size. There are three violations with 18 because 18 cannot be divided by each of the
three pairs of 2  2, so perfect orthogonality in the two-level factors will not be possible with 18 runs. Larger
sizes are reported as well. The macro prints orthogonal designs that are available from the %MktEx macro that
match your specification.
To see every size the macro considered, simply run PROC PRINT after the macro finishes. The output from this
step is not shown.
proc print label data=nums split=’-’;
id n;
run;
For 2 two-level factors, 2 three-level factors, 2 four-level factors, and 2 five-level factors specify:
%mktruns( 2 2 3 3 4 4 5 5 )
362 TS-677E Multinomial Logit, Discrete Choice Modeling

Here are the results:

Design Summary

Number of
Levels Frequency

2 2
3 2
4 2
5 2

Saturated = 21
Full Factorial = 14,400

Some Reasonable Cannot Be

Design Sizes Violations Divided By

120 3 9 16 25
180 6 8 16 25
60 7 8 9 16 25
144 15 5 10 15 20 25
48 16 5 9 10 15 20 25
72 16 5 10 15 16 20 25
80 16 3 6 9 12 15 25
96 16 5 9 10 15 20 25
160 16 3 6 9 12 15 25
192 16 5 9 10 15 20 25

Among the smaller design sizes, 60 or 48 look like good possibilities. The macro has an optional keyword
parameter: max=. It specifies the maximum number of sizes to try. The smallest design that is considered is
the saturated design. Usually you will not need to specify the max= option. For example, this specification tries
5000 sizes (21 to 5020) and reports that a perfect design can be found with 3600 runs.
%mktruns(2 2 3 3 4 4 5 5, max=5000)

Design Summary

Number of
Levels Frequency

2 2
3 2
4 2
5 2

Saturated = 21
Full Factorial = 14,400

Some Reasonable Cannot Be

Design Sizes Violations Divided By

3600 0
720 1 25
1200 1 9
1440 1 25
1800 1 16
The Macros 363

2160 1 25
2400 1 9
2880 1 25
4320 1 25
4800 1 9

Now consider again the problem with 3 two-level and 4 three-level factors, but this time we want to be estimable
the interaction of two of the two-level factors. Now, instead of specifying %mktruns( 2 2 2 3 3 3 3 ),
we replace two of the 2’s with a 4.
%mktruns( 2 4 3 3 3 3 )

Design Summary

Number of
Levels Frequency

2 1
3 4
4 1

Saturated = 13
Full Factorial = 648

Some Reasonable Cannot Be

Design Sizes Violations Divided By

72 * 0
144 0
36 1 8
108 1 8
18 6 4 8 12
24 6 9
48 6 9
54 6 4 8 12
90 6 4 8 12
96 6 9

* - 100% Efficient Design can be made with the MktEx Macro.

n Design Reference

72 2 ** 36 3 ** 13 4 ** 1 Hedayat, Sloane, and Stufken, 1999

72 2 ** 20 3 ** 24 4 ** 1 Wang, 1996
72 2 ** 13 3 ** 25 4 ** 1 Wang, 1996
72 2 ** 11 3 ** 24 4 ** 1 6 ** 1 Wang, 1996

Now we need 72 runs for perfect balance and orthogonality and there are six violations in 18 runs (4, 4  2, 4  3,
4  3, 4  3, and 4  3).
If you ever get errors running this macro, like invalid page errors, see “Macro Errors” on page 288.
364 TS-677E Multinomial Logit, Discrete Choice Modeling

%MktRuns Macro Options

The following options can be used with the %MktRuns macro. The %MktRuns macro has one positional
parameter, list, and several keyword parameters.

list
specifies a list of the numbers of levels of all the factors. For example, for 3 two-level factors specify ei-
ther 2 2 2 or 2 ** 3. Lists of numbers, like 2 2 3 3 4 4 or a levels**number of factors syntax like:
2**2 3**2 4**2 can be used, or both can be combined: 2 2 3**4 5 6. The specification 3**4 means
four three-level factors. You must specify a list. Note that the factor list is a positional parameter. This means it
must come first, and unlike all other parameters, it is not specified after a name and an equal sign.

n= n
specifies the design size to evaluate. By default, this option is not specified, and the max= option specification
provides a range of design sizes to evaluate.

max= n < m >

specifies the maximum number of design sizes to try. By default, max=200 2. The macro tries up to n sizes
starting with the saturated design. The macro stops trying larger sizes when it finds a design size with zero
violations that is m times as big as a previously found size with zero violations. The macro reports the best 10
sizes. For example, if the saturated design has 10 runs, and there are zero violations in 16 runs, then by default,
the largest size that the macro will consider is 32 = 2  16 runs.

options= options-list
specifies binary options. By default, none of these options are specified. Specify one the following values after
options=.

justparse
is used by other Mkt macros to have this macro just parse the list argument and return it as a simple list of
integers.

out= SAS-data-set
specifies the name of a SAS data set with the suggested sizes. The default is out=nums.

%PhChoice Macro
The %PhChoice autocall macro is used to customize the discrete choice output from PROC PHREG. Typically,
you run the following macro once to customize the PROC PHREG output.
%phchoice(on)

The macro uses PROC TEMPLATE and ODS (Output Delivery System) to customize the output from PROC
PHREG. Running this code edits the templates and stores copies in SASUSER. These changes will remain in
effect until you delete them. Note that these changes assume that each effect in the choice model has a variable
label associated with it so there is no need to print variable names. If you are coding with PROC TRANSREG,
this will usually be the case. To return to the default output from PROC PHREG, run the following macro.
%phchoice(off)

If you ever have errors running this macro, like invalid page errors, see “Macro Errors” on page 288. The rest
of this section discusses the details of what the %PhChoice macro does and why. Unless you are interested in
further customization of the output, you should skip to “%PhChoice Macro Options” on page 368.
The Macros 365

We are most interested in the ’Analysis of Maximum Likelihood Estimates’ table, which contains the parameter
estimates. We can first use PROC TEMPLATE to identify the template for the parameter estimates table and then
edit the template. First, let’s have PROC TEMPLATE display the templates for PROC PHREG. The source
[Link] statement specifies that we want to see PROC TEMPLATE source code for the STAT product and
the PHREG procedure.
proc template;
source [Link];
run;
If we search the results for the ’Analysis of Maximum Likelihood Estimates’ table we find the following code,
which defines the [Link] table.
define table [Link];
notes "Parameter Estimates Table";
dynamic Confidence NRows;
column Variable DF Estimate StdErr StdErrRatio ChiSq ProbChiSq HazardRatio
HRLowerCL HRUpperCL Label;
header h1 h2;
define h1;
text "Analysis of Maximum Likelihood Estimates";
space = 1;
spill_margin;
end;
define h2;
text Confidence BEST8. %nrstr("%% Hazard Ratio Confidence Limits");
space = 0;
end = HRUpperCL;
start = HRLowerCL;
spill_margin = OFF;
end;
define Variable;
header = "Variable";
style = RowHeader;
id;
end;
define DF;
parent = [Link];
end;
define Estimate;
header = ";Parameter;Estimate;";
format = D10.;
parent = [Link];
end;
define StdErr;
header = ";Standard;Error;";
format = D10.;
parent = [Link];
end;
define StdErrRatio;
header = ";StdErr;Ratio;";
format = 6.3;
end;
define ChiSq;
parent = [Link];
end;
366 TS-677E Multinomial Logit, Discrete Choice Modeling

define ProbChiSq;
parent = [Link];
end;
define HazardRatio;
header = ";Hazard;Ratio;";
glue = 2;
format = 8.3;
end;
define HRLowerCL;
glue = 2;
format = 8.3;
print_headers = OFF;
end;
define HRUpperCL;
format = 8.3;
print_headers = OFF;
end;
define Label;
header = "Variable Label";
end;
col_space_max = 4;
col_space_min = 1;
required_space = NRows;
end;

It contains header, format, spacing and other information for each column in the table. Most of this need not
concern us now. The template contains this column statement, which lists the columns of the table.
column Variable DF Estimate StdErr StdErrRatio ChiSq ProbChiSq HazardRatio
HRLowerCL HRUpperCL Label;
Since we will usually have a label that adequately names each parameter, we do not need the variable column.
We also do not need the hazard information. If we move the label to the front of the list and drop the variable
column and the hazard columns, we get this.
column Label DF Estimate StdErr ChiSq ProbChiSq;

We use the edit statement to edit the template. We can also modify some headers. We specify the new column
statement and the new headers. We can also modify the Summary table ([Link])
to use the vocabulary of choice models instead of survival analysis models. The code is grabbed from the
PROC TEMPLATE step with the source statement. The overall header ’Summary of the Number of Event and
Censored Values’ is changed to ’Summary of Subjects, Sets, and Chosen and Unchosen Alternatives’, ’Total’ is
changed to ’Number of Alternatives’, ’Event’ is changed to ’Chosen Alternatives’, ’Censored’ is changed to ’Not
Chosen’, and ’Percent Censored’ is dropped. Finally Style=RowHeader was specified on the label column.
This sets the color, font, and general style for HTML output. The RowHeader style is typically used on first
columns that provide names or labels for the rows. Here is the code that the %phchoice(on) macro runs.
proc template;
edit [Link];
column Label DF Estimate StdErr ChiSq ProbChiSq;
header h1;
define h1;
text "Multinomial Logit Parameter Estimates";
space = 1;
spill_margin;
end;
The Macros 367

define Label;
header = " " style = RowHeader;
end;
end;
edit [Link];
column Stratum Pattern Freq GenericStrVar Total
Event Censored;
header h1;
define h1;
text "Summary of Subjects, Sets, "
"and Chosen and Unchosen Alternatives";
space = 1;
spill_margin;
first_panel;
end;
define Freq;
header=";Number of;Choices" format=6.0;
end;
define Total;
header = ";Number of;Alternatives";
format_ndec = ndec;
format_width = 8;
end;
define Event;
header = ";Chosen;Alternatives";
format_ndec = ndec;
format_width = 8;
end;
define Censored;
header = "Not Chosen";
format_ndec = ndec;
format_width = 8;
end;
end;

run;
Here is the code that %phchoice(off) runs.
* Delete edited templates, restore original templates;
proc template;
delete [Link];
delete [Link];
run;
Our editing of the multinomial logit parameter estimates table assumes that each independent variable has a
label. If you are coding with PROC TRANSREG, this will be true of all variables created by class expansions.
You may have to provide labels for identity and other variables. Alternatively, if you want variable names
to appear in the table, you can do that as follows. This may be useful when you are not coding with PROC
TRANSREG.
%phchoice(on, Variable DF Estimate StdErr ChiSq ProbChiSq Label)

The optional second argument provides a list of the column names to print. The available columns are: Vari-
able DF Estimate StdErr StdErrRatio ChiSq ProbChiSq HazardRatio HRLowerCL
HRUpperCL Label. (HRLowerCL and HRUpperCL are confidence limits on the hazard ratio.) For very
detailed customizations, you may have to run PROC TEMPLATE directly.
368 TS-677E Multinomial Logit, Discrete Choice Modeling

%PhChoice Macro Options

The %PhChoice macro has two positional parameters, onoff and column. Positional parameters must come
first, and unlike all other parameters, are not specified after a name and an equal sign.

onoff
ON specifies choice model customization.
OFF turns off the choice model customization and returns to the default PROC PHREG templates.
EXPB turns on choice model customization and adds the hazard ratio to the output.
Upper/lower case does not matter.

column
specifies an optional column list for more extensive customizations.
Concluding Remarks 369

Concluding Remarks
This report has illustrated how to design a choice experiment; prepare the questionnaire; input, process, and code
the design; perform the analysis; and interpret the results. All examples were artificial. We would welcome
any real data sets that we could use in future examples. This report has already been revised many times, and
future revisions are likely. If you have comments or suggestions for future revisions write Warren F. Kuhfeld,
([Link]@[Link]) at SAS Institute Inc. Please direct questions to the technical support division. For
more information on discrete choice, see Carson et. al. (1994) and the papers they reference. For information on
designing experiments for discrete choice, see Lazari and Anderson (1994), and see Kuhfeld, Tobias, and Garratt
(1994) on page 25.
I hope you like the new macros. In particular, I hope you find the new %MktEx macro to be very powerful and
useful. My goal in writing this book is to help you do better research and do it more quickly and more easily. I
would like to hear what you think.
For Those Who Like a Challenge
What do most of the random number seeds used in the Multinomial Logit, Discrete Choice Modeling re-
port and all of the seeds used in the Conjoint Analysis Examples report have in common? Send answers to
[Link]@[Link]. I will send a small prize to the first person to send me the answer that I have in mind.
Hints: Ignore seed 7654321; it has nothing in common with the others. Seeds 201 and 155 almost but not quite
fit with the others. Seeds 446, 538, and 543 are part of a still larger group. Answers like “they are all less than
619,” while true, are not what I have in mind.
370 TS-677E Multinomial Logit, Discrete Choice Modeling

References
Addelman, S. (1962), “Orthogonal Main-Effects Plans for Asymmetrical Factorial Experiments”, Technometrics,
4, 21 46.
Bose, R.C. (1947), “Mathematical Theory of the Symmetrical Factorial Design”, Sankhya, 8, 107 166.
Carson, R.T., Louviere, J.J, Anderson, D.A., Arabie, P., Bunch, D., Hensher, D.A., Johnson, R.M., Kuhfeld,
W.F., Steinberg, D., Swait, J., Timmermans, H., and Wiley, J.B. (1994), “Experimental Analysis of Choice,”
Marketing Letters, 5(4), 351 368.
Cook, R.D. and Nachtsheim, C.J. (1980), “A Comparison of Algorithms for Constructing Exact D-optimal De-
signs”, Technometrics, 22, 315 324.
Dey, A. (1985), Orthogonal Fractional Factorial Designs, New York: Wiley.
Fedorov, V.V. (1972), Theory of Optimal Experiments, translated and edited by W.J. Studden and E.M. Klimko,
New York: Academic Press.
Finney, D.J. (1982), “Some Enumerations for the 6x6 Latin Squares”, Utilitas Mathematics, 21, 137 153.
Hadamard, J. (1893), “Resolution d’une question relative aux determinants”,Bull. des Sciences Math, (2), 17,
240 246.
Hedayat, A.S., Sloane, N.J.A., and Stufken, J. (1999), Orthogonal Arrays, New York: Springer.
Huber, J., and Zwerina, K. (1996), “The Importance of Utility Balance in Efficient Choice Designs,” Journal of
Marketing Research, 33, 307 317.
Kuhfeld, W.F., Tobias, R.D., and Garratt, M. (1994), “Efficient Experimental Design with Marketing Research
Applications,” Journal of Marketing Research, 31, 545 557.
Lazari, A.G. and Anderson, D.A. (1994), “Designs of Discrete Choice Set Experiments for Estimating Both
Attribute and Availability Cross Effects,” Journal of Marketing Research, 31, 375 383.
Louviere, J.J. (1991), “Consumer Choice Models and the Design and Analysis of Choice Experiments,” Tutorial
presented to the American Marketing Association Advanced Research Techniques Forum, Beaver Creek,
Colorado.
Louviere, J.J. and Woodworth, G (1983), “Design and Analysis of Simulated Consumer Choice of Allocation Ex-
periments: A Method Based on Aggregate Data,” Journal of Marketing Research, 20 (November), 350 67.
Manski, C.F., and McFadden, D. (1981), Structural Analysis of Discrete Data with Econometric Applications.
Cambridge: MIT Press.
Meyer, R.K., and Nachtsheim, C.J. (1995), “The Coordinate-Exchange Algorithm for Constructing Exact Opti-
mal Experimental Designs”, Technometrics, 37, 60 69.
Paley, R.E.A.C (1933), On Orthogonal Matrices, J. Math. Phys, 12, 311 320.
Rao, C.R. (1947), “Factorial Experiments Derivable from Combinatorial Arrangements of Arrays”, Journal of
the Royal Statistical Society, Suppl., 9, 128 139.
Suen, C.-Y. (1989), “Some Resolvable Orthogonal Arrays with Two Symbols”, Communications in Statistics,
Theory and Methods, 18, 3875 3881.
Suen, C.-Y. (1989), “A Class of Orthogonal Main Effects Plans”, Journal of Statistical Planning and Inference,
21, 391 394.
Taguchi, G. (1987), System of Experimental Design: Engineering Methods to Optimize Quality and Minimize
Costs. White Plains, NY: UNIPIB, and Dearborn, MI: American Supplier Institute.
Wang, J.C., and Wu, C.F.J. (1991), “An Approach to the Construction of Asymmetrical Orthogonal Arrays”,
Journal of the American Statistical Association, 86, 450 456.
References 371

Wang, J.C., (1996), “Mixed Difference Matrices and the Construction of Orthogonal Arrays”, Statist. Probab.
Lett., 28, 121 126.
Wang, J.C., (1996), A Recursive Construction of Orthogonal Arrays, Preprint.
Williamson, J. (1944), “Hadamard’s Determinant Theorem and the Sum of Four Squares”, Duke Math. J., 11,
65 81.
Zhang, Y.S., Lu, Y., and Pang, S., (1999), “Orthogonal Arrays Obtained by Orthogonal Decompositions of
Projection Matrices”, Statistica Sinica, 9, 595 604.

SAS and SAS/STAT are registered trademarks or trademarks of SAS in the USA and other countries. indicates
USA registration.
372 TS-677E Multinomial Logit, Discrete Choice Modeling

Multinomial Logit Models

Ying So and Warren F. Kuhfeld
SAS, Cary, NC

ABSTRACT Multinomial logit models are used to model relationships between a polytomous response variable
and a set of regressor variables. The term “multinomial logit model” includes, in a broad sense, a variety of
models. The cumulative logit model is used when the response of an individual unit is restricted to one of a finite
number of ordinal values. Generalized logit and conditional logit models are used to model consumer choices.
This article focuses on the statistical techniques for analyzing discrete choice data and discusses fitting these
models using SAS/STAT R software.
Introduction Multinomial logit models are used to model relationships between a polytomous response variable
and a set of regressor variables. These polytomous response models can be classified into two distinct types,
depending on whether the response variable has an ordered or unordered structure.
In an ordered model, the response Y of an individual unit is restricted to one of m ordered values. For example,
the severity of a medical condition may be: none, mild, and severe. The cumulative logit model assumes that the
ordinal nature of the observed response is due to methodological limitations in collecting the data that results in
lumping together values of an otherwise continuous response variable (McKelvey and Zavoina 1975). Suppose
Y takes values y1 ; y2 ; : : :; ym on some scale, where y1 < y2 < : : : < ym . It is assumed that the observable
variable is a categorized version of a continuous latent variable U such that

Y =y i , i 1 <U  ; i = 1; : : :; m
i

where 1 = 0 < 1 < : : : < m = 1. It is further assumed that the latent variable U is determined by the
explanatory variable vector x in the linear form U = 0 x + ; where is a vector of regression coefficients
and  is a random variable with a distribution function F . It follows that

PrfY  y jxg = F ( + 0 x)
i i

If F is the logistic distribution function, the cumulative model is also known as the proportional odds model.
You can use PROC LOGISTIC or PROC PROBIT directly to fit the cumulative logit models. Although the
cumulative model is the most widely used model for ordinal response data, other useful models include the
adjacent-categories logit model and the continuation-ratio model (Agresti 1990).
In an unordered model, the polytomous response variable does not have an ordered structure. Two classes of
models, the generalized logit models and the conditional logit models, can be used with nominal response data.
The generalized logit model consists of a combination of several binary logits estimated simultaneously. For
example, the response variable of interest is the occurrence or nonoccurrence of infection after a Caesarean
section with two types of (I,II) infection. Two binary logits are considered: one for type I infection versus no
infection and the other for type II infection versus no infection. The conditional logit model has been used in
biomedical research to estimate relative risks in matched case-control studies. The nuisance parameters that
correspond to the matched sets in an unconditional analysis are eliminated by using a conditional likelihood
that contains only the relative risk parameters (Breslow and Days 1980). The conditional logit model was also
introduced by McFadden (1973) in the context of econometrics.
In studying consumer behavior, an individual is presented with a set of alternatives and asked to choose the most
preferred alternative. Both the generalized logit and conditional logit models are used in the analysis of discrete
choice data. In a conditional logit model, a choice among alternatives is treated as a function of the characteristics
of the alternatives, whereas in a generalized logit model, the choice is a function of the characteristics of the
individual making the choice. In many situations, a mixed model that includes both the characteristics of the
alternatives and the individual is needed for investigating consumer choice.

 This paper was presented at SUGI 20 by Ying So and can also be found in the SUGI 20 proceedings.
Multinomial Logit Models 373

Consider an example of travel demand. People are asked to choose between travel by auto, plane or public
transit (bus or train). The following SAS R statements create the data set TRAVEL. The variables AUTOTIME,
PLANTIME, and TRANTIME represent the total travel time required to get to a destination by using auto, plane,
or transit, respectively. The variable AGE represents the age of the individual being surveyed, and the variable
CHOSEN contains the individual’s choice of travel mode.
data travel;
input AutoTime PlanTime TranTime Age Chosen $;
datalines;
10.0 4.5 10.5 32 Plane
5.5 4.0 7.5 13 Auto
4.5 6.0 5.5 41 Transit
3.5 2.0 5.0 41 Transit
1.5 4.5 4.0 47 Auto
10.5 3.0 10.5 24 Plane
7.0 3.0 9.0 27 Auto
9.0 3.5 9.0 21 Plane
4.0 5.0 5.5 23 Auto
22.0 4.5 22.5 30 Plane
7.5 5.5 10.0 58 Plane
11.5 3.5 11.5 36 Transit
3.5 4.5 4.5 43 Auto
12.0 3.0 11.0 33 Plane
18.0 5.5 20.0 30 Plane
23.0 5.5 21.5 28 Plane
4.0 3.0 4.5 44 Plane
5.0 2.5 7.0 37 Transit
3.5 2.0 7.0 45 Auto
12.5 3.5 15.5 35 Plane
1.5 4.0 2.0 22 Auto
;
In this example, AUTOTIME, PLANTIME, and TRANTIME are alternative-specific variables, whereas AGE is
a characteristic of the individual. You use a generalized logit model to investigate the relationship between the
choice of transportation and AGE, and you use a conditional logit model to investigate how travel time affects
the choice. To study how the choice depends on both the travel time and age of the individual, you need to use a
mixed model that incorporates both types of variables.
A survey of the literature reveals a confusion in the terminology for the nominal response models. The term
“multinomial logit model” is often used to describe the generalized logit model. The mixed logit is sometimes
referred to as the multinomial logit model in which the generalized logit and the conditional logit models are
special cases.
The following sections describe discrete choice models, illustrate how to use SAS/STAT software to fit these
models, and discuss cross-alternative effects.
Modeling Discrete Choice Data Consider an individual choosing among m alternatives in a choice set. Let jk
denote the probability that individual j chooses alternative k , let Xj represent the characteristics of individual j ,
and let Zjk be the characteristics of the k th alternative for individual j . For example, Xj may be an age and each
Zjk a travel time.

The generalized logit model focuses on the individual as the unit of analysis and uses individual characteristics
as explanatory variables. The explanatory variables, being characteristics of an individual, are constant over the
alternatives. For example, for each of the m travel modes, Xj = (1 age)0 , and for the first subject, X1 = (1 32)0 .
The probability that individual j chooses alternative k is
0
 = P exp(exp(X0 X
) =P
k j 1
)0 X )]
) =1 exp[(
jk m m
=1 l l j l l k j
374 TS-677E Multinomial Logit, Discrete Choice Modeling

1 ; : : :; m are m vectors of unknown

P regression parameters (each of which is different, even though Xj is
constant across alternatives). Since m k=1
jk = 1, the m sets of parameters are not unique. By setting the last
set of coefficients to null (that is, m = 0), the coefficients k represent the effects of the X variables on the
probability of choosing the k th alternative over the last alternative. In fitting such a model, you estimate m 1
sets of regression coefficients.
In the conditional logit model, the explanatory variables Z assume different values for each alternative and the
impact of a unit of Z is assumed to be constant across alternatives. For example, for each of the m travel modes,
Zjk = (time)0 , and for the first subject, Z11 = (10)0 , Z12 = (4:5)0 , and Z13 = (10:5)0 . The probability that the
individual j chooses alternative k is

0 Z )
 = P exp(exp(
jk

m
1
0 Z ) = P exp[0 (Z
jk
m
Zjk )]
=1 l =1 jl l jl

 is a single vector of regression coefficients. The impact of a variable on the choice probabilities derives from
the difference of its values across the alternatives.
For the mixed logit model that includes both characteristics of the individual and the alternatives, the choice
probabilities are
0 0
 = Pexp(exp(X 0 X+ +Z0 Z) )
jk m
k j jk

=1 l l j jl

1 ; : : :; m 1 and m
 0 are the alternative-specific coefficients, and  is the set of global coefficients.
Fitting Discrete Choice Models The CATMOD procedure in SAS/STAT software directly fits the generalized
logit model. SAS/STAT software does not yet have a procedure that is specially designed to fit the conditional or
mixed logit models. However, with some preliminary data processing, you can use the PHREG procedure to fit
these models.
The PHREG procedure fits the Cox proportional hazards model to survival data (refer to SAS Technical Report
P-229). The partial likelihood of Breslow has the same form as the likelihood in a conditional logit model.
Let zl denote the vector of explanatory variables for individual l. Let t1 < t2 < : : : < tk denote k distinct
ordered event times. Let di denote the number of failures at ti . Let si be the sum of the vectors zl for those
individuals that fail at ti , and let Ri denote the set of indices for those who are at risk just before ti .
The Breslow (partial) likelihood is

L () =
Y k
exp(0 s )
P i
B

=1
i
[ 2Ri exp(0 z )] i
l l
d

In a stratified analysis, the partial likelihood is the product of the partial likelihood for each individual stratum.
For example, in a study of the time to first infection from a surgery, the variables of a patient consist of TIME
(time from surgery to the first infection), STATUS (an indicator of whether the observation time is censored,
with value 2 identifying a censored time), Z1 and Z2 (explanatory variables thought to be related to the time to
infection), and GRP (a variable identifying the stratum to which the observation belongs). The specification in
PROC PHREG for fitting the Cox model using the Breslow likelihood is as follows:
proc phreg;
model time*status(2) = z1 z2 / ties=breslow;
strata grp;
run;
Multinomial Logit Models 375

To cast the likelihood of the conditional logit model in the form of the Breslow likelihood, consider m artificial
observed times for each individual who chooses one of m alternatives. The k th alternative is chosen at time 1; the
choices of all other alternatives (second choice, third choice, ...) are not observed and would have been chosen
at some later time. So a choice variable is coded with an observed time value of 1 for the chosen alternative and
a larger value, 2, for all unchosen (unobserved or censored alternatives). For each individual, there is exactly
one event time (1) and m 1 nonevent times, and the risk set just prior to this event time consists of all the m
alternatives. For individual j with alternative-specific characteristics Zjl , the Breslow likelihood is then

L () = P
exp(0 Z ) jk
0
=1 exp( Z )
B m
l jl

This is precisely the probability that individual j chooses alternative k in a conditional logit model. By stratifying
on individuals, you get the likelihood of the conditional logit model. Note that the observed time values of 1 and
2 are chosen for convenience; however, the censored times have to be larger than the event time to form the
correct risk set.
Before you invoke PROC PHREG to fit the conditional logit, you must arrange your data in such a way that there
is a survival time for each individual-alternative. In the example of travel demand, let SUBJECT identify the
individuals, let TRAVTIME represent the travel time for each mode of transportation, and let CHOICE have a
value 1 if the alternative is chosen and 2 otherwise. The CHOICE variable is used as the artificial time variable
as well as a censoring variable in PROC PHREG. The following SAS statements reshape the data set TRAVEL
into data set CHOICE and display the first nine observations:
data choice(keep=subject mode travtime choice);
array times[3] autotime plantime trantime;
array allmodes[3] $ _temporary_ (’Auto’ ’Plane’ ’Transit’);
set travel;
Subject = _n_;
do i = 1 to 3;
Mode = allmodes[i];
TravTime = times[i];
Choice = 2 - (chosen eq mode);
output;
end;
run;

proc print data=choice(obs=9);

run;

Trav
Obs Subject Mode Time Choice

1 1 Auto 10.0 2
2 1 Plane 4.5 1
3 1 Transit 10.5 2
4 2 Auto 5.5 1
5 2 Plane 4.0 2
6 2 Transit 7.5 2
7 3 Auto 4.5 2
8 3 Plane 6.0 2
9 3 Transit 5.5 1

Notice that each observation in TRAVEL corresponds to a block of three observations in CHOICE, exactly one
of which is chosen.
376 TS-677E Multinomial Logit, Discrete Choice Modeling

The following SAS statements invoke PROC PHREG to fit the conditional logit model. The Breslow likelihood
is requested by specifying TIES=BRESLOW. CHOICE is the artificial time variable, and a value of 2 identifies
censored times. SUBJECT is used as a stratification variable.
proc phreg data=choice;
model choice*choice(2) = travtime / ties=breslow;
strata subject;
title ’Conditional Logit Model Using PHREG’;
run;

Conditional Logit Model Using PHREG

The PHREG Procedure

Analysis of Maximum Likelihood Estimates

Parameter Standard Hazard

Variable DF Estimate Error Chi-Square Pr > ChiSq Ratio

TravTime 1 -0.26549 0.10215 6.7551 0.0093 0.767

To study the relationship between the choice of transportation and the age of people making the choice, the
analysis is based on the generalized logit model. You can use PROC CATMOD directly to fit the generalized
logit model (refer to SAS/STAT User’s Guide, Vol. 1). In the following invocation of PROC CATMOD, CHOSEN
is the response variable and AGE is the explanatory variable:
proc catmod data=travel;
direct age;
model chosen=age;
title ’Multinomial Logit Model Using Catmod’;
run;

Response Profiles

Response Chosen
-------------------
1 Auto
2 Plane
3 Transit

Analysis of Maximum Likelihood Estimates

Function Standard Chi-

Parameter Number Estimate Error Square Pr > ChiSq
-------------------------------------------------------------------
Intercept 1 3.0449 2.4268 1.57 0.2096
2 2.7212 2.2929 1.41 0.2353
Age 1 -0.0710 0.0652 1.19 0.2762
2 -0.0500 0.0596 0.70 0.4013

Note that there are two intercept coefficients and two slope coefficients for AGE. The first INTERCEPT and the
first AGE coefficients correspond to the effect on the probability of choosing auto over transit, and the second
intercept and second age coefficients correspond to the effect of choosing plane over transit.
Multinomial Logit Models 377

Let Xj be a (p +1)-vector representing the characteristics of individual j . The generalized logit model can be cast
in the framework of a conditional model by defining the global parameter vector  and the alternative-specific
regressor variables Zjk as follows:

2 3
2 3 2
X
3 0 2
0
3 2 3
1 j 6 X 7 0
6 0 7
6 7 6 6 j 7 6
6 ... 7
 = 66
2
..
7
7 Zj
7
1 = 6 .. 7 Zj
6
2 =6
0 77 ::: Zj;m 1 =6
7
7 Zjm = 4 ... 75
6
4 5 4 . 5 6 .. 7 4 0 5
.
0
4 . 5 0
m 1 0 Xj

where the 0 is a (p +1)-vector of zeros. The probability that individual j chooses alternative k for the generalized
logit model is put in the form that corresponds to a conditional logit model as follows:

0
 jk = P exp(exp(X0 X) )
m
k j

=1 l l j

exp( 0 Z )
= P exp( m
0 Z )
jk

l=1 jl

Here, the vector Xj representing the characteristics of individual j includes the element 1 for the intercept
parameter (provided that the intercept parameters are to be included in the model).
By casting the generalized logit model into a conditional logit model, you can then use PROC PHREG to analyze
the generalized logit model. In the example of travel demand, the alternative-specific variables AUTO, PLANE,
AGEAUTO, and AGEPLANE are created from the individual characteristic variable AGE. The following SAS
statements reshape the data set TRAVEL into data set CHOICE2 and display the first nine observations:
data choice2;
array times[3] autotime plantime trantime;
array allmodes[3] $ _temporary_ (’Auto’ ’Plane’ ’Transit’);
set travel;
Subject = _n_;
do i = 1 to 3;
Mode = allmodes[i];
TravTime = times[i];
Choice = 2 - (chosen eq mode);
Auto = (i eq 1);
Plane = (i eq 2);
AgeAuto = auto * age;
AgePlane = plane * age;
output;
end;
keep subject mode travtime choice auto plane ageauto ageplane;
run;
proc print data=choice2(obs=9);
run;
378 TS-677E Multinomial Logit, Discrete Choice Modeling

Trav Age Age

Obs Subject Mode Time Choice Auto Plane Auto Plane

1 1 Auto 10.0 2 1 0 32 0
2 1 Plane 4.5 1 0 1 0 32
3 1 Transit 10.5 2 0 0 0 0
4 2 Auto 5.5 1 1 0 13 0
5 2 Plane 4.0 2 0 1 0 13
6 2 Transit 7.5 2 0 0 0 0
7 3 Auto 4.5 2 1 0 41 0
8 3 Plane 6.0 2 0 1 0 41
9 3 Transit 5.5 1 0 0 0 0

The following SAS statements invoke PROC PHREG to fit the generalized logit model:
proc phreg data=choice2;
model choice*choice(2) = auto plane ageauto ageplane /
ties=breslow;
strata subject;
title ’Generalized Logit Model Using PHREG’;
run;

Generalized Logit Model Using PHREG

The PHREG Procedure

Analysis of Maximum Likelihood Estimates

Parameter Standard Hazard

Variable DF Estimate Error Chi-Square Pr > ChiSq Ratio

Auto 1 3.04494 2.42682 1.5743 0.2096 21.009

Plane 1 2.72121 2.29289 1.4085 0.2353 15.199
AgeAuto 1 -0.07097 0.06517 1.1859 0.2762 0.931
AgePlane 1 -0.05000 0.05958 0.7045 0.4013 0.951

By transforming individual characteristics into alternative-specific variables, the mixed logit model can be ana-
lyzed as a conditional logit model.
Analyzing the travel demand data for the effects of both travel time and age of individual requires the same data
set as the generalized logit model, only now the TRAVTIME variable will be used as well. The following SAS
statements use PROC PHREG to fit the mixed logit model:
proc phreg data=choice2;
model choice*choice(2) = auto plane ageauto ageplane travtime /
ties=breslow;
strata subject;
title ’Mixed Logit Model Using PHREG’;
run;
Multinomial Logit Models 379

Mixed Logit Model Using PHREG

The PHREG Procedure

Analysis of Maximum Likelihood Estimates

Parameter Standard Hazard

Variable DF Estimate Error Chi-Square Pr > ChiSq Ratio

Auto 1 2.50069 2.39585 1.0894 0.2966 12.191

Plane 1 -2.77912 3.52929 0.6201 0.4310 0.062
AgeAuto 1 -0.07826 0.06332 1.5274 0.2165 0.925
AgePlane 1 0.01695 0.07439 0.0519 0.8198 1.017
TravTime 1 -0.60845 0.27126 5.0315 0.0249 0.544

A special case of the mixed logit model is the conditional logit model with alternative-specific constants. Each
alternative in the model can be represented by its own intercept, which captures the unmeasured desirability of
the alternative.
proc phreg data=choice2;
model choice*choice(2) = auto plane travtime / ties=breslow;
strata subject;
title ’Conditional Logit Model with Alternative Specific Constants’;
run;

Conditional Logit Model with Alternative Specific Constants

The PHREG Procedure

Analysis of Maximum Likelihood Estimates

Parameter Standard Hazard

Variable DF Estimate Error Chi-Square Pr > ChiSq Ratio

Auto 1 -0.11966 0.70820 0.0285 0.8658 0.887

Plane 1 -1.63145 1.24251 1.7241 0.1892 0.196
TravTime 1 -0.48665 0.20725 5.5139 0.0189 0.615

With transit as the reference mode, the intercept for auto, which is negative, may reflect the inconvenience of
having to drive over traveling by bus/train, and the intercept for plane may reflect the high expense of traveling
by plane over bus/train.
Cross-Alternative Effects Discrete choice models are often derived from the principle of maximum random
utility. It is assumed that an unobserved utility Vk is associated with the k th alternative, and the response function
Y is determined by

Y = k , V = maxfV ; 1  l  mg
k l

Both the generalized logit and the conditional logit models are based on the assumption that V1 ; : : :; Vm are
independently distributed and each follows an extreme maxima value distribution (Hoffman and Duncan, 1988).
An important property of such models is Independence from Irrelevant Alternatives (IIA); that is, the ratio of
the choice probabilities for any two alternatives for a particular observation is not influenced systematically by
any other alternatives. IIA can be tested by fitting a model that contains all the cross-alternative effects and
examining the significance of these effects. The cross-alternative effects pick up a variety of IIA violations and
other sources of error in the model. (See pages 179, 185, 192, and 383 for other discussions of IIA.)
380 TS-677E Multinomial Logit, Discrete Choice Modeling

In the example of travel demand, there may be separate effects for the three travel modes and travel times. In
addition, there may be cross-alternative effects for travel times. Not all the effects are estimable, only two of
the three intercepts and three of the six cross-alternative effects can be estimated. The following SAS statements
create the design variables for all the cross-alternative effects and display the first nine observations:
* Number of alternatives in each choice set;
%let m = 3;
data choice3;
drop i j k autotime plantime trantime;

* Values of the variable CHOSEN;

array allmodes[&m] $
_temporary_ (’Auto’ ’Plane’ ’Transit’);
* Travel times for the alternatives;
array times[&m] autotime plantime trantime;
* New variables that will contain the design:;
array inters[&m]
Auto /*intercept for auto */
Plane /*intercept for plane */
Transit; /*intercept for transit */
array cross[%eval(&m * &m)]
TimeAuto /*time of auto alternative */
PlanAuto /*cross effect of plane on auto */
TranAuto /*cross effect of transit on auto */
AutoPlan /*cross effect of auto on plane */
TimePlan /*time of plane alternative */
TranPlan /*cross effect of transit on plane*/
AutoTran /*cross effect of auto on transit */
PlanTran /*cross effect of plane on transit*/
TimeTran; /*time of transit alternative */
set travel;
subject = _n_;
* Create &m observations for each choice set;
do i = 1 to &m;
Mode = allmodes[i]; /* this alternative */
Travtime = times[i]; /* travel time */
Choice = 2 - (chosen eq mode);/* 1 - chosen */
do j = 1 to &m;
inters[j] = (i eq j); /* mode indicator */
do k = 1 to &m;
* (j=k) - time, otherwise, cross effect;
cross[&m*(j-1)+k]=times[k]*inters[j];
end;
end;
output;
end;
run;
proc print data=choice3(obs=9) label noobs;
var subject mode travtime choice auto plane transit
timeauto timeplan timetran autoplan autotran planauto
plantran tranauto tranplan;
run;
Multinomial Logit Models 381

subject Mode Travtime Choice Auto Plane Transit

1 Auto 10.0 2 1 0 0
1 Plane 4.5 1 0 1 0
1 Transit 10.5 2 0 0 1
2 Auto 5.5 1 1 0 0
2 Plane 4.0 2 0 1 0
2 Transit 7.5 2 0 0 1
3 Auto 4.5 2 1 0 0
3 Plane 6.0 2 0 1 0
3 Transit 5.5 1 0 0 1

Time Time Time Auto Auto Plan Plan Tran Tran

Auto Plan Tran Plan Tran Auto Tran Auto Plan

10.0 0.0 0.0 0.0 0.0 4.5 0.0 10.5 0.0

0.0 4.5 0.0 10.0 0.0 0.0 0.0 0.0 10.5
0.0 0.0 10.5 0.0 10.0 0.0 4.5 0.0 0.0
5.5 0.0 0.0 0.0 0.0 4.0 0.0 7.5 0.0
0.0 4.0 0.0 5.5 0.0 0.0 0.0 0.0 7.5
0.0 0.0 7.5 0.0 5.5 0.0 4.0 0.0 0.0
4.5 0.0 0.0 0.0 0.0 6.0 0.0 5.5 0.0
0.0 6.0 0.0 4.5 0.0 0.0 0.0 0.0 5.5
0.0 0.0 5.5 0.0 4.5 0.0 6.0 0.0 0.0

PROC PHREG allows you to specify TEST statements for testing linear hypotheses of the parameters. The
test is a Wald test, which is based on the asymptotic normality of the parameter estimators. The following
SAS statements invoke PROC PHREG to fit the so called “Mother Logit” model that includes all the cross-
alternative effects. The TEST statement, with label IIA, specifies the null hypothesis that cross-alternative effects
AUTOPLAN, PLANTRAN, and TRANAUTO are 0. Since only three cross-alternative effects are estimable and
these are the first cross-alternative effects specified in the model, they account for all the cross-alternative effects
in the model.
proc phreg data=choice3;
model choice*choice(2) = auto plane transit timeauto timeplan
timetran autoplan plantran tranauto planauto tranplan
autotran / ties=breslow;
IIA: test autoplan,plantran,tranauto;
strata subject;
title ’Mother Logit Model’;
run;

Mother Logit Model

The PHREG Procedure

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 46.142 24.781

AIC 46.142 40.781
SBC 46.142 49.137
382 TS-677E Multinomial Logit, Discrete Choice Modeling

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 21.3607 8 0.0062

Score 15.4059 8 0.0517
Wald 6.2404 8 0.6203

Analysis of Maximum Likelihood Estimates

Parameter Standard Hazard

Variable DF Estimate Error Chi-Square Pr > ChiSq Ratio

Auto 1 -0.73812 3.05933 0.0582 0.8093 0.478

Plane 1 -3.62435 3.48049 1.0844 0.2977 0.027
Transit 0 0 . . . .
TimeAuto 1 -2.23433 1.89921 1.3840 0.2394 0.107
TimePlan 1 -0.10112 0.68621 0.0217 0.8829 0.904
TimeTran 1 0.09785 0.70096 0.0195 0.8890 1.103
AutoPlan 1 0.44495 0.68616 0.4205 0.5167 1.560
PlanTran 1 -0.53234 0.63481 0.7032 0.4017 0.587
TranAuto 1 1.66295 1.51193 1.2097 0.2714 5.275
PlanAuto 0 0 . . . .
TranPlan 0 0 . . . .
AutoTran 0 0 . . . .

Linear Hypotheses Testing Results

Wald
Label Chi-Square DF Pr > ChiSq

IIA 1.6526 3 0.6475

The 2 statistic for the Wald test is 1:6526 with 3 degrees of freedom, indicating that the cross-alternative effects
are not statistically significant (p = :6475). A generally more preferable way of testing the significance of
the cross-alternative effects is to compare the likelihood of the “Mother logit” model with the likelihood of the
reduced model with the cross- alternative effects removed. The following SAS statements invoke PROC PHREG
to fit the reduced model:
proc phreg data=choice3;
model choice*choice(2) = auto plane transit timeauto
timeplan timetran / ties=breslow;
strata subject;
title ’Reduced Model without Cross-Alternative Effects’;
run;

Reduced Model without Cross-Alternative Effects

The PHREG Procedure

Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.

Multinomial Logit Models 383

Model Fit Statistics

Without With
Criterion Covariates Covariates

-2 LOG L 46.142 27.153

AIC 46.142 37.153
SBC 46.142 42.376

Reduced Model without Cross-Alternative Effects

The PHREG Procedure

Testing Global Null Hypothesis: BETA=0

Test Chi-Square DF Pr > ChiSq

Likelihood Ratio 18.9886 5 0.0019

Score 14.4603 5 0.0129
Wald 7.3422 5 0.1964

Analysis of Maximum Likelihood Estimates

Parameter Standard Hazard

Variable DF Estimate Error Chi-Square Pr > ChiSq Ratio

Auto 1 1.71578 1.80467 0.9039 0.3417 5.561

Plane 1 -3.60073 3.30555 1.1866 0.2760 0.027
Transit 0 0 . . . .
TimeAuto 1 -0.79543 0.36327 4.7946 0.0285 0.451
TimePlan 1 0.12162 0.58954 0.0426 0.8366 1.129
TimeTran 1 -0.42184 0.25733 2.6873 0.1012 0.656

The chi-squared statistic for the likelihood ratio test of IIA is (27:153 24:781) = 2:372, which is not statistically
significant (p = :4989) when compared to a 2 distribution with 3 degrees of freedom. This is consistent with
the previous result of the Wald test. (See pages 179, 185, 192, and 379 for other discussions of IIA.)
Final Comments For some discrete choice problems, the number of available alternatives is not the same for
each individual. For example, in a study of consumer brand choices of laundry detergents as prices change, data
are pooled from different locations, not all of which offer a brand that contains potash. The varying choice sets
across individuals can easily be accommodated in PROC PHREG. For individual j who chooses from a set of
mj alternatives, consider mj artificial times in which the chosen alternative has an event time 1 and the unchosen
alternatives have a censored time of 2. The analysis is carried out in the same fashion as illustrated in the previous
section.
Unlike the example of travel demand in which data for each individual are provided, choice data are often given
in aggregate form, with choice frequencies indicating the repetition of each choice. One way of dealing with
aggregate data is to expand the data to the individual level and carry out the analysis as if you have nonaggregate
data. This approach is generally not recommended, because it defeats the purpose of having a smaller aggregate
data set. PROC PHREG provides a FREQ statement that allows you to specify a variable that identifies the
frequency of occurrence of each observation. However, with the specification of a FREQ variable, the artificial
event time is no longer the only event time in a given stratum, but has ties of the given frequency. With proper
stratification, the Breslow likelihood is proportional to the likelihood of the conditional logit model. Thus PROC
PHREG can be used to obtain parameter estimates and hypothesis testing results for the choice models.
384 TS-677E Multinomial Logit, Discrete Choice Modeling

The TIES=DISCRETE option should not be used instead of the TIES=BRESLOW option. This is especially
detrimental with aggregate choice data because the likelihood that PROC PHREG is maximizing may no longer
be the same as the likelihood of the conditional logit model. TIES=DISCRETE corresponds to the discrete
logistic model for genuinely discrete time scale, which is also suitable for the analysis of case-control studies
when there is more than one case in a matched set (Gail, Lubin, and Rubinstein, 1981). For nonaggregate choice
data, all TIES= options give the same results; however, the resources required for the computation are not the
same, with TIES=BRESLOW being the most efficient.
Once you have a basic understanding of how PROC PHREG works, you can use it to fit a variety of models
for the discrete choice data. The major involvement in such a task lies in reorganizing the data to create the
observations necessary to form the correct risk sets and the appropriate design variables. There are many options
in PROC PHREG that can also be useful in the analysis of discrete choice data. For example, the OFFSET=
option allows you to restrict the coefficient of an explanatory variable to the value of 1; the SELECTION= option
allows you to specify one of four methods for selecting variables into the model; the OUTEST= option allows
you to specify the name of the SAS data set that contains the parameter estimates, based on which you can easily
compute the predicted probabilities of the alternatives.
This article deals with estimating parameters of discrete choice models. There is active research in the field of
marketing research to use design of experiments to study consumer choice behavior. If you are interested in this
area, refer to Carson et al. (1994), Kuhfeld et al. (1994), and Lazari et al. (1994).
Multinomial Logit Models 385

References

Agresti, A. (1990) Categorical Data Analysis. New York: John Wiley & Sons.
Breslow, N. and Day, N.E. (1980), Statistical Methods in Cancer Research, Vol. II: The Design and Analysis of
Cohort Studies, Lyon: IARC.
Carson, R.T.; Louviere, J.J; Anderson, D.A.; Arabie, P.; Bunch, D.; Hensher, D.A.; Johnson, R.M.; Kuhfeld,
W.F.; Steinberg, D.; Swait, J.; Timmermans, H.; and Wiley, J.B. (1994), “Experimental Analysis of Choice,”
Marketing Letters, 5(4), 351-368.
Gail, M.H., Lubin, J.H., and Rubinstein, L.V. (1981), “Likelihood calculations for matched case-control studies
and survival studies with tied death times,” Biometrika, 68, 703-707.
Hoffman, S.D. and Duncan, G.J. (1988), “Multinomial and conditional logit discrete-choice models in Demog-
raphy,” Demography, 25 (3), 415-427.
Kuhfeld, W.F., Tobias, R.D., and Garratt, M. (1994), “Efficient Experimental Design with Marketing Research
Applications,” Journal of Marketing Research, 31, 545-557.
Lazari, A.G. and Anderson, D.A. (1994), “Designs of Discrete Choice Set Experiments for Estimating Both
Attribute and Availability Cross Effects,” Journal of Marketing Research, 31, 375-383.
McFadden, D. (1973), “Conditional logit analysis of qualitative choice behavior,” in P. Zarembka (Ed.) Frontiers
in Econometrics, New York: Academic Press, Inc.
McKelvey, R.D. and Zavoina, W. (1975), “A statistical model for the analysis of ordinal level dependent vari-
ables,” Journal of Mathematical Sociology, 4, 103-120.
SAS Institute Inc. (1989), SAS/STAT User’s Guide, Vol. 1, Version 6, Fourth Edition, Cary: NC: SAS Institute
Inc.
SAS Institute Inc. (1992), SAS Technical Report P-229. SAS/STAT Software: Changes and Enhancements,
Release 6.07, Cary, NC: SAS Institute Inc.

SAS and SAS/STAT are trademarks or registered trademarks of SAS in the USA and other countries. indicates
USA registration.
 386 TS-677E Multinomial Logit, Discrete Choice Modeling

Index

@@ 85 block= 312
accept defined 337 BLOCKED data set 313
accept option 281 285-286 333 336-339 blocking 209 238
A-efficiency 77 blocks 162 209
Age variable 230-231 blocks= defined 325 353
aggregate data 171-174 185-188 225-226 243 247- blocks= 135 353
249 383 blue bus 192
aliased 76 brand choice (aggregate data) example 173
aliasing structure 207 Brand variable 102-106 175 179-180 215 218-222
allcode defined 317 231 242-246 298 303 356-359
allocation study 237-247 branded defined 324
ALLOCS data set 305 branded 324
Alt variable 110 310-312 Breslow likelihood 191
- Alt- variable 359 brief 107-109 138 236
alt= defined 312 359 bundles of attributes 252 280
alt= 102-103 312 359 bus 192
alternative-specific effects 146-148 169 177 180 184 c = 2 - (i eq choice) 87
194 218 221 293 373-379 c variable 85-88 105 109 113 135 179-180 225 246
Ann 160 304 353-354
anneal= defined 340 c*c(2) 88 107
anneal= 281 340 c*c(3) 88
annealfun= defined 344 Can 160 198
annealing 124 cand= defined 315
anniter= defined 340 cand= 317
anniter= 340 candidate set 123 194 198 254 258 262-264 280 285-
arrays 92 100 112-114 131 164-166 173 185 223 242 288 292-293
264 375-377 380 canditer= defined 340
artificial data 75 166 223 canditer= 340-341
asymmetry 153 192 candy example 83
augmenting an existing design 269 canonical correlation 98
autocall macros 287 CB data set 326
availability cross effects 192-195 206 226 chair (generic attributes) example 252
available, not 347 check defined 317 337
bad variable 195 338 check the data entry 89
balance 77 95 163 214 check option 126 339
balance= defined 336 Chi-Square statistic 90
balance= 204-206 336-337 choice design 78-79 95 102-104
balanced and orthogonal 77 80-82 95 choice design efficiency 77 217-223 253-262 266 283
bestcov= defined 300 288 291-299
bestout= defined 300 choice design generation 255 258-262 266 289 292
bestout= 301 295-298 320-321
beta= defined 300 choice model, coding 105 136 140-144 148 169 175-
beta= 217 255 289 296 300 177 180 185-188 226 230 246
big designs 153 choice model, fitting 107 138-142 145 149 170-172
big= defined 314 343 175-177 183 187-188 226 234 247-249 374-
big= 314 317 343 382
bin of attributes 78 102 131 choice probabilities 92
binary coding 105-106 136 143 169 175 choice sets, minimum number 252
blank header 128 Choice variable 87
Block variable 133 239 247 307 312 349 choice-based conjoint 74
block= defined 312 %ChoicEff macro 74-75 215-220 226 252-267 275-
 Index 387

277 280 283 287-299 307 310-312 320-321 design 106 136 169 175
351 387-389 design, differences 303 307 313 318 339
%ChoicEff macro documentation 288-303 design, methods compared 267
%ChoicEff macro versus the %MktEx macro 267 design
%ChoicEff macro, alternative swapping 261 267 evaluation 97 124 129 154 160 200-206 239 305
%ChoicEff macro, set swapping 264 267 generation 95-96 110 119 126-128 154-157 162
Choose variable 115 195 199 202-204 238 252-254 261 264 268-
choose 343 269 273 289 292 296-297 307 310 320-323
chosen alternative 87 327 332 345 348-352 357-358
class statement 206 315 318 saturated 82
class 105-106 136 139 143-144 148 169 175-177 size 94 117 153 156 202 237 253 336 360-363
180 206 216-218 225-226 244 252 255 296 testing 215
367 design= defined 354 359
classopts= defined 315 design= 103-105 133 353 356 359
Client variable 347 Dest variable 131
coded defined 302 detail defined 302
coding down 157 315 342 detfuzz= defined 344
coding the choice model 105 136 140-144 148 169 different designs 303 307 313 318 339
175-177 180 185-188 226 230 246 diminishing returns on iterations 198
coding dolist= defined 351
binary 105-106 136 169 175 dolist= 351-353
effects 143 dollar format 239
coding= defined 315 D-optimality 327
Color variable 357-358 drop= defined 301
column defined 368 drop= 295 301
column statement 366 dropping variables 106 136
confounded 76 duplicate runs 333
constant alternative 94 110 173 edit statement 366
converge= defined 301 effects coding 143
converge= 293-295 effects 143 296
coordinate-exchange algorithm 123 efficiency 77
CORR data set 326 efficiency of a choice design 77
Count variable 242-243 303 Efficiency variable 257
cov= defined 301 eigenvalues 77
cross effects 175 179-185 192-195 206 218 222 225- errors in running macros 288
226 229 233-234 %EvalEff macro 206
customizing PHREG output 79 364-367 examine= defined 315 336
customizing the multinomial logit output 79 examine= 98 126 315 336
cyclic design 291 examining the design 97 124 129 154 160 200-206
data entry 85-87 102 114 133 167 173 185 224 229 239 305
242 373 example
data entry, checking 89 brand choice (aggregate data) 173
data processing 115 144 147 215-216 242 245 249 candy 83
264 375-377 380 chair (generic attributes) 252
data, generating artificial 166 223 fabric softener 94
data= defined 301 304 312 325 351-353 food product (availability) 192
data= 88 104-105 109 217 239 245 255 289 301 304 prescription drugs (allocation) 237
312 324 347-348 351-354 vacation 116
D-efficiency 77 198-199 207 vacation (alternative-specific) 152
D-efficiency, 0 to 100 scale 77 80-82 206 exchange= defined 343
degree= 142 exchange= 338 341-343
demographic information 229 existing design, improving 268
DESIGN data set 306 317 332 339 experimental design
design key 102 defined 76
Design variable 118 257 evaluation 97 124 129 154 160 200-206 239 305
 388 TS-677E Multinomial Logit, Discrete Choice Modeling

generation 95-96 110 119 126-128 154-157 162 Hadamard matrices 331-332 337 345-346
195 199 202-204 238 252-254 261 264 268- header, blank or null 128
269 273 289 292 296-297 307 310 320-323 holdouts= defined 343
327 332 345 348-352 357-358 holdouts= 269 273
saturated 82 host differences 303 307 313 318 339
size 94 117 153 156 202 237 253 336 360-363 HRLowerCL 367
testing 215 HRUpperCL 367
external attributes 229 (i eq choice) 87
extreme value type I distribution 192 i variable 338
f variable 269 id statement 106 136 169 175
fabric softener example 94 id= defined 312
facopts= defined 315 identity 106 140 175-177 180 218 226 367
factors 76 IIA 179 185 192 379-383
factors statement 317 imlopts= defined 344
factors= defined 312 315 325-326 improving an existing design 268
factors= 315-316 321 Income variable 230-231
failed initialization 333 independence 107
Federov, modified 123 289 independence from irrelevant alternatives 179 185
file statement 100 192
FINAL data set 352 Index variable 257 295
fitting the choice model 88-90 107 138-142 145 149 information matrix 77
170-172 175-177 183 187-188 226 234 247- init= defined 301 339
249 374-382 init= 126 217 268-269 275 291 295 301 337-339
fixed choice sets 269 343
fixed= defined 301 343 initblock= defined 312
fixed= 269 initialization failed 333
flags= defined 300 initialization switching 333
flags= 255 266 289 299-301 initvars= defined 301
food product (availability) example 192 initvars= 291 295 301
Form variable 110 135 171 input data 85
format statement 147 354 input statement 85
format= defined 326 input function 105 134
formats 92 95 99 112 167-169 173 185 212 216 int= defined 351
&forms variable 110 int= 254 289
fractional-factorial designs 76 interact= defined 316 337
FREQ data set 326 interact= 317
freq statement 172 187-188 226 247-249 interactions 76 146 155 171 193-194 208 214
Freq variable 187 intiter= defined 301
- FREQ- variable 171-172 225-226 intiter= 217 291 301-302
freq= defined 304 invalid page errors 288
freq= 246 304-305 iter= defined 302 306 312 316 341
freqs= defined 326 iter= 283 314-316
freqs= 326 iteration history 334
frequencies, n-way 326 j1 variable 338
frequency variable 171-173 185-188 j2 variable 338
FSUM data set 326 j3 variable 338
full-factorial design 194 justparse defined 364
G-efficiency 77 keep= defined 316 359
generate statement 316 keep= 215 316
generate= defined 316 KEY data set, %MktLab 212 239 346-351
generic attributes 138 KEY data set, %MktRoll 102-103 133 147 167 215
generic defined 324 244 264 292 296-297 310 320-322 344 357-
generic design 252-257 261-267 359
generic 324 key= defined 351 359
geometric mean 77 key= 103 211-212 215 239 287 344-350 353 356-
 Index 389

359 287-288 306 316 336 360-364 390

knots= 142 %PhChoice 74 79-80 88 107 138-140 170 176
label statement 354 226 247 287-288 364 367-368 391
label, variable 88-90 99 105-106 128 136 140-144 macros, autocall 287
169 175-177 180-188 212 218 226 230 350- main effects 76 193-194 208
352 366-367 &main variable 285-286
labels= defined 352 match- all 128
labels= 350 mautosource 287
large data sets 171 185 max= defined 364
levels 76 max= 156 362-364
likelihood 79 85-88 107 151 172 179 188-191 372- maxdesigns= defined 341
376 382-384 maxdesigns= 281
linear defined 324 maxiter= defined 302 306 312 316 341
linear design 78-79 95 102-104 maxiter= 199 255 302 305-306 341
linear 324 maxstages= defined 341
linesleft= 100 maxstarts= defined 306
LIST data set 326 maxstarts= 305-306
list defined 306 336 345 364 maxtime= defined 341
list 364 maxtime= 124 160 199 341-342
list= defined 326 maxtries= defined 306
Lodge variable 133 136 148 168-169 memory, running with less 171
lprefix= 106 136 169 175-177 218 method= defined 316
machine differences 303 307 313 318 339 Micro variable 215-218 221
macro errors 288 minimum number of choice sets 252
macro variables 95 110 missing statement 211
macro missing 212
%ChoicEff 74-75 215-220 226 252-267 275-277 %MktAllo macro 74 245-246 287 303-304 389
280 283 287-299 307 310-312 320-321 351 %MktAllo macro documentation 303-305
387-389 %MktBal macro 74 206 287 305-306 336 389
%EvalEff 206 %MktBal macro documentation 305-307
%MktAllo 74 245-246 287 303-304 389 %MktBlock macro 74 162 205 209-210 287 307
%MktBal 74 206 287 305-306 336 389 310-313 349-350 389
%MktBlock 74 162 205 209-210 287 307 310- %MktBlock macro documentation 307-313
313 349-350 389 %MktDes macro 74 287 314-318 340 389
%MktDes 74 287 314-318 340 389 %MktDes macro documentation 314-319
%MktDups 74 280 283 287 302 319-324 389 MKTDESCAT data set 356
%MktEval 74 97-98 124 129 154 160 200-205 MKTDESLEV data set 356
239 287 305-307 313 325 350 389 %MktDups macro 74 280 283 287 302 319-324 389
%MktEx 74 77-79 95-97 102 110 118-129 154- %MktDups macro documentation 319-325
157 162 195 199 202-206 238-239 252-256 %MktEval macro 74 97-98 124 129 154 160 200-
261 264 267-269 273 278-281 285-289 292- 205 239 287 305-307 313 325 350 389
293 296-297 305-307 312-314 318-336 343- %MktEval macro documentation 325-326
358 361 369 387-390 %MktEx macro 74 77-79 95-97 102 110 118-129
%MktKey 133 244 264-265 283 287 344-345 356 154-157 162 195 199 202-206 238-239 252-
390 256 261 264 267-269 273 278-281 285-289
%MktLab 74 105 110 129 211-212 239-240 254 292-293 296-297 305-307 312-314 318-336
275 287-289 292-293 307 321 332 345-352 343-358 361 369 387-390
388-390 %MktEx macro algorithm 123-124
%MktMerge 74 87 104 135 147 168 224 229 287 %MktEx macro documentation 327-344
353 390 %MktEx macro notes 333
%MktOrth 74 118 287 354-356 390 %MktEx macro versus the %ChoicEff macro 267
%MktRoll 74 102-103 133 147 168 211 215 244 %MktEx macro, common options explained 95 119
252 258 264-265 283 287 292 296-297 307 126
310-312 320-322 344 353-359 388-390 %MktKey macro 133 244 264-265 283 287 344-345
%MktRuns 74 94 117 153 156 194 202 237 253 356 390
 390 TS-677E Multinomial Logit, Discrete Choice Modeling

%MktKey macro documentation 344-345 nosort option 269 339

%MktLab macro 74 105 110 129 211-212 239-240 not available 347
254 275 287-289 292-293 307 321 332 345- notes defined 302 360
352 388-390 notes, %MktEx macro 333
%MktLab macro documentation 345-353 notests defined 302
%MktMerge macro 74 87 104 135 147 168 224 229 notruncate 249
287 353 390 nowarn defined 360
%MktMerge macro documentation 353-354 nozeroconstant 106 136 169 175
%MktOrth macro 74 118 287 354-356 390 nsets= defined 300 354
%MktOrth macro documentation 354-356 nsets= 104 135 217 255 289 299 353
%MktRoll macro 74 102-103 133 147 168 211 215 null header 128
244 252 258 264-265 283 287 292 296-297 number of choice sets, minimum 252
307 310-312 320-322 344 353-359 388-390 NUMS data set 364
%MktRoll macro documentation 356-360 n-way frequencies 326
%MktRuns macro 74 94 117 153 156 194 202 237 ODS 80 364
253 287-288 306 316 336 360-364 390 ods output 128
%MktRuns macro documentation 360-364 onoff defined 368
%MktRuns macro errors 288 options defined 324
%MktRuns macro, with interactions 156 options= defined 302 306 317 337 360 364
model comparisons 151 179 190 382-383 options=accept 281 285-286 333 336-339
model statement 88 106-107 136-138 169 175-177 options=allcode 317
180 206 264 289 299 317-318 options=branded 324
model= defined 299 options=check 126 317 337-339
model= 255 283 289 299-301 options=coded 302
morevars= defined 302 options=detail 302
mother logit 179 185 193 226 381-382 options=generic 324
multinomial logit 83 87-88 107 175-177 192 373 options=justparse 364
multiple choices 237 options=linear 324
mutate= defined 342 options=nocode 302 317
mutate= 281 340-342 options=nodups 98 302 337 340
mutations 124 options=nofinal 337
mutiter= defined 342 options=nohistory 337
mutiter= 342 options=noprint 306 324
n variable 118 257 options=nosort 269 337-339
.N special missing value 347 options=notes 302 360
n= defined 302 306 316 336 364 options=notests 302
n= 95 110 157 194 289 296-297 302 316 327-328 options=nowarn 360
nalts= defined 300 305 312 324 354 options=orthcan 302
nalts= 104 135 217 224 245 266 292 299-301 304 options=progress 306
312-313 324 353 optiter= defined 342
nblocks= defined 313 optiter= 281 340-342
next= defined 313 order= 169
nknots= 142 order=data 106 136
nlev= defined 316 orthcan defined 302
nlev= 315 318 orthogonal 76-77
nocode defined 302 317 orthogonal and balanced 77 80-82 95
nodups defined 302 337 orthogonal array 76
nodups option 98 337 340 orthogonal coding 80-82
nofinal defined 337 otherfac= defined 317
nohistory defined 337 otherint= defined 317
None alternative 194 215 226 229 234-236 out= defined 303-306 313 317 325 339 352-354 360
noprint defined 306 324 364
noprint 324 out= 103-106 136 169 175 239 245-246 304 313 321
norestoremissing 106 136 144 148 169 175 332 337-339 347-348 351-360
nosort defined 337 outall= defined 339 356
 Index 391

outall= 337 PROC LOGISTIC 372

outcat= defined 356 PROC MEANS 109
outcb= defined 326 PROC OPTEX 123 198-199 206 318
outcorr= defined 326 PROC PHREG 79-80 87-90 106-107 136-145 149
OUTDUPS data set 325 170-177 183 187-188 226 234 247-249 364
outest= 88 374-384
outfreq= defined 326 PROC PHREG, common options explained 88
outfsum= defined 326 PROC PLAN 123
outlev= defined 356 PROC PROBIT 372
outlev= 355-356 PROC SCORE 109
outlist= defined 325-326 PROC SORT 92 114
Output Delivery System 80 364 PROC SUMMARY 171-172 225 243
output statement 106 136 169 175 PROC TEMPLATE 80 364-367
outr= defined 339 PROC TRANSPOSE 112-114
outr= 332 337-339 346 PROC TRANSREG 105-107 136-148 169 175-177
page errors 288 180 183-188 226 230 246 264 367
page, new 113 PROC TRANSREG, common options explained 106
param=orthref 206 169
parameters 83 88-92 141-143 190 193-194 372-374 procopts= defined 317
377 381-384 progress defined 306
partial profiles 274 278-281 337 proportional hazards 79 87 188 374
partial= defined 337 proportions, analyzing 249
partial= 275 278-280 286 337-339 pseudo-factors 315
part-worth utility 83 91 143 167 171 pspline 142
&pass variable 285-286 put statement 167
Pattern variable 358 put function 105 134
permanent SAS data set 110 quadratic price effects 141-142 145 194
persist 128 quantitative factor 109 140-141 171
%PhChoice macro 74 79-80 88 107 138-140 170 questionnaire 100-101 110-114 131 164
176 226 247 287-288 364 367-368 391 Ran 160
%PhChoice macro documentation 364-368 random mutations 124 160
%PhChoice macro errors 288 random number seeds 95 119 126 198 205 209 255
PHREG output, customizing 79 364-367 289 303 307 313 318 339
Place variable 133 136 147-148 168-169 randomization 99 110 164 214
point= 87 113 RANDOMIZED data set 332 339
Pre 160 red bus 192
prefix= defined 352 reference level 91 139 143 194
prescription drugs (allocation) example 237 Reference variable 118
Price variable 102-106 109 133-136 140 148 168- resolution 76
169 175 179-180 193 211 215 218-222 246 restrictions 195 199 202-204 274 278-281 285 338
298 303 356-358 restrictions not met 333
price, assigning actual 105 134 141 147 168 restrictions= defined 338
PriceL variable 141-142 restrictions= 195 336-339
print= defined 313 326 RESULTS data set 303
print= 98 ridge= defined 313 344
Prob variable 257 RowHeader 366
probability of choice 83-85 92-93 109-110 192 373- rowname variable 128
377 Run variable 312-313 349
PROC CATMOD 376 run= defined 317
PROC FACTEX 123 runs 76
PROC FORMAT 92 95 128 134 167 173 185 212 saturated design 82 94-95 117 153
PROC FREQ 320 Scene variable 133 136 148 168-169
PROC GLM 207-208 score= 109
PROC GPLOT 84 199 second choice 85 88
PROC IML 256 seed= defined 303 307 313 318 339
 392 TS-677E Multinomial Logit, Discrete Choice Modeling

seed= 95 255 289 303 307 313 318 339 subdesign 194 206
separators= 169 177 180 218 Subj variable 85 88-89 105-107 173-174 179-180
sequential algorithm 206 226
set statement 87 113 subject attributes 229
Set variable 85 88-89 104-107 110 113-115 171-173 submat= defined 303
179-180 187 217 245 257 275 312 360 subsequent choice 85-88 135 173
Set 313 summary table 88-89 188 229
set= defined 313 360 survival analysis 79 87 374
setvars= defined 354 switching initialization 333
setvars= 104 135 353 symsize= 344
Shape variable 357-358 Tab 160
Shelf variable 215-218 221 tabiter= defined 342
shelf-talker 192 211 214-215 229 tabiter= 281
Side variable 168-169 tabsize= defined 344
simulated annealing 124 160 target= defined 344
Size variable 357-358 - temporary- 100
size= defined 318 ties=breslow 79 87-88 107 188
source statement 366 time (computer), saving 171
source [Link] statement 365 trace 77
statement &- trgind variable 107-109 138-142 145 149 170-
class 206 315 318 172 175-177 180 183 187-188 226 234 247-
column 366 249
edit 366 -2 LOG L 90 179 188-190 383
factors 317 type= 109
file 100 types= defined 303
format 147 354 types= 303
freq 172 187 249 typevar= defined 303
generate 316 typevar= 303
id 106 136 169 175 unbalanced= defined 342
input 85 vacation (alternative-specific) example 152
label 354 vacation example 116
missing 211 values= defined 353
model 88 106-107 136-138 169 175-177 180 206 values= 350-353
264 289 299 317-318 variable label 88-90 99 105-106 128 136 140-144 169
output 106 136 169 175 175-177 180-188 212 218 226 230 350-352
put 167 366-367
set 87 113 variable name 367
source [Link] 365 variable
source 366 Age 230-231
strata 88 107 173 187 Alt 110 310-312
where 206 225 247 319 - Alt- 359
statements= defined 352 bad 195 338
step= defined 318 Block 133 239 247 307 312 349
step= 317-318 Brand 102-106 175 179-180 215 218-222 231
stmts= defined 354 242-246 298 303 356-359
stmts= 147 c 85-88 105 109 113 135 179-180 225 246 304
stopearly= defined 343 353-354
stopearly= 333 Choice 87
stopping early 333 Choose 115
Stove variable 211 Client 347
strata 88-89 107-109 171-173 185 189-191 374-376 Color 357-358
383 Count 242-243 303
strata statement 88 107 173 187 Design 118 257
structural zeros 91 143 151 Dest 131
Style=RowHeader 366 Efficiency 257
Index 393

f 269 Side 168-169

Form 110 135 171 Size 357-358
&forms 110 Stove 211
Freq 187 Subj 85 88-89 105-107 173-174 179-180 226
- FREQ- 171-172 225-226 &- trgind 107-109 138-142 145 149 170-172
i 338 175-177 180 183 187-188 226 234 247-249
Income 230-231 w 216-217 225
Index 257 295 x 338
j1 338 x1 338
j2 338 x[j] 338
j3 338 xmat 338
Lodge 133 136 148 168-169 vars= defined 305 312 325-326 353
&main 285-286 vars= 129 246 304
Micro 215-218 221 view= 206
n 118 257 w variable 216-217 225
&pass 285-286 weight= defined 303
Pattern 358 weight= 217
Place 133 136 147-148 168-169 where statement 206 225 247 319
Price 102-106 109 133-136 140 148 168-169 where= defined 319
175 179-180 193 211 215 218-222 246 298 where= 109
303 356-358 With Covariates 90 151 179
PriceL 141-142 worksize= 344
Prob 257 x variable 338
Reference 118 x1 variable 338
rowname 128 x[j] variable 338
Run 312-313 349 xmat variable 338
Scene 133 136 148 168-169 zero= 106 136 139-140 143-144 169 175-177 185
Set 85 88-89 104-107 110 113-115 171-173 179- 218 226 275 298
180 187 217 245 257 275 312 360 zero=list 140 169
Shape 357-358 zero=’ ’ 169 293
Shelf 215-218 221