NOTICE
This document is disseminated under the
sponsorship of the Department of Trans¬
portation in the interest of information
exchange. The United States Government
assumes no liability for its contents or
use thereof.
1. Raport No.
D0T/RSPA/DPB-50/79/39
2. Government Accession No.
3. Recipient's Cotolog No.
4. Title and Subtitle
"A Method for Understanding and Predicting
Destination Choices"
5. Report Dote
December 1979
6. Performing Organizotion Code
8. Performing Organization Report No. !
7. Author's)
Peter R. Stopher, Principal Investigator
9. Performing Organizotion Noma and Address
Northwestern University
Transportation Center
2001 Sheridan Rd - Evanston, 111 60201
10. Work Unit No. (TRAIS)
11. Contract or Grant No.
DOT - OS - 40001
13. Type of Report and Period Covered
Final Report
May 1974 - May 1977
12. Sponsoring Agency Nome and Address
IJ. S. Department of Transportation
Research and Special Programs Administration
Qffiçe of University Res
dashinaton. D. C. 20690
earch
14. Sponsoring Agency Code 1
DPB-50
15. Supplementary Notas
Technical Monitor: Edward Weiner, P-12
The principal goal of the research reported here is to develop a proto¬
type model of destination choice from the basis of individual-choice hypo¬
theses and the application of attitudinal enquiry. The secondary goal is to
determine new kinds of measures of site attractiveness for destination-choice
models, substituting for conventional measures of size, such as employment
or floor area. The next task of the research was to develop a data base
appropriate for testing the hypotheses of the research. A variety of models
were constructed from the data and subjected to a number of tests. The re¬
sulting model structure consists of three integrated components. These de¬
scribe the individual perceptions of shopping locations; the individual
ratings of shipping-location attractiveness, based on relative preferences
for perceived characteristics of the location; and choice of shopping loca-
tion based on attractiveness ratings and accessibility. Four methods were
used to describe individual perceptions: fundamental attributes, non-metric
scaling, factor analysis, and discriminant analysis. Similarly, four pref¬
erence models were also tested: preference regression, first-preference
logit, expectancy value, and unit weights. The choice model tested was a
multinomial-logit model, applied to reported choice data. Subject to con¬
firmation by further empirical studies, it is recommended that factor analy¬
sis be used to identify perceptions, preference regression or first-prefer¬
ence logit be used to identify importance weights, and that revealed prefer¬
ence and intermediate preference models be used to identify choice behavior.
The resulting models indicate that attractiveness of alternative non-gro¬
cery shopping destinations is based on quality, variety, satisfaction, and
parking. Quality is the most important attractiveness aspect of shopping
destination choice.
17. Key Words
transportation
destination choices
18. Distribution Statement I
Document is available to the U. S.
Public through the National Technical
Information Service, Springfield,
Virginia 22161
19. Security Clossif. (of this report)
Uhclassified
20. Security Clossif. (of this page)
Unclassified
21» No. of Pages
22. Price
Form DOT F 1700.7 (8—72) Reproduction of completed poge outhorized
Approxiaoto Coovoitioai I» Mouic Moosmot
« yabol
Wkeu Y to Ksmv
Multiply by
Tc tin*
Syaibal
• —
LENGTH
la
inchaa
•2.1
cant ima tars
cm
h
faal
30
cant ima tara
cm
*
yards
0.»
matara
m
al
miles
14
kilometers
km
AREA
in»
square inchaa
6.6
aquara cantimatara
cm2
n1
square faat
0.09
aquara matara
at2
yd»
square yards
0.1
aquara matara
m2
1»
aquara ailaa
2.1
aquara kilomstara
km2
actus
0.4
hac taras
ka
MASS (wuight)
M
ouncaa
21
grama
0
lb .
pounda
0.46
kilograms
bg
abort (ona
0.0
tonnas
i
(2000 lb)
VOLUME
tap
taaapoona
6
millilitars
ml
Tba#
tablaapooaa
16
millilitara
ml
ft OS
fluid ouacaa
30
millilitars
ml
c
cupa
0.24
litsrs
1
P*
pinta
0.47
litara
1
4«
quarts
0.96
litars
1
«al
gallons
3.a
litara
1
fi»
cubic faal
0.03
Cubic matara
m2
yd»
cubic yards
0.76
cubic matara
ai»
TEMPERATURE (oxact)
Fahrenheit
6/9 (attar
Calaiua
•c
temperature
subtracting
tamparatura
METRIC CONVERSION FACTORS
- *c
'I Mi ' 2 A4 (eaaciiyl. Fu» oUmn «OCI convwimtu and mturn oaiiilad uUm, mm NftS Uk. Putol. 2M.
Uium oI Mamma and Umwh. Pi»cm *2.26. SO Catalog Mo. CU.iO.2a4.-
Approxiaolo Coovoixiooi liaai Manic Mi
Multiply by
Tu fiad
Symbul
LENGTH
mi II ima tar a
0.04
inches
u
cantimatara
0.4
inchaa
to
matera
3.3
faat
ft
matara
1.1
yards
H
kilomstara
0.6
miles
ml
AREA
aquara cantimatara
0.16
aquara inchaa
to»
aquara matara
14
aquara yarda
aquara kilomstara
0.4
aquara milaa
ml»
kKUtu 110.000 m'|
2.6
icm
MASS (weight)
grama
0.036
ounces
M
kilograms
24
pounds
to
tonnas (1000 kg)
1.1
short lou
VOLUME
millilitara
0.03
fluid ouncaa
flM
litars
2.1
pints
m
litara
1.06
quarts
litara
0.26
gal Ian a
®ï
cubic matara
36
cubic last
tt»
cubic matara
1.3
cubic yarda
vd»
TEMPERATURE (oxocl)
Cslaius 9/6 (thon Fahranhsit
tsmparatura add 32) tamparatura
•P
-40
Hnf-
-40
•C
f»<
140 «0 I
' I ' l' 'I1 f
20 U
•P
SI
■» ' 1 *.»'
I
EXECUTIVE SUMMARY
The principal goal of the research reported here,is to develop a
prototype model of destination choice from the basis of individual-choice
hypotheses and the application of attitudinal enquiry. The secondary goal
is to determine new kinds of measures of site attractiveness for destina¬
tion-choice models, substituting for conventional measures of size, such as
employment or floor area.
For pragmatic reasons, the research effort was limited to a study of
nongrocery shopping trips. Having made this decision, a wide-ranging
literature search was undertaken in a number of disciplines to determine
measures of store attractiveness and to review extant models of destination
choice. This literature search is described in detail in chapter 2 of the
report. It produced a number of appropriate measures of site attractiveness,
but failed to produce any suitable models or data sets.
Because of the lack of suitable, existing data, the next task of the
research was to develop a data base appropriate for testing the hypotheses
of the research. After substantial pilot testing, a survey instrument was
developed for drawing a choice-based sample of nongrocery shoppers and deter¬
mining various facts about the shopping trip and the individual, together
with extensive data on preferences and perceptions of shopping-center attri¬
butes. Questionnaires were returned by nearly 8,000 respondents, 7,362 being
sufficiently complete to permit substantial analysis. The characteristics of
the sample were found to be similar, in most respects, to those that would be
expected from a random sample of households in the catchment area of the
shopping locations used. Two further surveys were carried out subsequently.
The first was designed to collect data on the choice of travel mode for non-
grocery shopping trips and the second was an extension of the original survey
on preferences and perceptions, designed to permit more extensive analyses
and some limited policy tests of the prototype models. Unfortunately, within
the contract terms for the final year's effort, it was not possible to make
full use of these later two data sets. Full details of the survey processes
and the samples obtained, together with some details of the responses, are
provided in chapter 3 of the report.
A variety of models were constructed from the data and subjected to a
number of tests. The resulting model structure consists of three integrated
components. These describe the individual perceptions of shopping locations;
the individual ratings of shopping-location attractiveness, based on relative
preferences for perceived characteristics of the location; and choice of
shopping location based on attractiveness ratings and an accessibility
measure (distance). Four methods were used to describe individual percep¬
tions, these being fundamental attributes, non-metric scaling, factor analysis,
and discriminant analysis. Similarly, four preference models were also tested:
preference regression, first-preference logit, expectancy value, and unit
weights. Finally, the choice model tested was a multinomial-logit model,
applied both to historical choice data and to the observed choices of the
sample. Details of these various models are provided in chapter 6 of the
report.
Subject to confirmation by further empirical studies, it is recommended
that factor analysis be used to identify perceptions, preference regression or
first-preference logit be used to identify importance weights, and that
E-2
revealed preference and intermediate preference models be used to identify
choice behavior.
Attempts were also made to identify market segments within the population
as a means to improve both the goodness of fit of the various models and the
predictive power of the models. Extensive efforts were made to segment the
population by socioeconomic characteristics on various measures of preference,
perception, and combined preferences and perceptions. These efforts are des¬
cribed in chapter 5 of the report. It was concluded that no satisfactory
segmentation of the population could be achieved within the limitations of the
data, methods for segmentation, and available statistical tests for determin¬
ing the usefulness of any particular segmentation scheme. Therefore, the
perception, preference, and choice models were developed on an unsegmented
sample, of nongrocery shoppers.
Finally, it should be noted that the resulting prototype models of choice,
which were based on four compound measures of attractiveness and distance,
were found to be most sensitive to measures relating to service characteristics
and least sensitive to measures of the transportation system. As a result,
sensitivity tests on transportation-related policy issues were found to show
evidence of only limited changes in choice behavior. This finding may well
be indicative of the real short-term behavior of nongrocery shoppers and is
not necessarily a shortcoming of the models.
ACKNOWLEDGEMENTS
A substantial number of people have contributed to this project over
its life. The number of these seem too great to indicate them all as
co-authors, and the principal investigator, Professor Peter R. Stopher,
takes primary responsibility for the final report on the project. Acknowl¬
edgements are made to all of those who have assisted in the work on this
project over the three years that it took. First, we would like to acknowl¬
edge the help of Professor Peter L. Watson, who was co-principal investigator
for the first year of the project. We would also like to acknowledge the
assistance of Professor Richard B. Westin, who worked on much of the early
material in the development of theory and methods. Several segments of the
work were contributed by Professors Jean Blin and George L. Peterson, both
of whom assisted considerably in the first year of the project in helping to
develop the directions in which the project would go and in developing the
data collection efforts. We would like to acknowledge the assistance of
Professor Frank S. Koppelman and Professor John R. Hauser, who took primary
responsibility for the development of the preference and perception models
and the final choice models reported in the later chapters of this report.
In addition to these faculty members, a number of graduate students contributed
to the project quite extensively. Major contributions were provided by
Joseph N. Prashker and James Clark, and also by Tom Zlatoper and Bruce Bagamery.
Without the very major contributions that these four students made to the
project, it could not have been completed. We would also like to acknowledge
the assistance of a number of students who assisted with the data collection
efforts in each of the two summers that they were carried out. The students
assisting are too numerous to number.
The principal direction of the project was provided by Professor Peter
R. Stopher, who also takes responsibility for any omissions and errors in
this final report. Many of those listed above contributed to the writing
of the final report, but it has not seemed practical to identify their in¬
dividual contributions at the specific parts of the report that would be
appropriate.
- i -
TABLE OF CONTENTS
flje
1. INTRODUCTION 1
2. LITERATURE REVIEW 5
2.1 Background 5
2.2 Economie and Psychological Choice Foundations for
Predicting Travel Demand 6
2.3 Destination Choice Work in Marketing 11
2.3.1 Retail Location Choice Factors 12
2.3.2 Specification Problems 15
2.3.3 Predictive Models 16
2.4 Destination Choice Work in Geography 19
2.5 Hypotheses to be Tested 20
3. DATA COLLECTION AND DESCRIPTION 23
3.1 Introduction 23
3.2 The First Destination-Choice Survey 24
3.2.1 Development of the Attribute Set 24
3.2.2 Execution of the Primary Survey 28
3.2.3 Description of the Sample 37
3.3 The Second Destination-Choice Survey 52
3.3.1 Survey Design 52
3.3.2 Characteristics of the Second-Survey Sample 55
3.4 The Mode-Choice Survey 66
3.4.1 Survey Design and Execution
4. METHODOLOGY OF THE APPROACH 75
4.1 A Marketing Approach 75
4.1.1 Perceptual Models 77
4.1.2 Preference Models 81
4.1.3 Segmentation Analysis 84
4.1.4 Choice Models 86
4.2 Multidimensional Scaling Methods 87
4.2.1 Proximities Data 88
4.2.2 Dominance Data 90
4.2.3 Profile Data 92
4.2.4 Conjoint Measurement Data 92
4.3 Multidimensional Scaling of Perceived
Attractiveness 93
4.3.1 Principles of Reduction 93
4.3.2 Theory of Multidimensional Scaling 95
4.3.3 The INDSCAL Model 103
4.3.4 Identification of Dimensions 107
4.4 The Factor-Analysis Model 109
- iii
Table of Contents (Continued)
Page
MARKET SEGMENTATION
118
5.1
Introduction
118
5.2
Methods of Grouping
118
5.2.1
Prior Classification
120
5.2.2
Search for Classification
120
5.2.3
Tests for Similarity of Groups
122
5.3
Segmentation on Perceptions
122
5.3.1
General
122
5.3.2
Correlation Analysis
133
5.3.3
Cluster Analysis of Perceptions
139
5.4
Identification of an Attractiveness Space
144
5.5
Preference Segmentation
145
5.6
Preference and Perception Segmentation
157
5.7
Conclusions
170
MODELS OF PERCEPTIONS, PREFERENCE, AND CHOICE
175
6.1
Introduction
175
6.2
Objectives of the Research and Approach
176
6.2.1
Insight into Shopping Location Choice Behavior
178
6.2.2
Comparison of Model Structures
178
6.2.3
Modeling Consumer Perceptions
179
6.2.4
Modeling Consumer Preferences
181
6.2.5
Modeling Consumer-Choice Behavior
185
6.2.6
Linked Perception, Preference, and Choice Models
185
6.3
Empirical Setting and Experimental Design
187
6.4
Results of the Analysis
188
6.4.1
Perceptions Models
188
6.4.2
Preference Models
194
S
6.4.3
Choice Models
197
6.4.4
Linked Model Structure
199
6.4.5
Summary
200
6.5
Predictive Ability
200
6.5.1
Prediction Formation
200
6.5.2
Tests of Preference Prediction
202
6.5.3
Tests of Choice Prediction
202
6.5.4
Preference Prediction Results
202
6.5.5
Choice Prediction Results
204
6.5.6
Summary of Predictive Ability Analysis
204
6.6
Ease of Use and Cost
204
6.6.1
Ease of Use and Cost of Perception Model
Development
204
6.6.2
Ease of Use and Cost of Preference Model
Development
206
6.6.3
Ease of Use and Cost of Choice Model Development
206
6.6.4
Ease of Use and Cost of Model Sets
207
6.7
Reliability and Extendability
207
6.8
Sensitivity Tests
212
6.8.1
Description of the Tests
212
6.8.2
Results of the Sensitivity Tests
216
6.8.3
Conclusions
216
- iv -
Table of Contents (Continued)
Page
7. SUMMARY AND RECOMMENDATIONS FOR FUTURE RESEARCH 218
7.1 Summary 218
7.1 Future Research Directions 219
REFERENCES 221
APPENDIX A: PRETEST CROSSTABULATIONS A-l
APPENDIX B: 1974 QUESTIONNAIRE
1975 QUESTIONNAIRE B-l
APPENDIX C: ADDITIONAL TABULATIONS AND CROSS-
TABULATIONS FOR THE FIRST OLD
ORCHARD SURVEY C-l
APPENDIX D: DERIVATION OF PSYCHOLOGICAL DISTANCES
FOR MULTIDIMENSIONAL SCALING D-l
APPENDIX E: CLASSIFICATION WITH RESPECT TO
DIFFERENCES IN PERCEPTION OF
DISSIMILARITIES E-l
- v -
LIST OF TABLES
Page
3-1 38 Attractiveness Measures 25
3-2 Store and Shopping-Center Characteristics 27
3-3 Revised Store and Shopping Center Characteristics 29
3-4 Percentage Breakdown of Socioeconomic Characteristics for
the Pretest Survey 34
3-5 Sample Income Distribution 40
3-6 Distribution of Income by Sex 41
3-7 Distribution of Length of Residence 41
3-8 Distribution of Length of Residence by Sex 42
3-9 Distribution of Length of Residence by Income 42
3-10 Distribution of Occupation 44
3-11 Age Distribution for the Sample 44
3-12 Distribution of Next Destination 45
3-13 Distribution of Previous Origin 45
3-14 Origin by Destination 46
3-15 Length of Residence by Age 47
3-16 Income by Age 48
3-17 Mode of Travel Used to the Shopping Center 50
3-18 Items Purchased 50
3-19 Familiarity Reported by Respondents to Centers Used in
Psychological Scaling Questions 51
3-20 Income Distribution for Second-Survey Sample 56
3-21 Length of Residence Distribution for Second Sample 57
3-22 Distribution of Income by Sex for Second Sample 58
3-23 Distribution of Length of Residence by Sex for Second Sample 58
3-24 Occupation Distribution for Second Sample 59
3-25 Age Distribution for Second Sample 60
3-26 Educational Distribution for Old Orchard Sample 62
3-27 Life Cycle Distribution for Old Orchard Sample 62
3-28 Mode of Travel Used to Old Orchard 63
3-29 Distribution of Origin by Destination 63
3-30 Distribution of Previous Origin for Old Orchard Respondents 64
3-31 Distribution of Next Destination for Old Orchard Respondents 64
3-32 Main Items Shopped for by Old Orchard Respondents 65
3-33 Familiarity Information on Main Subject Set 65
3-34 Age Distribution for Evanston Survey 69
3-35 Stage in Family Life-Cycle 69
3-36 Distribution of Education Levels for the Evanston Survey 70
3-37 Distribution of Length of Residence for the Evanston Survey 70
3-38 Distribution of Family Income for the Evanston Survey 70
3-39 Cross-tabulation of Travel Mode and Shopping Day for the
Evanston Survey 72
3-40 Cross-tabulation of Income and Shopping Day for the Evanston
Survey 72
3-41 Distribution of Chosen and Alternative Modes for the Evanston
Survey 73
3-42 Cross-tabulation of Chosen and Alternative Modes for the
Evanston Survey 73
3-43 Cross-tabulation of Mode and Income for the Evanston Survey 74
- vii -
List of Tables (continued)
Page
4-1 Importance Weights for Reduced Dimensions 85
4-2 Weight and Height of 3 Stimuli 99
5-1 Socioeconomic Groups for First-Cut Analysis 119
5-2 Selected Dimensionalities for One-Way Groupings 132
5-3 Pearson Correlations Among Age Groups 134
5-4 Pearson Correlations Among Occupational Groups 134
5-5 Pearson Correlations Among Income Groups 135
5-6 Pearson Correlations By Length of Residence 135
5-7 Pearson Correlation Between the Sexes 136
5-8 Pearson Correlations Among Age Groups 136
5-9 Pearson Correlations Among Occupational Groups 137
5-10 Pearson Correlations Among Income Groups 137
5-11 Pearson Correlations by Length of Residence 138
5-12 Pearson Correlations for Income and Occupation Subgroups
Having 2-Dirnensional Solutions 138
5-13 Clustering of Four-Dimensional Solutions Within Socioeconomic
Variables 140
5-14 Clustering of Three-Dimensional Solutions Within Socio¬
economic Variables 141
5-15 Clustering of All Solutions in Three and Four Dimensions 143
5-16 Preference Rankings of Population Groupings According to Age 150
5-17 Preference Rankings of Population Groupings According to
Occupation 151
5-18 Preference Rankings of Population Groupings According to
Household Income • 152
5-19 Preference Rankings of Population Groupings According to
Length of Residence 153
5-20 Preference Rankings of Population Groupings According to Sex 154
5-21 Results of Friedman Tests on Age Groupings 155
5-22 Results of Friedman Tests on Income Groupings 155
5-23 Results of Friedman Test on Sex 155
5-24 Results of Friedman Test on Length of Residence 156
5-25 Results of Friedman Tests on Occupation Groups 156
5-26 Preference Rankings of Population Groupings According to
Distance to Chosen Shopping Center 158
5-27 Preference Rankings of Population Groupings According to
Closest Shopping Center 159
5-28 Preference Rankings of Population Groupings According to
Most Preferred Shopping Center 160
5-29 Preference Rankings of Population Groupings According to
Shopping Center Choice 161
5-30 Preference Rankings of Population Groupings According to
Methods of Survey Completion 162
5-31 Results of Friedman Test on Distance to Chosen Shopping Center 163
5-32 Results of Friedman Test on Distance to Closest Shopping
Center 163
5-33 Results of Friedman Tests on Most Preferred Shopping Center 164
5-34 Results of Friedman Tests on Chosen Shopping Center 164
5-35 Results of Friedman Tests on Method of Survey Completion 164
viii -
List of Tables (continued)
Page
5-36 One Way ANOVA of Cluster Solutions 166
5-37 Test of the Differences of Cluster Means from Population
Means for Each of the Four Variables 167
5-38 Cluster Means for All Cluster Solutions 168
5-39 Preference Rankings of the Population Grouped According to
the Four Clusters Solution 169
5-40 Preference Rankings of the Population Grouped According to
the Five Cluster Solution 169
5-41 Multivariate and Univariate Tests of Significance for
Attribute-Importance Ratings for Clusters of 4 and 5 171
5-42 Contingency Tests on 4- and 5-Cluster Solutions 172
5-43 Summary of Market Segmentation Results 173
6-1 Theoretical Constructs Underlying Four Models of Consumer
Perceptions 177
6-2 Theoretical Constructs Behind Five Models of Consumer
Preferences 183
6-3 Structure Matrices for Three Dimensional Perception Models 190
6-4 Structure Matrices for Four Dimensional Perception Models 192
6-5 Normalized Importance Weights 195
6-6 Normalized Importances for Fundamental Attributes 196
6-7 Choice Analysis Importance Weights 198
6-8 Preference Prediction Tests 203
6-9 Choice Prediction Tests 205
6-10 Rankings of Models in Terms of Least Cost and Ease of Use 208
6-11 Cost Index for Developing Perception/Preference Structure
and Choice Model Estimation 209
6-12 Stability of Importance Weights in Different Choice Situations 211
6-13 Basis of Sensitivity Tests on Choice Models 213
6-14 Changed Distribution of Free Parking in the Loop 214
6-15 Parking Fees at Old Orchard 214
6-16 Reduced Parking at Old Orchard 215
6-17 Pedestrian Mall in the Loop 215
6-18 Results of Sensitivity Tests on the Factor-Analysis, Revealed-
Preference Model 217
6-19 Results of Sensitivity Tests on the Factor-Analysis, First-
Preference Logit Model 218
6-20 Results of Sensitivity Tests on the Fundamental-Attributes
Model 219
ix -
LIST OF FIGURES
Page
3-1 Uni dimensional Scale for Store Characteristics Based on
Entire Sample _ 30
3-2 Unidimensional Scales for Store Characteristics Based on Sex 31
3-3 Unidimensional Scale for Shopping Center Characteristics
Based on Entire Sample t 32
3-4 Unidimensional Scales for Shopping Center Characteristics
Based on Lowest and Highest Income Groups 33
3-5 Locations of the Subject Shopping Centers for the First Survey 38
4-1 Integrated Marketing Approach to Transportation Service 76
4-2 Consumer Response Process 78
4-3 Average Ratings for Four Shopping Centers on the Underlying
Perceptual Scales 79
4-4 Example Measurement of Ratings for Shopping Centers 80
4-5 Factor Loadings for Perception of Shopping Centers 82
4-6 Average Perceptions on Reduced Dimensions 83
4-7 Interval Scale of Preference for Five Stimuli 97
4-8 Scatter Diagram Displaying a Monotone Relation 101
4-9 Fitting of Attribute into Perceptual Space 108
4-10 Assumed Components of an Attribute 110
4-11 Rotation of the Common Factors 114
4-12 Factor loadings before and after rotation of a factor
analytical solution. 116
5-1 Direct Similarity Measurement 124
5-2 Indirect Similarity Measurement 125
5-3 Two-Dimensional Space for Clerical Workers 126
5-4 Two-Dimensional Space for Incomes Over $50,000 127
5-5 Two-Dimensional Space for Incomes of $10,000-$15,000 128
5-6 Plot of Dimensionality against Stress for Two-, Three-, and
Four-Dimensional Solutions 131
5-7 146
5-8 147
5-9 148
5-10 Preference-Ranking Question for the First Survey 149
5-11 Total Sum-of-Squares Within Clusters for Each Cluster
Solution 165A
6-1 Linked Models of Choice Responses 186
6-2 Map of Fundamental Attributes Ratings for Seven Shipping
Locations 189
6-3 Perceptual Maps for the Four Models of Consumer Perceptions 193
6-4 Prediction Process 201
- x -
1 . INTRODUCTION
This report documents the work carried out during two-and-one-hal f years
on project D0T-0S-40001 entitled "A Method for Assessing Pricing and
Structural Changes on Transport Mode Use". The report is organized in the
order of tasks undertaken during the project period. Chapter 2 provides a
review of the current literature in travel demand and consumer behavior,
including the disciplines of economics, marketing, psychology, and engineering.
Chapter 3 describes the various data collection efforts undertaken to execute
the research. Three surveys were undertaken: two choice-based samples of
individuals making shopping trips to determine their perceptions and prefer¬
ences for alternative shopping sites and attributes, and a choice-based sample
of shoppers' choice of travel mode. Chapter 4 describes the general approach
and methods adopted in this research to develop procedures for predicting
destination choices.
An important way to understand any choice process is to determine whether
or not different population segments have different choice processes or if
they base their choices on substantially different perceptions or preferences.
Some attempts to determine whether the population can be segmented are reported
in chapter 5. Chapter 6 summarizes the results of the analysis that identifies
choice processes and their structures. This chapter reports on perception models
developed as inputs to the choice processes. These perception models attempt
to determine what importance people attach to different attributes. This is
accomplished by using data on individuals' reported preferences for alternative
sites and their perceptions of the attributes possessed by those sites. Chapter
6 also details the resulting choice procedures and documents some sensitivity
tests on the "best" models that were selected. Finally, chapter 7 provides a
summary of the principal findings, and recommendations for future research and
development of the processes developed in this study.
The original purpose of this research was "... to fill the gap which exists
in the set of tools available to urban and regional planners, notably the lack
of a mechanism for assessing the complex effects of changes in transportation
systems and policy. The objective of the research [is]... to develop a mechan¬
ism which is capable of examining a policy change, for example, a central busi¬
ness district parking surcharge, and of tracing out the effects of such a change,
not only on the relative utilization of alternative modes, but also on the
spatial distribution of travel. In other words, we will examine the shifts in
the patterns of destinations which arise from changes in the transportation
system as well as changes in modal usage. Such a development is extremely
important, given current concerns over our urban structure. The basic approach
[will be]... based on recent developments in disaggregate, behavioral, stochas¬
tic methods of travel-demand analysis. Econometric and psychometric techniques
(e.g., logit analysis and psychological-sealing methods) will be used to
develop, estimate and test the models developed. "[Abstract of the original
proposal]. "In short, the objective of this research is to develop a mechanism
which is capable of examining policy changes not only by tracing out the effect
of a change on the relative utilization of alternative modes, but also on the
spatial distribution of travel." [Statement of Work, September 19, 1973].
1
In order to place the remainder of this report into perspective, it is
useful to reproduce here the statements of tasks embodied in the contracts
between the U.S. Department of Transportation and Northwestern University
for this research.
For the first contract period (May 1974 through September 1975), the
work to be performed was stated as follows:
"To successfully complete the project the contractor shall:
1. conceptualize a model
2. establish a choice set
3. develop attractiveness measures
4. estimate the models
5. develop and test the predictions
"The output of this project shall be a prototypical planning tool, a model
which shall concisely display the modal choice and destination decisions of
the riding public. Also, the model shall have independent variables which
are closely related to possible policy tools. The consumer's behavior shall
be related to possible policy options. In the process of conceptualizing
the model the contractor shall pay particular attention to the establishing
of choice sets related to destination decisions. The decision as to the
ultimate modeling method to be selected will be made after consultation with
the contracting officer.
"The contractor shall develop measures of attractiveness for various
destinations. These measures shall be correlated with the various physical
attributes which are of importance to transportation planners. This process
shall permit a direct linkage between travelers perceptions of destinations
and the physical attributes of destinations.
"With the model formulated and measures of attractiveness determined
the contractor shall then estimate the parameters of this model. The con¬
tractor shall use a multiple choice version of logit analysis if at all
practical. The last step to be accomplished by the contractor is an anal¬
ysis of the predictive capabilities of the model in terms of the riders
behavioral responses to various policy changes. Normal econometric testing
procedures shall be used in addition to reverse prediction tests."
To this state of work, the third-year contract added the following:
"A prototype model was successfully structured and calibrated in the
second year of this project. Due to limitations in the data attributed to
the lack of variability regarding socioeconomic characteristics among respon¬
dents to the original survey it was necessary to augment the original data
set. Thus, additional data was collected at the close of the second year of
this project.
2
"B. Work Tasks
Task 1 : The contractor shall recalibrate the prototype model by utilizing
the initial and augmented data sets. In addition to utilizing more appropriate
data, the contractor shall seek to improve upon the first model in the areas of
definition of choice sets of destinations and definition of the travel and
choice utility functions.
Task 2: The contractor shall perform sensitivity testing of the model
response to likely alternative development (shopping centers) situations."
Given the multitude of alternative trip purposes requiring specific des¬
tination-attractiveness definitions, it is necessary to be restrictive on trip
purpose. Thus, an early task of the research was to limit the trip purposes
that would be considered. It became clear that the research would have to
concentrate on what has been defined within the transportation-planning liter¬
ature as a single trip purpose. For this task, a number of trip purpose cate¬
gories can be considered. These include:
1)
Work
2)
School
3)
Shopping soft goods
4)
Shopping hard goods
5)
Social
6)
Recreational
7)
Personal business
8)
Employer's business
9)
Eat meal
These trip purposes represent the most common categories encountered in tradi¬
tional transportation-planning studies, although alternative trip purposes
could be considered if this appeared desirable. It was not considered that a
lengthy research effort into appropriate categorization of trip purposes should
be a part of this research project.
The work trip was rejected as an alternative purpose because it is a
long-run product of the type and location of job, home location, choice of
life-style, and other similar considerations. Hence, it does not represent
a useful starting point for research into destination choices. Rather, it
becomes an element of the total decision process that leads to a specific
home choice, job choice, etc. Second, the standard purposes of social-recre¬
ational trips and personal-business trips contain too many and too varied
sub-trip purposes. These may include such complex issues as temporal and
seasonal instabilities that would exist in cases such as trips to cinemas,
theaters, outdoor recreational facilities, etc. These problems render such
trip purposes inappropriate for research into destination attractiveness.
The employer's business trip is not one in which the individual's choice
is generally exercised. Again, therefore, this would not appear to be an
appropriate type of trip to consider. The problems encountered for trip to
work apply also to trip to school. Little choice is exercised on a dynamic
basis outside the decision on home location and life-style. Trips to eat a
meal represent a relatively small proportion of all trips undertaken within
3
the urban area and, therefore, do not appear to be a main concern for an
investigation of this type. The remaining trip types are the two categories
of shopping trips. Most of the work that has been conducted in destination
choice has examined shopping trips. However, this previous work has focused
primarily on shopping trips for groceries and other perishable household
iterns.
The researchers on this project felt that regular shopping trips for
groceries and other consumables could be considered, in many respects, a
habitual trip. The trip is thus based upon an early decision on home location,
and the habit patterns of shopping at a particular supermarket, or of a choice
between one or two supermarkets, probably based upon the specials offered in
any given week. Since one hypothesis of this research was that a choice must
take place prior to each trip, a trip that entails so much habitual behavior
as the regular shopping trip for consumables does not appear to be an appro¬
priate candidate for this research.
Thus, the final trip purpose to be considered in this set — the trip
for the purchase of consumer durables — becomes the logical trip purpose to
consider for developing notions of destination attractiveness and a choice
model for destination choices. This kind of trip includes all shopping trips
for such items as clothing, shoes, furniture, appliances, luxury items and
housewares that are not included in the regular consumables shopping.
In total, shopping trips of all types (for perishables and consumer
durables) comprise some 20% to 30% of vehicular trips within urban areas.
The precise percentage that consumer durables shopping trips comprise is not
well documented in the literature. However, it appears to represent some¬
where between 5% to 15% of total trip making. Since trips returning to home
comprise about 45% of all trip making, and trips to work comprise another
20% or so, it is apparent that the shopping trip for consumer durables
represents an important trip in the urban area. Thus, this trip purpose
is a reasonably important one from a planning perspective as well as having
characteristics which tend to be more responsive to the research aims of
this project. It may also be noted that pragmatic concerns of soliciting
attitudes and preferences for shopping locations support a restriction to
trips that would be made more often than not to shopping centers, since these
are more readily identified and are usually known over a wider geographic area
than individual stores or local strip developments, catering primarily to the
perishable commodities trip. For these reasons, the analysis was directed to
the definition and measurement of attractiveness of destinations for non-
grocery shopping trips, and the development of models of destination choice
for this purpose.
4
2. LITERATURE REVIEW
2.1 Background
The foundations for this research are to be found in the series of models
of transport mode choice developed by Warner (1962), Quarmby (1967), Lisco
(1967), Stopher (1969), Watson (1973), inter alia. These models represent the
earliest-empirical statements of a new departure in travel-demand modeling, in
which the models were constructed from the data on individual travelers instead
of zonal averages or totals, and where the models expressed a probability of a
specific mode choice being made by the individual. Subsequently, numerous
statements have been made of the advantages, both realized and potential, of
this approach (for example, Stopher and Lisco, (1970); Brand (1973); inter alia).
The research into this approach has demonstrated that disaggregate, probabilistic
models of mode-choice behavior can both explain and predict travel behavior and
can do this, generally, more accurately and with greater policy-responsiveness
than the more traditional aggregate, deterministic models. Indeed, models of
this type are being used by planners in Chicago, New York State, Los Angeles,
San Francisco, and Stockholm, Sweden, to name a few.
However, the usefulness of the models is limited because the choice of
travel mode is not the only element in the travel-decision process which is
affected by changes in the transport and land-use systems. In fact, all aspects
of the travel-decision process may be affected by such changes. Consider for
example, the imposition of a downtown parking surcharge, aimed at shifting
travelers from automobiles to public transport. Such a measure might equally
well shift travelers to suburban destinations, and may reduce total trip-making
in some instances. The first of these two alternative consequences would be a
change in trip destination. Given the past research into mode choices, and
under the assumption of separability of the travel-decision process (an assump¬
tion made for analytical convenience rather than on the basis of hypothesized
reality), it appears to be appropriate to attempt to extend the disaggregate,
behavioral approach to the destination-choice process. The research presented
in this report is the first step in a program of research aimed at increasing
the usefulness of this modeling approach by making that extension to destination
choice.
Using the separability assumption and assuming a sequential model set in
the pattern of the standard urban-transportation planning process, the initial
hypothesis can be stated in the form of equation (2.1)
P(Dn |D], Dg, . . •» D^) = f(A-|, Fi ; A£, . . ., A^; F£, . . ., F^) (2.1)
where P(D,|D, D ) is the conditional probability of choosing
destination 1 from the set D1, . . ., Dn
A, A are measures of the attractiveness of the
destinations
and F,, . . ., F are measures of the difficulty of reaching the
destinations (i.e. the friction).
The probability is conditional upon the decision to make the trip.
5
The traditional transportation-planning approach would be to consider that
the friction measures relate to the automobile trip, but then follow this model
with a modal-split model. Alternatively, an "average" friction measure might be
used. Neither of these approaches is conceptually satisfying, however. A simul¬
taneous model of mode and destination choice would contain separate friction
factors for all alternative modes for each destination, operating with the
attractiveness measures to produce joint probabilities of choice. This, however,
generates an extremely large set of alternatives, n, and results in a model that
is cumbersome to use and difficult to calibrate as is discussed later in this
chapter.
Regardless of the postulated form of the model in terms of the inclusion
of different travel modes, there has been extensive work on travel friction,
whether couched in terms of distance, time, cost, generalized cost, or gener¬
alized time. Little, however, is known about what constitutes the attractive¬
ness of a destination. The traditional UTP procedure has been to use the
numbers of produced and attracted trips (output from the trip-generation model)
as a measure of attractiveness. This may be satisfactory for a trend analysis,
but leaves much to be desired for forecasting and prediction. Some studies
have used the number of employees in a zone as a measure of attractiveness of
the non-home end of the trip. Again, although the exigencies of research with
little data and the pressures under which most transportation studies are con¬
ducted may demand such a procedure, it is clearly inadequate. Destinations are
highly heterogeneous, even within the same zone, and these definitions of attrac¬
tiveness do not address causality nor the fundamental mechanisms of destination
choice. Thus, this study begins by approaching destination choice within the
framework of the above hypothesis and in the face of considerable ignorance
regarding perceptions of destination attractiveness.
The first specific task was to review the literature on the topic of destin¬
ation choice and associated concerns. We have identified the existence of rele¬
vant work not only in the transportation literature, but also in the economics,
literature, geography, and marketing, inter alia.
2.2 Economic and Psychological Choice Foundations for Predicting Travel Demand
The past ten years have shown an intensive effort by transportation-demand
analysts to clarify and put onto a rigorous basis the theoretical foundations of
the models they use. This subsection sketches briefly the relevant work that
has been done using economic theory and psychological choice theory to derive
statistical models for the analysis of travel demand. This area has recently
been surveyed in depth by two papers: "Travel Demand Forecasting: Some Founda¬
tions and a Review," by Daniel Brand, presented at the Williamsburg H.R.B./
U.S.D.O.T. Conference on Travel Demand Forecasting, December 1972, focusing on
the economic and psychological foundations of travel demand; and "Quantal Choice
Analysis: A Survey," by Daniel McFadden, presented to the NSF-NBER Conference
on Decision Rules and Uncertainty, March 1974, focusing on extension of psycho¬
logical choice models and the statistical problems involved in their estimation.
Given the extensive literature surveys contained in these papers and other docu¬
ments such as Domencich and McFadden (1975), Stopher and Meyburg (1975), etc.,
this chapter will contain only an introduction to this important field as it
relates to the work undertaken by this project; readers interested in a treat¬
ment of more depth are referred to the documents cited above.
6
Travelers do make choices between transportation modes and destinations;
moreover, different travelers who face similar choices do not necessarily make
the same choice. This much is obvious; the difficulty arises when we must
reconcile these facts with the requirement that some regularity in actions
across travelers must exist if we are to have any confidence in our empirically
determined models of choice behavior. Two reconciliations have been proposed.
The first theory postulates that travelers make choices randomly but that times,
costs, and attractiveness of destinations do influence the probabilities that
they will make one choice or another. The second reconciliation, on the other
hand, postulates that travelers do make deterministic choices based on the attri¬
butes of the alternatives, but that observed differences in choices between
travelers facing similar choices must be assigned to factors unobservable to an
outside investigator. The difference in the models is therefore in the interpre¬
tation of the selection probabilities of alternatives; in the first model, prob¬
abilities are intrinsic to travelers who decide their choices randomly, while in
the second model, probabilities must be interpreted in a relative frequency sense
across individuals, and the notion of random choice has no meaning for individual
behavior. In both these models, the dependence or selection probabilities on the
characteristics of alternatives and of travelers is summarized by postulating a
utility function of the form of equation (2.2)
U = V(s,x) + e(s,x), (2.2)
where V is nonstochastic and represents the representative tastes of the indivi¬
dual as they depend on s, the personal characteristics of the traveler, and x,
the characteristics of the alternatives being considered, and e is stochastic
and represents the idiosyncratic tastes of the individual. Using this function,
the first model discussed above can now be termed the strict utility model, as
the s component of the utility function is assumed degenerate and choices are
made randomly; while the second model discussed above can be termed the random
utility model as the e component of utility is the result of a drawing from a
random distribution across individuals, but individual's choice behavior is
determined uniquely as that choice which maximizes utility once the e component
is given.
Both the models given above can be used to generate explicit functional
relationships between s and x and the selection probabilities of alternatives.
Although there is an endless number of alternative representations that can be
generated, the most important statistically is the multinomial logit (conditional
logit) model because of its attractive mathematical representation. If in the
psychological derivation the V(s,x) component of the utility function is assumed
to be written as an exponential function of a linear combination of the variables
x and s, as shown in equation (2.3)
V(s,x) = exp(sa + xji), (2.3)
where a and l are vectors of unobservable constants, then the multinomial logit
model can be derived by invoking Luce's Axiom of Independence of Irrelevant
Alternatives (Stopher & Meyburg, 1974), to give equation (2.4)
exp{sia|< + Xjj3}
P.. = J (2.4)
J E exp{s.a. + x. ,
k=l 1 K K_
7
where P.. is the selection probability of the 1th individual for the jth alter-
' J
native and J is the number of alternatives. On the other hand, if we accept the
random utility model, the multinomial logit form results if we assume the e(s,x)
component of utility is distributed as a stochastically independent Weibull dis¬
tribution across alternatives. McFadden (1974) has shown that this condition on
the e(s,x) is both necessary and sufficient to yield the multinomial logit model.
By now, the strong implications of the Axiom of Indepdendence of Irrelevant
Alternatives (and correspondingly, the assumption of stochastic independence of
e across alternatives for the random utility model) are well documented, e.g.
Brand (1973); CRA (1976). Since in our research, the attitudes of travelers
toward shopping centers in a confined geographic area are quite likely not to
be stochastically independent, the use of a choice model that incorporates such
a strong assumption must be questioned. On the other hand, the cost of giving
up this assumption is also very high and possibly prohibitive. As Brand empha¬
sizes in his paper, the independence assumption implies a separabi1ity property
on the selection probabilities, in that the relative odds of two alternatives
are assumed to be determined independently of what other alternatives are avail¬
able. In our research, we have explicitly let travelers specify their own
destination choice sets rather than determine these separately and impose them
on the analysis. In particular, travelers were provided with an exhaustive list
of shopping centers in the same urban area, and they were asked which other
centers they would have gone to if the center at which they obtained the ques¬
tionnaire was not available. In addition, travelers were also queried about
their familiarity and the number of times they have visited each shopping center
in the exhaustive list so we can investigate the determinants of individual
choice sets. With respect to the estimation of our choice models, however,
since each traveler specifies his own choice set, we will have a great variety
of choice sets with only limited overlapping between individuals. If we pursue
models that do. not assume independence of the e terms and therefore do not have
the separability property, we must specify an interaction principle that can
handle many different choice sets. At the present time, a tractable model for
handling this problem does not exist.
Because travel is a choice problem, McFadden has emphasized that the rela¬
tive odds of choosing different alternatives should depend on differences in the
attributes of the alternatives and not on the characteristics of the traveler.
This criticism defines a major difference in McFadden's "conditional logit" model,
where the coefficients of the model are defined without reference to specific
alternatives (the 3. coefficients of our equation 2.3) while the independent vari¬
ables are measured specific to the alternatives (the Xj variables in equation 2-3),
as this model is contrasted to the "multinomial logit" model of Nerlove and Press
(1973), where model coefficients vary across alternatives (the .& coefficients of
equation 2.4) while independent variables are measured specific to the traveler and
not the alternatives (the s. variables of equation 2.4). In this sense, the multi¬
nomial logit model of Nerlove and Press is a generalization of multigroup dis¬
criminant analysis, while McFadden's model emphasizes the choice basis of travel
decision. We have formulated equation 2
since our destination-choice model will involve a large number of unranked alter¬
natives, alternative-specific coefficients are difficult to interpret and esti¬
mate and will generally not be included.
8
Although the multinomial logit specification can be derived from either the
strict utility or the random utility model, the interpretation of the selection
probabilities depends on which model is used. These models are indistinguishable
in cross-sectional analysis however, and they can be distinguished in time-series
analysis only under strong assumptions. For example, if we regard the choice
process in the strict utility model as stochastically independent over time, we
should observe individual traveler's choosing alternatives with relative fre¬
quencies given by the selection probabilities; on the other hand, if we regard
the stochastic portion of the utility function as fixed for individuals over
time for the random utility model, we would observe that no traveler ever changes
his original choice. These predictions rely heavily on the particular assumptions
of stochastic independence for one model and a fixed distribution for individuals
for the other model, however; and if we allow more flexible time-dependent sto¬
chastic processes for individuals, either model can give the same time-series
predictions. In this sense, therefore, we must consider the strict and random
utility hypotheses as indistinguishable representations of the marginal choice
probabilities.
The statistical properties of the multinomial logit model have been devel¬
oped by both McFadden and Nerlove and Press, and we will not survey these
results here.
Turning to the problem of destination choice, the multinomial logit model
has been applied to this problem by Domencich and McFadden (1975) and by Ben-
Akiva (1973). The Domencich and McFadden study included analysis of four trans¬
portation choices: mode, time of day, destination, and trip frequency. We will
direct our review to the interrelations between the mode and destination-choice
decisions.
Because the number of possible combinations of mode and destination choices
can become very large if we consider the choices simultaneously, Domencich and
McFadden factored the demand model into its component decisions. In particular,
they assume that for a given trip purpose, the marginal rates of substitution
between modal attributes in determining choice behavior are the same regardless
of the time of day of the trip, the trip destination, or the daily trip frequency.
Similarly, they assume that the marginal rates of substitution between the attri¬
butes of destinations are independent of mode, time of trip, and trip frequency
decisions. These assumptions allow the authors to 1) estimate separate models
for mode and destination choices, and 2) to summarize the effects of transporta¬
tion characteristics of the other decisions by the use of a single measure called
"the inclusive price of travel" that is an output of the mode-choice model. As
an example, we consider a destination-choice model where the traveler's decision
would depend on the time and cost of getting to each destination as well as the
attributes of the destinations themselves. Entering the time and cost of reach¬
ing each destination by each mode would greatly expand the number of explanatory
variables used, however, and would ignore the assumptions on marginal rates of
substitution giyen earlier. Domencich and McFadden show that under these assump¬
tions it is valid to define the inclusive price of travel to a given destination
for a given mode for a given purpose as equation (2.5)
9
Ôij = I Vijk' (2.5)
where c.. is the inclusive price for a trip to destination j by mode i
^ J
th
g. is the parameter estimate from the mode-choice model for the k
attribute of the mode, and
x... is the value of the k^ attribute of mode i for a trip to
1J destination j.
Now, for each destination, we define equation (2.6)
c. EÊ-j P,J (2.6)
th
where c. is the inclusive price of travel to the j destination,
J
c.. is given above for each mode, and
' J
P. • is the predicted probability that the traveler will choose mode
J i to destination j obtained from the mode choice model.
Under Domencich and McFadden's assumptions, this single inclusive price completely
summarizes the effects of transportation characteristics on the destination-choice
decision and is the only explanatory variable representing the transportation modes
that need be entered into the destination-choice model. This result certainly
allows a great reduction in the complexity of the separate models, and Domencich
and McFadden argue that the required assumptions are not unreasonably severe on
both intuitive and empirical grounds. Nevertheless, this aspect of their study
requires further investigation before it can be generally accepted.
Ben-Akiva's study basically criticized Domencich and McFadden's factoring
of the demand model. Arguing that a simultaneous mode and destination-choice
model was more attractive intuitively than the recursive structure, he estimates
and compares models of each type with some alternative definitions of inclusive
prices. His major result is that empirical results are sensitive to the model
specification and he advocates the use of the simultaneous choice model. His
findings, although indicative, cannot be considered conclusive, however, without
further work on broader theoretical models linking the choices and on the empir¬
ical costs of the various required assumptions.
The work of Domencich and McFadden and Ben-Akiva has been very useful in
defining the theoretical issues involved in applying disaggregate choice models
to destination choices, but the empirical work on destination choice in both
studies must be deemed deficient. Both studies focus on the shopping trip, but
neither has data on individual perceptions of the characteristics of destinations
and individual perceived choice sets. Both studies define individual choice sets
in an ad hoc fashion by dividing the urban area into zones; for travelers living
in a given zone, destinations in another zone were regarded as being in their
choice set if at least one traveler from the origin zone had gone to that desti¬
nation zone in their sampled trips. On the characteristics that determine the
10
i
attractiveness of destinations, the only measure specific to the destination
that either study used (besides a generalized price of travel variable) was an
index of employment in the destination zone (Ben-Akiva also used a dummy vari¬
able for the CBD). The question of attractiveness of destinations is examined
extensively in our research, and this should produce much more satisfactory
models.
In addition, Ben-Akiva's study was the only one that attempted to study
empirically interrelated mode and destination choices; Domencich and McFadden
used only auto-made trips in their destination-choice model. Finally, since
both studies used small samples (63 for Domencich and McFadden, 123 for Ben-
Akiva), it is fair to say that neither study has explored adequately the empir¬
ical structure of destination-choice models.
The final topic we treat in this part of the literature survey is the problem
of using the estimated models for predictions for travel demand. This problem has
been studied for the logit model by Talvitie (1973), Westin (1974), and Koppelman
(1974) and results obtained by Watson and Westin (1973) and Liou, Cohen, and
Hartgen (1974) indicate that disaggregate models can give very good predictions,
at least for aggregate modal splits, with data requirements far below those
necessary for current aggregate models. Of the studies cited, only Talvitie and
Koppelman consider the problems of predictions from a multinomial logit model,
Westin's study is restricted to the binary logit case. The Taylor series approx¬
imation used by Talvitie can be shown to be non-robust, however, and Westin's
method can be extended to the multinomial case easily.
In summary, the theoretical development of disaggregate choice models has
now defined the theoretical restrictions on these models and indicated the prob¬
lems that must be solved before they become a practical planning tool. The indi¬
cations from modal-choice applications are that these models are potentially very
valuable in their precision, flexibility, and reduced data requirements. Work on
the extension of these models to destination choice, however, has proceeded only
to very limited empirical applications. The extensive information on destination
choice collected in this research should go far toward defining these models as
practical planning tools.
2.3 Destination Choice Work in Marketing
Marketing literature has dealt with the consumer-patronage decision on three
levels of retail aggregation: the city, the intraurban shopping center, and the
individual store. Pioneering retail patronage studies such as those conducted by
Reilly (1929) concerned themselves with the consumer choice among cities. More
recently, attention has been focused on the choices among shopping centers and
among individual stores. The intent of this subsection is threefold: first, to
delineate some of the factors which authors in the fields of marketing and geo¬
graphy note as having an influence on consumer choice of shopping centers and
stores; second, to outline some of the difficulties involved in attempting to
specify the appropriate forms of these factors for use in predictive models; and
third, to present some of the models which have been used to forecast consumer
patronage decisions.
11
2.3.1 Retail Location Choice Factors
Factors specified as influencing the consumer-patronage process fall into
one of two categories: characteristics of the retail location or characteris¬
tics of the consumer. The former classification consists of features attribut¬
able to either shopping centers or stores which influence the consumer patronage
choice. The literature suggests that many of these measures are common to both
levels of retail aggregation. It is the interaction between the characteristics
of the retail location and of the consumer which leads to the ultimate patronage
decision.
Retai1-Location Characteristics. Factors included in the category of retail-
location characteristics can be conveniently classified according to five of the
dimensions of image which the Wharton Studies found to be relevant to consumer
patronage. These dimensions are locational convenience, merchandise suitability,
value for price, sales efforts and store services, and congeniality of retail
location. (Fisk, 1961-62).
Characteristics of the retail location which comprise the dimension of loca¬
tional convenience include the factors of accessibility, parking availability and
spatial parameters such as distance. Accessibility can be measured in terms of
road networks and modes of transportation servicing retail locations. Little has
been done to determine the impact of these measures on the consumer-patronage
decision.
Providing parking is an expensive undertaking. Still, most retail locations
make sure of having an ample amount of it which suggests that parking is thought
to have a signficant influence on retail destination choice. There are guidelines
which recommend the number of parking spaces to allocate for different types of
shopping areas, Kelley (1956). However, thus far few empirical studies have been
undertaken to test the hypothesis that this factor does have a significant effect
on retail-location choice.
The importance of spatial parameters has been particularly noted. Yuill
(1967) for instance, empirically verified the importance and usefulness of a
distance or spatial parameter in approximating the resultant behavior of con¬
sumers. Articles specifically related to shopping center and store choice
provided support to the hypothesis that spatial parameters bear a significant
influence on the patronage decision. Employing the technique of discriminant
analysis, Bucklin (1967) discovered distance factors to be dominant in deter¬
mining the choice of shopping centers. In a survey of marketing literature on
consumer-patronage behavior, Forbes (1968) found distance to be one of the vari¬
ables exerting a significant influence on store choice. In addition, he noted
that it is the only influencing factor which research has quantified to any
considerable degree.
The hypothesized distance effect is that consumers farther away from a
retail location are less likely to patronize it than are consumers situated
closer to it. Some attempts have been made to pinpoint the exact distance
effect. Appelbaum (1966) reported empirical distance effects on store patron¬
age in supermarkets. He compared the percentage of customers to the distance
from the store and found an inverse relationship to hold.
12
Contrary to the Nearest Center Hypothesis of Central Place Theory, distance
is not the only factor influencing patronage choice. The hypothesis says that
consumers purchase goods and services at the nearest center offering them. Clark
(1968) in a study concerned with the orderliness of spatial behavior found that
less than half of the sample of consumers he examined went to the nearest center.
The fact is that in many instances consumers do travel farther than is essential
to purchase goods and services.
There is some debate as to whether or not pure distance is the spatial
factor which most influences the patronage decision process. Brunner and Mason
(1968) maintained that driving time rather than distance is the most significant
spatial factor since the effort required to reach a shopping center is inversely
associated with the driving time to reach the center. Aside from disagreement
on their appropriate form, there is a general consensus that some spatial factors
significantly influence the consumer patronage decision.
Factors related to the goods and services offered at retail locations such
as quality, type, and variety comprise the dimension of merchandise suitability.
The literature has especially considered the effect of variety on the consumer-
patronage choice. The effect appears to be significantly positive: as width
and depth of product assortment increase so does the level of patronage.
The variety associated with a retail outlet is not necessarily limited to
the assortment of goods and services it offers. The outlet's variety may also
encompass the goods selection of enterprises located in close proximity since
even in cases when ownership of retailing units is diverse there is drawing
power in product assortment (Forbes, 1968). Thus, the variety attributed to a
store located in a shopping center is effectively greater than that associated
with an isolated, identical store.
The importance of variety in the retail location choice is related to shop¬
ping trip purpose. If the trip purpose is to purchase either a high-order good
or a large number of commodities, the significance of the factor increases.
Consumers spend more time, and travel further, in searching for high-order goods
which are characterized by high price, considerable service needs, durability,
and low replacement rates (Aspinwall, 1962). Given that consumers seek to mini¬
mize overall costs in the selection of a location at which to purchase high-order
goods, variety will be a primary criterion, since wider selection allows greater
comparison and thereby reduces search costs. Similarly, variety is important in
shopping for many commodities at one time since it reduces travel costs by enabl¬
ing the combination of trips in a more compact area.
Even in the selection of a site for the purchase of low-order convenience
goods such as food and drugs, variety has an effect. Skinner (1969) found in
his analysis of supermarket choice that nearness to other services is one of the
most important determinants. Using a model which explained the variation in
distance traveled for intra-urban grocery purchases, Bishop and Brown (1969) per¬
ceived a statistically significant positive association between distance travel¬
led for grocery purchases and the number of services available in close proximity
to supermarkets.
The value for price dimension consists of various price-related factors
which influence retail destination choice. Absolute prices and the difference
13
in price at various retail locations for the same product or substitute products
are examples of pertinent price considerations. That price factors significantly
influence patronage decisions was verified in the study by Bucklin (1967). The
importance of price has been noted, particularly with regard to the selection of
a specific store. Thompson (1969) presented evidence that price is one of the
primary determinants in the supermarket choice process. On the basis of his
findings, Brown (1968) concluded that price consideration is the second most
important patronage motive in supermarket selection.
The sales efforts and store services dimension is made up of factors such
as advertising, courtesy and helpfulness of clerks, credit arrangements, deli¬
very, and eating facilities. Advertising has been shown to have a statistically
significant influence on consumer retail location decisions (Bucklin, 1967).
The other factors have not been extensively tested so their importance is unde-
termined.
The dimension shown as congeniality of retail location consists of compo¬
nents such as size, layout, decor, class of customers, traffic within the loca¬
tion, and congestion. Of these factors only size has been treated in the liter¬
ature to any considerable degree. This factor is closely related to the notion
of variety since scale economies necessitating large units for success are
limited unless innovations in product offerings occur (Forbes, 1968). The
influence of size on retail location choice is similar to that of variety as
suggested by the results of a study of Cox and Cook (1970). In a regression
analysis, they found size of shopping-center store area to have a significantly
positive effect on the percentage of customers visiting various centers.
Consumer Characteristics. Various consumer characteristics play a role in
the ultimate retail destination choice. It appears that their influence on the
decision stems from the manner in which they affect the significance consumers
attach to the aforementioned destination characteristics. Demographic measures
such as income, age, and education fall into the category of consumer character¬
istics which influence retail location decisions. They appear to affect strongly
the importance consumers attach to the retail location characteristics of dis¬
tance and type of merchandise.
The distance effect diminishes for those in certain demographic groupings.
An analysis of shoppers by Herrmann and Beik (1968) concluded that the range of
shopping movement within a metropolitan area increases with consumer's income
and status. The psychographic results of Reynolds and Darden (1972) also sug¬
gested that shoppers who travel greater distances for purchases tend to be better
educated and have a higher income level. That demographic factors influence the
type of merchandise and hence the type of store chosen by a consumer is inferred
by Simmons (1964) who postulated that higher income people are willing to pay for
the increased services of specialty stores.
Although it is generally agreed that demographic factors influence consumer
patronage decisions, these variables apparently are not the only consumer char¬
acteristics having an effect on the choice. The results of the regression anal¬
ysis of Mason and Moore (1970-71) indicated that households comparable in socio¬
economic characteristics such as income, occupation and education do not neces¬
sarily display comparable shopping travel behavior. These writers suggested
that additional consumer characteristics which can account for more of the dis¬
crepancy may fall into the domain of psychological or attitudinal differences.
14
Huff (1960) has indicated that psychological characteristics may indeed
influence the retail location choice. He presented a model designating the
elements affecting consumer space preference. Using the techniques of linear
graph theory and matrix algebra, he found that elements exerting the most influ¬
ence in affecting space preference include demographic factors (e.g., age, sex,
education, occupation and income) and psychological characteristics such as
personality and mental synthesizing abilities.
In summary, the ultimate patronage decision is influenced by a combination
of characteristics of both retail locations and consumers. The intent of this
section has been to outline some of these characteristics. It should be noted
that these factors do not have the same influence in all cases. The weight of
each element's effect varies among individuals and situations.
2.3.2 Specification Problems
Although many retail location and consumer characteristics which influence
retail destination choice have been identified in marketing and geographic lit¬
erature, appropriate forms for their use in predictive models have not been
determined. Marble and Bowl by (1968) stated that, to use a model which predicts
travel behavior, those attributes of establishments which travelers consider to
be of importance in the selection of a destination must be specified and scaled.
Hence, this matter of specification is one of the primary tasks researchers must
perform if they wish to approximate consumer shopping behavior better. The
endeavor appears to be formidable, however, since objective measurements of
location characteristics do not appear to be sufficient and it is unclear on
what basis consumers should be aggregated for the purposes of retail study.
Objective measurements of easily quantified location characteristics such
as distance, driving time and price are not entirely appropriate inputs for
models designed to describe patterns of consumer patronage behavior. The find¬
ings of Thompson (1963) suggested that consumers tend to overestimate driving
time and distance travelled to retail outlets. The approximations are also
usually influenced by one's subjective feelings about the outlets. Brown (1969)
found that consumers' perceptions of price do not necessarily coincide with
actual price levels. Rather than attempt to observe actual price levels, con¬
sumers tend to estimate prices on the basis of certain indicators. For instance,
they generally feel that large-volume operations have lower prices and that
operations offering additional services (e.g. extra sales clerks) have higher
prices. Thus, it is not the objective measurement of driving time, distance and
price level which is the appropriate specification of these characteristics in
the consumers' view.
There is some question as to what criterion should be used in the grouping
of consumers for the purposes of retail analysis. That large groupings are
undesirable is suggested by the fact that the use of aggregate data averages in
trading-area studies does not reveal possible internal dissimilarities within
an area that may be of signficance in terms of marketing policy and investment
decisions (Mason and Moore, 1970-71). The inadequacy of aggregating on a demo¬
graphic basis alone has already been noted.
15
The results of a factor-analytic study of women's apparel shoppers by
Phi 1 pot, Reizenstein and Sweeney (1972) indicated that the determinants of a
retail location choice are somewhat unique to stable groups of consumers. The
task remains to determine what characterizes these stable groups of consumers.
It appears that significant progress can be made in defining determinants of
retail location choice if the analyst focuses on the interrelationships among
a variety of different types of retail attributes and consumer characteristics
variables, using analytical tools capable of reflecting the multiple dimensions
of the consumer patronage decision process.
2.3.3 Predictive Models
Various models have been used to predict consumer patronage choice. Models
based on the conceptual properties of the retail gravity model have been employ¬
ed in the estimation of both shopping center and store patronage, while certain
other formulations have been utilized specifically to predict store choice. It
is the intent of this section to outline some of these models.
Originally the retail gravity model as developed by Reilly (1929) was used
to predict the point between two cities where trade between them would be divi¬
ded. More recently the form of the model has been modified to predict the
probability that a customer would patronize one of two or more retail areas
given various factors such as distance to and size of the areas. Bucklin (1967)
provided a theoretical basis for the form of the retail gravity model currently
employed for the purpose of predicting patronage behavior in his general model
for the retail trade areas. This latter model consists of three factors: the
shopping utility to the potential patron of the retail facility in question;
the cost to that person of reaching the facility and the strength of competing
retail centers. Mathematically, these elements are combined as shown in equa¬
tion (2.7)
PfHAl*,) - r4- (2.7)
E U./C.
where P(HA|)
= probability that
= shopping utility
= cost of reaching
consumer X-| will
of facility j
facility j.
choose facility A
The numerator of the above probability statement represents the drawing-power
of retail area a, while the denominator stands for the sum of the drawing powers
of all g centers under consideration.
In Bucklin's formulation (1967), the shopping utility of a retail area is
derived from two sources: mass and image. The mass component which defines the
range of goods and services available at the location provides utility by reduc¬
ing the time necessary for individual transactions. The image component depends
upon consumer perception of factors such as the area's price level, physical
plant and social-class orientation. The cost element of the model is contingent
upon consumer expenditure in time, money, and the effort needed to reach the
area.
16
As pointed out earlier in this chapter, the utility and cost that a consumer
associates with a retail area are each based on a multitude of factors. However,
for the purposes of prediction via the gravity model, a smaller number of rele¬
vant variables has been utilized as proxies for these two concepts. The attrac¬
tiveness of centers and distance from the consumers to the respective areas are
sometimes used as proxies for utility and cost, respectively. Algebraically
this model is shown in equation (2.8).
Vik
Pik = n >k,j--l n (2.8)
z (A/D*)
i=l J 1J
J
where Pik = the probability that a consumer located at point i will visit
a retail area at point k, given a set of n competing markets
of which k is a member.
A. = the attractiveness of j,
J
D.j = the distance of j from i.
A = a measure of the strength of the distance variable (Buck!in, 1971).
Other researchers have been even more specific than those using the above
formulation in their choice of variables to be used in the retail gravity models.
Empirical evidence of Carrothers (1956) suggested that two variables exert such
an influence on a consumer's choice of a retail area that they may be the only
variables needed to predict behavior using the model. These variables are:
(1) the number of items of the kind a consumer desires that are carried by var¬
ious retail areas and (2) the travel time involved in getting from a consumer's
travel base to alternative retail areas. Huff (1963) approximated the number
of items desired by using the square footage of selling space devoted to the
sale of such items. He also measured actual travelling time. His model is
shown in equation (2.9).
Sk/TikX
Pik - n \ (2'9>
jî/W
where P-k = probability of a consumer at a given point of origin i
1 travelling to a given shopping area k.
S. = square footage of selling space devoted to the sale of
J a particular class of goods by shopping area j.
T.. = travel time involved in getting from a consumer's travel
1J base i to shopping area j; and
A = a parameter which reflects the effect of travel time on
various kinds of shopping trips.
Although the retail gravity model has achieved a certain degree of predic¬
tive success, from both a theoretical and empirical perspective the model has
major defects. Its conceptual foundations are not well conceived because they
make implicit assumptions about what constitutes consumers' opportunity costs
for travel. In particular, the comparison of distances to alternative centers
17
versus respective utilities in the form of the model which utilizes distance as
a proxy for cost is insufficient evidence upon which to estimate consumer's des¬
tination choice since time is a major element in the cost of a shopping decision
(Bucklin, 1971).
Examples of behavioral models of store choice are the formulations of
Baumol and I de (1956), Aaker and Jones (1971) and MacKay (1972). The first
is a decision model; the second is a formulation which portrays store choice
as an extension of grand choice; and the third is a spatially-defined model of
store selection that accounts for multistop shopping trips.
Baumol and I de (1956) presented a linear model which portrayed the decision
to shop or not to shop at a given retail outlet as resulting from a weighing of
the probability that the consumer finds some set of items in the store which
makes his trip successful against the costs of shopping. They maintained that
a consumer will not shop at a particular store unless equation (2.10) is satis¬
fied.
f(N,D) = wp(n) - v(CdD + Cn N + Ci) > 0 (2.10)
where f(N,D) = the net benefit from entering the store, N is the number
of items offered for sale,
D = the consumer's distance from the store,
p(n) = the probability of successfully finding the desired item
at the store,
w and v = subjective weights assigned by the consumer,
Cd = a constant,
Cn = the cost of getting to where desired items are kept,
C.j = fixed shopping costs.
There are certain economic implications of the Baumol-Ide model. Increased
variety is an advantage to a consumer only up to a point. The minimum number of
items necessary to induce a consumer to shop at a store increases with his dis¬
tance from the store. Also, the optimum variety from the viewpoint of the con¬
sumer is independent of the distance from the retailer.
The linear learning model has been employed by Aaker and Jones (1971) for
estimating the probability of choosing a particular type of store for a parti¬
cular purchase at time n given previous probabilities and information concern¬
ing last period's store choice. The specific form of their model is shown in
equation (2.11).
p(t) = a + BD + Xp(t-l) (2.11)
where p(t) = probability of choosing a particular type of store at time t
p(t-l) = probability of choosing a particular type of store at timet-1
D = dummy variable with value 1 if the type of store being con¬
sidered was chosen in period t-1, value 0 otherwise
18
The model fitted relatively well when applied to data obtained from the
Chicago Tribune on the purchase history of paper products, toothpaste and coffee
although it fitted better with respect to a particular chain rather than type of
store.
MacKay (1972) criticized the two previous store-choice models for failing
to take account of multistop shopping trips. He allowed for these shopping
trips in spatially defining a model by means of discriminant analysis and Monte
Carlo simulation. The framework for his formulation portrayed consumers as
going through a three-stage, sequential process of store selection: (1) the
decision of whether or not to go shopping, (2) the decision of how many stops
to make, and (3) the decision of which establishment to visit on each stop.
Good predictive results were acquired with this model which suggests that a
sequential consumer-decision process may be more common than a holistic decision
process.
2.4 Destination Choice Work in Geography
Most of the research in geography has concentrated on shopping-center
choices, where the trips are primarily concerned with non-grocery shopping and
are therefore of considerable relevance to this research, the earliest work of
this type, studied in this project, was that of Burnett (1973), who attempted
to find out the parameters of shopping-center choices through multidimensional
scaling (MDS). Burnett hypothesized that length of residence would affect the
choice parameters. Consequently, she split the sample of housewives into two
groups — short and long residence periods. The MDS analysis was then carried
out separately for each group. In each case, dimensionality could be reduced
to two, but one of the dimensions for each group was different and scores, rela¬
tive to the common axis were also different. For short-term residents, the axes
were labelled "value for time, effort, and money" and "similarity to CBD shop¬
ping;" while for long-term residents the axes were labelled "value for time,
effort, and money" and "ease and convenience of parking." It was concluded that
the differences in the MDS results confirmed the hypothesis that a learning pro¬
cess is involved in shopping-center choices, which is temporally related to
length of residence.
Pursuing this idea further, Burnett (1974a) then investigated the effect
of learning on the spatial distribution of user origins for a specific destin¬
ation. This was investigated by examining the patronage of a new Savings and
Loan institution over a period of 5 years. She hypothesized that the distribu¬
tion of user origins should be circular normal* and should grow, in terms of
numbers and distance as an inverse exponential. Based on the data, Burnett
found that the ciruclar-normal distribution hypothesis was confirmed for each
of the five years. However, she also found that growth was continuing at a rate
which was not consonant with an asymptotic growth pattern after five years, thus
concluding that the 5 year time period may be too short to establish the tena-
bility of this hypothesis.
* A circular-normal probability distribution is a 3-dimensional normal
distribution on a circular plane, see Figure 1, Burnett (1974a).
19
Following this work, Burnett (1974a) attempted to construct a linear learn¬
ing model based upon a history of store visits to a grocery store. The model
was based upon two postulates: that the probability of choosing a store in¬
creases as time goes by; and that the probability of a store choice increases
more rapidly if the store is visited. Burnett obtained a satisfactory fit of
the data to this model. However, it should be noted that the model becomes
extremely complex when the number of visits exceeds three, and that the records
on each individual were used in blocks of three visits as comprising separate
choice histories. These limitations raise some question about the usefulness
of these results. Building on this work, Burnett then hypothesized (Burnett
1974b) a three-states-of-1earning model, from which a Markov model could be
constructed. The three states are based on an assumption that there are two
units of learning. Initially an individual is in state . From here he may
gain one unit of learning, or_may remain in the initial state. When he gains
that unit, he moves to state S^. One further unit of learning may then be gain¬
ed, which will place the individual in the equilibrium state, S. A transition
matrix can then be defined, in which the probabilities represent the probability
i. t_
of being in any specified state at the end of the i time period. Probabilities
of the choice of a specific destination can then be defined as a function of the
learning state of the individual, using the model form put forward by Burnett
(1973).
The most recent research of Burnett (1974c) involved the testing of a null
hypothesis concerning grocery store choices. The null hypothesis from the trans¬
portation planner's point-of-view, is that choices are totally random and are,
therefore, not susceptible to modelling. Using data from Uppsala, Sweden,
Burnett shows that grocery store choices are not random in the majority of cases.
She does find, however, that for three subgroups of the population the null
hypothesis of random choices cannot be rejected.
The final piece of relevant work in this area is contained in a paper by
Hanson (1974). In this paper, Hanson puts forward a hypothetical model of
destination choice, which is developed in part from the findings of Burnett.
However, Hanson postulates a model in which learning is introduced directly into
the model, rather than using the two stage model process of Burnett (1974c).
This approach requires the definition and evaluation of a "level of information"
variable which is then assumed to be a multiplier of the satisfactions of a
shopper with each destination. Hence, the satisfactions are modified, or weight¬
ed, by the level of learning the individual possesses for each destination con¬
sidered. Hanson then discusses the data needs for calibrating such a model, but
no calibration is reported.
2.5 Hypotheses to be Tested
Based upon the literature search and the general considerations of model
structure and type, a set of specific hypotheses can be developed that form the
backbone of this research project.
From the initial theoretical and empirical research that was done on the
project, and examined in the literature review, two candidate model structures
were considered in broad terms. The first of these is a logit model (See McFad-
den, 1972; Berkson, 1943; Stopher and Meyburg, 1974), which is shown in equation
(2.12).
20
,1t. .wlvl'
J
I
(2.12)
where P1^ = the probability of individual i choosing destination j
J for a specific activity, t.
A1^ = the attractiveness of destination j to individual i for
J the activity t
C1. = the cost for individual i to travel to destination j
J
A^'Ck = the attractiveness and cost of the other destinations
available to individual i for activity t.
The second candidate model structure is that of a Markov model, (See Peterson,
et al. 1975). In particular, a version of the Markov model in which the trans¬
ition matrix of probabilities are exogenously defined through probabilistic
models was considered relevant to examine as a potential model structure for
this work. These model structures are discussed in greater detail elsewhere in
this report.
Regardless of model structure, it is necessary to define a choice set, from
which it is assumed an individual chooses his destination. The definition of
choice set is an important element in both the logit model and the Markov model,
and will influence significantly the calibration of such models. Two hypotheses
need to be set up here. First, it is necessary to define the way in which a des¬
tination is to be characterized. Second, it is necessary then to define the set
of shopping destinations from which the choice is to be made. After considering
a number of alternative hypotheses, it was decided to hypothesize that a choice
of destination would focus upon a generic type of shopping destination, rather
than upon a geographical location. This, unfortunately, means that the destin¬
ation-choice model will have to become a two-stage model, if the desired output
is the geographical or spatial location of trip ends. The generic types that
were considered as the basis of the choice process are determinants of the size
and function of the shopping center. Thus, the following might comprise a
choice set: a major regional shopping center, a medium-sized local shopping
Center, a small neighborhood shopping center, an urban or suburban downtown area,
a strip development, and a neighborhood shopping area. These definitions are by
no means exhaustive nor necessarily the ideal ones for a choice set, but serve
to illustrate the type of choice set hypothesized.
Next, it is hypothesized that each individual chooses a shopping location
from among this complete set. Having made his choice, the individual will then
choose that specific location that is the least costly to access within the pre¬
scribed choice set. It is realized that this description of the choice process
is somewhat unrealistic, but it is found to be necessary for the purposes of
this study. The requirements for the development of a geographically specific
model appeared to be too great to be surmountable in a prototypical study such
as this.
21
As discussed in the preceding section, the next hypotheses relate to the
attractiveness indices. Three hypotheses are to be tested in this respect.
First, it is hypothesized that measures of attractiveness can be obtained by
means of psychometric techniques, i.e. scaling methods. Second, it is hypothe¬
sized that these attractiveness measures can be formulated in terms of attributes
of shopping centers, shopping locations, and the stores within them alone, and
may be defined separately from ideas of accessibility. Third, it is hypothe¬
sized that these attractiveness indices can be input into a decision model of
destination choice, together with measures of accessibility, and jointly used
to predict the choices that will be made by individuals engaged in non-grocery
shopping trips.
The above paragraph has alluded to ideas of accessibility. It is therefore
necessary to include within the model some measures of the utility of travel.
The next hypotheses relate to this utility. It is hypothesized that the destin¬
ation-choice decision is based upon the same relative utilities of attributes
of travel as would be found in a shopping-trip mode-choice study. It is hypoth¬
esized that these utilities will be constructed from attributes of travel, such
as travel costs, travel times, convenience, and ability to carry packages.
In constructing models of mode and destination choice, it is hypothesized
that individuals will have varying utilities of both accessibility elements and
attractiveness elements. As such, it is hypothesized that the population may be
split into market segments, within which choice processes are relatively homo¬
geneous. It is further hypothesized that such market segments can be defined in
terms of the social and economic status of the individuals within them. Thus, a
further task of the research is to develop appropriate subgroups, based on socio¬
economic criteria, within which the destination-choice decision can be considered
to take place in a relatively homogeneous fashion.
Finally, all of this research presupposes that a clear link exists between
attitudes and behavior. The scaling techniques referred to earlier are based
upon measures of people's attitudes and preferences. These attitudes and pref¬
erences relate to specific detailed attributes of shopping locations and stores.
In order to construct models of the type described in this chapter, it is neces¬
sary to assume that attitudes and preferences are, in part, determinants of
behavior. It is a well-known fact that attitudes and preferences are formed,
or modified, by experience and learning. Clearly, experience and learning come
from the exercise of specific choices. As such, the link between behavior and
attitudes may be somewhat weak. However, it is hypothesized here that the cog¬
nitive link, i.e. the link between behavior and attitudes, is strong enough to
permit the construction of meaningful and plausible models of travel choices.
These hypotheses form the basic set around which the research is structured.
It is necessary to obtain relevant data sets, in order to test the various hypoth¬
eses. Only limited testing is possible in a prototypical sense, since resources
do not exist to obtain extensive before-and-after-data, which would form the best
tests of dynamic stability of these hypotheses. However, most of the hypotheses
are susceptible to at least limited testing through the use of a single set of
cross-sectional data.
22
3. DATA COLLECTION AND DESCRIPTION
3.1 Introduction
The preceding chapter has alluded to the need for data as a means to test
the various hypotheses advanced by this research project. It is apparent from
these hypotheses that the data required for a rigorous testing of any of the
hypotheses do not currently exist from any source. Indeed, an early part of
the literature search of this project was aimed at attempting to determine
whether any such data sets existed. None were found. In order to be able to
test the various hypotheses, it is necessary to specify the details of the data
that would be needed for adequate hypothesis testing.
As was discussed in the last chapter, the primary hypotheses to be tested
revolve around the attractiveness of shopping locations, and the structuring of
a model that represents a trade-off between attractiveness and accessibility.
As a subsidiary task, data were required on mode choices for shopping trips, in
order to provide estimates of the travel utility for destination choices. In
all, three surveys were undertaken: Two to obtain data on perceptions of shop¬
ping centers and other shopping opportunities and one to obtain data on mode
choices for shopping trips. These three surveys are described in this chapter
and details are provided on the profiles of the respondents.
The research concentrated initially upon the definition of attractiveness,
and on some limited testing of choice sets and on the pattern of shopping trip-
making. In addition, it was felt necessary to develop a data base that would
permit extensive testing of the market-segmentation ideas as a part of the
initial definition of attractiveness. Therefore, the first survey was set up
primarily to test the hypotheses relating to the image of attractiveness of
alternative shopping centers, the variation of this image among individuals, and
to provide enough data to make preliminary tests upon a model structure.
Based upon the data from the first survey, attractiveness variables were
analyzed for the seven shopping centers used in that survey. These shopping
locations were the Chicago Loop, Edens Plaza, Golf Mill, Korvette City, Old
Orchard, Plaza del Lago and Woodfield. Apart from the Chicago Loop, all of the
shopping locations can be classified as shopping centers. That is, they are
retail locations whose stores are compactly located within well-defined geo¬
graphic boundaries. For the second survey, it appeared to be desirable to
determine how, in terms of attractiveness, individuals compare shopping centers
with retail locations which are not shopping centers.
The second survey was also viewed as a means of collecting data for the
purpose of testing the predictive capability of the project's non-grocery shop¬
ping destination-choice model. There is a plan to convert the State Street
shopping area in the Chicago CBD into a pedestrian mall. Given attitudinal
data on the State Street shopping area, the destination-choice model could pre¬
dict changes in shopping behavior resulting from pedestrianization. The first
survey collected data on shopping in the Chicago Loop, but this information is
too gross to determine the specific attractiveness of State Street as a shopping
area since it includes other major shopping areas, such as North Michigan Avenue.
By including State Street as one of the shopping opportunities in the attitudinal
?3
questions, the second survey could collect the appropriate data for the testing
procedure. Also, it was felt that the survey could be used to supplement the
attitudinal, socioeconomic and shopping trip information gathered on shoppers
in the first survey.
The overall goal of this project was to develop a combination destination-
choice and mode-choice model for shopping trips. In order to fulfill this goal,
information on mode-choice behavior on shopping trips must be obtained to go
with the data on destination-choice behavior previously gathered. The data
gathered on mode-choice behavior included information on the modes chosen and
not chosen by individual travelers on shopping trips, information on those
characteristics of the modes that are hypothesized as being determinants of
travel behavior, and information on those characteristics of individual trav¬
elers that are hypothesized as being determinants of travel behavior-
As an early stage in the research, decisions had to be made about the way
in which attractiveness would be measured and applied in a modeling context.
After reviewing the existing literature, it was determined that the most promis¬
ing approach would be to use methods of attitude and preference scaling from
psychology (Torgerson, 1958). This decision generated some extremely specific
requirements for the collection of data. Primarily, it required that a number
of potential attributes of attractiveness be defined by the researchers; that
preferences for these attributes be determined, in relation to specific shopping
opportunities; and that importances of these attributes be established through
the survey procedure. Given the probable difficulty that people would have in
responding to the types of detailed perception and attitude questions, it was
also necessary to build in a reasonable set of checks and balances on the atti¬
tudinal information, so that similar results could be derived from at least two
different and partially independent procedures. Such dual derivations would
provide a check on the validity of the conclusions reached, irrespective of
possible intransitivities and illogicalities among the respondents. Again, the
desire to be able to check results from alternative measurement techniques and
analytical procedures dictated a large proportion of the data collection effort.
In summary, the research determined that no current data sets existed that
would be adequate for the testing of the hypotheses proposed in this research,
and therefore dictated that specific data sets be collected for testing those
hypotheses. The testing of hypotheses relating to choice sets, the existence
and measurement of attractiveness, market segmentation, and prototypical model
structure for destination choice, required that data be collected upon the per¬
ceptions of individual shoppers and the characteristics of those shoppers them¬
selves, together with limited information on the trip to which those data relate.
This information on the trip should include the types of commodities purchased,
whether the trip is a part of a chain of trips, the mode of travel used, and the
origin from which the trip was made.
3.2 The First Destination-Choice Survey
3.2.1 Development of the Attribute Set
The literature search, particularly in the marketing literature, produced
a list of candidate attributes that might comprise attractiveness. The initial
list, derived from the marketing literature, is shown in Table 3-1. After
24
LOCATIONAL CONVENIENCE
MERCHANDISE SUITABILITY
Access Routes
Traffic Barriers
Traveling Time
Parking—Availability and Cost
Available Modes of Transport
Proximity to Home
Proximity to Other Stores
Number of Brands Stocked
Quality of Stock
Breadth of Assortment
Depth of Assortment
Number of Outstanding Departments
within Store
PRICE CONSIDERATIONS
Price of a Particular Item (Z)
in a Particular Store
Price of Item Z in Competing
Store
Price of Item Z in a Particular
Store on Sale Day
Prices of Substitute Products in
Substitute Stores
SALES EFFORTS AND STORE SERVICES
Courtesy of Sales Clerks
Helpfulness of Sales Clerks
Advertising
Billing Procedure
Credit Arrangements
Delivery Promptness
Services Rendered
Store Hours
Style of Operation
Eating Facilities
CONGENIALITY OF STORE
Layout
Decor and Attractiveness of
Merchandise Displays
Store Traffic and Congestion
Class of Customers
Store Reputation
Store Age
Store Size
POST-TRANSACTION SATISFACTION
Satisfaction with Merchandise in
Use
Satisfaction with Returns and
- Adjustments
Satisfaction with Price Paid
Satisfaction with Shopping
Experience in Store
Satisfaction with Accessibility
to Store
TABLE 3-1
38 Attractiveness Measures
25
excluding those attributes that describe transportation related accessibility
attributes, an initial list of candidate attributes was compiled. This list
was derived primarily from those sections of Table 3-1 labelled price consid¬
erations, congeniality of store, merchandise suitability, sales efforts and
store services, and post-transaction satisfaction. The attributes under loca-
tional convenience were considered to be attributes of accessibility. However,
the attribute of parking was considered to be an attribute of the site rather
than an attribute of accessibility. This assumption may be somewhat question¬
able. However, it is clear that the lack of parking at a shopping location
would most probably deter those shoppers who have a preference for using their
automobiles for a shopping trip. Likewise, a shopping location which has ade¬
quate free parking that is located close to stores would be likely to attract
trips from those who prefer to use their automobiles. Thus, it appears that
the parking attribute is a somewhat mixed attribute that may be considered as
both an attribute of the site and an attribute of accessibility. For this
research, it was assumed to be an attribute of attractiveness.
The list of potential attributes was split into two sections, one section
of attributes that describe a specific store and a second section of attributes
that describe an entire shopping center. These attributes are shown in Table
3-2. This set of candidate attributes was set up in part on the basis of the
original listing in Table 3-1, and in part by a consideration of the potential
combinations of attributes from the original list into new compound attributes
that may still be appropriate for consideration as elements of attractiveness.
Since Table 3-2 still contains 21 attributes, and since these were generated
from other sources, it was considered necessary to further refine the list of
attributes. In terms of the size of the list, this was necessary in order to
design a survey that would be small enough to encourage completion. In addi¬
tion, it was not considered that this set of attributes was sacrosanct. Some
experimental work was deemed advisable to determine whether or not these attri¬
butes were seen by local shoppers to be of importance and to attempt to deter¬
mine whether any important attributes were missing.
A small pilot survey was undertaken to refine and expand this candidate
list of attributes, using an importance scaling of the attributes. In addition,
respondents were asked to add any attributes that were important to them that
had not been included, indicating their importance. Limited information was
also collected on social and economic characteristics of the respondents as
well as some data upon the shopping trip that they were undertaking. For any
attitude data on importances to be meaningful, it is necessary that they be
collected in relation to a specific trip that is currently in the mind of the
respondent. Therefore, the survey device was designed to permit respondents
to examine these attributes in relation to a current shopping trip for consumer
durables. Given time and money limitations, a home interview process was ruled
out as being impractical. Such a process would necessitate a prior contact to
establish whether or not the individual had undertaken a shopping trip for con¬
sumer durables within the last 24 hours, with the rejection of any respondent
who did not qualify in this way. Such a procedure would involve a very large
number of home interview contacts with a relatively small return of usable
responses. The alternative procedure is to select, as respondents, people
actually engaged in a consumer-durable shopping trip. This could be done most
26
STORE CHARACTERISTICS
CENTER CHARACTERISTICS
Layout of Store
After Sales Service
Store Opening Hours
Status of Store
Value for Money
Availability of Rest Rooms
Variety of Range of Merchandise
Availability of Store Credit
Availability of Sale Items
Courteous and Helpful Sales Assistants
Freedom to use Major Credit Cards
Store Atmosphere
Free Parking
Availability of Eating Facilities
Covered Walkways between Stores
Stores Located in Compact Area
Availability, of a Specific Store
Architecture and Design of Center
Auto-free Mal 1
Ability to find Convenient Parking
Number and Variety of Stores
TABLE 3-2
Store and Shopping-Center Characteristics
27
easily by intercepting people in a shopping center, or other shopping location,
and obtaining their responses at that time. The research team felt that an inter¬
view would still be impractical, since most shoppers would have too little time
available to be willing to stop and answer a set of questions. The decision was,
therefore, made to carry out this survey as a self-administered questionnaire that
would be handed out at shopping centers to potential respondents. It was recognized
that such a survey procedure has a number of inherent disadvantages, such as a lack
of control on the sampling, the traditionally high nonresponse rate, and the pros¬
pects of bias in the responses that are obtained. However, the advantages of the
procedure were seen to outweigh the disadvantages by a substantial margin and the
pilot survey proceeded as a choice-based, self-administered survey.
The survey achieved a 12% response rate, compared with the anticipated one
of 10%, and from this a set of revised characteristics were developed. These
characteristics comprised the deletion of some attributes from the original
list, a change of the wording of some attributes, the combination of some attri¬
butes, and the addition of some attributes. A list of the revised characteris¬
tics is shown in Table 3-3 together with a comparison with the original set of
attributes used in the survey. The survey procedure achieved a reduction in
the number of attributes to 16. The mechanism by which these characteristics
were developed may be seen most readily by examining the results of unidimen-
sional scales developed from the importance rankings. These scales are shown
in Figures 3-1, 3-2, 3-3, and 3-4, including a stratification by sex for each
set of store and shopping center characteristics. It was also found that other
stratifications of the data produced virtually identical results. The deleted
attributes were found to be located consistently at the bottom of the scales
for all socioeconomic stratifications. While this does not imply lack of impor¬
tance, it does imply that those attributes are always considered to be less
important than any others. Similarly, the scales implied no value to separating
store credit from major credit cards. The interpretation of this was that cre¬
dit availability may be important but it is unimportant whether it is made
available through the store or a major credit card.
It is also worthwhile to examine the distribution of social and economic
characteristics across the population that responded. These characteristics
are shown in Table 3-4. In addition, some cross tabulations of the population
are shown in Appendix A of this report.
3.2.2 Execution of the Primary Survey
The next step in the procedure was to develop a major survey to provide
estimates of attractiveness indices, based upon the attributes that were devel¬
oped in the pilot survey. Once again, it was determined that the preferred
survey method would be a repetition of the method used in the pilot survey,
i.e., the use of respondents who were then engaged in a shopping trip, and who
could be persuaded to accept a questionnaire to administer to themselves sub¬
sequently. In addition, however, it was decided that it would be worthwhile
to attempt to monitor the effectiveness of the self-administered responses by
carrying out limited in situ interviewing. Thus the major survey was designed
around a dual strategy, in which most respondents were asked to accept the
28
Original Characteristics
Revised Characteristics
Al.
Layout of Store
CI.
Layout of Store
A2.
After-sales Service
C2.
Ease of Returning or Servicing
Merchandise
A3.
Store Opening Hours
(deleted)
A4.
Status of Store
C3.
Prestige of Store
A5.
Value for Money
C4.
C5.
Reasonable Price
Quality of Merchandise
A6.
Availability of Rest Rooms
(deleted)
A7.
Variety or Range of Merchandise
C6.
Variety or Range of Merchandise
A8.
Availability of Store Credit
C7.
Availability of Credit
A9.
Availability of Sale Items
C8.
Availability of Sale Items
("Specials")
A10.
Courteous and Helpful Sales
Assistants
C9.
Courteous and Helpful Sales
Assistants
All.
Freedom to Use Major Credit
Cards
(see C7)
Al 2.
Store Atmosphere
CIO.
Store Atmosphere (Heating, Cooling,
noise, crowds, etc.)
Bl.
Free Parking
Cll.
Free Parking
B2.
Availability of Eating Facilities
(deleted)
B3.
Covered Walkways between Stores
CI 2.
Shopping Center Atmosphere (pedestrian-
only area, flowers and shrubs, covered
walkways, etc.)
B4.
Stores Located in Compact Area
C13.
Stores Located in Compact Area
B5.
Availability of a Specific Store
C14.
Availability of a Specific Store
B6.
Architecture and Design of Center
(see CI2)
B7.
Auto-free mall
(see C12)
B8.
Ability to find convenient parking
CI 5.
Ability to Park where you want
B9.
Number and Variety of Stores
CI 6.
Number and Variety of Stores
Table
3-3
Revised Store and Shopping Center Characteristics
29
1.00 Variety or Range
.66 Value for Money
.48 Specials
.48 Sales Assistants
.48 After-Sales Service
.41 Status
.41 Layout
.36 Store Credit
.35 Atmosphere
.24 Opening Hours
.04 Rest Rooms
0.00 Major Credit Cards
FIGURE 3-1
Uni dimensional Scale for Store Characteristics
Based on Entire Sample
30
Female
Male
Variety or Range
1.00
Value for Money
After-Sales Service
Specials
Sales Assistants
Sta tus
Store Credit
Layout
Atmosphere
Opening Hours
Rest Rooms
Major Credit Cards
.66
.53
.53
.48
.46
.41
.38
.26
.20
.03
0.00
Variety or Range
Value for Money
Sales Assistants
Status
Layout
Specials
After-Sales Service
Atmosphere
Opening Hours
Store Credit
Rest Rooms
Major Credit Cards
FIGURE 3-2
Unidimensional Scales for Store Characteristics
Based on Sex
31
1.00
.95
.94
.82
.81
.46
.30
.13
0.00
Variety of Stores
Free Parking
Compactness
Convenient Parking
Specific Store
Pedestrian Mall
Covered Walkways
Architecture
Eating Facilities
FIGURE 3-3
Unidimensional Scale for Shopping Center Characteristics
Based on Entire Sample
32
SlO.OOO-and-under
Over $50,000
1.00
.93
,77
,68
.58
,39
.34
.03
0.00
Free Parking
Variety of Stores
Convenient Parking
Compactness
Specific Store
Pedestrian Mall
Covered Walkways
Architecture
Eating Facilities
1.00
.92
.90
.88
.81
.37
.28
.20
0.00
Variety of Stores
Free Parking
Compactness
Specific Store
Convenient Parking
Pedestrian Mall
Covered Walkways
Architecture
Eating Facilities
FIGURE 3-4
Unidimensional Scales for Shopping Center Characteristics
Based on Lowest and Highest Income Groups
33
Socioeconomic Characteristics
Percentage
Male
14%
Female
81
Unspecified
6
21 and under
17
22 - 29
18
30 - 39
15
40 - 49
22
50 - 59
14
60 and over
10
Unspecified
3
Salesman/Buyer
4
Professional/Technical/Managerial
8
Craftsman/Mechanic/Factory Worker
1
Clerical/Secretarial/Office Worker
6
Teacher/Professor
10
Student
17
Housewife
41
Governmental
1
Retired
6
Other
4
Unspecified
2
$10,000 and under
10
$10,001 - $15,000
15
$15,001 - $20,000
14
$20,001 - $25,000
14
$25,001 - $50,000
24
Over $50,000
14
Unspecified
8
TABLE 3-4
Percentage Breakdown of Socioeconomic Characteristics
for the Pretest Survey
34
questionnaire, fill it out at home and mail it back, while a few were requested
to fill out the questionnaire at the shopping center, with the assistance of an
interviewer, and leave the questionnaire there. It was assumed that comparisons
of the results of these two alternative procedures would provide a reasonably
good control mechanism on the effectiveness and biases of the self-administered
procedure. Clearly, however, the extent to which this process can provide a
check is limited, in that both types of respondent are likely to be subject to
the same inherent biases, in terms of cooperation in completing the question¬
naire. Nevertheless, some concern might be appropriate over the extent to which
an individual can self-administer a questionnaire which is designed to determine
a fairly complex set of preferences and attitudes and this mechanism permits the
checking of this aspect.
As discussed in chapter 2, the attractiveness measures were to be developed
through the technique of multidimensional scaling. This technique, in part,
dictates the form in which data are collected for subsequent analysis. In order
to be able to develop an estimate of the dimensionality of the perceptual space,
in which attractiveness is perceived, and to be able to label that space in
terms of compound attributes, a number of questions are necessary. First, it is
necessary to obtain data on the comparative importances that people associate
with each of the candidate attributes for the concept of attractiveness. Second,
it is necessary to obtain estimates of the location of a number of real or hypo¬
thetical shopping locations on scales describing a range of each of the attri¬
butes between common polar points. Third, it is necessary to determine, in some¬
what abstract terms, the proximity or distance between the shopping locations in
the perception of the individual, relative to the concept of attractiveness. In
other words, it is necessary to find out how similar or dissimilar shopping cen¬
ters are perceived to be in relation to some idea of the attractiveness, or
appropriateness of those shopping locations for the specific shopping trip.
These questions provide the basis for the development of psychological scales
that can be analyzed to produce a representation of the perceptual space. They
also provide some degree of validity checking, in the sense that similar results
can be inferred from at least two different sets of questions. If these results
are consistent with each other, it may be concluded that an individual has an¬
swered the questionnaire consistently and with reasonable understanding of the
survey's requirements. In addition, data would be required on the choice set of
shopping locations that the individual perceived was open to him. While this
question provides interesting and important information relating to questions
of the formulation of a choice set, it is also very important as a means of
establishing whether or not respondents have familiarity with the set of shop¬
ping locations that are the subject of the attitude questions. Finally, a ques¬
tion on the rank order of preference for shopping at the subject shopping loca¬
tions was included to provide a simple check on the more complex attitude ques¬
tions and as a potential measure of choice. Inconsistencies between the response
to the ranking question and the implied ranks derived from the more complex
questions would suggest that the response in question be considered very circum¬
spectly for further analytical work, since it may contain either inconsistencies
or intentional deception.
The remaining questions on the questionnaire related either to
specific aspects of the shopping trip, or to the individual responding
to the survey. First, questions were asked about the commodities pur¬
chased, stores visited and mode of travel of the shopping trip, together
35
with details on the previous origin and next destination. Respondents were
also asked to indicate what other shopping locations they had considered for
this trip, or alternatively would consider were the location where they re¬
ceived the questionnaire unavailable for the shopping trip in question. In
the second set, questions were asked on the home address of the shopper, occu¬
pation, income, age, and sex of the shopper. Respondents were also asked to
indicate how long they had resided in the geographic area within which the
survey was being undertaken. This question, in particular, was suggested by
the work of Burnett (1973) in which the length of residence was found to be an
important variable discriminating between perceptual spaces of shoppers.
An extensive design process was then undertaken to refine the form of the
questions in the survey. This design process included pilot testing and exten¬
sive group testing and criticism of the survey instrument. The result of this
process was the development of a final questionnaire format that was selected
for use. A detailed description of the design process is given in the interim
report (Transportation Center, 1974) and is not repeated here. The questionnaire
is shown in Appendix B.
The final task that remained was to select the sites for the survey. The
selection of survey sites is closely intertwined with the selection of a set of
shopping locations, about which the attitudinal questions would be posed. It
had already been decided that the set of shopping locations to be used should
be drawn from the area in the north shore suburbs of Chicago. For this initial
survey, it was decided to concentrate on shopping center locations, for two
reasons. First, shopping centers can be identified more readily on a self-
administered questionnaire by use of the shopping center's name. Other types
of shopping developments generally are not responsive to such easy identifica¬
tion, and could result in serious ambiguities in the mind of the respondent.
Second, it was considered, at this stage of the research, that the definition
of an attractiveness concept for shopping centers would be likely to be an
easier job than the definition of a similar concept for a less well-defined
and well-oriented shopping location.
The selection of shopping centers as the subjects of the first survey
raised a new problem, that had not been anticipated originally. This problem
was that of obtaining permission from a shopping center developer or agent to
conduct the survey at the center. It was found that the majority of shopping
centers in the Chicago area maintain a policy of discouraging any form of soli¬
citation of shoppers in the shopping center. The jurisdiction of the agents for
the shopping center covers all areas outside the actual stores themselves. In
the early pilot testing work, this problem had been circumvented by requesting
permission from individual stores to undertake the survey within the store.
These pilot surveys had been conducted, therefore, by handing out questionnaires
to shoppers entering and leaving a store. For this, the survey team was sta¬
tioned at the doors of the store, and not in the shopping-center malls. However,
this procedure was felt to be less than optimal for the full-scale survey. A
long period of negotiations was therefore undertaken with the management of each
of the shopping centers that were desired as survey sites. The result of these
negotiations was the obtaining of permission to undertake the survey at the Plaza
del Lago Shopping Center in Wilmette, Edens Plaza Shopping Center in Wilmette,
and Old Orchard Shopping Center in Skokie. A fourth site, Golf Mill Shopping
Center in Niles was also desired, but permission could not be gained from the
shopping center management. In the end, it was decided to undertake the survey
36
at that site by obtaining permission from the major store owners in the shop¬
ping center to conduct the survey on their premises. This was duly done.
Three additional sites were chosen for inclusion in the total set of shopping
locations, the Chicago Loop, Korvette City Shopping Area in Morton Grove,
Illinois, and Woodfield Shopping Center in Schaumburg, Illinois. There was no
intent to undertake a survey at the Chicago Loop, and this was included pri¬
marily to provide a reference point that would be appropriate for any other
follow-up surveys in the Chicago area, and also to provide specific informa¬
tion relating to the potential of pedestrianization in the downtown area.
An attempt was made to gain permission from each of Korvette City and Wood-
field, but without success. Under the circumstances, it was decided not to
pursue these locations further as survey sites, but still to use them as
additional locations for the attitude questions.
The set of selected survey sites embraces a reasonably large range of
size and type of shopping center. Old Orchard Shopping Center represents a
large regional shopping center, Golf Mill is a moderate size shopping center
while Edens Plaza is a small center, of more local significance, built around
a single large department store. Plaza del Lago represents a local shopping
center, comprising small specialty stores with no large department store.
The other three sites represent the largest covered shopping center in the
country (Woodfield), a large metropolitan central business district (Chicago
Loop), and a small discount center containing a single department store and
very few specialty stores (Korvette City). Thus, the survey embraced a very
wide range of shopping opportunities, while yet restricting consideration to
shopping centers. The approximate locations of these shopping centers is
shown 1n Figure 3-5.
The survey was undertaken for one week at each location. In total, about
37,000 questionnaires were distributed, with the expectation that the response
rate would be in the region of 5%. This response rate was based upon experi¬
ence with the first survey where a 10% response rate was anticipated and a 12%
rate achieved. The primary survey was much longer and more complex than the
original pilot survey, and thus led the research team to lower their estimate
of response rate. The survey took place during three weeks of late July and
early August of 1974.
3.2.3 Description of the Sample
The final response rate achieved from the distribution of questionnaires
at the four shopping centers was a rate of approximately 22%, comprising 7,362
completed questionnaires and an additional 400 or so incomplete questionnaires.
The incomplete questionnaires have not been subjected to analysis. A number
of reasons may be put forward for the extraordinarily high response rate
achieved by the survey. First, it seems plausible to suppose that the survey
itself generated some degree of interest in the respondents, such that a larg¬
er number than might otherwise be expected completed the questionnaire and
returned it. Second, the cooperation received at three of the four shopping
centers in the form of sponsorship or permission to use the shopping center
for the survey may have improved the perception of potential respondents of
the worthwhileness and value of the survey itself.
37
©
Northwestern
university
1. CHICAGO LOOP
2. EDENS PLAZA
3. golf mia shopping center
4. corvette city
5. OLD ORCHARD SHOPPING CENTER
6. PLAZA DEL LAGO
7. WOODFI ELD SHOPPING C0-JTER
T
FIGURE 3-5
Locations of the Subject Shopping Centers for the First Survey
38
The form of the survey, i.e., a self-administered survey with no control
on sampling or response, does not permit any form of a representative sample
to be achieved. Given the exploratory nature of the research being under¬
taken, this lack of representation was not deemed to be a consideration in the
design of the survey. Therefore, the profiles generated for the survey re¬
spondents cannot be considered to be representative, necessarily, of the
north-shore residential areas of Chicago. However, if these profiles corres¬
pond reasonably with other known data, e.g., census data, it may be assumed
that serious biases have not been introduced into the survey by the adminis¬
tration of it.
The income of the suburbs from which the sample is most probably drawn
is known to be relatively high. In fact, within the residential areas from
which shoppers are drawn to the shopping centers, there are 13 contiguous
census tracts with incomes, in the 1970 census, that are more than twice the
median income of the Chicago area. This is a very significant group, and
would suggest a high income bias in the area. Such a bias was indeed found
in the survey results, as shown in Table 3-5. Thus, it would appear that
the sample is reasonably typical of what might be expected in the north shore
area of Chicago, but is clearly a highly biased sample for the generalization
of survey findings to any metropolitan area. The other principal bias deter¬
mined in the survey is by sex. The breakdown was found to be 20% male and
76% female, with 4% respondents not answering the question. While this may
not represent a bias in terms of shoppers, it clearly does represent a bias
with respect to overall population. The distribution of income by sex is
shown in Table 3-6. This table shows no significant difference in the incomes
by sex.
The length-of-residence distribution was found to be more skewed towards
long periods of residence than was expected a priori. Given the mobility of
the U.S. population, this seems a somewhat surprising result. Table 3-7
shows the distribution of length of residence. The table shows a surprising¬
ly large proportion (62%) that report being resident in the area for more
than ten years. However, it should be noted that the residency question was
asked in terms of residence within the suburban area covering the majority
of the north and northwest suburbs of Chicago. Hence, it does not indicate the
length of residence at a single address, but rather length of time settled
in the area. This was intentional, since the length-of-residence question
was designed to determine the degree of familiarity that a respondent would
be likely to have with the shopping opportunities in the area.
Given the biases in income, sex, and length of residence, it might be
expected that some correlations would exist within these variables. Tables
3-8 and 3-9 show the distributions of length of residence by each of sex and
income. In neither of these tables is it apparent that there is any strong
correlation between the biased variables. Indeed, length of residence and
sex appears to yield no significant variation, while the distribution of
length of residence by income shows only the expected variations that are
most probably correlated as much with age as with length of residence. For
example, there is a higher proportion of people with incomes under $10,000
who have resided in the area less than one year than any other income group.
In all of the higher income groups, the percentages increase with increasing
length of residence. This is an expected result. In the lower income groups,
there is a tendency for the distributions to peak initially in the short
39
Income Group
Frequency (%)
$10,000 and under
832 (11.3)
$10,001 - $15,000
1150 (15.6)
$15,001 - $20,000
1215 (16.5)
$20,001 - $25,000
114'8 (15.6)
$25,001 - $50,000
1479 (20.1)
Over $50,000
517 (7.0)
No Response
1021 (13.9)
TABLE 3-5
Sample Income Distribution
40
Income
SEX
Under $10,000
$10,001-$15,000
$15,001-$20,000
$20,001-$25,000
$25,001-$50,000
Over $50,000
Female
13.5%
17.4%
19.0%
18.1%
23.5%
8.6%
Male
12.4%
20.5%
19.6%
17.8%
23.2%
6.6%
TABLE 3-6
Distribution of Income by Sex
Length of Residence
Frequency (%)
Less than 1 year
215 (2.9)
1-3 years
690 (9.4)
4-6 years
674 (9.2)
7-10 years
777 (10.6)
More than 10 years
4558 (61.9)
No response
448 (6.1)
TABLE 3-7
Distribution of Length of Residence
Length of Residence
SEX
Less than
1 year
1 - 3
years
4-6
years
7-10
years
Over 10
years
Female
3.0%
9.7%
10.0%
11.2%
66.1%
Male
3.7%
11.2%
8.9%
11.0%
65.3%
TABLE 3-8
Distribution of Length of Residence by Sex
INCOME
Length of Residence
Less than
1 year
1 - 3
years
4-6
years
7 - 10
years
Over 10
years
Under $10,000
6.1%
11.8%
6.9%
6.5%
68.8%
$10,001-$15,000
4.2%
13.1%
9.1%
9.6%
64.0%
$15,001-$20,000
2.9%
11.1%
10.4%
11.2%
64.4%
$20,001-$25,000
2.9%
9.8%
11.6%
11.0%
64.7%
$25,001-$50,000
2.2%
8.7%
9.7 %
13.1%
66.3%
Over $50,000
2.0%
7.3%
8.8%
13o5%
68.4%
TABLE 3-9
Distribution of Length of Residence by Income
42
lengths of residence and peak again in the longest residence category.
This final peak, observed in all income groups, is in conformance with the
length of residence distribution observed over the whole population.
The other two variables used to describe respondents were occupation
and age. The distribution of these over the sample population is shown in
Tables 3-10 and 3-11. No conclusions can be drawn from the distribution of
occupation with respect to the representativeness of the sample, since data
of this type are not generally available from other sources. In terms of the
age distribution, it may be noted that there is some skewing towards younger
groups, with approximately half of the respondents being drawn from ages of
30 or under. This is not a very surprising result, given that the North-
Shore area contains a relatively high number of students. However, it seems
possible that some of the preponderance of lower age groups may be a specific
bias generated by the survey itself. It is probable that a more uniform
distribution should be expected from a totally representative sample, and the
tailing off in the higher age groups may indicate a lower level of tolerance
of the complexity of the questionnaire, together with a lesser inclination
to respond to questionnaires of this type.
Finally, it is interesting to examine the degree to which people reported
that they were on a linked trip of some type. That is, the previous origin
and the following destination would indicate whether an individual is in the
process of a multiple purpose trip, or alternatively is making multiple shop¬
ping visits. The extent of such trips appears relatively low for this sample.
On the basis of Table 3-12, it can be seen that only 11% of trips were des¬
tined for another shopping center, while 76% of trips were then going home.
Similar figures are shown in Table 3-13 for the origins. Table 3-14 shows
the last origin-first destination pattern for the sample. From this, it can
be seen that over 64% of trips are between home and the shopping center only,
while a further 5% are trips between home and work. Only 12.4% of trips are
made between multiple shopping destinations and home, and these comprise the
majority of the remaining trips. From this, it may be concluded that the
occurrence of multiple destination trips for shopping is relatively low in
the north shore area at this time. Two possible conclusions may be drawn
from this. First, it may be postulated that the effects of the fuel short¬
ages of the winter of 1974 have receded sufficiently in the memories of the
majority of residents of the area that little attempt is being made to con¬
serve energy on trips. Alternatively, it may be postulated that a shopping
center represents an opportunity for a multiple-destination trip, without
the attendant travel movement that would normally be associated with multiple-
destination trips.
Cross-tabulations of age and length of residence, and age and income are
of interest. These are shown in Tables 3-15 and 3-16. As may be expected,
the longest residence groups are associated with both the youngest and oldest
age groups. These indicate the members of families that have matured in the
area. Similarly, the shorter periods of residence are associated with the
intermediate age groups, particularly in the 22-29 year-old age group. The
cross-tabulation of sex and age does not indicate any very marked variation
between these.
43
Occupation
Number (%)
Military
11 (.1)
Salesman
305 (4.1)
Teacher
641 (8.7)
Professional
1182 (16.1)
Craftsman
108 (1.5)
Clerical
777 (10.6)
Student
1700 (23.1)
Housewife
1889 (25.7)
Governmental
85 (.7)
Retired
146 (2.0)
Other
274 (3.7)
No Response
%
274 (3.7)
TABLE 3-10
Distribution of Occupation
Age
Number (%)
Under 16 years
357 (4.8)
16 - 21 years
1588 (21.6)
22-29 years
1647 (22.4)
30 - 39 years
1148 (15.6)
40 - 49 years
1122 (15.2)
50 - 59 years
840 (11.4)
Over 59 years
404 (5.5)
No Response
256 (3.5)
TABLE 3-11
Age Distribution for the Sample
44
Next Destination
Percentage
Other shopping locations
11.1
Work
2.5
Home
75.7
Other
10.6
TABLE 3-12
Distribution of Next Destination
Previous Origin
Percentage
Other shopping locations
6.9
Work
5.6
Home
80.8
Other
6.7
TABLE 3-13
Distribution of Previous Origin
45
—.^Ori gi n
Desti natiofr,*,^^><><^
Other
Shopping
Work
Home
Other
Other shopping
89(1.3)
15(.2)
628(9.2)
47(.7)
Work
1(0)
84(1.2)
82(1.2)
4(0.1)
Home
217(3.2)
275(4.0)
4,393(64.2)
283(4.1)
Other
28(.4)
24(.4)
534(7.8)
136(2.0)
TABLE 3-14
Origin by Destination
46
Length of
^«^Residence
Age
Less than
1 year
1-3
years
4-6
years
7-10
years
Over
10 years
Under 16
2.6
4.1
10.2
21.9
61.1
16 - 21
2.5
5.9
7.4
11.9
72.3
22 - 29
7.0
19.9
12.8
5.7
54.5
30 - 39
2.7
14.6
15.9
19.1
47.7
40 - 49
1.1
4.6
7.4
12.0
74.9
50 - 59
1.5
5.0
5.0
7.9
80.7
60 and Over
1.0
4.5
5.5
4.0
84.9
TABLE 3-15
Length of Residence by Age
(All entries are percentages of row totals)
47
Age
Less than
$1 OK
$10K-
$15K
$15K-
$20K
1
O IT) ;
CVI CM
$25K-
$50K
Over
$50K
Under 16
12-7
15.8
19.7
14.9
23.2
13.6
16 - 21
19.5
14.6
17.1
16.0
21.4
11.5
22 - 29
21.8
27.1
22.4
15.2
10.8
2.7
30 - 39
3.5
17.2
21.1
21.5
29.1
7.6
40-49
3.8
10.9
16.7
22.1
36.0
10.5
50 - 59
7.7
14.1
18.4
20.0
30.3
9.5
60 and Over
20.4
24.8
14.7
15.0
16.9
8.2
TABLE 3-16
Income by Age
(All entries are percentages of row totals)
48
The most prevalent occupation of the under 22 year-olds is student, as
might be expected, accounting for 78.6% of that age group. The 22-29 year-
olds are more scattered across the occupations, with 26.6% professionals,
19.2% housewives. For ages 30-59, the most frequently reported occupation
is housewife. In the over-59 age group, 30.8% are retired and 33.3% are
housewives.
Expected biases appear in the cross-tabulation of sex and occupation.
Housewife accounts for 33.6% of the females but only 0.6% of males. Most
males report being professionals, with 42.9% in this category, followed by
20.2% who are students and 10.2% salesmen. Similarly, 25% of females are
students, followed by 13.1% in secretarial and clerical occupations.
Additional statistics are relevant, with relation to the trips under¬
taken, referring to the modes of travel, and the items purchased. Table 3-17
shows the distribution of modes used to get to and from the shopping centers
used in the survey. Based upon these figures, average auto occupancy appears
to be 1.32, which is close to the usual average auto occupancy reported for
major metropolitan areas, such as Chicago. In total, 93.2% of the sample
came by car, as either driver or passenger, and only 2% came by bus. Inter¬
estingly, almost the same number walked as came by bus, while most of the
"other" category comprises bicycles, which also account for almost as many
trips as bus. Table 3-18 indicates the items purchased. Clothing was the
item purchased most frequently. On average, respondents reported 2.65 diff¬
erent items purchased, which is consistent with the relatively low proportion
of multiple-destination trips. The actual average may be somewhat higher
than this, since the questionnaire permitted only 5 responses on items pur¬
chased. However, it was noted that a rapidly decreasing number of people
reported each of 4 and 5 separate items, suggesting that the number purchas¬
ing more than 5 was probably insignificant. Several other tables and cross-
tabulations are to be found in Appendix c.
Finally, questions were asked on the familiarity and frequency of use of
a large selection of shopping locations in the North Shore suburbs. Of part¬
icular interest is the number of people who have never heard of those shop¬
ping centers used in subsequent questions on the attitudes and perceptions of
people about specific shopping opportunities. The reported familiarity is
shown in Table 3-19. From this, it appears that less than 5% of the popula¬
tion were unaware of five of the seven shopping centers. However, nearly 20%
were unaware of one shopping opportunity and over 25% were unaware of the
seventh center. These results suggest that substantially less than all of
the responses will be usable for the psychological scaling analysis.
These results serve to provide a profile of the sample achieved from the
survey. In summary, a biased sample of a metropolitan area was obtained, in
which the respondents were predominantly females from high income households,
who have lived in the north shore area of Chicago for more than six years.
The sample is also biased strongly towards younger people (under 30) and to
students and housewives. In terms of attitude biases, there is no indication
that this necessarily represents a poor sampling of shoppers. Little has been
reported in the literature on the demographic characteristics of the average
shopper, particularly for a suburban shopping center.
49
Mode Used
Number in
Sample (%)
Auto driver
5064 (70.4)
Auto passenger
1641 (22.8)
Bus
144 (2.0)
Taxi
10 (0.1)
Walk
135 (1.9)
Other
195 (2.7)
TABLE 3-17
Mode of Travel Used to the Shopping Center
Item
Number in
Purchased
Sample (%)
Clothing
5243 (26.8)
Gifts
1854 (9.5)
Books and Stationery
1512 (7.7)
Window Shopping
1505 (7.7)
Home Furnishings
1199 (6.1)
Housewares
1145 (5.9)
Drugs & Cosmetics
1120 (5.7)
Eating Out
946 (4.8)
Other
4998 (25.6)
TABLE 3-18
Items Purchased
50
SHOPPING LOCATION
FAMILIARITY
EDENS
PLAZA
OLD
ORCHARD
GOLF
MILL
WOOD-
FIELD
CHICAGO
LOOP
KORVETTE
CITY
PLAZA
DEL
LAGO
Never
heard of
4.4
0.7
2.2
3.0
1.7
19.2
26.9
Heard of,
not visited
18.0
2.5
7.7
16.8
5.5
36.7
42.8
TABLE 3-19
Familiarity Reported by Respondents to Centers
Used in Psychological Scaling Questions
3.3 The Second Destination-Choice Survey
3.3.1 Survey Design
As discussed at the beginning of this chapter, the principal goals of
the second destination-choice survey were to extend the measurement techni¬
ques to retail locations that are not structured as a shopping center, and
to provide a clearer distinction between the Chicago Loop shopping opportu¬
nities. In addition, several refinements were made to the questioning pro¬
cedures, based upon experience with the first survey. As before, the selected
process for conducting the survey was to hand out questionnaires to shoppers,
for them to self-administer the survey, and mail it back in a reply-paid
envelope. It was determined that a minimum of 500 completed questionnaires
was necessary for the type of analysis anticipated. Based on the mail back
rate of the 1974 survey, it was projected that at least 500 responses could
be received entirely by the self-administered questionnaire method if 5,000
questionnaires were distributed to shoppers.
Design of the questionnaire was begun in mid-June and completed by mid-
July, 1975. The questionnaire from the 1974 survey was used as the basis for
the new questionnaire. The final form of the 1975 questionnaire resembles
the 1974 questionnaire in many respects. Comparison of the two questionnaires,
which can be found in Appendix B, bears this out.
The list of retail locations included in the familiarity question of the
1974 questionnaire was altered for the 1975 survey. It was decided to drop
from the list those locations never heard of by a significant proportion of
the 1974 respondents and to add a few of the locations frequently written in
by respondents. The subject set used in questions relating to multidimension¬
al scaling was also revised. This was done in order to gain the information
necessary for placing more shopping opportunities in the attractiveness space
already containing the subject set from the 1974 survey.
The new subject set included four shopping locations from the 1974 sub¬
ject set: Edens Plaza, Golf Mill, Old Orchard and Woodfield. Results of the
1974 survey indicated that respondents were highly familiar with each of these
locations (See Table 3-19). Because the four sites were well known, they
served as reference points for the comparison of attractiveness information
from the 1974 and 1975 surveys.
The Chicago Loop was replaced in the new subject set by two locations:
North Michigan Avenue and State Street. These are two separate shopping
opportunities in downtown Chicago and their inclusion would give more detailed
information on CBD shopping opportunities than would the general designation
of Chicago Loop. Also, as was mentioned earlier, a pedestrian shopping mall
has been proposed for the State Street area. Therefore, any information
gathered on shoppers' current perceptions of the State Street shopping area
would help in predicting what effect the construction of the proposed mall
would have on shopping patterns.
Three suburban shopping locations -- Dempster Street (at Skokie Swift),
Downtown Skokie (Oakton Street and Niles Center Road) and Downtown Evanston
-- were added to the new subject set. Their inclusion increased the variety
52
of shopping opportunities represented in the set. Unlike the four shopping
locations maintained from the 1974 subject set, these three retail opportun¬
ities have neither well-defined geographic boundaries nor compact shopping
areas. Dempster Street and Downtown Skokie both represent strip developments
along major thoroughfares, while Downtown Evanston is a suburban CBD spread¬
ing over several blocks and is also the site of the mode-choice survey, des¬
cribed later in this chapter-
The final subject set consisted of the following nine shopping locations:
Dempster Street (at Skokie Swift): strip development
Edens Plaza: small suburban plaza around one department store
Downtown Evanston: vehicle-pedestrian suburban downtown area
Golf Mill: large suburban cénter; two department stores
North Michigan Avenue (Chicago): vehicle-pedestrian downtown area
Old Orchard: large suburban center; two department stores
Downtown Skokie (at Oakton Street and Niles Center Road): strip
development
State Street (Chicago): vehicle-pedestrian downtown area
Woodfield: world's largest indoor center
Although direct similarities questions were included in the first ques¬
tionnaire, these questions were deleted from the second survey. This decision
was reached after it was determined that the benefits from omitting the ques¬
tions outweighed the costs in terms of lost information. The benefits includ¬
ed the higher quality of responses which could be expected on the remainder of
the questionnaire. Expanding the number of shopping opportunities in the sub¬
ject set from seven to nine increased the number of direct similarities ques¬
tions from 21 to 36. This increase, almost double the number of questions,
would place additional strain on respondents and might precipitate poor re¬
sponses on these questions and others. Besides, analysis of the 1974 data
indicated that the information lost by omitting the direct similarities could
be minimized. Results suggested that, given a good set of shopping location
characteristics, almost the same information had been derived from indirect
similarities as was obtained from direct similarities.
Two new questions were included in the second questionnaire to determine
satisfaction levels of the respondents. One question asked shoppers how
satisfied they are with respect to shopping at the various locations in the
subject set. In the other question they were asked how satisfied they are
with respect to each of the 16 retail location characteristics at the shop¬
ping center where they received the questionnaire. Since elsewhere in the
questionnaire shoppers were asked for preference and importance rankings, it
was deemed informative to determine how satisfaction correlates with prefer¬
ence and importance.
The second questionnaire elicited more information on the make-up of
shoppers than did the first survey. In addition to the socioeconomic ques¬
tions of the first survey, the second questionnaire included questions which
requested the last year of school completed, marital status, and the ages of
all children living with the respondent. The latter two questions were asked
to determine respondents' positions in the family life-cycle.
53
The questionnaire also requested more information on the trip shoppers
made when they received their surveys. For example, it asked respondents to
estimate the time at which they left the location from which they travelled
directly to the shopping center and the time at which they arrived at the
shopping center. Also, it requested that respondents estimate their trip
cost. The responses to these questions would give an indication of how mem¬
bers of the sample perceived their shopping trip and would permit the use of
the mode-choice utilities in a destination-choice model.
Revisions were made in the answer categories of some of the questions
carried over from the first questionnaire. The number of possible responses
in the familiarity question was expanded from five to seven; the income cate-
tory of $25,001 to $50,000 which was used in the first survey was broken up
into five separate categories; the years of residence answer categories were
further subdivided. These changes were made in order to acquire more precise
information. In some cases, answer categories were deleted or combined with
other categories. For instance, the military category was dropped from the
list of occupations in the second survey because an insignificant number of
respondents reported this category in the first survey. In the second ques¬
tionnaire the age categories of under 16 and 16 to 22 were combined into one
category, because the 1974 survey indicated that respondents from the two
categories were quite similar in terms of characteristics other than age.
The second questionnaire used seven-point semantic scales instead of
the five-point scales of the first survey. This change was prompted in part
by the fact that with the new subject set nine instead of seven shopping
opportunities had to be ranked. With respect to filling out the question¬
naire, respondents were asked to circle numbers corresponding to the appro¬
priate answers rather than place checks in boxes or on scales. This change
was made primarily to facilitate coding. The second questionnaire was print¬
ed on light green paper to improve its readability.
It was decided that, since only 5,000 questionnaires were to be distri¬
buted, one shopping site would be sufficient. It was hoped that one of the
four distribution sites of the previous year could again be utilized since
a sample similar in character to that of the first survey was desired. After
negotiation, permission was obtained to use Old Orchard, in return for infor¬
mation acquired from the survey.
All questionnaires were distributed on Thursday, July 31, 1975. A team
of twelve people handed out 5,000 questionnaires throughout the mall area of
Old Orchard Shopping Center. A one-month deadline was set for the return of
questionnaires, by which time a total of 647 questionnaires had been returned,
representing a response rate of 13%. Of this total, 602 were considered to
be sufficiently complete to be satisfactory for analysis. Questionnaires
were considered complete if they contained responses for all of the socio¬
economic questions and if they had at least an arbitrarily specified number
of answers for the remaining questions. The information from the 602 accept¬
able questionnaires was keypunched onto computer cards and the SPSS package
of computer programs used to compute frequency distributions and cross-tabu¬
lations of the data. The remainder of this section describes the salient
socioeconomic and shopping trip characteristics of the sample.
54
3.3.2 Characteristics of the Second-Survey Sample
The income distribution in the second sample is even more heavily skewed
toward high incomes than is the distribution of the first survey. Table 3-20
shows that when nonresponses are deleted more than 60% of the sample list
household incomes exceeding $20,000. In the first survey just under 50% of
the sample reported incomes greater than $20,000. Higher reported incomes in
the more recent survey may in part be accounted for by the nature of shopping
opportunities at the distribution site. Old Orchard has department stores
and specialty shops which are more prestigious and expensive than those of
the other three distribution sites used in the first survey — Edens Plaza,
Golf Mill and Plaza del Lago. Hence, it is likely that Old Orchard attracts
a wealthier clientele.
Similar to the first sample, the population of the second survey is
heavily biased in terms of sex. The respondents are 84.2% female and 14.6%
male (1.2% did not designate their sex). The female, male, no response per¬
centages in the 1974 sample are 76%, 20% and 4%, respectively.
The length of residence information, shown in Table 3-21, indicates a
highly stable sample population. Excluding nonresponses, more than 83% of
those who answered the questionnaire have lived in the north suburban' Chicago
area for over six years. In the first survey, just over 75% of the respon¬
dents reported having resided in the same area for more than six years. Thus,
the population of the Old Orchard survey is even more stable than that of the
earlier survey.
Tables 3-22 and 3-23 demonstrate that sex has little correlation with
either income or length of residence. In each table, entries are expressed
as percentages of row totals and nonresponses have been omitted. The first
survey data also indicated a lack of correlation among the biased variables
of sex, income, and length of residence.
The breakdown of the Old Orchard sample on the occupational variable is
not entirely analogous to the job distribution of the 1974 sample. Table 3-24
shows that, in the second survey, housewife is the most prevalent occupation
(31.6%) followed by professional (19.3%) and student (18.1%). In the first
survey, housewife is the primary occupation of 25.7% of the sample; student
and professional follow at 23.1% and 16.1%, respectively. The decline in per¬
centage of students in the second sample may be accounted for in part by the
already-mentioned fact that Old Orchard is more expensive than the other dis¬
tribution sites used in the first survey.
With respect to age, the second survey is skewed toward younger groups,
though not quite as heavily as is the first survey. Table 3-25 shows that the
largest age concentration is in the 22-29 year category. 43.7% of the total
respondents are under 30 years of age compared with a figure of 50.5% for the
first survey. The drop in percentage of younger respondents in the second
survey relative to the earlier survey is consistent with the decline in the
proportion of students noted earlier.
55
Income Group
Frequency (%)
$10,000 and under
43 (7.1)
$10,001 - $15,000
73 (12.1)
$15,001 - $20,000
87 (14.5)
$20,001 - $25,000
83 (13.8)
$25,001 - $30,000
71 (11.8)
$30,001 - $35,000
42 (7.0)
$35,001 - $40,000
32 (5.3)
$40,001 - $45,000
21 (3.5)
$45,001 - $50,000
12 (2.0)
Over $50,000
79 (13.1)
No Response
59 (9.8)
TABLE 3-20
Income Distribution for Second-Survey Sample
56
LENGTH OF RESIDENCE
FREQUENCY (%)
Not a resident
11 (1.8)
Less than 1 year
9 (1.5)
1 year
15 (2.5)
2 years
23 (3.8)
3 years
22 (3.7)
4 years
14 (2.3)
5 years
25 (4.2)
6 years
19 (3.2)
7 years
21 (3.5)
8 years
17 (2.8)
9 years
12 (2.0)
10 or more years
412 (58.4)
TABLE 3-21
Length of Residence Distribution for Second Sample
57
INCOME
SEX
Under
$10,000
$10,001-
$15,000
$15,001-
$20,000
$20,001-
$25,000
$25,001-
$30,000
$30,001-
$35,000
$35,001-
$40,000
$40,001-
$45,000
$45,001-
$50,000
Over
$50,000
Female
7.7
12.8
15.2
14.8
13.9
8.4
6.6
3.5
2.4
14.6
Male
7.2
18.1
20.5
18.1
9.6
3.6
2.4
4.8
1.2
14.5
TABLE 3-22
Distribution of Income by Sex for Second Sample
LENGTH OF RESIDENCE
SEX
Not a
Resident
Less than
1 year
1
year
2
years
3
years
4 | 5
years I years
6
years
7
years
8
years
9
years
10 years
and over
Female
1.6
1.2
2.4
3.6
3.6
2.4 j 4.6
2.2
3.2
2.4
2.2
70.9
Male
3.4
3.4
3.4
5.7
3.4
2.3 ! 2.3
I
i...
8.0
4.5
5.7
1.1
56.8
TABLE 3-23
Distribution of Length of Residence by Sex for Second Sample
OCCUPATION
FREQUENCY (%)
Salesman
18 (3.0)
Teacher
71 (11.8)
Professional
116 (19.3)
Craftsman
2 ( .3)
Clerical
46 (7.6)
Student
109 (18.1)
Housewife
190 (31.6)
Governmental
2 ( .3)
Retired
19 C3.2)
Other
26 (4.3)
No Response
3 ( .5)
TABLE 3-24
Occupation Distribution for Second Sample
59
AGE
FREQUENCY (%)
Under 22 years
117 (19.4)
22 - 29 years
145 (24.1)
30 - 39 years
117 (19.4)
40 - 49 years
111 (18.4)
50 - 59 years
68 (11.3)
60 - 65 years
23 (3.8)
Over 65 years
19 (3.2)
No response
2 (.3)
TABLE 3-25
Age Distribution for Second Sample
60
Cross-tabulations on the second-survey data show that 78.4% of those
under 22 years of age are students. The most prevalent occupation among 22
to 29-year-olds is professional with 26.9% falling into this job category.
For ages 30 through 65, housewife is the prominent occupation. Over 65 years
the majority of respondents are retired.
There appears to be a correlation between sex and occupation in the
second sample. Housewife is the principal occupation of females with 37.7%
of them falling into this category. With 52.3% of the male responses, pro¬
fessional ranks as the largest occupational category for men. Student is
the second most frequent occupation designated by both females (18.8%) and
males (15.9%).
The second sample is well-educated. As shown in Table 3-26, more than
81% of the respondents have received schooling beyond the high school level.
A breakdown of the sample in terms of stages of the life-cycle is given in
Table 3-27. In this table the category young consists of all ages up to and
including age 39, while old includes ages 40 and over. As shown, the highest
percentage of respondents are in the life-cycle category of young and unmar¬
ried.
All trip information gathered in this survey pertain to the particular
shopping trip respondents made when they received their questionnaires at Old
Orchard. As shown in Table 3-28, there is a heavy bias towards the use of
the automobile. 93.3% of the sample made the trip to and from the shopping
center by car (either as driver or passenger). This percentage is almost
identical with the analogous figure — 93.2% -- for the first survey. The
second most prevalent mode used to reach Old Orchard was bus. 3.5% of the
sample used this means of travel compared to 2% bus usage in the previous
year.
As was the case with the first survey, the Old Orchard survey yielded
little evidence of multiple-destination trips. As indicated in Table 3-29,
59.2% of the respondents traveled directly to the shopping center from home
and returned immediately to home upon leaving the center. This is close to
the figure of 64.2% for the previous survey. Table 3-30 shows that home was
the immediate origin of the shopping trip in 76.7% of the cases in this sur¬
vey. This compares with a figure of 81% in the previous year. It can be
seen in Table 3-31 that home was the immediate destination upon leaving the
shopping center in 75% of the cases in 1975. The comparable figure for the
previous year is 76%.
The Old Orchard data lend support to the hypothesis that a shopping
center serves as a multiple-destination opportunity. However, the support
given by this data was not quite as strong as that coming from the first
survey data. 42% of the Old Orchard shoppers purchased items from at least
three categories, while almost 50% of the shoppers .did so in the previous
year.
As shown in Table 3-32, clothing was the most popular shopping item among
the Old Orchard sample. More than 91% of the respondents listed it as one of
the major items they came to purchase during their shopping trip. The next
most popular items were window shopping and drugs-cosmetics which were sought
by 22.9% and 16.6%, respectively, of the respondents.
61
Highest Level
of Education
Frequency {%)
High School Degree or Less
Some College
College Degree
Masters Degree
Doctorate
Other
No Response
111 (18.6)
185 (30.9)
192 (32.1)
68 (11.4)
12 (2.0)
30 (5.0)
4 (.7)
TABLE 3-26
Educational Distribution for Old Orchard Sample
STAGE IN LIFE CYCLE
FREQUENCY (35)
Young Unmarried
195 (32.4)
Old Unmarried
35 (5.8)
Young Married, No Children
44 (7.3)
Old Married, No Children
63 (10.5)
Married, Oldest Child 0-5 Years
68 (11.3)
Married, Oldest Child 6-11 Years
56 (9.3)
Married, Oldest Child 12-18 Years
62 (10.3)
Married, Oldest Child 19-22 Years
68 (11.3)
No Response
11 (1.8)
TABLE 3-27
Life Cycle Distribution for Old Orchard Sample
62
Mode Used
Frequency (%)
Auto Driver
469 (77.9)
Auto Passenger
93 (15.4)
Taxi
1 (-2)
Bus
21 (3.5)
Bicycle
4 (.7)
Walking
14 (2.3)
TABLE 3-28
Mode of Travel Used to Old Orchard
Origin
Des ti n at iofr*^-^.^
Other
Shopping
Work
Home
Other
Other
Shopping
3( • 5)
3 ( - 5)
41(6.9)
8(1.3)
Work
0(0)
11(1.9)
9(1.5)
0(0)
Home
32(5.4)
38(6.4)
351(59.2)
24(4.0)
Other
4( .7)
4( .7)
54(9.1)
11(1.9)
TABLE 3-29
Distribution of Origin by Destination
(Figures in Parentheses are percentages)
63
Previous Origin
Pecentage
Other Shopping Locations
6.6
Work
9.6
Home
76.7
Other
7.1
TABLE 3-30
Distribution of Previous Origin for Old Orchard Respondents
Next Destination
Percentage
Other Shopping Locations
9.3
Work
3.4
Home
75.0
Other
12.3
TABLE 3-31
Distribution of Next Destination for Old Orchard Respondents
64
Item Shopped For
Frequency {%)*
Clothing
550 (91.4)
Window Shopping
138 (22.9)
Drugs and Cosmetics
100 (16.6)
Eating Out
85 (14.1)
Books and Stationery
81 (13.4)
Housewares
69 (11.5)
Home Furnishings
68 (11.4)
Jewelry, Clocks, Watches
57 (9.5)
China, Glass, Flatware
50 (8.3)
Other
129 (21.4)
TABLE 3-32
Main Items Shopped For by Old Orchard Respondents
♦Percentages refer to the portion of the
sample which shopped for the various items.
SHOPPING LOCATIONS
FAMILIARITY
Edens
Plaza
Old
Orchard
Golf
Mill
Downtown
Evans ton
Wood-
field
No. Mich¬
igan Ave.
State
Street
Downtown
Skokie
Dempster
Street
Never heard
of
1.7
0
.5
.7
.7
2.0
0
5.2
11.6
Heard of,
not visited
6.4
0
5.0
9.7
10.6
4.7
2.7
27.9
30.3
TABLE 3-33
Familiarity Information on Main Subject Set
65
Responses to the familiarity question indicate that members of the Old
Orchard sample are well-acquainted with all of the shopping opportunities in
the main subject set except for the two strip developments. For each of the
nine locations in the subject set, Table 3-33 gives the percentage of the
sample that never heard of the location and the percentage that never visited
the location. The table shows that respondents are least familiar with the
two strip developments. 27.9% of those responding to the familiarity query
on Downtown Skokie have never visited that location; 30.3% of those who
answered the question on Dempster Street never visited that location. Still,
only 5.2% and 11.6% of the respondents never head of Downtown Skokie and
Dempster Street, respectively. Thus, most of the sample had some basis on
which to compare all of the shopping opportunities in the subject set. This
suggests that the responses on the questions pertaining to multidimensional
scaling are reasonably reliable.
3.4 The Mode-Choice Survey
3.4.1 Survey Design and Execution
The location chosen for the mode-choice survey was downtown Evanston,
Illinois. Evanston is a suburb of Chicago, located immediately north of the
city on Lake Michigan, and has a population of approximately 80,000. The
downtown area of about .75 square miles is typical of older American cities,
with shop-lined streets rather than shopping malls. There are two large
department stores and numerous specialty shops, along with several restaur¬
ants and office buildings.
Downtown Evanston was chosen as the site of the mode-choice survey for
several reasons. Unlike the large shopping mall where the destination-choice
information was obtained, the downtown Evanston area is well served by many
modes of transportation; the shopping malls are accessible principally by car
alone. In addition to having convenient parking, Evanston is served by a
number of bus routes, by the Chicago Transit Authority's elevated rapid tran¬
sit trains, and by the Chicago and Northwestern commuter railroad.
Another reason for choosing downtown Evanston for the shopping mode-
choice survey is that most pedestrians in the downtown Evanston area are
shoppers, rather than workers; other than store employees, there are rela¬
tively few jobs in the shopping area of Evanston. An early candidate for the
shopping mode-choice survey location, the downtown Chicago (Loop) area, was
rejected because too many of the shoppers in that area also work there; it
would not be possible to separate out shopping-trip mode-choice decisions
from work-trip mode-choice decisions.
Lastly, an additional reason for choosing downtown Evanston as the survey
site was its proximity to Northwestern University; it is about a five minute
walk from the Transportation Center to the survey area. This proximity greatly
reduced the cost and complexity of the survey effort.
The basic information needed from each respondent to the mode-choice
survey is the identification of the transport mode used by the respondent
for his shopping trip to Evanston, the identification of the mode the re¬
spondent would have used if he did not use the mode actually chosen, and the
66
characteristics for the chosen and alternative modes for the trip to downtown
Evanston of time and cost. Both time and cost should be obtained in such a
way as to make possible trip disaggregation into access, wait, in-vehicle and
egress segments. Information on the respondent's individual characteristics,
and information other than times and costs for the respondent's shopping trip
were also requested in order to build more sophisticated models and to enable
various hypotheses about shopping behavior to be tested.
Much research has shown that travel behavior varies with the character¬
istics of the individual traveler. In order to test this hypothesis, socio¬
economic data such as sex, age, stage in family life cycle, educational level,
and income are desirable. An additional variable that has been shown to be a
significant determinant of travel behavior in some cases is the car competi¬
tion ratio which is determined from number of cars in household divided by
number of licensed drivers in household. These two items of data were,
therefore, also requested.
It has been hypothesized that mode-choice behavior, especially for shop¬
ping trips, may be determined by modal factors other than times and costs.
One factor that has been found to be significant is the ability to carry pur¬
chases home on various modes; if present, this factor would be expected to
generate an increased preference for the automobile over public transit.
Another factor likely to affect mode-choice behavior is the size of the group
traveling together. It has been hypothesized that small groups traveling to¬
gether are more likely to choose the automobile, since the marginal cost of
an additional occupant in an automobile is close to zero until capacity is
reached. Information on the size and composition of the traveling group is
needed to test this hypothesis.
As with the previous surveys, it was decided to use a hand-out, mail-back
questionnaire and administer it to a choice-based sample. The reasons for
this choice are basically those described for the destination-choice surveys,
together with the increasing familiarity and success achieved by the staff
with this form of survey.
The questionnaires were distributed in downtown Evanston on Saturday,
June 28, 1975 and on Monday, June 30, 1975. A weekday and a weekend day were
chosen to allow for the possibilities that the characteristics of the shopping
population might differ between weekdays and weekends, or that the same shop¬
pers might behave differently on weekdays and weekends. Distribution was
accomplished by having project employees approach shoppers on the street,
describe briefly the aims of the questionnaire, and request that the shoppers
take a questionnaire and fill it out at home. Approximately 2,500 question¬
naires were distributed each day.
67
All questionnaires distributed had a postpaid envelope attached. Returned
questionnaires started arriving at the project office on the second working day
following questionnaire distribution. The third and fourth working days follow¬
ing distribution saw the greatest number of returns; by a week after distribution
the daily number of returned questionnaires was down to a trickle. When collec¬
tion was terminated two weeks after distribution, a total of 1,560 questionnaires
had been returned (777 from the Saturday distribution and 783 from the Monday
distribution), giving and overall response rate of 31.2%, considerably above
our planning estimate of 15%.
Returned questionnaires were required to meet certain criteria before
any coding could begin. Questionnaires were rejected when the information
given on the chosen mode was incomplete, unless the missing information was
such as could be obtained from other sources such as fare tables or schedules.
Questionnaires were also rejected when the information on the alternative mode
was incomplete and unobtainable, unless the respondent gave an address from
which a public transit alternative could be constructed.
A total of 1,159 questionnaires (599 from Saturday and 560 from Monday)
were coded and keypunched. Those questionnaires with incomplete but obtain¬
able mode information were coded and keypunched as received; no effort was
made at this time to generate synthetic mode information. Of the 1,159 ques¬
tionnaires coded and punched, 840 (439 from Saturday and 401 from Monday) had
complete information on chosen and alternative modes; this is the sample used
in project work based upon the downtown Evanston survey.
Of those responding, 33% were male and 67% were female. When cross-tab¬
ulated against the other socioeconomic variables, there were no noticeable
differences between the characteristics of the male population and the char¬
acteristics of the female population. The distribution of ages in the sample
is shown in Table 3-34. When cross-tabulated with income, age shows the
expected positive correlation. When cross-tabulated with years of residence,
length of residence shows a steady increase with age, with the exception of the
youngest group, who show a relatively long length of residence. Again, this
is an expected result.
Eight possible stages are defined for the family life-cycle variable.
These stages are listed in Table 3-35, along with the population distribu¬
tion .
The distribution of the educational level within the sample is given in
Table 3-36 and shows a similar distribution to that found in the second des¬
tination-choice survey. The distribution of the respondents' lengths of
residence are given in Table 3-37 and show again a fairly stable population,
with almost 70% of the residents reporting having lived in the area for more
than four years. Table 3-38 summarizes the distribution of family incomes
for the sample. This distribution is, as expected, rather different from
that of the destination-choice surveys, with a much smaller bias towards
higher income groups.
68
Age *
Under
22
22 -
29
30 -
39
40 -
49
50 -
59
60 -
65
Over
65
Frequency
20.2%
34.3%
12.9%
12.1%
9.5%
5.6%
5.4%
TABLE 3-34
Age Distribution for Evanston Survey
Stage
Young*
Unmarried
Old
Unmarried
Young, Married
No children
Old, Married
No children
Married,
oldest**
child 0-5
Married,
oldest
child 6-11
Married,
oldest
child 12-18
Married
oldest
child 19-22
Unknown
Freq.
41.3%
11.9%
14.4%
8.6%
6.4%
4.3%
7.4%
4.4%
1.3%
TABLE 3-35
Stage in Family Life-Cycle
*The dividing line between young and old is set at 40 years.
**"01dest child" refers to the oldest child currently living at home.
Level
Less than
high school
High
School
Some
College
BA/
BS
MA/ | Ph.D.
MS j
Other
degree
Freq.
7.0%
-
10.2%
25.1%
31.6%
18.0% j 6.7%
1
2.0%
î
!
TABLE 3-36
Distribution of Education Levels for the Evanston Survey
Length of
residence
Less than
4 years
4-6
years
More than
6 years
. - T
Non-resident
Freq.
23.1%
12.6%
39.9%
24.7%
I
TABLE 3-37
Distribution of Length of Residence
for the Evanston Survey
Income
Under
$10,000
$10,000-
$15,000
$15,000-
$20,000
$20,000-
$25,000
$25,000-
$50,000
Over
$50,000
No
Answer
Freq.
23.5%
18.0%
14.0%
12.6%
21.5%
4.0%
6.3%
TABLE 3-38
Distribution of Family Income for the Evanston Survey
70
Contrary to the original hypothesis, there does not seem to be any sig¬
nificant difference between the population shopping on weekends and on week¬
days. Cross-tabulations of the day handed out with the chosen and alternative
modes, and with the set of socioeconomic variables show a close match between
the weekday and weekend sample characteristics. Examples are shown in Tables
3-39 and 3-40.
Nine possible modes were defined for travel to downtown Evanston. Tables
3-41 and 3-42 show the distribution of the modes chosen and the alternative
modes for the shopping trip to Evanston. Table 3-42 shows a cross-tabulation
for the chosen mode against the alternative mode. To increase readability,
the set of mode alternatives has been collapsed by grouping all public transit
options together and grouping auto driver with auto passenger. The entries
represent the number of observations in each element of the table.
Cross-tabulations were also performed for the mode chosen against various
socioeconomic variables as a prelude to the use of socioeconomic variables in
the mode-choice modeling. When cross-tabulated against sex, mode use shows a
rather uniform distribution, with the exception of auto users. About 50% of
the males and 50% of the females used autos, but of these, only 6% of the
males were passengers, while 30% of the female auto users were passengers and
only 70% drivers.
The cross-tabulation of mode used against stage in family life cycle
shows a dichotomy between those with children at home and those without.
Individuals without children at home were distributed rather evenly across
all modes, while those with children at home were much more concentrated in
the auto-user modes. This may be a result of the hypothesis on group behav¬
ior discussed above.
Family income is one of the most frequently used socioeconomic variables
in travel analysis. The cross-tabulation of income against mode used is given
in Table 3-43. The entries represent the number of observations in each cell
of the table, with the column percentage in parentheses below.
71
Mode
Auto
Driver
Auto
Passenger
Taxi
Walk
Bicycle
Bus
El
Railroad
Mi xed
Chosen mode-
Saturday
frequency
43%
13%
12%
23%
2%
11%
6%
0
1%
Chosen mode-
Monday
frequency
38%
8%
.7%
26%
3%
13%
6%
3%
2%
Al ternative
mode-
Saturday
frequency
22%
13%
4%
14%
1%
25%
15%
2%
4%
Alternative
mode-
Monday
25%
15%
2%
11%
2%
25%
15%
2%
4%
frequency
TABLE 3-39
Cross-tabulation of Travel Mode and Shopping Day
for the Evanston Survey
Income
under
$10,000
$10,000-
$15,000
$15,000-
$20,000
$20,000-
$25,000
$25,000-
$50,000
Over
$50,000
No
Answer
Saturday
frequency
23%
18%
14%
13%
21%
4%
7%
Monday
. frequency
24%
18%
15%
12%
22%
4%
6%
TABLE 3-40
Cross-tabulation of Income and Shopping Day
for the Evanston Survey
72
Mode
Auto
Driver
Auto
Passenger
Taxi
Walk
Bike
Bus
El
Rail road
Mi xed
Mode*
Chosen
mode
frequency
40.5%
10.7%
.5%
24.5%
3.5%
12.2%
6.0%
1 .4%
1 .71
Alterna¬
tive mode
frequency
23.6%
14.0%
2.9%
12.5%
1 .3%
25.2%
14.5%
1 .8%
4.0%
TABLE 3-41
Distribution of Chosen and Alternative Modes
for the Evanston Survey
*Mixed-mode refers only to a combination
of public transit modes, e.g. bus and el.
>sAlternative
mode
Chosen
mode
Pub!ic
Transit
Auto
Taxi
Wal k
Bike
TOTAL
Public transit
63
79
7
26
2
177
Auto
250
100
7
68
5
430
Taxi
3
1
0
0
0
4
Walk
62
125
10
6
4
207
Bike
5
11
0
5
0
21
Total
383
316
24
105
11
839
TABLE 3-42
Cross-tabulation of Chosen and Alternative Modes
for the Evanston Survey
73
Income
ModeV.
Used
Under
$10,000
$10,000-
$15,000
$15,000-
$20,000
$20,000-
$25,000
$25,000-
$50,000
Over
$50,000
No
Answer
Total
Driver
41
(21%)
61
(40%)
62
(53%)
43
(41%)
93
(51%)
17
(50%)
23 i
(43%)
340
Passenger
12
(6%)
17
(11%)
15
(13%)
16
(15%)
22
(12%)
5
(15%)
3
(6%)
90
Taxi
2
(1%)
0
0
0
(1%)
0
1
(2%)
4
Walk
82
(42%)
29
(19%)
23 •
(20%)
22
(21%)
33
(18%)
8
(24%)
10
(19%)
207
Bicycle
6
(3%)
7
(5%)
2
(2%)
3
(3%)
3
(2%)
0
0
21
Bus
36
(18%)
21
(14%)
7
(6%)
12
(11%)
15
(8%)
3
(9%)
8
(15%)
102
El
12
(6%)
13
(9%)
5
(4%)
4
(4%)
10
(6%)
1
(3%)
5
(10%)
50
Railroad
4
(2%)
(1%)
0
3
(3%)
(1%)
0
2
(4%)
12
Mixed
(1%)
(il)
4
(3%)
3
(3%)
3
(2%)
0
1
(2%)
14
Total
197
151
118
106
181
34
53
840
TABLE 3-43
Cross-tabulation of Mode and Income
for the Evanston Survey
74
4. METHODOLOGY OF THE APPROACH
4.1 A Marketing Approach
Innovations in transportation services are meant to improve service, or
reduce costs, or both. The success of such innovation depends upon how con¬
sumers respond by changing their travel habits. A successful innovation will
attract and retain riders by providing them with options they view as superior.
To insure success planners and managers must design new services based upon an
understanding of consumer needs and desires, and must implement the service with
an integrated marketing strategy.
This paper presents a methodology for innovation in transportation services.
This marketing approach for design, implementation, and monitoring of service
draws upon state-of-the-art knowledge and experience in marketing and travel
demand forecasting. The design phase uses measures and models of transportation
consumer behavior to provide managers w.ith.(l) diagnostic information which
suggests how best to improve the mix of transportation services and (2) predic¬
tions of consumer response needed to evaluate the suggested improvements. The
implementation phase then sets the marketing mix (service strategy, service
availability, fares, advertising, promotions, etc.) and evaluates the success
of the implemented strategy in enhancing ridership or in meeting other objec¬
tives. This phase allows dynamic modification of the marketing and service
strategies so that managerial goals can best be realized. Finally, the moni¬
toring phase identifies changes in travel needs and in the environment. This
enables the transportation manager to react rapidly to the new developments
in the social, political, economic, and cultural environments.
The discussion in this paper is adapted from approaches that have been
proven successful in the design of innovative health maintenance organizations,
management education programs, financial services and banking services [22],
in more than seven separate categories of new consumer products [51], and in
both consumer products and durables [35]. This success is based on the con¬
sumer-oriented common sense approach which carefully considers all important
effects influencing market acceptance. The approach is not one of simply
selling the service, or of choosing appropriate advertising and promotion.
These components are important in getting people to try a new service, but
they cannot stand alone. Success requires repeat riders and this comes only
through design of the service to meet the needs of consumers.
The marketing approach complements both the standard planning approach
[2, 9, 40, 45, 50, 51, 52] and recent developments in transportation attitu-
dinal research [6, 7, 11, 33, 41]. The emphasis is not on particular models,
but rather on the solution of managerial problems through the consumer orien¬
tation of marketing research.
This paper presents a conceptual approach illustrated with particular
models and some empirical experience. It draws on previous research in trans¬
portation and marketing and integrates these developments into a single pack¬
age. Although each component of the approach is illustrated with a single
model, the technical references give many related models of varying complexity
and accuracy.
75
4.1.1 Overview of the marketing approach
If a community introduces a "better" public transportation service
chances are the public will not immediately adopt it for much of their
daily travel. First, they must become aware of the service, they must
be convinced that it is a superior option for them, and they must deter¬
mine that it satisfies their particular trip needs. Only then will they
try it. Once they use the new system, their perceptions of it may be altered
and their future behavior (e.g. repeat or not) will depend upon their
experience. That is, the consumer response process is not a simple one
step process, but rather a complex series of stages. The better a manager
or planner understands this process, the better they can design and imple¬
ment service changes. The consumer-oriented approach consists of three
phases; design, implementation and monitoring; which represent chrono¬
logical steps in the evolution of a service strategy and fare structure
based on analysis of consumer responses to alternative transportation
services. See Figure 4-1.
The design phase begins with qualitative studies which identify
opportunities for service improvement, identify design characteristics
which are important to the consumer, and give the manager and planner
a basic understanding of the riding public. Qualitative studies are a
necessary first step but innovation requires models and measures which
can help the manager make specific decisions. Thus the design phase
next provides quantified diagnostic information on market structure,
consumer desires, and segmentation. This diagnostic information tells
the manager which service attributes to concentrate on, what trade-offs
to make in the design of service, and how to set varied service strategy
for segments of the public. Even with the best design the manager must
make an initial GO/NO GO decision on implementation. The final step in
the design phase is to predict consumer response to the proposed changes.
The design phase leads to the development of a superior service for
the needs of the target population and a better estimate of expected
performance. The next phase, implementation, establishes the advertising
and promotion strategy and measures consumer reaction to the service
innovation. The advertising and promotion strategy is designed to make
consumers aware of the system and induce them to try it. Measures of
consumer awareness, availability, initial trial, repeat usage, and satis¬
faction allow the manager to control the implementation and to immediately
identify when consumer responses differ substantially from those predicted.
This information enables the operator to respond quickly to improve his
service or marketing mix. This process helps insure that innovations will
meet their stated objectives.
The final phase in the process is monitoring service operations and
traveler responses on a continuing basis. Once ridership has fully
responded to changes, the intensive consumer measurement of the implemen¬
tation phase is no longer necessary. This does not mean that the operator
76
DESIGN
L
[
Qualitative
Measures
Diagnostic
Models
Prediction/Evaluation
Models
&
Idea
Generation
Potential Service
Strategies
Best initial service
strategy
)
I
j
1
IMPLEMENTATION
[
Ridership Growth
Models
M
I Measure
I Aval1abi1
Land Post-
Awareness , n c
ility, Trial Ratel P <
t-purchase I 1
Behavior V
Set initial service
Strategy and
Marketing Mix
Modify Strategy and
Marketing Mix 1f
Necessary
]
MONITORING
1
Periodic Consumer
Measures
V Improve System or )
f Marketing Mix if I
V. Necessary /
*
FIGURE 4-1
Integrated Marketing Approach to
Transportation Service
77
can ignore consumer behavior. The consumer-oriented marketing approach
continues to monitor consumer behavior so that the transit manager can stay
in touch with the changing needs and desires of the riding public. This
will maintain a quality service with sufficient patronage.
The remainder of the paper develops these basic concepts with discussions
of the managerial issues. Simple examples are given to illustrate the con¬
cepts. Readers are directed to the technical references for details and more
complete examples.
4.1.2 The Design Phase
The purpose of the design phase is to develop transportation service
innovations which satisfy community objectives such as increasing revenues
by increasing ridership. In other cases, a wider range of objectives such
as provision of service to special groups or reduction in road congestion,
fuel consumption, and air pollution may be relevant. To design such service
strategies, a transportation operator must have an understanding of how
consumers respond to transportation service characteristics. Figure 4-2
adapted from models in marketing and in transportation [15, 19, 34],
describes this consumer response process, the boxes represent elements of
consumer response including, how they acquire information, form perceptions,
evaluate service, make transportation choices, and acquire improved information
by experience. The circles represent measurable characteristics of the
transportation service, advertising, and promotion and public image of the
transportation system. Other important measures include individual percep¬
tions of service, tastes, goodness ratings, service availability for trip
needs, and choice.
To better understand Figure 4-2, imagine that a community introduces a
new bus route connecting your neighborhood to the downtown area. The transit
company sets the system characteristics (circle 1) and advertising, promotion,
etc. (circle 2). You hear about the service by radio, newspaper, brochure,
or by word-of-mouth (box A). You now have information about the service
(circle 3) and form perceptions (box B) of the attributes (comfort, travel
time, safety, cost) of the service. You consider other alternatives such
as auto or bicycle and evaluate them (box C) based on your personal tastes
(circle 4). (For example, do you want a fast but expensive service or a slow
inexpensive one.) Then if the route serves your particular trip needs
(circle 6) you might choose it for your next trip downtown (box D). Finally,
the experience improves your information about the system (box E) and feeds
back to influence each aspect of your choice process.
Each step is modeled separately because each step provides important
information to the transportation manager, and because service and marketing
strategies can be designed to influence each step. For example, it is impor¬
tant for a manager to know what dimensions, e.g. reliability, influence con¬
sumers' evaluations of transportation systems. Furthermore, it is important
to measure how consumers perceive each existing or potential system relative
to these dimensions and to know how these dimensions influence consumer
behavior. Armed with this information a manager can select those system or
marketing strategies that best serve consumers. Finally, predictions of
78
Consumer Response Process
79
consumer behavior coupled with cost considerations enable managers to eval¬
uate strategies and choose the strategy which best fulfills their managerial
goals.
Thus the design process is a set of consumer measurements and consumer
models that give managers the marketing information to make strategic deci¬
sions. In particular, the design process gives outputs in the form of
measures of public image, consumer perceptions, and preferences by consumer
segment (circle 3, 4, 5, and 7) and predictions of consumer choice behavior
(circle 6 and 8).
4.1.3 Public Image
There are both quantifiable (survey) and qualitative measures of public
image. The qualitative studies provide a necessary subjective input to iden¬
tify issues and to look at the service through the eyes of the consumer.
Focus groups [24, 29] have proven particularly efficient and useful for
qualitative studies. Focus groups are groups of 6-8 consumers brought
together and encouraged by a moderator to discuss their attitudes toward
existing transportation alternatives and to indicate how they now make
choices. Group dynamics become important and skillful moderation is essen¬
tial. In addition to focus groups, questionnaires requiring open-ended
responses, intercept interviews, citizens groups, and library research to
uncover past studies are useful qualitative techniques. These techniques
are limited only by time, cost, and imagination.
In interpreting these studies, managers must recognize that these quali¬
tative measures are not from a representative sample and tend to favor the
more outspoken and articulate consumers. The emphasis of the qualitative
analysis is on breadth of ideas and identification of important attributes.
These are then quantitatively analyzed in the next step of the design phase.
An important output of the qualitative analysis is a set of questions which
can be used to measure the attributes relevant to the consumers' choices.
These attributes, or image measures, are then quantified through consumer
surveys [37, 38]. For example Figure 4-3 is a comparative "map" of consumers'
images of four shopping centers in the Chicago area.
4.1.4 Consumer Perceptions
Image "maps" such as Figure 3 are useful to describe how consumers view
detailed aspects of a transportation related service, but for designing
systems it is hard to get clear insights from so much complex information.
Thus to understand the true perceptual process, to gain managerial insight,
and to enhance creative strategy development, the design phase uses models to
identify the cognitive structure of consumer perceptions. These perceptual
models either reduce the set of attributes through factor analysis [43],
discriminant ability [23, 39], or they independently uncover the dimensions
based on measures of dissimilarity [12]. Of these models, factor analysis
seems to predict best and cost the least to use [17].
80
1. Layout of store
2. Ease of returning or
servicing merchandise
3. Prestige of store
4. Variety or range of
merchandise
5. Quality of merchandise
6. Availability of credit
7. Reasonable price
8. Availability of sale items
("specials")
9. Free parking
10. Stores located in compact
area
11. Store atmosphere (heating,
cooling, noise, crowds, etc.)
12. Ability to park where you
want
13. Shopping center atmosphere
14. Courteous helpful sales
assistants
15. Availability of a specific
store
16. Number and variety of stores
£
• Woodfield
O Chicago Loop
□ Plaza del Lago
X Korvette City
FIGURE 4-3
Average Ratings for Four Shopping Centers on
the Underlying Perceptual Scales
81
For example, the factor loadings matrix In Figure 4-4 Indicates that
there are four basic factors underlying perceptions of shopping centers:
variety, shopping satisfaction, price/value, and parking. Variety includes
variety of stores, variety of merchandise, and the availability of specific
stores. Shopping satisfaction Includes ease of purchasing (.layout, return
and service, courteous sales assistants], atmosphere, quality of merchandise,
and prestige of store. Price/value includes reasonable price, specials,
credit, and to some extent ease of returning. Finally, parking includes
availability, parking cost, and layout of the shopping center- The factor
analysis also produces measures of how each consumer perceives each trans¬
portation alternative on the factors. Figure 4-5 gives average perceptions
for the same four shopping centers mapped in Figure 4-3. Maps such as Figure 4-5
provide managers with insight on how their services are perceived relative
to their competitors. Gaps in the market as well as their service's
strengths and weaknesses can be readily identified from these maps. For
example, Plaza del Lago is a North Shore prestige shopping center with
Spanish architecture and exclusive stores. As expected, our map shows that
it is enjoyable to shop there but variety and value are better at other
locations.
4.1.5 Preference Models
The perceptual maps identify market structure but they do not describe
the relative importances of the perceptual factors in the consumer's eval¬
uation process. Joint analysis of preference rankings and perceptions (both
collected in a mailed survey) identifies the effect that each perceptual
factor has in the consumer's evaluation process. There are a number of
methods available to accomplish this task including preference regression
[46], expectancy value [10, 42], direct consumer utility assessment [20],
trade-off analysis [22], conjoint analysis [13], and logit analysis [40].
These models share the common property that they represent the evaluation
of an alternative as a function of the perceptions of the factors. They
produce scalar goodness ratings (circle 7 in Figure 4-2), representing each
consumer's evaluation of each transportation alternative. Many of these
models identify weights which indicate to managers the relative importances
of the perceptual dimensions. These weights indicate which attributes shoe
be improved to obtain maximum consumer impact. For example, suppose the
importance weight for reliability is significantly larger than the impor¬
tance weight for comfort. If the costs of improving reliability and comfort
are equal, the manager should concentrate on improving reliability.
Empirical tests [4, 17, 28] have shown many preference models to be
robust in that they give comparable predictions or importance weights. For
example, both preference regression and logit analysis select satisfaction
as the most important perceptual dimension and parking as the least important
dimension in Figure 4-5 [17]. (Actual normalized weights for logit are variety,
.38, satisfaction, .54, parking, .01, and price/value, .07.) These weights
partially explain why Woodfield has the major share of preference even though
Korvette City is slightly superior in both parking and price/value.
82
In this question, we would like you to rate each of the shopping centers
on these characteristics. We have provided a range from good to poor for
each characteristics. We would like you to tell us where yew feel each
shopping center fits on this range.
For example:
Good
Q-
O
O
O
(O
O
•r—
SZ
o
Eating
Facilities
Poor
<D
CD
>
i-
2
s
o
CD
to
O)
"O
to
N
fO
v'
FIGURE 4-4
Example Measurement of Ratings
for Shopping Centers
83
SHOPPING
VARIETY
SATISFACTION
PARKING
PRICE/VALUE
Layout of store
.267
.583
.200
.156
Return and service
.095
.528
.255
.343
Prestige of store
.338
.822
-.058
-.001
Variety of merchandise
.665
.327
-.185
.309
Quality of merchandise
.307
.810
o
•
i
.037
Availability of credit
.159
.337
.049
.487
Reasonable price
.067
-.063
.113
.599
"Specials"
.223
.074
.008
.739
Free parking
-.150
.068
.811
.043
Center layout
.030
.308
.560
.074
Store atmosphere
.080
.658
.400
.034
Parking available
.145
.105
.841
.108
Center atmosphere
.244
.694
.404
-.040
Sales assistants
.173
.560
.319
.147
Store availability
.619
.320
.034
.204
Variety of stores
.829
.288
-.173
.160
FIGURE 4-5
Factor Loadings for Perception of Shopping Centers
84
4.1.6 Segmentation
Different individuals have different preferences for transportation
attributes. Some people favor fast, reliable, premium service over low
cost adequate service, while others favor the low cost service over the
premium service. Merging these groups may lead to estimated importances
which imply equal trade-offs between speed and reliability versus cost.
A service designed to satisfy the "average" consumers may actually satisfy
none of the consumers. Thus, we must identify market preference segments
whenever they exist. Most classification schemes fbr mode preference have
been based on socioeconomic characteristics, trip purpose, or prior trans¬
portation behavior [11, 49] or based on multiple dimensions including
socioeconomic and travel characteristics [33].
Each of these schemes serves a valid purpose because each acts as a
proxy for segmentation by preferences. For example, age is often suggested
because most analysts feel that the young have different preferences from
the elderly. An improved technique for segmentation is benefit segmentation
[14, 19]. Benefit segmentation is operationalized by first searching for
segments by prior beliefs [such as age) or by more sophisticated search
procedures and then testing these segments to see if the segments really
capture differences in preference. The criteria for the tests are (1) signif¬
icantly different importance weights between segments and (2) significantly
better explanatory power when the segments are used. Alternatively, some
models such as direct utility assessment, conjoint analysis, and expectancy
value produce importance weights for each individual. In this case the best
segmentation approach is to identify segments by these weights without regard
to demographics or travel characteristics [6, 7, 19].
Identification of these preference groups is essential to the development
of improved service strategies. This information is useful even if it is not
possible to put demographic labels on the different groups. For example, in
considering the adoption of a proposed dial-a-ride service, it is important
to know that there are distinct groups in the population who cannot readily
afford taxi fares but are prepared to pay a moderate premium over normal bus
fares for door-to-door service. It is important to do this even if it is not
possible to identify these groups in terms of their demographic or travel
characteristics.
Together the perceptual, preference, and segmentation models provide
key diagnostic information to the manager. This information helps him iden¬
tify high potential opportunities and design consumer-oriented transportation
service. But before he can select a best initial design, he needs to predict
how many consumers will actually use the system he plans to implement.
4.1.7 Choice Models and Prediction Process
Use of new transportation alternatives results from choice decisions by
numerous consumers acting independently [25]. Each consumer decides whether
or not to use the proposed service for a particular trip based on its avail¬
ability for that trip and its preference rating compared to other available
85
alternatives. If the indivldueil1 s. goodness ratings are known without error,
his predicted choice is the available alternative with the highest prefer¬
ence rating. In practice, we do not know preference ratings with certainty.
We can only predict the probability that the individual will choose each
available alternative [5, 36]. These predictions differ from those used
in the standard planning approach by the inclusion of individual perceptions
of underlying cognitive factors in addition to or in place of objective
measures of service characteristics. We evaluate a choice model in terms
of its ability to explain observed choice behavior in the sample population.
For example, on "saved data" the models discussed earlier could correctly
predict 55% of consumer preferences and 32% of the "information" in choice
behavior [16, 17, 28]. Although far from perfect these individual predic¬
tions aggregate to predict market shares within a few percentage points.
These predictions are then sufficient to decide among alternative design
strategies.
To predict system usage we must link the sequence of models (public
image, perceptions, preference, segmentation, and choice) together to obtain
estimates of changes in ridership based on changes in service character¬
istics. Although the model structure is based on analysis of individual
behavior, the system operator needs aggregate predictions of ridership.
Aggregate predictions are made by simply adding together individual predic¬
tions within the sample and projecting these to the entire population [27].
But how are individual predictions made? Predictions of changes in
behavior must be based on changes in design characteristics and their
expected influence on consumer perceptions. Consumer perceptions of the
underlying factors for a new alternative may be obtained by (1) remeasure-
ment, (2) linkage from underlying attributes and (3) operator judgment.
Remeasurement presents a sample of consumers with a description of the new
service (ranging from written descriptions to trial usage) and obtaining
ratings of the new service with respect to the same attributes used in
developing the original perception models. Linkage to underlying attributes
is based on estimating changes in individual attributes and calculating their
effect on factor scores using factor score coefficients [43]. Finally,
perception values may be obtained by judgmental 1 y comparing the proposed
alternatives to existing alternatives. In this case, factor scores which
apply to each new alternative are judgmentally estimated. Once factor scores
have been estimated prediction follows by direct substitution in the prefer¬
ence, choice models, and aggregation models.
Based on these predictions an initial design strategy is selected for
implementation. This design strategy will be the one which most nearly meets
the objectives of the operator.
Parts of section 4.1 have appeared in J.R. Hauser and F. K. Koppelman,
"Designing Transportation Services: A Marketing Approach" - Transportation
Research Forum Proceedings, 1977, 638-652.
86
4.2 Multidimensional Scaling Methods
In the past twenty to thirty years mathematical psychologists, economists
and other social scientists have developed a number of models to represent
psychological relations among stimulus objects as geometrical relations between
points in a multidimensional space. These methods have come to be known as
"multidimensional-scaling techniques". Also, researchers have attempted to
examine the effect of several independent variables interacting according to
some well specified mathematical function on a dependent variable and the
ordering of its values. In the psychological area, however, it is often diff¬
icult to measure these independent variables, let alone the dependent variable.
As a result researchers have sought to solve the measurement and interaction
mode (i.e. the functional form involving the explanatory variables) jointly,
by finding scales which closely follow the chosen functional form. These
methods are usually referred to as "conjoint-measurement methods". Actually,
as will be pointed out later, these two groups of techniques are closely rela¬
ted; specifically, Tversky (1967) and Young (1969) have shown how all so-called
nonmetric-scaling methods can be interpreted as cases of conjoint measurement.
Given the purpose of multidimensional scaling, namely to provide a geome¬
tric representation in a multidimensional space of the relationships between
a set of stimulus objects under study, it seems natural to examine first the
kinds of scaling data that can be used for such methods. Shepard (1972) pro¬
vides a very lucid and clear classification of the models and algorithms of
multidimensional scaling (MDS). Specifically, he distinguishes four basic
data classes:
(1) Proximity data: they usually consist of an n x n square matrix
in which each cell provides the measure of the dissimilarity, affin¬
ity, substitutability, correlation, congruence or interaction between
a pair (i,j) among the n stimuli. If two sets of stimuli are being
compared, the matrix is no longer necessarily square and the interpre¬
tation of each cell is that of a measure of dissimilarity, affinity
(etc.) between two stimuli from two different sets. The within-sets
relations are no longer directly observed.
(2) Dominance data: they are normally given as an n x n square matrix
of the "tournament" type (in mathematical terminology); i.e., the
(i,j)th entry measures the extent to which the ith row stimulus is
preferred (or chosen over, or "dominates" in some way) the jth column
stimulus. These dominance matrices can be replicated over the same
subject under different circumstances or over various subjects under
similar circumstances.
(3) Profile data: they are usually given in matrix form with n rows
and m columns. The rows correspond to stimuli and each column corres¬
ponds to some attribute (variable) on which the stimuli are being
judged. Thus, each row provides a stimulus profile over a given set
of criteria.
(4) Conjoint-measurement data: each entry of a p x q matrix measures
the magnitude of an effect resulting from the simultaneous interaction
of two explanatory variables taking on two specified values over their
domain of variation.
87
It is clear that a given data matrix can be treated differently depending
upon the goals of the researchers. For instance from a profile data matrix,
one can compute some measure of pairwise interprofile similarity across the
whole attribute set, and then derive a similarity matrix. The converse is also
possible since a proximity matrix can be mapped into a profile matrix of n
stimuli in k dimensions. It is also clear that many other such transformations
from one data type to another are feasible. The nature and extent of the appli¬
cable transformations is primarily determined by the experimental design adopted
in preparing the questionnaire, and the willingness of the researcher to take
the raw data obtained from these questionnaires and pre-process through a number
of preliminary steps. The models and algorithms are now discussed which have
been developed to handle the various data formats listed above.
4.2.1 Proximities Data
As explained previously the term proximities covers numerous measures of
"association" between pairs of stimuli. In some context "association" is meant
as a measure of similarity (or dissimilarity), distance, correlation, etc.
between stimulus pairs drawn (i) from the same set, or (ii) from different sets.
The measures used as input data for scaling purposes may have been derived
directly from subjective judgments of similarity between stimuli
indirectly from pairs of stimulus profiles pre-processed to obtain a single
measure. From the standpoint of analyzing the joint mode-destination choice
in transportation, we are primarily interested in finding the main perceptual
dimension of the stimuli as perceived by the respondents, their relative
salience and the general subjective processing effected by the respondents
to make an overall similarity (or dissimilarity) judgment. Before discussing
the available scaling models and their outputs for the processing of such data,
various experimental designs are reviewed briefly, from which it would be con¬
ceivable to derive the data.
(1) Alternative Experimental Designs:
a. An obvious possibility is to ask the respondent to rank all
distinct stimuli pairs according to their relative similarity (or
dissimilarity). The drawback in this method is that it may involve
too many pairs for effective processing by the respondents. Trying
to force this design on them may result in data distorted by the
confusion and fatigue of the subjects.
b. A somewhat stronger set-up consists in having the subjects
rate all distinct stimuli pairs on a similarity-dissimilarity semantic
scale. One issue is to determine how fine the scale ought to be. If
a 7-point scale, say, is used as opposed to a 5-point one, it may well
be that the discrimination ability of the subjects is not so fine as
to allow them to give reliable answers on the larger scale. As a
result, itmightbe postulated that certain levels are clustered to¬
gether as indistinguishable and the choice of a rating in each cluster
is purely random. The clusters may well be unevenly spaced over the
scale which further distorts the data. The semantic-scale method may
be applied globally, without specifying the attributes to be used by
the respondents or locally, by considering a specific list of attri¬
butes considered likely as perceptual dimensions of the respondents
answer.
88
c. A simple method also applicable globally or locally in the
above sense, is to ask the subjects to group the stimuli in similar¬
ity (or dissimilarity) clusters.
d. The "triples-method" considers all distinct triples of stimuli
and requires that the subject states in each triple that pair which is
most similar and that pair which is least similar.
e. Choose each stimulus in turn as a reference point and ask
the subject to rank the remaining (n-1) stimuli in order of increasing
dissimilarity (decreasing dissimilarity). Again the method can be
applied globally or locally.
f. An alternative clustering method is to ask the subject to
pick some subset of the stimuli (say h of them) most similar to some
reference stimulus and do this for all stimuli, one at a time. The
cluster size may or may not be prespecified by the experimenter.
All of the above methods obtain direct similarity measures. Alternatively,
indirect measures can be obtained from multiattributed ratings of the stimuli
(or rankings) given by the subjects on a prespecified list of attributes.
Such "profile data" can be processed via correlation coefficients calculations
or distance calculations. An extensive review of many alternative data collec¬
tion procedures is given in Green and Rao (1972).
(2) Scaling Proximities Data:
Here, a basic distinction must be drawn between metric and nonmetric
model s.
a. Metric Models: Here, it is assumed that the input data is
ratio or at least, interval-scaled. Also, the output of these algo¬
rithms is assumed to be metrically related to the input, i.e., pro¬
portional in the ratio-scale case, or linear in the interval-scale
case. These "classical" scaling models are best explained in Tor-
gerson (1952) and (1960). A two-stage procedure is followed: a
functional form is just specified and used to compute distance esti¬
mates and its eigen-values and eigen-vectors are computed to determine
the dimensionality of the space and obtain the coordinates of the
points in that space. An alternative to these factor-analytic methods
was devised by Kruskal and Carmone (1969) to allow the researcher to
vary parametrically the functional form (in the class of polynomials
up to the 4th degree). Optimization of fit is obtained through
iterative gradient procedures.
b. Non-metric Models: These differ from the metric ones in that
they only use the ordinal properties of the input data. They specify
a goodness-of-fit function (e.g. Kruskal's "stress" measure) to relate
proximity data to distances between the points in a "to-be-recovered
space". This function is monotone. In practice the solution involves
an iterative coordinate adjustment of the points starting from an initial
configuration in a space of pre-specified dimension. Examples are:
Kruskal (1964), MDSCAL V; Young and Torgerson (1967) TORSCA; McGee (1968);
Gleason (1969); Roskam (1968); de Leeuw (1969). These procedures are
discussed at greater length later in this chapter.
89
Alternatively, Carroll (1972) has devised a model seeking to
optimize a "smoothness" criterion (PARAMAP). A first group of models
can be found to analyze individual differences.
c. Models for disaggregate-level analysis: in this class,
Carroll and Chang's INDSCAL model (1970), is the most representative.
An important difference with the previous models is that the configur¬
ation of points can be differentially elongated along the axes. Also,
the orientation of the axes is unique. Practically, coordinates are
iteratively adjusted to minimize a "badness-of-fit" measure.
Somewhat different in purpose but well suited for processing of proximity
data, we might also mention hierarchical clustering procedures both metric
(Ward, 1963) and nonmetric (Johnson, 1967; Sneath, 1957; Sorensen, 1947). In
practice a tree is constructed by relating proximity data to "ultrametric"
distances between its terminal modes; a monotonie function is used to effect
this relation in the nonmetric case. These methods are useful when they are
combined with the spatial representation models derived by the previous methods
a, b, and c.
4.2.2 Dominance Data
The data are usually in matrix form where the (i,j)th entry measures the
extent of dominance (preference) of the ith stimulus over the jth stimulus.
In the mathematical literature the class of such matrices characterizing a
complete asymmetric, irreflexive (and possibly weighted) graph has been studied
under the name of "tournament matrices" (tournament graphs or relations). If
there are N respondents in a sample, there is a set of such matrices from which
multidimensional scales are to be extracted in a (possibly joint) space of
stimuli (and judges). If we wish to limit the study to the recovery of unidi-
mensional scales a number of procedures are available. These methods are
reviewed briefly before dealing with multidimensional scales.
(1) Unidimensional Analysis: It is usually applied to a single
dominance matrix (although Torgerson (1960) has developed several
procedures to derive a single aggregate dominance scale from the
data of N judges.)
The Thurstonian model (1927) is metric and is based on relating
dominance data with the signed differences between points on a uni¬
dimensional scale. The function effecting this relation is derived
from the normal density function.
A non-metric method has also been devised by Klemmer and Shrimpton
(1963). It uses a function which is simply monotonie, and adjusts
points iteratively to minimize non-monotonicity.
(2) Multidimensional Analysis: following Carroll (1972) it seems
useful to distinguish between two types of analyses, external and
internal.
(a) External analysis of preference data refers to the case
where a stimulus configuration is available at the outset and
has been obtained from another source besides the preference
data, e.g., proximity data (see 4.2.1 above).
90
Carroll and Chang's PREFMAP model is one of the most
versatile tools for this sort of analysis. The algorithm
consists of 4 phases corresponding to four different models
of increasing complexity.
The vector model is the simplest. Carroll (1972) has
shown how it is a special case of the "ideal point" model
when the distance between the (externally determined) stim¬
ulus configuration and the ideal point becomes infinite. A
"preferred direction" is determined in the multidimensional
space obtained independently. Each subject preference order¬
ing is then represented by projections of the stimuli along that
direction. The program finds the best-fitting representation
for both the metric (Slater, 1960) and the nonmetric case.
In the latter case the "preferred direction" is chosen so as
to result in an ordering of the stimulus projections which
is as monotonically related as possible to the original stim¬
ulus ranks. Also, an interpretation of the corner of the
angles between this direction and the external reference axes
is possible by viewing them as a measure of the relative impor¬
tance of the factors (corresponding to each axis) in the overall
preference judgment. Phase IV of the algorithm deals with the
vector model.
The simple ideal-point model is applied in Phase III.
In its nonmetric form, it finds the position of the subject's
ideal point so that preferences for the actual stimuli repre¬
sent nonincreasing monotone functions of their squared distances
from the (unknown) ideal point. Anti-ideal points can also be
found, in which case preference increases with increasing
squared distance from such points.
The weighted ideal-point models are most complex inasmuch
as they allow stretching of the axes of the stimulus space for
some "extreme" individuals (Phase II) and even different ori¬
entation of the axes of the stimulus space for different indi¬
viduals (Phase I).
(b) Internal Analysis of Preference Data is applied whenever
the investigator lacks an "external" stimulus configuration at
the outset. Typically, in such cases, the available data con¬
sists of N individual rankings (or ratings) of the stimuli.
Here the researcher's goal is to use this preference data alone
to find a joint multidimensional space to locate both stimuli
and subjects. The same distinction, as in external analysis,
is useful to classify models under this rubric.
Vector model : Carroll and Chang's MDPREF is a well known
algorithm to represent judges as vectors and stimuli as points
in a common multidimensional space. The same interpretation
applies here as in external analysis. Shepard and Kruskal
(1964), Roskam (1968) and Lingoes and Guttman (1967) have
developed nonmetric analogues to factor analysis; and thus
they can be used to fit vector models.
Ideal point model: Kruskal's MDSCAL V can be used to
fit such a model. A variety of functions (non-metric or poly¬
nomial) can be specified for this purpose.
91
In the same category Lingoes (1965) "non-metric unfolding"
method may be mentioned. ["Unfolding" consists of the simul¬
taneous determination of a space containing M ideal points and
m stimulus points in such a way that the rank order of all
stimuli from each ideal point preserves as closely as possible
(for a given dimension) the original N rankings.]
Finally, Carroll's PARAMAP algorithm can also be used for
ideal point-type analyses.
4.2.3 Profile Data
They are usually associated with classical factor analysis. In factor anal¬
ysis one usually starts from an n x m profile matrix where the (i,j)th entry
represents the scores of the ith stimulus on the jth variable.
(1) Metric Case: In R-type factor analysis a measure of asso¬
ciation product-moment correlation -- is then computed between the
(^) pairs of variables and the resulting square correlation matrix
is then factored. The variables are then referred to a common set of
factor axes. Finally, one can calculate factor scores for each stim¬
ulus in factor space, as a linear combination of the original stimulus
profiles in variables space.
In Q-type factor analysis, correlation across pairs of stimuli
are computed and summarized in an (^) x (£) matrix. There we try to
find the variables as linear combinations of the factors. Both of
these approaches are metric.
(2) Nonmetric Cases: This has been studies by a number of
authors especially Shepard and Kruskal (1964) and Roskam (1968).
In these models, point positions for the stimuli and vector direc¬
tions for the attributes are simultaneously determined in such a
way that the orders of the stimulus projections on each attribute
vector in a (previously determined) reduced space best fits the
original rank orders in the data matrix. Finally, Carroll and
Chang's PARAMAP algorithm (see 4.2.1 above) also generalizes class¬
ical factor analysis.
4.2.4 Conjoint Measurement Data
Under this heading are methods that seek to measure the joint effect of
several (two or more) independent variables on the ordering of a dependent
variable. They all posit a simple functional representation for this effect
and try to fit it to the original data matrix via some algorithm. Usually
these algorithms are similar to many used in MDS inasmuch as they use itera¬
tive adjustment procedures for the independent variables -- starting from some
initial trial values — to minimize departures from the "theoretical" values
derived from the functional form used to express the interaction between the
independent variables. A general class of functional representations is the
class of polynomial models. A very important subclass consists of the additive
models.
92
(1) Additive Conjoint-Measurement Models
This model implies that a respondent's total utility for a
multiattributed stimulus is the sum of the stimulus "part-worths"
(utilities on a given attribute scale). It determines the uni-
dimensional utilities such that the ordering of the component
sums preserves as nearly as possible the ordering of the depen¬
dent variable. The unidimensional utilities are interval-scaled
with a common unit. Many algorithms belong to this class: Lingoes
(1967), Pennell (1970), and Tversky and Zivian (1966). Kruskal's
M0NAN0VA is probably the most well-known procedure for this kind
of analysis. Also in cases where the respondents' responses are
not ordinal but simply categorical, Carroll's categorical conjoint-
measurement model (1969) is available for the same type of analysis.
(2) Polynomial Conjoint-Measurement Models
Here, the interaction mode between independent variables is
extended to take account of combinations of sums, differences,
products or subsets of them. In this class, we refer to Carroll
and Chang's PREFMAP (see 4.2.2 above). By considering factorial
design data as dummy predictor variables in a sort of monotone
regression, this model can accommodate interaction and product
terms in conjoint measurement.
In conclusion, as pointed out by Shepard (1972) it should also be mentioned
that, once a spatial representation of the data has been found in terms of one
of the MDS models discussed above, a variety of methods are available to help
in interpreting these spatial configurations. We mentioned earlier the use of
Cluster Analysis to discover groupings of the data once imbedded in the appro¬
priate space. Hierarchical clustering in two-dimensional space lends itself
to an easily interprétable graphic display as clusters are nested in each other.
4.3 Multidimensional Scaling of Perceived Attractiveness
4.3.1 Principles of Reduction
The concern of this research is to identify and measure the performance
measures (dimensions) of attractiveness of shopping centers. In this research,
16 attributes were chosen to represent the notion of attractiveness of shopping
locations. It is assumed a priori that individuals evaluating attractiveness
will not simultaneously and explicitly consider the 16 attributes in making a
choice. Rather, they will evaluate attractiveness based on a few performance
measures (dimensions), each of which consists of a group of the attributes.
The identification of those underlying dimensions is the main concern of this
section.
This problem can be formulated as follows: Given a set of L attributes,
Y, Y = {Yl9 Y2... Yl>, find the mapping Y into X, where X = {Xx XK>
is the set of underlying performance measures, so that L > K. This process
was called reduction by Hauser (1975). Following the above notation observe
that the analysis starts with a full set of ratings made by the individuals in
the sample. Let A be a set of J stimuli of interest so that A = {Ax, A2,...,A,}
and let there be I individuals in the sample. Define a vector Y^j = (y^ji»
y..2 y... ) to be the ratings of alternative j by individual i with respect
I J I J L
93
to the L stated attributes. Observe, however, that there is one vector Y.. for
' \J
each alternative. Thus, the basic information consists of a rectangular matrix
of J x L ratings for each individual. This type of information is the profile
data described in section 4.2. The reduction process as stated above has to be
identified by a formal model mapping Y into X (i.e. R: Y -> X). Ideally, the
resulting mapping process should be neither alternative- nor individual-specific.
Define as a reduction process for an "individual iel and alternative
jeJ. If a reduction process is the same for all alternatives for a given indi¬
vidual, then the alternatives can be assumed to be homogeneous for a given indi¬
vidual. If the reduction process is the same for all individuals for a given
alternative, then the individuals are homogeneous with respect to that alter¬
native. However, in order to use the reduced set of performance measures in a
choice model estimated across individuals and alternatives, as in the multi¬
nomial logit model, the reduction process should be homogeneous across indivi¬
duals and alternatives.
This requirement of homogeneity raises a new dimension of the importance
of segmentation of the population. Individuals who have the same reduction
process with respect to the total set of alternatives define a homogeneous
segment with respect to their perception of the performance measures. This
definition of homogeneous groups with respect to perception does not mean that
all the individuals in the group perceive the alternatives to be the same. It
means that those individuals employ the same reduction process to identify the
reduced set of performance measures.
Formally stated, let S be a homogeneous group of individuals with respect
to their perception of alternative j then if Y^-j = Xjj for i»leS then X^ = X-jj.
This property was defined as encompassing characteristics of reduction (Hauser,
1975). The importance of these characteristics is twofold. First, if different
individuals have different reduced performance measures, no efficiency in data
handling can be gained by reducing the initial set of attributes and we might
have been better off using the initial attribute set. At an extreme case, if
there are I individuals, K individual- and alternative-specific reduced dimen¬
sions, and J alternatives we might end up with IxKxJ "reduced" performance
measures compared with L original attributes. The second reason to require
homogeneity is that the whole modeling process of reduction is geared towards
an easier and more meaningful presentation of the perception process to the
decision makers and investigators. Thus, instead of dealing with L attributes
they can manipulate only K performance measures, where K < L, in an investiga¬
tion of changes in a system being designed. However, if we end up with a rather
large set of performance measures, many of them being individual- or alternative-
specific we might have complicated the process rather than simplified it.
A crucial property of the reduction process is identification of the per¬
formance measures produced by the model. Not much insight is gained if
the reduction process produces a set of performance measures which are not
related to any physical variable of the system that can be influenced by the
planner. Thus, two properties of the reduction process which are necessary
for the model to be useful are: A simple identification of the reduced set
of the performance measures and a clear association of the reduced set with
defined variables of the system. Some models, such as factor analysis, assume
94
a priori a functional relation between the original set of attributes and the
reduced set of performance measures. However, they do not assume the functional
relation between the performance measures and the underlying causal variables
of the system. Let V = {Vl5 ..., V^} be a set of causal variables of perfor¬
mance measures X-. The relation between the Y attributes and X performance
J
measures was defined as reduction R: Y -* X. Define another process F which
maps the causal variables into the performance measures (i.e. F: V •* X).
The identification of the reduced dimension is probably the most diffi¬
cult and least defined step of the reduction process. As mentioned earlier,
the reduced dimensions are the end product of the process and thus the founda¬
tion for the modeling of choice behavior. Depending on the model used for the
reduction process, the problems are somewhat different but in the two models
that will be discussed later -- factor-analysis and INDSCAL -- the investiga¬
tor's intuition plays an important role in the identification of the dimen¬
sions. However, INDSCAL has a unique property compared to all the other re¬
duction methods inasmuch as the reduced performance measures are completely
and uniquely defined. But the meanings of the dimensions defined by the model
still need to be interpreted by the investigator.
4.3.2 Theory of Multidimensional Scaling
In recent years, new methods of scaling perceptions and preferences
emerged. In a general classification one should probably include factor-
analysis in this group. However, for historical reasons, this group of
scaling procedures known as Multidimensional Scaling (MDS) is treated in the
literature as a separate family of scaling models. The discussion is
limited to the scaling of perceptions, although a large body of the MDS liter¬
ature deals with preference modeling. Some investigations attempt to model
simultaneously preference and perception, (see for example, the paper by
Burnett (1973), where, in one model, the perceptual dimensions as well as the
preferences for shopping alternatives were identified). In the general con¬
text of individual choice modeling, each alternative of the three mentioned
above assumes a different underlying process of choice. However, since the
objective of this research is to identify and measure the perceived attrac¬
tiveness of shopping locations, only the perceptual models are of interest.
The basic input data for multidimensional-perceptual scaling are proximity
data. As discussed earlier, each cell of the J x J matrix contains some mea¬
sure of the similarity, substitutability, affinity
correlation, or interaction between the two objects corresponding to that row
and column. The measure can be direct in the sense that it arises from the
pair of objects immediately, or indirect in the sense that it is calculated
on the basis of other data, such as a profile matrix. Formally stated, let
A1 be a proximity matrix of a set of J stimuli for individual i. An entry in
matrix A1, ô] » represents the proximity between stimulus j and stimulus r as
J '
perceived by individual i. Proximity between stimuli j and r can be expressed
as similarity s. or dissimilarity ds. . Assume, without loss of generality,
J ' \J ■
that the proximity measure is mapped into the closed interval [a,b]. The
similarity of a stimulus to itself s.. = b, while the dissimilarity ds.. = a.
95
The similarity between j and r is a decreasing function of the perceptual
distance of the two stimuli while dissimilarity is an increasing function of
the perceptual closeness. Intuitively, similarity is analogous to a correla¬
tion measure (disregarding the linearity assumption) and dissimilarity is
analogous to distance. One can easily suggest several functional forms rela¬
ting similarity to dissimilarity. For example, assume that Se [a,b] then
DS = b - s, also defined over [a,b]. Since some confusion can easily be
caused by the above terms, proximity will be used as a general term and simi¬
larity or dissimilarity where it is appropriate. Up to this point, proximity
data have been defined as a square matrix. Two remarks are appropriate with
respect to these data. In most cases, proximity data form a symmetric square
matrix, since it is assumed that pi = p^., (i.e. the perceptual distance be-
J ' 'J
tween stimulus j and stimulus r). Although the basic structure of the proxi¬
mity data is a square matrix, depending on the. specific model the input may be
a rectangular matrix. If the model is applied across I individuals, the full
input data for J stimuli will consist of a rectangular matrix of the order
((I x J) x J).
As mentioned earlier, there are two basic types of proximity information
-- direct and derived. A short discussion of these two types of data is appro¬
priate at this point. The direct proximity measure is collected by asking
individuals to indicate in some way how they perceive the similarity between
all the pairs of J stimuli. Usually this is done without specifying in the
question what the investigators mean by the term similarity, so as not to in¬
fluence the respondent's answers. However, it should be emphasized that simi¬
larity is related to perception rather than to preference. An example will
clarify this point.
Assume that an individual has the following order of preference with
respect to one characteristic of five stimuli: A>B>C>D>E. Assume further
that the preference is expressed on an interval scale as shown in Figure 4-7.
This individual if asked to rank all pairs of stimuli with respect to
their similarity, starting from most similar to least similar pair, would list
ten pairs and it would look as follows: CD, AB, DE, ... AE. Observe that this
similarity information is not enough to construct his order of preference with
respect to the five stimuli. Even if it is known that the phenomenon measured
is unidimensional and the preference scale could be constructed, it would still
not be known without further information, whether A>E or E>A.
The great advantage of direct proximity is that the investigator does not
influence the individual's rating through his perception of the stimulus space.
Its disadvantage is that people are asked to rate a set of stimuli with respect
to a vaguely defined term such as similarity. It is doubtful whether individuals
can actually distinguish between perception and preference of stimuli. Thus,
instead of recovering a perceptual space, we might be obtaining a mixture of
perception and preference.
The process of constructing derived proximity measures consists of mapping
a profile matrix into a proximity matrix. The profile matrix of individual i
for J stimuli consists of the ratings of those stimuli on L attributes. Thus,
y^ represents the ratings of stimulus j on attribute 1 for individual i. The
96
FIGURE 4-7
Interval Scale of Preference for Five Stimuli
97
rhe standardization is needed to remove the influence of the measurement
Dn the dissimilarity value. An example will clarify this point. Assume
ie have three stimuli AH, A , A , and two attributes weight and height or
J i 5
these attributes are rated. The dissimilarity between the stimuli is del
as the Euclidian distance between the stimuli in the two-dimensional spac
the attribute. Assume that the weight is expressed in pounds and the hei
in feet as specified in table 4-2. The Euclidian distances are as follov
J?r = 25.25, d?s = 10.00, d^r = 4.25. However, if height is measured in
these distances become d?r = 50, d?s = 153, d£s = 53. Observe that the r
jrder of the distances is not maintained when the measurement unit of hei
vas changed. However, if the values are standardized before calculating
distances, the ratio between the distances will be maintained and thus tt
rank order will also be maintained. Since the units of measurement perce
jy the respondents in the sample cannot be known a priori, standardizatic
lecessary to offset the biases caused by the fact that different people n
perceive the attribute scales differently.
Two problems are encountered with the use of derived proximity date
:irst, the investigator may not have picked all the relevant attributes 1
people might evoke in making overall proximity judgments. Second, there
question of what weights to assign to each scale in computing the overall
issociation measures. The way to overcome the first problem is to conduc
preliminary survey to test whether the set of attributes is complete and
jnderstood. As stated, this was one of the objectives of our first surve
The second problem has no satisfactory solution, for lack of any better n
it is customary to wpigh all attributes equally.
A basic assumption underlying all the MDS models is that a set of C
ili can be represented in a K dimensional space whose axes are the perfor
neasures of these stimuli, where K<J. It is further assumed that the pre
98
Stimuli
Weight (lb)
Height (ft)
Aj
60
3.0
Ar
65
3.5
As
63
4.0
TABLE 4-2
Weight and Height of 3 Stimuli
99
measures between the stimuli should be related through some distance function
defined over the K dimensional space. The distance function can be any Min¬
kowski p metric defined as equation (4.4).
V
£ (xjk -
k=l
Vp
p>1.0 (4.4)
where d. = the perceptual distance between stimulus j and
Jr stimulus r in the K-dimensional space
x,k,x .= coordinates of stimuli j and r on performance
J measure k of the K-dimensional space
When p=2 the above equation defines a Euclidian distance which is invariant
under orthogonal rotation or translation of orthogonal axes. For p=l the
distance is the familiar "city block" distance. Observe, however, that for
p^2 the distance is not generally invariant under an orthogonal rotation of
the axes, although it is still invariant to the translation of the axes. The
condition of p>_l is needed to maintain the triangular relations: Consider
three stimuli j, r, and s. The following three conditions define the trian¬
gular property i) d.. = 0 and d. > 0, ii) d. = d . i.e. distance is symme-
J J J' J ■ 'J
trie, and iii) djr £ djs + d$r.
The scaling model can be either nonmetric or metric, depending on the
assumptions of the specific model. By far the greatest contribution to the
MDS field was made by Shepard (1962) with the introduction of the concept of
monotone regression, which is the basis of all the nonmetric MDS models. A
monotone relation between two sets of numbers is defined as follows. Let A
be a dissimilarity data matrix consisting of J x (J-l)/2 ratings. Assume no
ties in this matrix. Let the j-th entry of the matrix be 6. . Assume that
jr
some scaling algorithm has been defined that estimates the location of the
stimuli in k-dimensional space. Let the estimated distance matrix be D and
its entry d. A monotone relation between D and A is defined by equation
J *
(4.5), illustrated in Figure 4-8.
< d^ whenever <5. < 6 (4.5)
jr — sr jr sr
Observe that this relation is valid when the 6. are defined only on ordinal
jr
scales, so that all that is needed is a rank order of the proximity data for
the set of stimuli, hence the name nonmetric is used. Kruskal (1964) opera-
tionalized this concept and defined a full nonmetric scaling procedure. He
introduced the measure of goodness of fit termed stress which the procedure
tries to minimize subject to a mononicity constraint. Let djr
be the distance between stimuli j and r in the k-dimensional space of the per-
A
formance measures and let d^ be a number as close as possible to djr, subject
to the original monotonicity of the dissimilarities <5jr. Kruskal suggested
minimizing equation (4.6) for stress.
100
Distance Measure d-
jr.
FIGURE 4-8
Scatter Diagram Displaying a Monotone Relation
101
J
min. S = z (d, - d, )2/ z (d. - d)2
j<r J J j<r J'
(4.6)
subject to
2jr i dsr whenever sjr < «sr
Observe that this function is completely analogous to the expression for R2
used in linear regression. Computationally, the algorithm starts with an
initial position of points in a space of predetermined dimensionality and
perturbs the configuration to move in a direction to minimize S. The solu¬
tion is repeated for several different dimensional spaces and the increase
in the stress value is compared to the reduction in the dimensionality of the
space. The objective is to represent the J stimuli in as small a dimensional
space as possible, keeping the value of the goodness-of-fit measure as small
as possible. As with factor analysis, the ease of interpretability of the
space plays an important role in the decisions of the appropriate dimension¬
ality, as is discussed later.
Several remarks are appropriate here, first observe that it is always
possible to represent J points with J(J-l)/2 ranked dissimilarities in a J-l
dimensional space with a stress value S=0. For a k-dimensional space there
are K x (J-l) unknowns for the general case (we can always choose the coor¬
dinate of one point arbitrarily). If we assume Euclidian distance relations
(and thus allow orthogonal rotation of the axes) there are KJ - K(K + l)/2
unknown. Against this number of unknowns there exist J x (J-l)/2 inequalities
for the nonmetric scaling. The uniqueness of the solution depends on the
specific dissimilarity data. However, experience with synthetic data, as
well as with real applications, indicates that, for both nonmetric and metric
algorithms (which are discussed later) there should exist a ratio of at least
3 between the number of stimuli and the number of dimensions to achieve a
stable solution, i.e. J/K 3 (Green and Carmone [1970], Green and Wind [1972]).
The nonmetric MDS models clearly represent a significant contribution to
multidimensional scaling. However, in many cases the problem, or the under¬
lying model is too complicated to be solved through a nonmetric solution.
Thus, in many instances a metric model is used. In concept the metric model
uses a similar philosophy to the nonmetric models and, historically, metric
models were used before nonmetric scaling procedures were defined. The metric
model assumes that there exists a functional relation between dissimilarity of
stimuli j and r and their distance in a K-dimensional space. It further
assumes that the input data is defined at least on an interval scale. The
functional relation between 6. and d. can be linear or some other prespeci-
J ' J '
fied function. The goodness-of-fit measure can be R2 or some other appropri¬
ate function. The appropriate dimensionality is chosen in the metric model
as before based on the trade off between the goodness-of-fit measure and the
dimensionality of the space.
Most of the nonmetric algorithms start with a metric linear solution as
the initial configuration so that, in a sense, the metric solution represents
a bound on the nonmetric solution. This is especially true when the set of
102
inequalities of the proximity data does not provide enough constraints to
define uniquely a nonmetric configuration. Some algorithms are considered
quasinonmetric, in the sense that they alternate during the iterative solu¬
tion process between one metric and one nonmetric iteration. The input data
are usually considered to be defined on an interval scale.
Before concluding the general description of MDS models another classi¬
fication of the scaling techniques is in order. There are three approaches
to the scaling of the perceptual performance measures (dimensions) of indivi¬
duals. The first approach is to identify a perceptual space for each indivi¬
dual based on his specific proximity data. This will be the most detailed
level of analysis capable of maintaining the analysis at the most disaggre¬
gate level. A second approach is to recover an aggregate perceptual space
based on a proximity matrix averaged across individual proximity matrices.
In this approach, every entry in the aggregate proximity matrix p. is de-
— i
fined as p. = z Pir/n. The third approach is to use the proximity data of
J I J '
I individuals. However, the model defines a common configuration for all the
individuals in the data set allowing for individual-specific differences to
be expressed in some functional form. The INDSCAL model utilizes this
approach and it is discussed in detail in the next section.
4.3.3 The INDSCAL Model
The INDSCAL model was suggested by Carroll and Chang (1970), it is a
metric MDS model. The model makes the following assumptions:
(1) The stimuli in the choice set are perceived by individuals
along K underlying performance measures (dimensions).
(2) The K underlying performance measures are common to all
individuals in the sample.
(3) The dissimilarity judgments for each individual are related
to the group similarity space by differential weighting of
the underlying dimensions.
More specifically the relationship of equation (4.7) is assumed
K
^ "lk(xjk " xrk':
4 =
(4.7)
where d1. = the distance between stimuli j and r as perceived
jr by individual i
x., and x . = the coordinates of stimuli j and r on dimension
jk aMU rk
k respectively
w.. = the weight which individual i places on performance
measure (dimension) 1
Assume for a moment, that equation (4.8) is true.
4 = 4 xjk (4-8>
103
Substituting equation (4.8) into equation (4.7) yields equation (4.9)
h
djk =
j, (v]k - "rk»2
(4.9)
This expression defines a simple Euclidian distance between stimuli j
and r as perceived by individual i. Equation (4.7) represents the relations
between the individual coordinate and the common perceptual space.
The scaling problem can be stated as follows: given the dissimilarity
judgments, , for J stimuli by I individuals estimate the coordinates x.. ,
J r J K
and individual weights w^ of J stimuli and I individuals in a space of
dimensionality k. The procedure consists of the following steps:
1) For each individual the dissimilarity ratings are converted into
distance values following a procedure suggested by Torgerson (1958).
This procedure consists of finding the smallest possible additive
a1 so that the dissimilarity ratings for every three stimuli in the
set will satisfy the triangle inequality (i.e. (<s] + a1)^! + a^
j 4 J • J 5
(6 s r + a ) for all j, r, and s). For an errorless data set, an exact
solution exists. For data with error, a lengthy estimation procedure
was defined by Abelson and Messick (Torgerson [1958]). However,
Torgerson has suggested a very simple procedure which gives a rather
good approximation of the Abelson and Messick method. He suggests
estimating these constants as shown in equation (4.10).
a1 = max {5jr - - ôsr> V-j, r, s (4.10)
This is used to define the estimated distance V. = a1 + 6. .
jr jr
2) The distance estimates are converted into scalar products between
the points represented as vectors issuing from an origin at the
centroid of all points. The details of the step are described in
Torgerson (1958) and Green and Wind (1972).
3) The scalar product matrix for each subject is standardized such
that the sum of squares of entries of the matrix is equal to one
and the mean equal to zero. Thus, this matrix can be looked upon
as the variance of the standardized distances between the J stimuli
of individual i. By standardizing this matrix we ensure that the
influence of each individual on the final common configuration is
equal.
4) The basic equation of the INDSCAL model is expressed in terms of
the scalar product matrix B1 defined above. Let bjr be an entry
in B1. Then equation (4.11) follows.
104
K K
bjr * kE/jkvrk + ejr * ^"ikVrk + *jr (4'n)
where e1. = a random error,
jr
Observe that only the values on the left-hand side of equation (4.11) are
known, the values on the right-hand side are estimated in a least-squares iter¬
ative procedure. Let us rewrite equation (4.11) as equation (4.12).
K
zijr = kz, wik XL XR + eijr <4',2>
where z1jr = bjr
L,R = subscripts to distinguish between Xjk and xrk
Let Z, W, X^, and XR be the corresponding matrices of the variables in
equation (4.12). Observe that the dimension of those matrices is (I x K) for W
and of XL, XR it is (J x K). In the final solution, XL = XR since both
represent the coordinates of the J stimuli in K-dimensional space. Let
s = J(j - 1) + r so that s varies from 1 to J2. Define gsk = Xjk • x^k and
zijk where the hats indicate estimation of the appropriate values. Then,
equation (4.12) can be written as equation (4.13).
^ a ~ ~-r
zis = z wil ^sk or in matn'x notation, £ = (4.13)
Knowing the values for G, this equation can be solved for W by a least-squares
estimation ft = Z* £ (£T £)-1- Substituting the values of w^ back into equa¬
tion (4.12), one can solve in a similar fashion for XL and then for XR, (for
more detail see the original paper by Carrol and Chang [1970]). Once W,
X^, XR are estimated, a new iteration can start evaluating the three matrices
again. This procedure continues until some goodness-of-fit measure is satis¬
fied. Observe that each iteration of the procedure consists of three least-
squares solutions, each for a different matrix, while holding the other two
constant. To start the process, initial values are needed for W, XL> and XR.
Usually W is set to have all entries equal to unity and XL = XR is generated
at random by the computer or set to some values specified by the user, hope¬
fully based on some good initial guess. Carrol and Chang (1970) state that
this procedure will always converge to a local minimum and based on their
experience and on simulation will "almost always" converge to a global minimum.
However, since the global convergence is not assured, it is recommended (J.J.
Chang - personal communication) to try at least two different starting con¬
figurations.
105
During the iterative process there is no constraint making
XL = XR. However, due to the basic symmetry of the data z..jr =
zirk for 15 J' anc' r at convergence so that the following rela¬
tion will hold X^ = CXp where C is any K x K diagonal matrix of
non-zero elements. This is due to the fact that multiplying the
columns of XL by any constant not equal to zero and dividing the
corresponding row of XR by the same constant will not alter the
result of the estimation process. Matrix C is not recovered in
the solution process, since XR is the last matrix estimated and
X^ is simply set equal to XR.
The goodness-of-fit measure used by the INDSCAL procedure is based
on simple correlation between the original z^r values and the pre¬
dicted ones z.jjr- Since the INDSCAL model is nonlinear, simple
relations do not exist between the correlation value and the R2
measure, however (1-p2) is treated approximately as the amount of
variance explained for individuals in the sample. As a byproduct
the program also produces the correlation coefficients for each
individual between z... and z.. as well as the average correlation
I J K 1 J I
coefficient for all the individuals in the sample.
5) The final step of the process is to scale the solution space, since
the solution is defined up to an arbitrary positive constant. The
solution space is standardized so that the variance of coordinates
of the stimuli on each axis is equal to one. This standardization
leads to certain interpretive niceties. It means that the square
of the Euclidian distance of the individual's weights from the origin
can be interpreted (approximately) as the total variance accounted
K
for, S., in his scalar-products matrix, (i.e. S. = Z w?.). If the
1 1 k=l 1K
axes are uncorrelated, then Si will exactly represent the amount of
variance explained, and it will be equal to unity if the total vari¬
ance for individual i was actually explained by the model. However,
the converse is not true. When the dimensions are correlated,
might be smaller than one, but all the variance for individual i may
nonetheless be accounted for.
The individual INDSCAL scores are not calculated by the program.
However, they can be easily calculated as x^.^ = w..^ Xjk, i.e. the'
performance measure k of stimulus j as perceived by individual i is
equal to the square root of the individual weight along the k-th
dimension times the coordinate of the j stimulus on the same dimen¬
sion.
106
4.3.4 Identification of Dimensions
The INDSCAL procedure defines the spatial configuration of stimuli
common to all individuals as well as the idiosyncratic perceptions of each
individual in the group. The axes are uniquely defined since the estimated
individual scalar product matrices are a function of the dimensions of the
solution. Conceptually we may use the INDSCAL result as a set of indices
for any further analysis without identifying the dimensions. However, iden¬
tification of the performance measures has a crucial managerial or engineer¬
ing importance. One approach to the identification of the dimensions is by
examination of the configuration of the stimuli in the perceptual space.
This examination identifies the important characteristics that differentiate
stimuli along each dimension. This approach must be used if the only data
collected are direct proximity data and thus there is no other basis for
determining the characteristics of the dimensions in the perceptual space.
Effective use of this approach depends on the researcher's knowledge of the
characteristics of the stimuli included and obviously this type of interpre¬
tation is highly subjective.
A much more objective method, although not free from subjective judgment,
is used in most cases to identify the dimensions. Let yl-j be the rating of
stimulus j on attribute 1 by individual i. By some averaging method we can
construct a représentative_unidimensional scale of the J stimuli on each
attribute. Let x-| = yi-|»y2]» be a vector, averaged across indivi¬
duals, of the ratings on attribute 1 for each stimulus j. Such a vector is
constructed for each attribute 1. The actual construction method can simply
be averaging the appropriate ratings across all individuals or, as it was
done in this research, throuoh a Thurstonian unidimensional-scaling method.
These vectors can be fitted one at a time into the perceptual space through
the origin by a metric or nonmetric method. The "closeness" of the set of
attributes to the axes of the space provides a clue to the underlying perfor¬
mance measures those axes represent. Figure 4-9 demonstrates this process
graphically. Algebraically, the linear metric model PROFIT (Carroll and Chang
[1970]) can be stated as follows: Let X] be a vector of stimulus ratings on
attribute 1, X (J x K) a matrix of J stimuli on K dimensions. The origin of
the K-dimensional space is assumed to be at the centroid of the N objects.
Let T be a K-component vector of direction cosines of the fitted vector and Y
a J-component vector of the value of projections of J stimuli on the fitted
vector. The objective of the fitting procedure is to minimize the squared
differences between the projection points of the stimuli and their ratings on
the attribute vectors. Thus, we are trying to position the vector in the
space so that (Y - Y)2 will be minimized. The relation between the original
ratings Y and the stimulus space are shown by equation (4.14).
Y=£[ + £ (4.14)
From equation (4.14), an estimate of T can be obtained, as shown in equation
(4.15).
T = (XT X)"1 XT Y (4.15)
107
9 Bus
*Auto
Auto
s Bus
■Taxi
Taxi
Train
'< Train
2-D Perceptual, Space
Attribute-Privacy
Fitted Attribute-
FIGURE 4-9
Fitting of Attribute into Perceptual Space
108
The stimuli ratings on the fitted vector are given by equation (4.16).
Y = Il = X(IT I)"1 IT Y (4.16)
A
The vector of direction cosines is obtained by normalization of the T vector
to unit length.
A nonmetric procedure of fitting the attribute space is called PREFMAP
(Green and Wind [1972]). This method consists basically of an iterative
application of the above-mentioned linear procedure. The first iteration is
a simple linear fit of the attribute vector, after which, the estimated pro-
a/ \
jections Yu; are compared with the original ones, Y. If the monotone rela¬
tions are preserved, the procedure is terminated. Otherwise the Y are
a / \
replaced with Y which are based on Yvw and a linear procedure is applied to
obtain new estimates of the projections Y^. This iterative procedure con¬
tinues until the monotone relations between Y and Y^ are satisfied based on
the stopping criterion for stress values.
4.4 The Factor-Analysis Model
The basic input to the factor-analysis model is a data matrix Y of the
value of the attributes on a number of alternatives. These profile data, y^
consist of ratings for each individual i in I, alternative j in J and attri¬
bute 1 in L. The reduction process of the attributes can be executed across
alternatives for each individual or across alternatives and individuals or
across individuals for each alternative. These three approaches are called
R-type factor analysis, the first being individual-specific and the latter
two are executed across individuals either for each alternative separately
or for all the alternatives simultaneously. In all three cases, the process
identifies a reduced set of performance measures. The reduction process can
also be executed across individuals for the set of attributes, for each alter¬
native or across individuals for attributes and alternatives (stimuli), given
a large enough number of attributes. This type of factor analysis is called
Q-type analysis. Observe that the reduced set defines, in this case, a clus¬
ter of individuals who are similar with respect to their perception of the
attributes. This type of analysis is a method of segmentation of the total
sample population and is not of interest in this investigation.
The R-type factor analysis assumes that there exists a common set of
factors-reduced performance measures, X, in the initial profile data matrix Y.
It further assumes that each attribute Y1 can be expressed as a linear combin¬
ation of the underlying factors X^, a specific component of the attribute U1,
and a random error term e1. Schematically, the above assumptions are presented
in Figure 4-10. Mathematically, the common factor model can be stated as
equation (4.17)
*n * rk fu xik+ si x*i+ eii (4-17)
109
Common Component -»• Unique ->•
Linear combination of the identified factors Part of the attribute
y | , j
<- -f- -V
Specific Random
Error
FIGURE 4-10
Assumed Components of an Attribute
110
where = loadings of factor scores on attribute
= factor scores of factor k for individual i
★
x., = unique factor scores associated with attribute y, and
individual i
s-j = the loadings of the unique factor scores on attribute y-j
e.i = random error
It is assumed that f^ and s-j are the same for all the individuals in the
sample and thus can be estimated across individuals. Observe that the mathe¬
matical problem as stated above seems similar to a linear-regression formula¬
tion. However, a fundamental difference exists between factor-analysis and
linear regression. Unlike linear regression, in factor analysis, values of
the factors on the right-hand side of equation (4.17) are not known. Thus,
the loadings as well as the factor values, which are called factor scores in
the factor-analysis terminology, are to be estimated.
It is further assumed that the unique part of the attribute U-j, given by
S.| + e-| » is uncorrelated with its common part or with the common or unique
part of any other attribute.
It is also assumed that the factor scores are orthogonal to each other.
The attribute ratings Y, and the factors S. are standardized across individuals
I i i K i i
and alternatives so that = 0 and Var (Y^) = Var(Xk) = 1.0 for all 1
and k.
I I
Let Y, and Y be the standardized ratings on two attributes, then
ill m
Y, • Y = rn is the correlation coefficient between these two attribute
I m. lm , i
since Y, and Y are both standardized. The standardized profile matrix Y
I m
consists of L attribute (columns) and N=IxJ cases (individuals x alternatives)
and is used to calculate the correlation matrix between all the attributes as
shown in euqation (4.18).
(4.18)
Equation (4.17) can then be rewritten as equation (4.19).
I = XL + I*U • (4.19)
The sizes of the matrices are as follows:
y' - [N x L], x" - [N x K], F - [K x L], X* - [N x L] & U - [L x L].
Premultiplying both sides of equation (4.19) by their transpose gives the
result shown in equation (4.20).
Ill
x'T I = £TX TX £ + uVx'e + FTx'TX*ii + (4.20)
Since we required the unique factor scores to be standardized and uncorrected
with the other factors and further assumed that the factors are orthogonal to
each other, the following relations hold, equations (4.21), (4.22), (4.23),
and (4.24).
T 1
I* I
= 0
(4.21)
'T
l 'x*
= 0
(4.22)
'T 1
I I
= N=
(4.23)
l*Tl*
-NÏ
(4.24)
Thus, equation (4.20) can be simplified to equation (4.25).
lJl = N£T£ + Nyju (4.25)
Dividing both sides of the equation by N and rearranging the terms yields
equation (4.26).
R - LL2 = £T£ (4.26)
Equation (4.26) is considered to be a fundamental theorem of factor analysis
(Rummel, 1970). Observe that the left-hand side of this equation has an inter¬
esting interpretation. Since U2 is a diagonal matrix of the size (L x L) we
can rewrite it as shown in equation (4.27).
R - II2 = R - 1+H2 (4.27)
H2 is called the "communality" matrix and is also a diagonal matrix of (L x L).
The communality of attribute L is the proportion of the variable's total vari¬
ance that is accounted for by the factor.
Equation (4.26) states that the factor loadings can be found by "factor¬
ing" the data correlation matrix with communalities h| replacing the unit
values on the main diagonal of R. This problem can be solved by determining
the eigenvalues and eigenvectors of the symmetric matrix R - U2- The eigen¬
values are real numbers and the eigenvectors span the common vector space.
The characteristics equation |(R - JJ2) - Ai| = £L will yield the eigenvalues
and the principle axes which span the vector space. Let the eigenvalues be
arranged from large to small in a diagonal matrix A(L x L), and let the corre¬
sponding eigenvectors be E = {ej, §2, ..., e^}, then the factoring of the
fundamental equation is given by equation (4.28).
R - U2 = £A£ = (éA^HÀV) (4.28)
Thus we have recovered the values of the factor since £ = £4*. The rank of
R - U2 will be the number of eigenvalues greater than zero and the eigenvectors
corresponding to the zero eigenvalues will be deleted to recover the principle-
component solution. Thus if there are S zero eigenvalues the number of factors
will be K = L - S.
112
However, equation (4.26) is not fully specified since the value of the
matrix U2 (or the communalities) is not known. Thus, the number of eigen¬
vectors different from zero that are needed to determine the factors cannot
be determined. Fortunately, there exist upper and lower limits for the
communalities values. The upper limit is simply 1 and the lower limit was
shown by Guttman (Rummel, 1970) to be the squared multiple correlation (r2 in
linear-regression terminology) of an attribute with all the other attributes.
Thus the usual procedure to estimate the factor values is an iterative one.
The lower bound of the communality values is substituted in the diagonal of
the correlation matrix. A predetermined number of factors are calculated
(see Rummel [1970] for methods to determine the desired number of factors),
and the communalities are recalculated from the factors according to equation
(4.29).
1 ■ l Vik <4-2"
The new communalities are substituted again into the diagonal of the correla¬
tion matrix and the iterative procedure continues until a satisfactory conver¬
gence is achieved. Observe that the trace of the original correlation matrix
is simply equal to the number of attributes, L. However, the trace of the
matrix R - U2 at the last iteration is smaller than L. Rummel (1970) has
shown that equation (4.30) is true.
tr (R - 1 + M2) = tr(R2) = tr(R2) = zh2 = ex.
1 ! 1 I
(4.30)
In other words, the sum of the eigenvalues is equal to the common variance and
the ratio EA./L determine the percent of the common variance explained by the
1 1
underlying factors. The factor loadings identified in this process have a
very useful descriptive interpretation. For orthogonal factors, these values
represent the correlation between an attribute and a factor score. For exam¬
ple, the factor loading f^ represents the correlation between the attribute
and the factor X^, f|k represents the proportion of the variance of attri¬
bute 1 accounted for by factor k.
Now that the factors underlying the attributes have been defined, has
any insight been gained into the data structure? Unless we are extremely
lucky, which is r-arely the case (Murphy's Law states that if something can
go wrong, it will), we started with L correlated but defined attributes and
at this stage K unidentified factors have been defined, that are correlated
with all the L attributes. Observe, however, that the identified underlying
factors can still be rotated to a position in space that will reveal a simple
and meaningful data structure (see Figure 4-11 for a geometrical representa¬
tion). This may be done becuase a similarity transform T can be applied to
the fundamental factor-analysis equation without changing the eigenvalues
(the eigenvector, of course, changes), as shown in equation (4.31).
JL - gz = ITFT£I (4.31)
Thurstone (Rummel, 1970) has identified a set of heuristic rules to recover the
simple structure. The essence of these rules calls for rotation of the dimen¬
sions in such a way that each factor will be highly correlated with attributes
113
f? •
\\
\
fix
^7/
V
///
/ /
x\
fl
b. After Rotation
FIGURE 4-11
Rotation of the Common Factors
114
in one group and has a low correlation with attributes in other groups. A
group of attributes might consist of only one attribute. Schematically, this
is shown in Figure 4-12. Thurstone's rules did not require the rotated axes
to be orthogonal to each other, however, all uniquely defined algebraic methods
of rotation assume orthogonality of the dimensions. There are many rotation
procedures but the one which comes closest to Thurstone's rules is called
Varimax Rotation and was defined by Kaiser (Harman, 1967). The Varimax cri¬
terion of rotation is a function of the variance of the columns of factor
loadings. As there are more high and low loadings on a factpr, the variance
of the squared factor loadings is larger. The highest variance will be ob¬
tained when the loadings are either zero or one. Thus, an orthogonal rotation
can be defined so as to maximize the variance of the squared loadings on all
dimensions. Thus the expression of equation (4.32) is to be maximized.
K L flk , K L flk
V = L E S (t-^)1* -l ( E tj1^)2 (4.32)
k=i l=i nl k=i l=i nj
The solution of this maximization is algebraically rather complicated
and is detailed in Harman (1967).
At this stage of the factor-analysis procedure, a simple structure has
been identified which constitutes the underlying performance measures (dimen¬
sions) of the attribute set. Let this factor matrix be F [K x L]. The prob¬
lem at hand now is to calculate the factor scores for each individual and
alternative. These values are needed for the estimation phase of the choice
model. However, the factor scores cannot be calculated directly from the
fundamental theorem of factor analysis, equation (4.26) ,-|-since, due to the
reduction process, F will not be a full-rank matrix and F F is singular. Thus,
a linear-regression model is used to estimate the factor scores. Equation
(4.33) is assumed to hold.
xik * fkiyn + eik <4-33)
I I
In matrix notation X = X â + £ where B is the matrix of unknown coefficients
and £ is an error matrix. Dividing both sides of the equation by /N yields
equation (4-34).
I , I , *
v'N X = (^1) & + (4-34)
The linear-regression estimate of B, is given by equation (4.35).
£ = J- (l'T X')"1 ^X'T l (4.35)
1 'T 1 _1 _1 1 'T 1 T
Observe that j^X X) =JL and n X i = E • Thus> the final equation
to estimate the factor scores is equation (4.36).
115
Simple Structure
Unrotated
Si S2
Factors
S3
After
*
Si
Rotation
* *
s2 s3
Attribute
1
X
X
X
Attribute
2
X
X
X
X
Attribute
3
X
X
X
X
X
Attribute
4
V
A
X
X
Attribute
5
X
X
X
Attribute
6
X
X.
X
Attribute
7
X
X
X
X
FIGURE 4-12
Factor loadings before and after rotation of a factor analytical
solution. The X indicates high factor loadings, thus high cor¬
relation between the attribute and the factor.
116
The factor-analysis procedure has two crucial steps. One is selecting
the appropriate number of factors-performance measures and the second is the
interpretation of the axes after rotation. Rummel (1970) suggests a set of
guide rules for the selection of the number of factors. However, it seems
that the best method for selecting the number of factors is to examine several
factor-analytical solutions in different dimensions. The increase in the
amount of variance explained as the number of factors increases provides a
basis for choosing the appropriate solution. The trade off should be between
the increase in the amount of variance explained against the complication
caused by increasing the dimensionality of the solution. Another criterion
for deciding on the appropriate dimensionality of the factor-analytic solution
is the subjective interpretability of the factors. If too many factors are
extracted, some of them will not be loaded by any attribute, indicating that
they represent spurious correlation in the data rather than any underlying
perceptual dimensions. If too few factors are extracted the interpretation
of the dimensions will be difficult, since it is highly unlikely that a really
simple structure can be found by the Varimax Rotation. Thus, the subjective
ease of interpretation of the recovered dimensions is an important criterion
for the determination of the number of dimensions.
117
5. MARKET SEGMENTATION
5.1 Introduction
The phase of analysis reported in this chapter has several purposes.
First, it is desired to segment the sample, derived from the first Old Orchard
survey, into homogeneous groupings with respect to perceptions of an attrac¬
tiveness space. Second, it is intended to identify and label the dimensions
of the attractiveness space and develop a quantitative measure for a shopping
center located in the space. Third, an attempt is made to segment the same
sample into homogeneous groups based upon preference rankings of the shopping
centers and compare those groupings with the groups based on perceptions of
the attractiveness space. Finally, the sample is segmented on the basis of
combined perceptions and preferences by using the perceptual measures for the
most preferred shopping location.
The underlying hypothesis of this research is that the socioeconomic char¬
acteristics, determined in the survey, form a reasonable basis for grouping
the population in order to understand better their choice behavior in selecting
shopping locations. In other words, it is assumed that people within a given
socioeconomic group are more likely to behave similarly to each other, than
those in random, diverse groups. Given this basic assumption, the tasks of
determining homogeneous groupings of the population require a determination
of whether the finest level of groupings obtained in the survey are necessary
to characterize homogeneity. The procedure adopted is, therefore, one of
attempting to combine the smallest groupings into larger groupings that yet
represent a level of homogeneity in perception. Clearly, questions can be
raised as to whether or not the socioeconomic groupings are appropriate group¬
ings to begin with. These questions are not addressed directly within this
research, although some attempt is made to determine whether or not within-
group variances are significantly smaller than between-group variances.
The available subgroupings of the population are shown in Table 5-1. The
basis of the first grouping process is to attempt to obtain a perceptual space
for each subgroup, and then to attempt to determine similarity of the spaces
between groups. The analysis is concerned with a single-dimensioned grouping
of the population. In other words, the analysis was carried out on the basis
of one socioeconomic variable at a time, without examination of two- or three-
way classifications of the population. The other two grouping processes are
based upon the use of direct preference rankings from the data and upon the
results of a factor analysis of the perceptual attributes. In both cases, the
analysis is still carried out on one socioeconomic variable at a time. Multi¬
ple classifications were not attempted.
5.2 Methods of Grouping
Increased understanding and improved modelling of travel behavior can
generally be obtained if the assessment of travel behavior is based on an
analysis of groups displaying similar behavior patterns. In this section,
strategies for identifying similar groups with respect to shopping-center
choice are identified. Socioeconomic characteristics are intended to provide
the basis for identifying the homogeneous groups. Two basic approaches to
118
Socio-Economic Variable
Sex:
Female
Male
Age:
Under 16 years
16-21 years
22 - 29 years
30 - 39 years
40 - 49 years
50 - 59 years
Over 59 years
Income:
Under $10,000
$10,001 - $15,000
$15,001 - $20,000
$20,001 - $25,000
$25,001 - $50,000
Over $50,000
Occupation:
Military
Salesman
Teacher
Professional
Craftsman
Clerical
Student
Housewife
Governmental
Retired
Other
Length of
Residence:
Less than 4 years
4-6 years
7-10 years
Over 10 years
TABLE 5-1
Socioeconomic Groups for First-Cut Analysis
119
group identification are discussed: a prior classification and a search for
classification. The adequacy of each grouping method is assessed through a
number of tests measuring the similarity within and/or differences between the
groups in terms of their choice functions, perceptions and preferences as
related to shopping centers.
5.2.1 Prior Classification
Prior classification methods identify similar groups directly according
to socioeconomic characteristics of the individual. Possible classification
characteristics include sex, age, income, occupation, and length of residence,
as shown in Table 5-1. The commonality of groups created by division with
respect to these characteristics can be tested with respect to shopping-center
choice models, preferences and perceptions.
It should be noted that this type of strategy would develop classifica¬
tions for variation within each socioeconomic variable individually. Such an
analysis, however, does not account for the effects of intercorrelated or
interacting socioeconomic factors which, if properly considered, could aid in
identifying groupings of individuals of a more homogeneous nature. A statis¬
tical technique known as the automatic interaction detector (AID) is based on
considering these factors in the grouping of observations. The basis of AID
is a sequential identification of subgroupings within a data set according to
explanatory variable variation with the aim of maximizing explained variation
in the dependent variable relative to the number of identified groups. The
technique has been applied successfully to a number of transportation problems
including mode choice. However, its use in shopping-center choice problems
would be difficult as a result of the number of alternative shopping-center
choices which define the dependent variable.
Another procedure for identifying similar groups is through a prior
classification according to reported or revealed preferences. Classification
according to reported preferences would use the ranked preference data collec¬
ted in the survey. Groupings could be developed according to the most prefer¬
red shopping center (< 7 groups), the two highest preferred shopping centers
in order (<_ 30 groupsj, or the three highest preferred regardless of order
(<^ 20 groups). Classification in terms of the shopping center chosen would
result in an analysis based on four groups. It also has been suggested that
the effect of distance of the shopping center used from home location be used
as a method of classifying the population. An examination of the possible
biases created by having two types of survey completion procedures -- mail-
back and interviewer assisted — may also be useful.
5.2.2 Search for Classification
The search for classification attempts to define common groups with
respect to preferences and perceptions. The relationships between these
groupings and socioeconomic categories is analyzed in an attempt to define
homogeneous socioeconomic groups. Contingency tables have been suggested as
a means of determining whether the common perception or preference groups are
significantly related to socioeconomic groups.
120
The search for common groups uses clustering techniques to perform an
aggregation of individuals according to similar preferences or perceptions.
Preference can be measured in terms of individual preference rankings of the
shopping centers of ideal-point location along relevant dimensions. Attribute
importance ratings may also provide a basis for assessing similarity. Deter¬
mination of common perceptions among individuals could be based on INDSCAL
weightings, factor scores, multidimensional-scaling distances between shopping
centers, or raw direct or indirect similarity data.
The technique suggested for use in performing the aggregation is cluster
analysis. Cluster analysis is used in problems generally involving the aggre¬
gation of n individuals to m groups so that the individuals in each group are
similar and are different from all other individuals not in the group. Simi¬
larity and difference in cluster analysis are defined in terms of a quantifi¬
able distance or similarity measure. The clustering process is usually hier¬
archical; that is, n clusters originally exist, which are then combined to
form n-1, n-2, and so forth until the desired level of aggregation is reached.
Many options exist with respect to the definition of distance and the hierar¬
chical clustering process. The options relevant to this study consist of those
available in the cluster analysis program implemented at the Northwestern Univ¬
ersity computing center.
The cluster analysis program, CLUSTAN 6000, has six basic clustering
procedures including hierarchic fusion, monothetic division, iterative-relo¬
cation mode analysis, Calinski-Harabasy dendrite, and Jardine-Sibson K-parti-
tion. In addition, 40 different distance/similarity measures can be used in
the clustering procedures. Principally, as a result of anticipated problem size,
hierarchic fusion and iterative relocation comprise the options available for
the proposed analysis.
Hierarchic fusion is a type of cluster analysis which when beginning with
n objects (individuals or groups) joins the two which are most similar, and
continues this fusion process until the desired level of aggregation is attain¬
ed. Similarity is defined by the similarity/distance measure. Iterative-re¬
location clustering differs from hierarchic fusion in that following each
cluster formation, an attempt is made to relocate every individual object among
existing clusters in order that the highest degree of similarity is achieved.
Information on the expected cost of using various clustering procedures or
similarity/distance options is not provided with the description of CLUSTAN.
A run with a population of about 100 objects has been estimated to cost about
$30. Populations of 200 and 500 will require from $70 to $200. It would seem,
from a conceptual point of view, that a procedure -- such as Ward's method,
which seeks to minimize within group (cluster) variation of distance measures
would be the preferred option as it, unlike many other methods, accounts di¬
rectly for group structure in the clustering of individual objects.
It should be noted that AID could also be used to identify groupings in
the search process. But, as stated earlier, proper consideration of the
dependent variable is not clear.
121
5.2.3 Tests for Similarity of Groups
Testing the similarity of groups identified by direct methods of classi¬
fication, can include analysis of homogeneity of preferences and perceptions
within groups, and determination of differences between groups. Testing of
differences between groups in terms of revealed preferences can be conducted
through a comparison of logit models of shopping-center choice estimated for each,
group and for the entire population. The models can be compared for significant
differences as a whole and by individual coefficients. Testing of differences
between groups can also be conducted with reported preferences with a modifi¬
cation of the Friedman test or a test based on an analysis of variance. The
difference tests of the modified Friedman statistic will indicate if the pref¬
erence rankings of shopping centers are different between groups. Analysis-
of-variance procedures may also be used to test these hypotheses concerning
ranking within and between groups, if problems concerning the ordinal nature
of the data and distribution characteristics of test statistics are resolved.
Groups identified as comprising homogeneous individuals by the two class¬
ification methods can also be analyzed with respect to their similarity of
perception of shopping centers. One means of testing the perception differ¬
ences between groups consists of comparing multidimensional scaling (MDS)
distances between shopping centers of various groups through correlation
analysis. This approach, however, cannot identify statistically significant
differences, but only provide an indication of variation among groups.
Another method of testing the similarity within groups would use either
form of similarity data which could be used in defining the MDS distance.
This approach would allow measurement of the significance of differences
between groups' perceptions of those distances. If the necessary normal
distributional assumptions can be accepted, a Hotelling test of the signifi¬
cance of differences in groups' mean similarity ratings of shopping centers
can be conducted. Another framework which can be utilized to test the signi¬
ficance of perception differences between groups with the similarity informa¬
tion is an analysis of variance. The proportion of total variance in those
ratings which can be accounted for by separating the total population into the
identified groups can then be established. These two tests, Hotelling and
analysis of variance, can also be conducted with the MDS distances if the
required normal distribution assumptions can be made. Similarity of MDS dis¬
tance can also be tested with the Friedman test when the distances are des¬
cribed by a rank order. Similarly, group differences in INDSCAL weightings
for relevant dimensions and factor scores could be used to test for perceptual
differences if distributional requirements can be assumed to be satisfied.
5.3 Segmentation on Perceptions
5.3.1 General
In order to understand the problems of seeking homogeneity of perceptual
spaces, some understanding is necessary of the multidimensional-sealing proce¬
dures and the results generated by these procedures. The perceptual spaces
are, in this case, generated as aggregate spaces for a preselected group or
subgroup of the population. In other words, for each socioeconomic group
identified in Table 5-1, one aggregate perceptual space is developed. The
122
aggregate multidimensional-sealing procedure involves the selection of a
dimensionality that is most efficient for representing the aggregate informa¬
tion obtained on perceived "distances" between the set of stimuli, (shopping
centers in this case). These distances may be obtained by questions that
request directly information on the similarity that people perceive between
alternative shopping centers vis-a-vis some prespecified metric or quality,
or may be derived by asking people to rate each of a set of shopping centers
on a number of different attributes, postulated as making up the quality or
metric to be used for judging similarity. The two types of questions from
which multidimensional scales might be developed, used in the first Old Orchard
surveys, are shown in Figures 5-1 and 5-2. If a set of n-stimuli are used in
either of these two types of questions, then the distances between the stimuli
may be represented uniquely in (n-1)-dimensioned space. For example, the first
survey used seven shopping-center locations for the two types of questions.
Thus, the interpoint distances may be represented uniquely in six-dimensional
space. The procedures for developing perceptual distances between stimuli are
detailed in Appendix D.
In the method used, average distances are computed for each of the iden¬
tified subgroups in the population. These distances are distances between
each of the seven shopping centers in the perceived space of attractiveness
to shop. The first task of the analysis if to find the most efficient dimen¬
sionality in which to express the perceptual space for the attractiveness
concept. Thus, with the seven shopping centers, the highest dimensionality
possible is a six-dimensional space. However, one would like to reduce this
space to as few dimensions as possible, without distorting the perceived dis¬
tances between the shopping centers. This is the procedure that the multi¬
dimensional-scaling program (MDSCAL) performs. In carrying out a collapsing
of the dimensionality of the space, the procedure requires that a monotonie
relationship be preserved between the original interpoint distances and those
in each successive reduced-dimensionality space. The monotonicity requirement
is placed upon the procedure, rather than a strict linear requirement, since
the data from which the information is derived is only ordinal in nature.
Thus, it would not be appropriate or correct to invest ratio properties in the
base data, nor to require preservation of the sizes of the intervals between
stimulus points in the space in the collapse process. In the process of devel¬
oping a perceptual space through the MDSCAL program, the orthogonal axes des¬
cribing the space are located arbitrarily. Thus, there is no ready mechanism
for comparing the final resulting multidimensional spaces with each other from
different socioeconomic subgroups of the population. The reasons such a com¬
parison is not possible is that the orientation of the spaces is completely
arbitrary over the entire set. Thus, no two spaces are necessarily located in
any common way. Both rotation and translation of the axes is possible from one
space to another. Thus it becomes quite impossible for the analyst to be able
to compare multidimensional solutions from alternative subgroups and derive any
information of the comparative values or differences between the spaces. Fig¬
ures 5-3, 5-4, and 5-5 show three solutions from the multidimensional scaling
process for different subgroups of the population. It is clear from these that
conclusions cannot be drawn, given that axes may be rotated or translated at
will from one to the next.
123
Again, if all the shopping centers were
equally easy to get to, how similar do you think
they are to each other? In answering this ques¬
tion, please think about your preference to shop
at them for the goods you came to buy. Check
the box which best describes how similar they
are. Please be sure to do this for all pairs
of shopping centers.
57 ^
<U **
ry O
Woodfield
Edens Plaza
Woodfield
Chicago Loop
Old Orchard
Golf Mill
Chicago Loop
Plaza del Lago
Xorvette City
Plaza del Lago
Cnicago Loop
Woodfield
Edens Plaza
Old Orchard
Edens Plaza
Old Orchard
Xorvette City
Ch icago Loop
Woodfield
Plaza del Lago
Edens Plaza
and
and
and
and
and
and
and
and
and
and
and
and
and
and
and
and
and
and
and
ajid
and
Chicago Loop
Golf Mill
Plaza del Lago
Coif Mill
Woodfield
Xorvette City
Old Orchard
Golf Mill
Old Orchard
Edens Plaza
Xorvette City
Golf Mill
Xorvette City
Golf Mill
Woodfield
Plaza del Logo
Plaza del Lago
Edens Plaza
Xorvette CiPy
Chicago Loop
Old Orchard
if
FIGURE 5-1
Direct Similarity Measurement
124
In this question, we would like you to rate each of the shopping centers on these characteristics. We have
provided a range from good to poor for each characteristic,
center fits on this range.
We would like you to tell us where you feel each shopping
For example:
good J
Eating
Facili¬
ties
poor
IS
m
£
S
J
se
-J
O
■o
o
Please do this for all the shopping centers on all the characteristics,
center, please guess where you think it would fit.)
h to
a.
o
J
n
3
3
3
è
s
•a
3
(If you are not familiar with a shopping
o°-
4
Sr
c,
Cj
o
Z
ej
•c
CO
4
3
■o
4-1
CJ
è
o
•tl
IS
«8
■o
a1
good
good
Layout
of
store
poor
Ease of
re turning
or
servicing
merchandise
poor
good
Prestige
of
store
Variety
or range
of
merchandise
good
poor
poor
good
Quality
of
merchandise
good
Availability
of
credit
poor
poor
good
Reasonable
price
Availability
of sale
items
("specials")
good
poor
poor
FIGURE 5-2
Indirect Similarity Measurement
125
Dim
Golf Hill
"0.5
1..qqp
xKorvettes
. 2
1.5
* Old Orchard
1.0
h •-dim, 1
-1.0 -0.5
0.5 1.0
x Edens Plaza
Plaza del Lago * ^.5
- -1.0
-1.5
x woodfield
FIGURE 5-3
Two-Dimensional Space for Clerical Workers
126
Din, 2
s >
"2.0
*0ld Orchard
-1.5
"1.0
-0.5
XEdens Plaza
. Loop „ * pla,za del lago.
1 1 ^— 1 1
-1.0 -0.5 0.5 X 1.0 Dim- 1
Korvettes
- r *Golf Mill
-0.5
-1.0
* Woodfield " -1.5
FIGURE 5-4
Two-Dimensional Space for Incomes Over $50,000
127
Dim, 2
i <
* Old Orchard
Edens Plaza * "0,5
-1.0
4-
-0.5
* Plaza del
Loop
Lago
X
1.0
x Korvettes
0,5
1.0
" -0.5
Golf Hill
-1.0
woodfield
-•-Dim. 1
FIGURE 5-5
Two-Dimensional Space for Incomes of $10,000-$15,000
128
In order to be able to draw comparisons between the spaces, it is necess¬
ary to find a means by which alternative spaces can be compared. Two such
processes appeared as candidates for this from the multidimensional-scaling
work. First, the multidimensional-scaling results in the production of a new
set of interpoint distances for the most efficient space determined. These
interpoint distances, which represent average distances for members of each
subgroup in the most efficient dimensionality space, may be considered as a
set of candidate values that describe each subgroup in terms of the perceptual
space. Thus, one may compute either a rank or metric correlation between the
sets of distances of one group with another. Since there are seven stimuli in
the space, there are 21 interpoint distances that are necessary to describe
each multidimensional solution. Thus one may obtain a correlation by taking
the 21 distance measures for each of two subgroups and computing either a rank
(Spearman) correlation or a linear (Pearson) correlation between them. Such
a measure is computed irrespective of the rotation and translation of the re¬
presentation of the multidimensional space. All that one is looking for here
is a correlation of the distances between each pair of points. Again, since
the original data from which the spaces were derived was ordinal in nature,
rather than cardinal, and since the procedure for developing the multidimen¬
sional space requires only the preservation of ordinal information, it may be
more appropriate to consider a rank correlation than a linear correlation.
In fact, both types of correlations were run for these data and comparisons
made between the results obtained. In general, it was found that the rank
correlations were higher than but consistent with the Pearson correlations.
Therefore, only the latter are reported.
A second procedure was devised within this research for determining the
comparisons between alternative attractiveness spaces. This procedure involved
the use of the average interpoint distances for each subgroup as an input to
the individual-sealing analysis method (INDSCAL). The individual- scaling
model is described in 4.3.3.
For this research, it was proposed that one might substitute average inter-
point distances, derived from MDSCAL, for the individual data that would nor¬
mally be the input to INDSCAL. In this manner, each of the multidimensional
spaces found for the socioeconomic subgroups could be fitted to a common space,
and the output weights on the various dimensions of the common space would pro¬
vide a metric that could be used in some type of correlation or cluster analysis.
Naturally, it is clear that such a process loses the information of variance
within each group. It is not clear how serious such a loss of information is
in this instance. However, there is no way in which the information can be
incorporated in the process. Subsequent analyses have attempted to develop
alternative procedures for determining the appropriate market segments in the
population.
Having fitted each of the socioeconomic subgroups into a common space, and
obtaining weights for each of the axes of that common space, the cluster anal¬
ysis is performed upon the weights, from which a hierarchy of groupings of the
original subgroups can be determined. It is important to note, however, that
neither of the methods proposed here have associated with them any statistical
measures of goodness of fit.
129
At a number of points in the preceding discussion, the selection of a
parsimonious space has been mentioned. So far, however, no discussion has
been addressed to the question of how parsimony and efficiency are determined.
As an aid to such selection, a statistic (Stress) has been developed by Kruskal
(1964), which measures the degree of distortion introduced by each solution
produced. Thus, as the dimensionality is reduced from the starting configura¬
tion of, say six dimensions, a value of Stress can be computed that can then
be used to determine whether the lower dimensionality solution is acceptable
or not. A set of empirical values have been determined for Stress, in terms
of specifying degree of goodness of fit to the original data. These values
are provided with descriptions, in the following form: perfect fit, excellent
fit, good fit, fair fit, poor fit. Ideally, a plot of the value of Stress
against the dimensionality will show a characteristic elbow, as shown in Figure
5-6. Conceptually, this figure indicates that reduction in dimensionality ini¬
tially causes no distortion in the interpoint distances. However, a point is
reached at which a further reduction of one dimension causes significant dis¬
tortion. It may therefore be assumed that the dimensionality immediately pre¬
ceding the substantial increase in Stress indicates the most efficient and
parsimonious multidimensional-scaling solution. The stress will not always
behave in this precise fashion. It will, however, either remain approximately
constant, exhibit a well-defined elbow, or will have a generally upward slop¬
ing curve for reducing dimensionality. In general, no other forms are possible.
In this research, all subgroups were run for four-, three-, and two-dimen¬
sional solutions. In each case, a plot was obtained of the stress against the
dimensionality and this was used to select the appropriate dimensionality for
that particular subgroup. In most instances, it was found that the change of
Stress with dimensionality followed the ideal plot shown in Figure 5-6. In
these cases, the decision of the most efficient dimensionality was obvious.
In some instances, however, the stress followed a more-or-less straight line
increasing with decreasing dimensionality. In these cases, a solution was
chosen that was based upon the interpretations of fit developed by Kruskal.
Where possible, the lowest dimensionality was chosen that was consistent with
the empirical range for good to excellent fit. In some instances, it was found
that the change in Stress was such that two or more dimensionality solutions
fell within the same region of fit as each other. In these instances, more
than one dimensionality was selected as a solution. The selected solutions are
shown in Table 5-2.
5.3.2 Correlation Analysis
As indicated earlier, the first item of analysis undertaken was to deter¬
mine both Pearson and Spearman correlations among the interpoint distances
from the MDS solutions for the entire survey sample of 7,362 observations.
Correlations were determined for subsets of the subgroups, determined by the
dimensionality of the solution selected. Thus, one set of correlations was
determined for four-dimensional solutions, a second for three-dimensional
solutions, and a third for two-dimensional solutions. It was not felt to be
valid to compute correlations between groups whose representations were in
different dimensionalities. The two types of correlations are distinguished
130
0.6 -■
0.5 --
0.4 +
FIGURE 5-6
Plot of Dimensionality against Stress for 2, 3, and 4 Dimensional Solutions
Socio-Fconomic Variable
Dimensionality
Sex: Female
3
Male
4, 3
Age: Under 16 years
4, 3
16 - 21 years
4, 3
22-29 years
4
30 - 39 years
3
40 - 49 years
3
50 - 59 years
4, 3
Over 59 years
4, 3
Income: Under $10,000
3
$10,001-$15,000
2, 3
$15,001-$20,000
3
$20,001-$25,000
4
$25,001-$50,000
4
Over $50,000
3, 2
Occupation: Military
*
Salesman
4
Teacher
3
Professional
4, 3
Craftsman
3
Clerical
3, 2
Student
4
Housewife
4, 3, 2
Governmental
4, 3
Retired
4, 3
Other
3
Length of Less than 4 years
4, 3
Residence: , ,
4-6 years
4, 3
7-10 years
4, 5
Over 10 years
1
4, 5
TABLE 5-2
Selected Dimensionalities for One-Way Groupings
132
by the fact that the Spearman correlations are correlations only on the rank
ordering of the interpoint distances, while the Pearson correlations are
linear-regression type correlations that are determined by assuming the dis¬
tances to be metric distances. The only correlations of interest are those
within a particular socioeconomic group. These are shown in Tables 5-3
through 5-6 for the four-dimensional solutions, Tables 5-7 through 5-11 for
three-dimensional solutions and all two-dimensional solutions are shown in
Table 5-12. In all cases, only the Pearson correlations are shown, since
these were found to be consistently lower than the Spearman rank correlations.
This is not completely surprising, since the Spearman rank correlations are
based upon less information, and are therefore weaker than the Pearson corre¬
lations.
Considering the tables, one may use as a rule of thumb that correlations
below .5 indicate relatively little association between the variables, while
those above .5 indicate a fairly substantial degree of correlation. On this
basis, one may conclude that there are relatively high correlations between
the sexes for the three-dimensional solutions, as shown by Table 5-7. On the
basis of the four-dimensional solutions, one could potentially group the under
16 year olds with the 16-21 year olds, and the 16-21 year olds with the 22-29
year olds. Since, however, the correlation between the under 16 year olds and
the 22-29 year olds is relatively low, one might conclude that an optimal
combination would be under 22 rather than breaking at 16. Similarly, Table 5-3
shows a high degree of correlation between the 50-59 year-old group and the
60 and over group. Relatively high correlations seem to be demonstrated
between the 60 and overs and all of the other age groups except the 16 to 21
year olds. It is not completely clear why this might be so, but may indicate
that this particular age category is not a useful one for discriminating per¬
ceptions of shopping opportunities. In contrast, Table 5-8 shows a very low
correlation between the under 16 year olds and 16 to 21 year olds in a three-
dimensional solution, and the only high correlations are to be found between
the age groups 40-49, 50-59, and 60 and over. In fact, the conclusion from
Table 5-8 would probably be that one age group from 40 and over would be
sufficient to describe age groups with respect to perception of shopping-center
destinations.
It does not appear to be very meaningful to look at major combinations of
occupational categories. An examination of Tables 5-4 and 5-9 does tend to
suggest that there are some quite strong correlations between certain occupa¬
tional categories, and very weak ones among others. For example there are high
correlations between clerical-secretarial workers and teachers, between house¬
wives and clerical and secretarial workers, and between housewives and retired
people. Precisely what conclusions can be drawn from this is not clear.
Nevertheless, the correlations are reported for completeness of the analysis.
Based upon the information on income, shown in Tables 5-5, 5-10 and 5-12, it
does not appear as though income is a good discriminator of perceptions of
shopping center destinations. Indeed there are no correlations in any of
these tables below .5, and some of the highest correlations to be seen are
found in these tables. Finally, Tables 5-6 and 5-11 indicate a very clear
polarization on length of residence. There is a high correlation between
those people who have lived in the area less than four years and those who
have lived in the area four to six years, and a similarly high correlation
between those having lived in the area seven to ten years and those over ten
years. Both tables, which are for different dimensionalities, exhibit the
133
Aqe j
Group
<15
15-21
22-29
50-59
60 and over
<16
^679
.500
.528
.574
16-21
<555
.312
.451
| 22-29
.51-6
.671
J 50-59
.726
60 and
over
TABLE 5-3
Pearson Correlations Among Age Groups
(4-dimensional solutions)
TABLE 5-4
Pearson Correlations Among Occupational Groups
(4-dimensional solutions)
134
Income
$20-25,000
$25-50,000~]
$20-25,000
.643
$26-50,000
TABLE 5-5
Pearson Correlations Among Income Groups
(4-dimensional solutions)
Lencsth
of
Residence
<4 yrs.
4-6 yrs.
7-10 yrs.
1 over 10 yrs.
<4 yrs.
.870
.565
I
.532 |
4-6 yrs.
.449 I
.383 j
7-10 yrs.
.876 j
over 10 yrs. |
a
1
e
~
TABLE 5-6
Pearson Correlations By Length of Residence
(4-dimensional solutions)
135
Sex
Male
Female
Male
.632
Female
TABLE 5-7
Pearson Correlation Between the Sexes
(3-dimensional solutions)
Age
<16
16-21
30-39
40-49
50-59
60+
<16
.23*
.209
.187
.306
.149
16-21
.455
.368
.422
.613
30-39
.443
.342
.644
40-49
.577
.684
50-59
.793
60+
\
TABLE 5-8
Pearson Correlations Among Age Groups
(3-dimensional solutions)
136
Occupation
Teach
Prof.
Crafts
CI er/
Sec
Hsv/fe
Govt.
Ret'd
Other
Teacher
.522
.608
.870
.852
.409
.755
.465
Professional
.310
.422
.499
.122
.578
.433
Crafts.
.636
.53 7
.318
.670
.483
Clér./Sec.
.922
.496
.799
.577
Housewi fe
.559
.816
.390
Government
.358
.174
Retired
.487
Other
TABLE 5-9
Pearson Correlations Among Occupational Groups
(3-dimensional solutions)
Income
A
OT-
O
7^
10-15 K
15-20K
>5 OK
<$1 OK
x^
.637
-0^
.901
.598
10-15K
.586
. 794
15-20K
.59Q
>50l<
TABLE 5-10
Pearson Correlations Among Income Groups
(3-dimensional solutions)
137
Length
of
Residence
<4 yrs.
4-6
7-10
Over 10
<4 yrs.
.723
.490
.490
4-5 yrs.
.447
.445
7-10 yrs.
1.00
Over 10 yrs.
TABLE 5-11
Pearson Correlations by Length of Residence
(3-dimensional solutions)
Subgroup
Cler.
Hswfe
$10-15K
>S50K
CI er.
.835
/
Hswfe
y
/
$10-15K
.784
>$50K
X'
TABLE 5-12
Pearson Correlations for Income and
Occupation Subgroups Having 2-dimensional Solutions
138
identical pattern. Correlations between the other pairs of groups are nowhere
near as high, and in Table 5-11 are all less than .5. One can conclude from
this that a grouping of length of residence with a break point at six years
would appear to be appropriate. This is by far the strongest result obtained
in this analysis.
5.3.3 Cluster Analysis of Perceptions
A cluster analysis was performed on the weights for each subgroup obtained
from the INDSCAL analysis using the hierarchic-fusion process.
The cluster analysis provided various hierarchical levels of clustering of the
subgroups. Generally, only the lowest level of clustering was considered to be
worthwhile examining. Results of the clustering of four-dimensional solutions
are shown in Table 5-13, and those from the three-dimensional solutions are
shown in Table 5-14. The two-dimensional solutions were not subjected to a
separate cluster analysis. On the basis of these two tables, it is again evi¬
dent that length of residence may be divided at between six and seven years,
based upon the original categorization in the questionnaire. This result
appears for both three- and four-dimensional solutions, and is completely con¬
sistent with the results determined in the correlation analysis. Again, some
groupings of occupations appear within the two tables, and these are generally
rather similar to those found in the correlation analysis. For example, the
correlation analysis found a high correlation between housewife and retired
for the four-dimensional solutions, which again shows up in Table 5-13. Simi¬
larly, one could group clerical and secretarial workers with housewives and
retired people in Table 5-9 and the same grouping appears in Table 5-14. How¬
ever, there is also one inconsistency in the occupational groupings, in that
the cluster analysis groups the professional, craftsman, and governmental
employees, while these are seen, in Table 5-9, to have very low correlations
with each other.
Both the correlation analysis and the cluster analysis on INDSCAL weights
show a possible grouping, at four dimensions, of the under 16 year-olds and the
16-21 year-olds. The cluster analysis did not show a grouping of those in the
age groups of 50-59 and over 59. The results from the three-dimensional solu¬
tions remain consistent in grouping the under 16 year-olds and the 16-21 year-
olds, while this was not found so in the correlation analysis. The cluster
analysis also groups the over 59 year-olds with the same group, a correlation
that is again not exhibited in Table 5-8 under the correlation analyses. In
the separate analyses, the cluster analysis shows no clustering of income groups,
while the conclusion drawn from a correlation analysis was that income was a
very weak determinant of perceptual differences within the population. Finally,
both the correlation analysis and the cluster analyses indicate that sex is a
poor discriminator of perceptual differences in the population.
In the correlation analysis, it was not considered to be appropriate
to run correlations across different dimensionality solutions. As a result,
the correlation analysis has a number of gaps, where solutions are not always
obtained in the same dimensionality for all subgroups. In contrast, it was
felt to be reasonable to attempt a cluster analysis of INDSCAL results combined
across all dimensionalities. In order to do this, the INDSCAL was run in a
four-dimensional and a three-dimensional mode, and all MDS results were input.
139
Ox-.ig.inal Characteristics
Cluster
Residence: Over 10 years
Over 6 years
7-10 years
J
4-6 years
\ Under 6 years
under 4 years
J
Occupation: Salesman
Salesman
Px-ofessional
Professional
Student
Student
Housewife
\ Housewife or Re-
Retired
J tired
Governmental
Governmental
Age: Under 16 years
^ Under 22 ycai'S
16 - 21 years
J
22 - 29 years
22 - 29 years
50 - 59 years
50 - 59 years
Over 59 years
Over 59 years
Income: $20,001 - $25,000
$20,001 - $25,000
$25,001 - $50,000
$25,001 - $50,000
TABLE 5-13
Clustering of Four-Dimensional Solutions Within
Socio-Economic Variables
140
Characteristics
Cluster
Sox: Female
Male
Age: Under 16 years
16-21 years
30 - 39 years
40 - 49 years
50 - 59 years
Over 59 years
Occupation: Teacher
Professional
Craftsman
Clerical
Housewife
Governmental
Retired
Other
Income:
Residence:
$10,000 and under
$10,001 - $15,000
$15,001 - $20,000
Over $50,000
Less than 4 years
4-6 years
7-10 years
Over 10 years
}
}
}
}
}
}
Combine Sexes
Under 22 years
and over 59 years
30 - 39
40 - 49
50 - 59
(see above)
Teacher
Professional, Crafts¬
man and Governmen¬
tal
Clerical, Housewife
and Retired
(see above)
(see above)
Other
$10,000 and under
$10,001 - $15,000
$15,001 - $20,000
Over $50,000
6 years and under
Over 6 years
TABLE 5-14
Clustering of Three-Dimensional Solutions Within
Socio-Economic Variables
141
The MDS results used for INDSCAL comprise only the interpoint distances. The
dimensionality of the solution does not, therefore, affect the number of inter-
point distances that are determined in any space. The results of these com¬
bined runs for INDSCAL are shown in Table 5-15. In general, it may be noted
that there are considerable consistencies across the three-dimensional and
four-dimensional solutions for the combined results, and similarly consistency
between these results and the ones for the separate dimensionality solutions
in Tables 5-13 and 5-14. In general, the differences that may be observed
between Table 5-15 and the results in Tables 5-13 and 5-14 are more consistent
with the results from the correlation analysis. This may be due to the fact
that the level of clustering is set arbitrarily in each instance, and the level
at which clusters are formed and reported in Table 5-15 may be a higher level
than that at which they are formed and reported in Tables 5-13 and 5-14. Unfor¬
tunately, as has been noted previously, there are no statistical measures that
can be used to define or assess levels of clustering in this type of exercise.
One or two of the notable results are that the results of Table 5-15 show a
clustering of incomes from $20,000 to $50,000, which is becoming more consis¬
tent with the results for the correlation analysis, as reported in Table 5-5
and Table 5-10. Similarly, occupational grouping of teachers, housewives,
clerical and secretarial workers, and the retired is also consistent with find¬
ings in Tables 5-4 and 5-9. The identification of student and other occupa¬
tional categories as having no strong grouping with any other group is also
borne out in both Table 5-15 and the earlier results of the correlation analysis.
Groupings of sex, ages, and length of residence are fairly consistent between
Table 5-15 and Tables 5-13 and 5-14, and again with the correlation analyses.
A further point of interest in Table 5-15 concerns the groupings of the
solutions for different dimensionalities of the same attribute. In general,
it may be concluded that where the two dimensionality solutions for the same
subgroup are clustered, the selection of the lower-dimensionality solution
would not introduce any biases into the process. In other words, one could in
these cases consider the lower dimensionality as being appropriate. This would
be the case, for example, for the age groups of under 16 and 16-21, 50-59 years,
and over 59 years. Similarly, it would be appropriate for the income group from
$10,000-$15,000 and for the occupational groups of teacher, professional, cler¬
ical, housewife, governmental, and retired. Likewise, it would be appropriate
for the length-of-residence variable to be considered only at a three-dimen¬
sional solution, rather than four. Interestingly, there does not appear to be
a close similarity between three-dimensional and four-dimensional solutions
for males. This would tend to suggest that a significant bias is introduced by
dropping from four dimensions to three dimensions and may, therefore, require
further analysis of whether or not sex is a good discriminating variable of
perception.
5.4 Identification of an Attractiveness Space
The preceding section has indicated the process by which multidimensional
scaling may be applied to the data obtained in order to develop representations
of the perceptual space that describes the attractiveness of each of the shop¬
ping centers of concern in this research. In addition to determining the space,
it is also desirable to be able to label the axes of the space. This is done
by using the information on the attributes that have been postulated as making
up the concept of attractiveness. The questions that provide the information
142
Original Characteristic
3D Cl us ter
40 Cluster
Sex:
Age:
Income:
Occupation:
Residence:
Female (3)
Male (3)
Male (4)
Under 16 years (3)
Under 16 years (4)
16 21 years (3)
16 - 21 years (4)
22 - 29 years (4)
30 - 39 years (3)
40 - 49 years (3)
50 59 years (3)
50 - 59 years (4)
Over 59 years (3)
Over 59 years (4)
Under $10,000 (3)
$10,001 $15,000 (2)
$10,001 - $15,000 (3)
$15,001 $20,000 (3)
$20,001 - $25,000 (4)
$25,001 - $50,000 (4)
Over $50,000 (2)
Over $50,000 (3)
Salesman (4)
Teacher (3)
Professional (3)
Professional (4)
Craftsman
Clerical (2)
Clerical (3)
Student (4)
Housewife (2)
Housewife (3)
Housewife (4)
Governmental (3)
Governmental (4)
Retired (3)
Retired (4)
Other (3)
ess than 4 years (3)
ess than 4 years (4)
6 years (3)
6 years (4)
10 years (3)
10 years (4)
Over 10 years (3)
Over 10 years (4)
}
}
}
>
}
}
}
}
>
}
Male (3) and
Female (3)
Male (4)
Under 16 (3)
Under 22 (4)
16 - 21 years (3)
Under 22 (4)
22 29 years (4)
30 - 39 years (3)
40 - 49 years (3)
50 - 59 years (3&4)
Over 59 years (384)
Under $10,000 (3)
$10,001 $15,000 (2&3)
$15,001 $20,000 (3)
$20,001 $50,000 (4)
Over $50,000 (2)
Over $50,000 (3)
Salesman
Teacher (3), Housewife
(2,3,4), Clerical (2,3),
Retired (3,4)
Professional (3&4),
Craftsman (3), Govern¬
mental (384)
(See Teacher)
Student (4)
(See Teacher)
(See Craftsman)
(See Teacher)
Other (3)
. 6 years and under
(3 8 4)
■ Over 6 years (3 6 4)
}
Male (3) and
Female (3)
Male (4)
Under 16 (3)
Under 22 (4)
16 21 years (3)
(See 16 21 (4))
22 - 29 years (4)
30 39 years (3)
40 49 years (3)
50 59 years (384)
Over 59 years (384)
Under $10,000 (3)
y $10,001 $15,000 (283)
$15,001 $20,000 (3)
J- $20,001 - $50,000 (4)
Over $50,000 (283)
Salesman (4) and
Professional (3,4)
Teacher (3), Housewife
(2,3,4), Clerical (2,3),
Retired (3,4)
}
See Salesman
Craftsman (3),
Government (384)
(See Teacher)
Student (4)
^(See Teacher)
y (See Craftsman)
^•(See Teacher)
Other (3)
►6 years and under
(3 8 4)
►Over 6 years (3)
7 10 years (4)
Over 10 years (4)
TABLE 5-15
Clustering of All Solutions in Three and Four Dimensions
143
on this are shown in Figure 5-2. Each of the individual semantic scales on an
attribute provide a mapping of the locations of shopping centers along that
attribute. Having determined a multidimensional space that describes attrac¬
tiveness as parsimoniously as possible, one may determine the most appropriate
projection in that space of each of the attributes with their positioning of
shopping centers along them. In other words, the desire is to find that orien¬
tation of each of the attributes that correlates best with the positionings of
the relevant shopping centers obtained for that attribute. This is determined
by projecting the shopping center positions in the perceptual space onto a line
representing each attribute. The process for fitting the properties into the
space may be carried out through two alternative procedures — PROFIT or PREF¬
MAP. Primarily, the PROFIT program fits attributes into the space by using a
linear regression between the projected positions of the stimuli in the space
and their positioning along an attribute, where the positioning is assumed to
have metric or ratio-scale properties. The PREFMAP procedure allows the pro¬
perties to be fitted through a monotonicity constraint only. In other words,
in PREFMAP the procedure may be either a normal linear regression or a mono-
tonic regression, where the latter preserves only the ordinal information and
no metric information. While the project staff felt that the monotonie fitting
of properties in the space was more appropriate for the available data, the
PREFMAP program was not available to the research team at the time that this
research was conducted. However, the PROFIT program was available and was
therefore used to locate attributes in the multidimensional space.
By examining the positions of the attributes in the space, by means of
the correlations between each attribute and each of the axes of the selected
space, together with the importances of each attribute, an interpretation can
be obtained of the meaning of each axis used to describe the space. The pro¬
cedure is analogous to that used for identifying the meaning of factors in a
metric factor-analysis procedure. In fact, as is reported in a subsequent
chapter, these data can be subjected to a factor analysis, provided that
the assumption that the data contain metric information is accepted. Again,
since the axis locations for the multidimensional-scaling solutions are arbi¬
trary, rotation and translation of axes is possible to obtain a better explan¬
ation or identification of the meaning of axes. In this research, it was not
felt to be relevant to attempt to identify the meanings of the axes for each
of the individual subgroups from the multidimensional-scaling output. Rather,
it was felt that the appropriate strategy would be to examine the common space
developed from the INDSCAL procedure and identify axes in that space. One of
the reasons for doing this is that the INDSCAL process has' already produced a
common space for the representation of all of the solutions for all socioeco¬
nomic subgroups. Thus, the number of degrees of freedom for defining sets of
axes is substantially reduced.
As previously indicated, the interpoint distances from the selected multi¬
dimensional representations were then used as inputs to an individual-scaling
(INDSCAL) procedure, from which weights were determined for each of the sub-
groups. These weights were subjected to cluster analysis, as previously indi¬
cated. Finally, the overall space developed for INDSCAL in both three and four
dimensions was identified by fitting the sixteen attributes into the space,
using the program PROFIT (Carroll and Chang. 1970). A visual examination of
the plots reveals that the first two dimensions of both the three- and four-
dimensional solutions are virtually identical, while the third dimension of the
three-dimensional space is significantly different from either the third or
144
fourth dimensions of the four-dimensional space, but where it appears that the
four-dimensional space provides two new dimensions that replace the third dimen¬
sion of the three-dimensional solution. The three-dimensional solution is shown
in Figures 5-7, 5-8, and 5-9, with the attribute vectors plotted in the space.
In general terms, the dimensions can be interpreted as quality and convenience
on dimension one, price on dimension two, and a measure of expedience on dimen¬
sion three. This last dimension appears to be a dimension that relates to the
experience of an individual in undertaking a shopping trip. Two attributes,
location of stores in a compact area and a number and variety of stores, did
not load significantly on any one dimension and may, potentially, relate to a
fourth dimension not defined in this solution. However, the use of this
attractiveness space is not recommended, since full analysis of the clusters
has not been completed.
5.5 Preference Segmentation
In this element of the study, the data used were of the preference rankings
on shopping centers, shown in Figure 5-10. As before, the purpose of the seg¬
mentation was to identify homogeneous groups on the basis of known socioeconomic
characteristics. The classification characteristics were the same as for the
previous segmentation exercise and are shown in Table 5-1. The values of these
preference ranks are shown in Tables 5-16 through 5-20. The difference between
this analysis and the one reported in section 5.3 needs to be emphasized. The
earlier analysis was based upon stated perceptions of the shopping locations.
Thus, the segmentation was carried out on the way in which people perceive the
shopping locations without any information on how much they like various attri¬
butes. For example, an individual may rate a shopping location as offering
considerable variety of stores. This is a perception measure. However, that
same individual may find many stores to be confusing, so that he may prefer to
shop at a smaller shopping center. This would be a measure of preference. In
this analysis, commonality of preferences is being sought, rather than common¬
ality of perception.
The procedure for determining the similarity of preference rankings was a
modified Friedman Test (see Appendix E). In this test, the null hypothesis is
that there are no meaningful groupings in the data. The Friedman T value is
distributed like x2 for large samples, such that a large value of T will tend
to lead to rejection of the null hypothesis, i.e., will support the assumption
that there are market segments. Tests were run on various groupings of each
socioeconomic variable and for adjacent pairs of groups to attempt to determine
whether or not significant groups were present and how various groups should be
combined. The results are shown in Tables 5-21 through 5-25.
The only significant factors found in this analysis are age and income.
In the age groupings, five subgroups were found to be optimal, consisting of:
under 22, 22-29, 30-39, 40-59, and 60 and over. Similarly, income was found
to be grouped optimally into the four subgroups: less than $10,000, $10,000-
$24,999, $25,000-$50,000, and over $50,000. None of the other variables pro¬
duced T values that were significant at 95% or better. Examination of the
preference rankings in Tables 5-17, 5-19, and 5-20 appear to be consistent
with these findings.
145
3 DIM
DLL GROUPS
DIM 3
FIGURE 5-7
146
3 DIM fil.L GROUPS
DIM 2
Figure 5-8
147
3 DIM DLL GROUPS
DIM 3
Figure 5-9
148
If all the following shopping centers
were equally easy to get to, which of
them would you prefer to shop at for the
goods you came to buy?
ate yo
numbe
ence by placing a number beside each
center. Start with number 1 for the
most preferred shopping center, number
2 for the second most preferred, and so
on down to the least preferred shopping
center.
Please rank all the shopping centers.
Chicago Loop [ ]
Edens Plaza (Wilmette) [ ]
Golf Mill Shopping Center
Korvette City (Dempeter &
Waukegan)
Plaza del Lago [ ]
Old, Orchard [ ]
Woodfield [ ]
FIGURE 5-10
Preference-Ranking Question for the First Survey
149
Age Group
Shopping
Centers
Less than
16 years
16-21
years
22-29
years
30-39
years
40-49
years
50-59
years
60 years
and Over
1
3.5
3.7
4.2
4.4
4.1
4.0
5.0
2
4.2
4.5
4.3
4.4
4.0
4.0
3.4
3
4.2
3.8
3.7
3.1
3.6
3.4
3.1
4
6.4
6.3
6.1
5.9
6.1
6.0
5.9
5
2.3
2.2
2.2
2.1
1 .8
1 .8
1.4
6
5.4
5.1
4.9
5.1
5.1
5.5
5.7
7
2.1
2.5
2.5
2.9
3.2
3.3
3.5
Total
Observations
19
107
143
89
83
46
11
TABLE 5-16
Preference Rankings of Population
Groupings According to Age
150
Occupation
Shopping
Centers
Sales
Teacher
Profes¬
sional
Craftsman
Clerical
Student
Housewife
Government
Retired
Other
1
4.1
4.1
4.3
5.8
4.1
3.7
4.4
4.0
4.0
4.1
2
4.2
4.2
4.3
4.3
4.2
4.4
4.1
2.0
4.3
4.3
3
3.9
3.4
3.4
2.3
3.4
3.8
3.4
6.0
4.0
3.6
4
6.2
6.1
6.1
5.8
6.1
6.3
5.9
6.0
7.0
6.1
5
2.0
2.3
1.9
2.5
2.0
2.2
2.0
2.3
1.7
2.1
6
5.0
5.2
5.2
5.8
5.1
5.1
5.1
4.7
4.3
5.1
7
2.5
2.9
2.7
1.8
3.0
2.4
3.0
3.0
2.7
2.8
Total
Observations
22
51
86
4
54
120
140
3
3
15
TABLE 5-17
Preference Rankings of Population
Groupings According to Occupation
Income Group
Shopping
Centers
Less than
$1 OK
$10-15K
$15-20K
$20-25K
$25-50K
$50K Plus
1
4.3
4.2
4.7
4.2
3.8
2.8
i 2
4.4
4.2
4.2
4.4
4.2
4.3
3
3.2
3.4
3.4
3.5
3.9
4.3
4
5.8
5.9
6.0
6.0
6.3
6.6
5
2.6
2.2
2.0
1 .8
2.2
2.0
6
5.6
5.1
5.0
5.4
4.9
4.5
7
2.1
3.0
2.7
2.8
2.7
3.5
Total .
Observations
44
84
108
90
137
35
TABLE 5-18
Preference Rankings of Population
Groupings According to Household Income
152
Length of Residence
Shopping
Centers
Less than
1 year
1-3 years
4-6 years
7-10 years
10 years
Plus
1
4.6
4.0
3.9
3.9
4.2
2
4.5
4.6
4.1
4.3
4.2
3
3.9
3.4
3.4
3.6
3.6
4
6.1
6.1
6.3
6.3
6.1
5
1 .9
2.2
2.1
2.1
2.1
6
4.8
5.0
5.2
4.9
5.1
7
2.2
2.7
3.0
2.9
2.8
Total
Observations
14
53
54
56
321
TABLE 5-19
Preference Rankings of Population Groupings
According to Length of Residence
153
Sex
Shopping
Centers
Female
Male
1
4.1
4.1
2
4.2
4.4
3
3.7
3.2
4
6.2
5.8
5 ,
2.1
2.2
6
5.0
5.3
7
2.7
2.9
Total
Observations
408
90
TABLE 5-20
Preference Rankings of Population
Groupings According to Sex
154
Grouping
T
d.f.
Significance
7 Subgroups
48.4
36
>99%
5 Subgroups
44.3
24
>99%
3 Subgroups
34.8
12
>99.5%
TABLE 5-21
Results of Friedman Tests on Age Groupings
Grouping
T
d.f.
Significance
6 Subgroups
60.8
30
99%
4 Subgroups
52.3
18
>99%
TABLE 5
-22
Results of Friedman Tests on Income Groupings
Grouping
T
d.f.
Significance
2 Subgroups
6.9
6
<70%
TABLE 5-23
Results of Friedman Test on Sex
155
Grouping
T
d.f.
Significance
5 Subgroups
8.2
24
<0.5%
TABLE 5-
24
Results of Friedman Test on Length of Residence
Grouping
T
d.f.
Significance
11 Subgroups
50.0
60
<25%
5 Subgroups
26.5
24
<70%
6 Subgroups
28.3
30
<50%
TABLE 5-25
Results of Friedman Tests on Occupation Groups
156
The groupings reported here were based upon a subsample of 500 individuals
drawn at random from a sample of 1600 persons who gave complete responses to
all questions and indicated familiarity with all of the shopping centers. In
contrast, the earlier segmentation work was done on the whole sample of 7,300
observations. It is apparent that the subsample of 500 is more heavily biased
towards long residence as a result of the requirement that these people must
be familiar with all shopping centers. .Hence, in this analysis, the strong
segmentation on length of residence as a proxy for learning is not evident.
In addition to these tests, segmentation was examined with respect to
distances to closest and chosen shopping centers, the most preferred shopping
center, the chosen shopping center, and the method of questionnaire completion
(self-administered or interviewer-assisted). The preference rankings for these
groupings are shown in Tables 5-26 through 5-30. An identical set of modified
Friedman tests were run on these groupings with the results shown in Tables
5-31 through 5-35. As can be seen from these tables, the only significant
results are achieved on most preferred shopping center and chosen shopping
center. These results are not unexpected. In respect to the most preferred
shopping center, the grouping may be expected to reveal significant differences
based upon consistency of preference rankings. Similarly, the grouping on
chosen shopping center is based upon consistency between preference and behav¬
ior, albeit limited by geographic location. It is encouraging to note that the
method of survey completion does not show significant grouping, indicating that
no bias has been introduced into the preference ranking by one or other methods
of survey completion.
5.6 Preference and Perception Segmentation
In the final effort on market segmentation, the procedure was based on the
four factor-analysis, scores for each individual's most preferred shopping cen¬
ter. Each individual had provided ratings of seven shopping centers on sixteen
attributes, where these ratings represent perceptions. These ratings were sub¬
jected to a factor-analysis procedure in an attempt to reduce the dimensional¬
ity from sixteen to a much lower number. Research showed (see chapter ) that
four factors were generally optimal for this data set. Thus, the four factor
scores represent information on perceptions. Such information was obtained on
all seven shopping locations used in the first survey. -However, this segmen¬
tation process used only the four factor scores for the most preferred shopping
location of each individual, thus attempting to incorporate a measure of pref¬
erence into the perceptual data.
In this instance, posterior classification was used. That is, individuals
were grouped on the basis of common factor scores and attempts were made subse¬
quently to determine whether or not these groups were identifiable with measured
socioeconomic groupings. To group individuals in the first place, Ward's method
of clustering was used (Evert, 1974). Initially, this method produced eight
clusters, which were subsequently reduced by iterative relocation to two clusters.
An important element of the cluster analysis is the determination of the
solution that provides the best grouping of individuals, that is, the number
of clusters which best represent a homogeneous grouping. A number of criteria
were used to determine the best grouping. One consideration is the degree of
similarity within each group. This can be measured by the total within groups
157
Distance to Chosen Shopping Center
Shopping
Center
Less than
1 mile
1-2 miles
2-4 miles
4-8 miles
9 miles
and more
1
4.4
4.3
4.1
4.3
3.4
2
4.2
4.2
4.1
4.3
4.6
3
3.4
3.4
3.5
3.7
3.8
4
6.1
5.9
6.2
6.0
6.3
5
2.2
2.0
2.0
2.0
2.4
6
5.0
5.3
5.2
5.0
4.9
7
2.7
2.8
2.9
2.7
2.7
Total
Observations
72
73
122
143
88
TABLE 5-26
Preference Rankings of Population Groupings
According to Distance to Chosen Shopping Center
158
Distance to Closest Shopping Center
Shopping
Centers
Less than
1 mile
1-2 miles 2-4 miles
4-7 miles
7-10 miles
10 miles
and more
1
4.4
4.3
4.1
3.8
3.2
4.2
2
4.2
4.3
4.1
4.4
4.8
3.8
3
3.4
3.3
3.5
4.0
4.0
4.0
4
6.1
5.9
6.3
6.0
6.4
6.7
5
2.1
2.1
2.0
2.1
2.4
2.0
6
5.1
5.2
5.3
4.9
5.0
4.3
7
2.7
3.0
2.8
2.8
2.3
3.0
Total
Observations
138
103
122
93
36
6
TABLE 5-27
Preference Rankings of Population Groupings
According to Closest Shopping Center
159
Shopping
Center
Most Preferred Shopping Center
1
2
3
4
5
6
7
1
1.0
5.2
5.4
7.0
4.9
4.2
4.5
2
4.9
1 .0
4.4
5.0
3.8
5.2
4.5
3
4.6
3.4
1.0
2.0
3.7
4.9
3.7
4
6.4
6.2
5.3
1.0
6.1
6.7
6.2
5
2.8
2.4
2.7
3.5
1.0
2.4
2.7
6
5.0
5.3
5.9
5.5
5.0
1 .0
5.3
7
3.2
4.6
3.4
4.0
3.6
3.7
1.0
Total
Observations
94
12
51
2
181
11
147
TABLE 5-28
Preference Rankings of Population Groupings
According to Most Preferred Shopping Center
160
Shopping-Center Choice
Shopping
Centers
Golf Mill
Plaza
del Lago
Edens
Plaza
Old
Orchard
1
3.9
3.8
3.9
4.7
2
4.2
4.1
4.1
4.4
3
3.9
4.8
3.4
2.8
4
6.2
6.6
6.4
5.8
5
1.9
2.3
2.2
2.4
6
5.0
3.3
5.2
5.5
7
2.9
3.0
2.9
2.4
Total
Observations
297
20
34
147
TABLE 5-29
Preference Rankings of Population Groupings
According to Shopping Center Choice
161
Method of Completion
Shopping
Center
Mail back
At-Site
1
4.2
3.8
2
4.3
4.1
3
3.5
4.0
4
6.1
6.2
5
M
2.0
6
5.1
5.1
7
2.8
2.8
Total
Observations
420
78
TABLE 5-30
Preference Rankings of Population Groupings
According to Methods of Survey Completion
162
Grouping
T
d.f.
Significance
16 Subgroups
56.8
90
<0.5%
5 Subgroups
27.0
24
<70%
6 Subgroups
28.0
30
<50%
6 Subgroups
24.1
30
<25%
TABLE
5-31
Results of Friedman Test
on
Distance to Chosen Shopping Center
Grouping
T
d.f.
Significance
16 Subgroups
49.1
90
<0.5%
6 Subgroups
28.2
30
<50%
6 Subgroups
32.4
30
<70%
TABLE 5-32
Results of Friedman Test on
Distance to Closest Shopping Center
163
Grouping
T
d.f.
Significance
7 Subgroups
685.8
36
»99.9%
TABLE 5-33
Results of Friedman Tests on
Most Preferred Shopping Center
Grouping
T
d.f.
Significance
4 Subgroups
82.4
18
>99.9%
TABLE 5-34
Results of Friedman Tests on
Chosen Shopping Center
Grouping
T
d.f.
Significance
2 Subgroups
5.7
6
60%
TABLE 5-35
Results of Friedman Tests on
Method of Survey Completion
164
sum of squares associated with the grouping of the population. As shown in
Figure 5-11, the rate of increase in sum of squares becomes greater as the
number of clusters in the solution decreases. The graphical presentation,
however, does not provide a clear indication of what the optimal level of
clustering is. Hence, alternative methods were also examined.
Another possible measure of similarity within groups is a one-way analysis
of variance of the cluster means. This test determines whether cluster means
are different from each other. In other words, the null hypothesis is that
cluster means are the same and, hence, that the clusters are not distinctive.
Such a test would be conducted on the means of all four factor scores for each
level of clustering. The results are shown in Table 5-36. From this table,
it can be seen that the only times that the null hypothesis cannot be rejected
at better than 95% confidence is for variable 2 and two or three clusters. A
similar test is an analysis of whether the cluster means on each factor are
significantly different from populations means. As can be seen in Table 5-37,
only with 8 clusters are all cluster means significantly different from the
population means. However, for all other levels of clustering, there is never
more than one cluster mean that is not significantly different from the popula¬
tion mean for any factor. Thus, these analysis of variance tests only provide
limited help in selecting the number of clusters for the analysis.
A final criterion that can be used is the size of the cluster. If market
segmentation is to be used to permit better models to be estimated, then there
is a need to be concerned with the number of observations in each segment.
Table 5-38 shows the individual cluster means and the number of observations
in each cluster for all seven levels of clustering being studied. From this,
it is apparent that there are likely to be small-sample problems for each of
6, 7, and 8 clusters. The analysis-of-variance tests and the graphical display
of total within-cluster variance suggest that 2 or 3 clusters are too few and
represent too heterogeneous groupings. Hence, the optimal appears to be 4 or 5
clusters. These groupings provide a high degree of homogeneity within and
differences between groups. In addition, a sufficient number of observations
were within each cluster. Moreover, a significant amount of variation within
the population was no longer explained by the cluster means when the number of
clusters was less than four.
The next step in the process is to determine if the clusters demonstrate
common preferences, perceptions, and attribute importances between groups.
This testing involved the use of the modified Friedman test. As shown in
Tables 5-39 and 5-40, preference rankings for each solution were different
between clusters. In fact, for each solution differences greater than a one
percent level of significance were observed (Four clusters: x2 = 251 and v = 18,
five clusters: x2 = 256 and v = 24). However, it should be noted that reason¬
ably similar preference rankings were observed between some cluster pairs in
each solution.
Testing of differences between clusters with respect to both perception
and attribute importance ratings was accomplished by a multivariate analysis
of variance. This test is a multivariate extension of univariate analysis of
variance which is a test of the equality of a number of group means. The
multivariate analysis of variance is a test of the equality of vectors of
group means. For both four and five cluster solutions, average attribute-
165
Within
Clusters
Sum of Squares
200 -
i
Number of Clusters
FIGURE 5-11
Total Sum-of-Squares Within Clusters
for Each Cluster Solution
165A
Number of
CIusters
Significance
for Factor 1
Significance
for Factor 2
Significance
for Factor 3
Significance
for Factor 4
8
99.95%
99.95%
99.95%
99.95%
7
99.95%
99.95%
99.95%
99.95%
6
99.95%
99.95%
99.95%
99.95%
5
99.95%
99.95%
99.95%
99.95%
4
99.95%
99.95%
99.95%
99.95%
3
99.95%
65.7%
99.95%
99.95%
2
99.95%
91.2%
99.95%
99.9%
TABLE 5-36
One Way ANOVA of Cluster Solutions
166
Number of Means
Not Significantly Different at 95%
Number of
CI usters
Factor 1
Factor 2
Factor 3
Factor 4
8
0
0
0
0
7
1
0
1
0
6
0
1
1
0
5
1
0
0
0
4
1
1
0
0
3
0
1
0
1
2
0
1
0
1
TABLE 5-37
Test of the Differences of Cluster Means from
Population Means for Each of the Four Variables
167
Cluster Level
Factor 1
Factor 2
Factor 3
Factor 4
N
2 clusters
1
.5
.7
.5
.2
388
2
-1.6
.6
.3
.4
110
3 clusters
1
.6
.7
.6
.6
226
2
.2
.7
.4
-.4
173
3
-1.7
.6
.3
.5
99
4 clusters
1
.6
.8
.6
.7
195
2
.1
-.2
.7
-.1
76
3
.4
1.0
.3
-.4
130
4
01.7
.6
.6
.5
97
5 clusters
1
.6
.8
.6
.7
191
2
.1
-.2
.7
-.1
75
3
.4
1.0
.3
-.4
128
4
-1.7
.5
.7
.8
61
5
-1 .5
.9
0
-.3
43
6 clusters
1
.6
.8
.6
.7
185
2
.2
-.3
.8
0
54
3
-.2
.7
.1
-.2
76
4
.6
1.0
.5
-.5
88
5
-1.7
.5
.7
.7
67
6
-1.7
1.0
-.5
.1
28
7 clusters
1
.7
.9
.6
.7
162
2
.4
-.1
.8
.4
61
3
-.3
-.1
.5
-.7
31
4
-.1
.9
0
0
66
5
.6
1.0
.5
-.5
84
6
-1.7
.5
.7
.8
62
7
-1.8
1.0
-.4
.1
32
8 clusters
1
.6
.8
.7
.8
124
2
.7
.9
.2
.3
73
3
.4
-.2
.8
.3
54
4
-.3
-.1
.5
-.7
31
5
-.2
.8
.1
-.1
56
6
.5
1 .0
.5
.5
66
7
-1.7
.5
.7
.8
62
8
-1.8
1.0
-.4
.1
32
TABLE 5-38
Cluster Means for All Cluster Solutions
168
Clusters
Shopping
Center
1
2
3
4
1
4.9
4.5
4.7
1.5
2
4.1
4.3
3.8
4.8
3
3.3
3.5
3.3
4.4
4
5.9
6.2
6.1
6.4
5
1.9
2.0
1.8
2.8
6
5.2
5.2
4.8
5.0
7
2.7
2.2
3.6
3.1
Total
Observations
195
130
76
97
TABLE 5-39
Preference Rankings of the Population Grouped
According to the Four Clusters Solution
Shopping
Center
CIusters
1
2
3
4
5
1
4.9
4.6
4.6
1.4
1.9
2
4.1
3.8
4.3
4.8
4.8
3
3.3
3.3
3.5
4.6
4.1
4
5.9
6.1
6.2
6.5
6.2
5
1.9
1.8
2.0
2.6
3.2
6
5.2
4.8
5.2
5.1
4.9
7
2.7
3.6
2.2
3.1
2.9
Total
Observations
191
75
128
61
43
TABLE 5-40
Preference Rankings of the Population Grouped
According to the Five Cluster Solution
169
importance ratings of the clusters were observed to differ at levels of sig¬
nificance of less than one-tenth of one percent, as shown in Table 5-41.
Univariate analyses of variance indicated that major differences between
clusters in attribute importance ratings primarily occurred for the following
attributes: availability of sale items, free parking, variety of merchandise,
number and variety of stores, parking location selection, and compact area for
stores.
The perception of shopping centers was measured by interpoint distances
between shopping centers. For each cluster solution the interpoint distances
of the clusters were different at levels of significance of less than one-
tenth of one percent. Thus, the analysis of the clusters obtained for both
four- and five-cluster solutions indicated that preferences, perceptions, and
attribute importance ratings were different between the clusters. Not tested
were possible differences in utility functions between clusters because appro¬
priate logit model formulations have not been identified.
The final element of this segmentation procedure is to determine whether
the clusters can be identified with specific socioeconomic groupings of the
population. This was determined by the use of contingency tables, utilizing
a x2 test, with the results shown in Table 5-42. Only sex and occupation
exhibit any significant relationship to the clusters. Age is significant only
at 85% confidence, while all other variables are significant at much lower
percentage levels. Again, the tests show no detectable variation between sur¬
veys completed by the respondent and those assisted by the interviewer.
When age and occupation were controlled by variation in sex in the popu¬
lation, the clusters were no longer significantly related to either age or
occupation. Therefore, it must be concluded that the significance, or near-
significance of these two variables is primarily a result of interrelationships
with sex. Hence, the clustering appears to be primarily a function of the one
socioeconomic variable of sex.
5.7 Conclusions
Taken together, the results of these various market-segmentation efforts
must be considered to be inconclusive in identifying the best variables to use
as a segmentation basis. The results are summarized in Table 5-43. The only
socioeconomic variable that appears as a relevant segmentation variable is age,
which also only appears in the perceptions of the most preferred location when
the criterion of significance is relaxed to 85%. In turn, it disappears com¬
pletely when interactions with sex are taken into account. Apart from this,
the results of the various procedures are frequently in direct conflict. The
analysis of the perceptual data indicated length of residence as an important
variable. The other two analyses, both of which were undertaken on a small
subsample of the data used for the first analysis, indicated the reverse. As
mentioned previously, it may be speculated that this result is caused, in part,
by the restriction of the subsample to those who completed all questions on
the questionnaire and who indicated knowledge, at least of the existence, of
all shopping opportunities. Assuming that length of residence is a proxy for
learning,, this is a consistent finding.
170
Univariate Level of Significance
Attribute
4-Cluster
Solution
5-Cluster
Solution
1. Store Layout
80%
81%
2. Store Prestige
49%
59%
3. Merchandise Quality
95%*
78%
4. Reasonable Price
45%
77%
5. Ease of Returns
86%
93%
6. Credit Availability
99.9%**
99.96%**
7. Availability of Sale Items
99.999%**
99.998%**
8. Free Parking
97%*
93%
9. Stores in Compact Area
97%
94%
10. Store Atmosphere
86%
76%
11. Center Atmosphere
66%
86%
12. Specific Store Availability
61%
89%
13. Ability to Park
99%**
95%**
14. Courteous Sales Assistants
77%
70%
15. Number & Variety of Stores
99%**
97%*
16. Variety of Merchandise
99.97%**
99.3%**
Multivariate Level of
Significance
99.994%**
99.93%**
TABLE 5-41
Multivariate and Univariate Tests of Significance for
Attribute-Importance Ratings for Clusters of 4 and 5
171
Socioeconomic
Variable
Contingency Test
x2
d.f.
Sex (4)t
11 .7**
3
Sex (5)t
8.8*
4
Age (4)
17.1
12
Age (5)
21 .7
16
Occupation (4)
34.5*
18
Occupation (5)
34.3
24
Income (4)
11.3
9
Income (5)
11 .8
12
Length of Residence (4)
5.8
12
Length of Residence (5)
13.3
16
Method of Completion (4)
2.9
3
Method of Completion (5)
8.9
4
Chosen Shopping Center (4)
11 .8
9
Chosen Shopping Center (5)
16.0
12
TABLE 5-42
Contingency Tests on 4- and 5-Cluster Solutions
* Significant at 95%
** Significant at 99% or better
t Signifies the number of clusters
172
Segmentation
Criterion
Socioeconomic
Variables
Perceptions of
Shopping Centers
Preferences for
Shopping Centers
Perceptions of most
Preferred Shopping Center
Length of Residence
Age
Age
Income
Most Preferred Shopping
Center
Sex
Occupation
TABLE 5-43
Summary of Market Segmentation Results
173
In all other respects, however, the results are fairly inconsistent. Sex
appears as a segmentation variable for perceptions in general and for percep¬
tions of the most preferred location. However, it was not found to be a use¬
ful segmentation variable for preferences. Income was significant for prefer¬
ences but showed very little evidence of being suitable as a segmentation
variable for perceptions or the combined preference-perception analysis.
In general, it appears to be appropriate to draw two principal conclusions
from this research. First, it may be concluded that distinct segments do
appear to exist with respect both to perceptions and preferences. However, it
appears possible that the population is segmented differently for preferences
than for perceptions. It also appears that socioeconomic variables are not
proven to be useful segmentation variables. Second, it must be concluded that
this research has provided an insufficient basis to segment the population with
respect to any measured criterion variables. Hence, estimation of models must
proceed on the total sample, rather than on a segmented sample.
It is evident that testing must be undertaken on more than one-way cluster¬
ing of the population. Thus, it would be appropriate to examine the possibility
that two or more socioeconomic variables are needed simultaneously to define
market segments in the population. Some limited exploration of this was under¬
taken for the combined preference-perception analysis, the results of which
confirm the need to undertake a more detailed analysis of this type.
174
6. MODELS OF PERCEPTION, PREFERENCE, AND CHOICE
6.1 Introduction
The purpose of this section of the research is to develop models of the
destination-choice process. This is done through the process of developing
several models that provide a basis for understanding and modeling the deci¬
sion process of the individual. The primary objectives of these models are
to identify the characteristics of alternatives which influence choice behav¬
ior and to evaluate the relative importance of the identified characteristics.
Travel choice behavior is typically represented by a simple evaluation
and selection process. Each individual evaluates each alternative which is
known and available to him and chooses that alternative which he/she values
most highly. The value of each alternative is an individually determined
function of the characteristics of the alternative. Normally, the function
takes on a relatively simple, usually linear, form. The alternatives are
described by measured values of selected characteristics. The importance of
the characteristics in forming the "value" of each alternative may be modified
by measured characteristics of the individual. Because the value of an alter¬
native to an individual cannot be specified precisely, the choice process is
represented by a probabilistic choice model in terms of those aspects of
"value" which can be identified. That is, based on a partial valuation of
each alternative, the model predicts the probability that the individual will
select each of the available alternatives. When a large group of homogeneous
individuals is faced with identical alternatives, the number of individuals
choosing each alternative is expected to be proportional to the individual
choice probabilities.
The structure of the consumer-response process used in this study differs
from the more conventional approach described above in two significant ways.
First, the characteristics of the alternatives are described in terms perceived
by the individual rather than engineering measures. This approach extends the
range of attributes to include those which cannot be measured by direct engin¬
eering means and it accounts for differences of individual perceptions of iden¬
tical alternatives.
Second, the substantive aspects of the destination-choice alternatives,
the characteristics of shopping locations which determine their attractiveness,
are treated separately from the situational aspects of the choice alternatives,
the spatial aspects of shopping locations which influence their accessibility
to individual residential locations (Hanson, 1974).
The resultant model structure consists of three integrated components which
describe (1) individual perceptions of shopping locations, (2) individual rat¬
ings of shopping location attractiveness based on relative preferences for per¬
ceived characteristics of shopping locations, and (3) choice of shopping loca¬
tion based on attractiveness ratings and accessibility measures.
Perception models describe the choice alternatives in terms of their under¬
lying cognitive dimensions. These cognitive dimensions are identified by a
variety of methods including non-metric scaling (Green and Wind [1972], Green
and Rao [1972]), factor analysis (Urban [1975]), and discriminant analysis
175
(Johnson [1970], Pessimier [1976]). Preference models identify the relative
importance of the cognitive dimensions in the formation of preference or
attractiveness ratings. Importances may be estimated by statistical techni¬
ques such as preference regression (Hauser and Urban [1977]), preference logit
(Luce and Suppes [1965])> and revealed preference choice logit (McFadden
C1970])• Self-reported importance weights may be used also (Fishbein [1967],
Rosenberg [1956], Wilkie and Pessimier [1973]). The choice models incorporate
preference or attractiveness ratings for shopping locations and accessibility
measures to form overall value ratings which determine choice behavior
(McFadden [1970]). These models are linked together to provide a unified
model of consumer-choice behavior based on individual perceptions and prefer¬
ences .
6.2 Objectives of the Research and Approach
The objective of the research reported herein is two-fold. The primary
objective is to develop increased understanding of the process by which indi¬
vidual consumers select locations for non-grocery shopping trips. This in¬
creased understanding provides a basis for improved insight into the general
process of destination-choice behavior and for development of improved plan¬
ning models of destination-choice behavior. The secondary objective is to
develop and evaluate critically alternative empirical models of the shopping
location process. The results of this evaluation will provide guidance in the
formulation of other models of consumer-choice behavior.
These objectives are achieved by developing and interpreting alternative
models of perception and preference and integrating them with a logit choice
model. The alternative models provide different perspectives into the consumer
process and contribute to an overall understanding of that process. The com¬
parison among models provides a basis for selecting those models which will be
most useful in particular situations. The primary criteria for model evaluation,
are their ability to (1) provide useful insights into consumer behavior and
(2) predict accurately consumer preferences and choice behavior. These criteria
are balanced by (3) ease of use, and (4) overall cost.
The model structures examined include four models of perceptions and four
models of consumer preference combined with the multimonial logit choice model.
The models of perception are (1) fundamental attributes, (2) factor analysis,
(3) non-metric scaling, and (4) discriminant analysis. Fundamental attributes
represent perceptions in terms of an extensive list of attributes. The other
perception models reduce this list to the underlying cognitive dimensions consumers
use to evaluate products or services (Bruner et al [1956]). This reduced set of
cognitive dimensions can be interpreted more readily. The primary characteristics
of each of these perceptual models are summarized in Table 6.1.
The four preference models considered are (1) preference regression, (2)
first-preference logit, (3) expectancy value, and (4) unit weights.* Prefer¬
ence regression and first-preference logit select relative weights for attributes
*This study deals with analytic models based on mailed questionnaires.
Techniques such as conjoint analysis, trade-off analysis, and direct
utility assessment which require personal interviews are not included.
A limited comparison between statistical and personal interview tech¬
niques is discussed by Hauser and Urban [1 977].
176
NON-METRIC SCALING
• dissimilarity ^ distance
• "position" stimuli to best
recover distance
• "fit" attribute ratings to
explain stimuli positions
FACTOR ANALYSIS
• search for common component
of scale rating
• rating = common + specific
+ error
• correlations identify
dimensions
DISCRIMINANT ANALYSIS
• can identify stimuli by
its ratings
• search for dimensions that
discriminate best
• discriminant weights identify
dimensions
FUNDAMENTAL ATTRIBUTES
• no reduction possible
without loss of important
information
• individual can simultan¬
eously evaluate multiple
attributes
TABLE 6-1
Theoretical Constructs Underlying Four Models
of Consumer Perceptions
177
which best explain rank order preferences or first preferences, respectively.
Expectancy value, which can be used only with fundamental attributes, weights
each attribute by self-reported importance ratings. Unit weights is a base
model which weights all attributes equally.
The choice model predicts behavior taking account of individual prefer¬
ences for alternatives and other measures which indicate the availability or
relative availability of the different alternatives. The results of the anal¬
ysis are summarized below.
6.2.1 Insight into Shopping Location Choice Behavior
The different perceptual models developed identify cognitive dimensions
which consumers use to represent the characteristics of shopping locations.
Although the cognitive dimensions identified are different for the different
perception models they are all based on a common set of cognitive aspects.
These aspects are variety, quality, satisfaction, value and parking. This
result indicates that commonly used measures of destination size alone (Ben-
Akiva 0973], Rushton [1969]) cannot represent the range of aspects which are
meaningful to trip makers in distinguishing between alternative destinations.
Preference analysis consistently identifies quality as the aspect which
is most important in forming attractiveness or preference ratings and in in¬
fluencing choice behavior. This result further emphasizes the need to include
other than size measures to represent alternàtive destinations." Not only
do size measures provide an incomplete description of alternative destina¬
tions, they fail to include the most important aspects of these alternatives.
The perceptions and preference ratings for shopping locations are con¬
sistently shown to be important relative to access measures in determining
shopping-location choice behavior.
These results indicate that models of shopping location choice, and gen¬
eral models of destination-choice behavior should include measures -- perceived
or engineering -- which describe a range of major destination aspects. The
practice of describing destinations in terms of spatial attributes alone or
spatial attributes and engineering measures of size excludes important behav¬
ioral information.
6.2.2 Comparison of Model Structures
The different perception models identify similar aspects of shopping loca¬
tions which consumers use to distinguish among them. Among these models, factor
analysis is preferred because it provides the clearest insights into the rele¬
vant aspects of shopping-location choice behavior, predicts well and is easy
and inexpensive to use. Fundamental attributes also predicts well and is easy
and inexpensive to use with reported importance weights and it provides insights
which augment those obtained by use of factor analysis. However, it does not
provide useful information about the importance of the cognitive dimensions.
Both fundamental attributes and factor analysis are superior to discriminant
analysis and non-metric scaling which are difficult to interpret and predict
less accurately. Non-metric scaling also is expensive and difficult to use
and requires additional data for development.
178
The various statistical preference models (preference regression, first-
preference logit, and revealed-preference logit) produce similar qualitative
interpretations when applied to cognitive dimensions, but are unstable when
applied to fundamental attributes. These models are superior to unit weights
in interpretability and predictive ability. The expectancy value model, used
only with fundamental attributes, provides useful interpretations and has good
predictive ability. Choice models including perception/preference information
are consistently superior to those which include distance alone in predicting
choice behavior.
Based on these results, subject to confirmation in other empirical studies,
we recommend that statistical analyses of consumer-response behavior be based on
(1) factor analysis to identify perception, (2) preference regression or first-
preference logit to identify importance weights, (3) revealed preference and
intermediate preference models to identify choice behavior, and (4) include
both the appropriate attractiveness variables and the appropriate accessibility
variables when predicting consumer choice.
The results described above are documented in detail in the balance of
this chapter.
6.2.3 Modeling Consumer Perceptions
Consideration of consumer perceptions rather than direct (engineering)
measures of alternatives makes it possible to include attributes or charac¬
teristics for which direct (engineering) measures do not exist and to account
for differences between consumer perceptions of alternatives and engineering
characterizations. The usefulness of incorporating non-engineering measures
in travel choice behavior has been demonstrated in studies by Spear [1974],
Nicolaidis [1975], and Dobson and Kehoe [1974]. Differences between percep¬
tions among individuals and/or differences between perceived and engineering
measures have been identified by Burnett [1973].
Focus groups, open-ended surveys and other qualitative measurement tech¬
niques identify elemental or fundamental attributes which consumers use to
describe a particular product or service. However, to understand the true
consumer response process, it is necessary to identify the few basic dimen¬
sions consumers use to reduce the cognitive strain in evaluating the product
or service (Bruner et al. [1956]). We examine four alternative perceptual
models in this study. These include fundamental attributes, which assumes no
reduction in the original list of attributes and factor analysis, non-metric
scaling, and discriminant analysis, which identify a reduced set of cognitive
dimensions. The primary characteristics of each of these perceptual models
are described below and summarized in Table 6-1.
Fundamental Attributes - The simplest and most obvious method of repre¬
senting consumer perceptions is by individual ratings for an exhaustive list
of attributes. These scales provide a complete description of consumer per¬
ceptions and are easy to use because no further data collection or analysis
is required. Use of the complete list assumes that no further reduction is
possible without loss of important information and that the individual simul¬
taneously evaluates a long list of attributes in formulating preferences among
179
alternatives. A number of problems arise in the analysis of fundamental attri¬
butes. First, a complete list of attributes often includes a large number of
partially redundant scales. Second, the sheer size of the list can provide too
much information for a planner to process mentally and, perhaps, thus prevent
insightful analysis. Finally, redundancy in attributes can lead to multicol-
linearity which makes the estimated coefficients in preference and choice
models unreliable and difficult to interpret. Finally, estimation cost for
some preference and choice models increases dramatically when the number of
attributes is large.
Factor Analysis assumes that underlying cognitive dimensions exist and
that consumer ratings of attributes include a common component attributable to
these cognitive dimensions, an attribute specific component, and some measure¬
ment error. The common components or cognitive dimensions can be found by
factor analysis of the attribute ratings across products and consumers (Rummel
[1970]). The structure of consumer perceptions and the names of the common
dimensions are determined by examining the correlations (factor loadings) be¬
tween the fundamental attributes and the common dimensions. The method is
described in detail in chapter 4.
The advantage of factor analysis relative to fundamental attributes is that
it identifies a simpler perceptual structure which can provide clearer insight
on how consumers perceive alternatives. Furthermore, since factor dimensions
can be made orthogonal, multicol linearity can be reduced so that we can obtain
stable coefficients in preference and choice models. Finally, since the number
of factors is small, estimation cost is reduced. Since all of these advantages
come at the cost of a reduction in information and predictability, we must exa¬
mine empirically the magnitude of these benefits and weight them against the
loss in predictability.
Discriminant analysis assumes that it is possible to identify a particular
product or service by knowing how consumers rate that produce or service on the
fundamental attributes. The model searches for the dimensions (combinations of
attributes) which best discriminate between products. The discriminant weights
are determined from this process and the relative weightings of the attributes
in the discriminant function identify the name of each dimensions. The appeal
of discriminant analysis is that the dimensions are specifically chosen to dis¬
criminate among products or services. Its drawbacks are that it implicitly
assumes that the attribute ratings are interval-scaled and it confounds differ¬
ences between stimuli due to cognitive reduction with actual differences.
Non-metrie Scaling identifies cognitive dimensions by analysis of perceived
similarities between products or services. Non-metric scaling "positions" alter¬
natives in n-dimensional space so that the distance between pairs of alternatives
corresponds as closely as possible to the reported similarity between them
(Kruskal [1964], Torgerson [I960]) as was described in chapter 4.
It would be desirable to estimate perceptual maps for each individual.
However, limitations on the number of alternatives for which dissimilarity
judgments can be collected for a single individual suggest the development of
a common space representation. The INDSCAL procedure (Carrol and Chang [1970])
estimates such a common space (such that x.. = x. for all i) but estimates
i j y* j K*
180
individual weights, w^r. Thus, the effective measure of an individual's per¬
ception of a stimulus along a selected dimension becomes w^x^. This con¬
struction creates problems in the development of preference and choice models
because it implicitly assumes that all individuals have a common rank ordering
of shopping locations along each dimension.
It is necessary to use additional techniques to relate dimensional indices
to the fundamental attributes in order to identify the cognitive dimensions. A
regression technique conceptually similar to estimating factor loadings, PROFIT
( Chang and Carroll [ 1970 ] estimates "directional cosines" which indi¬
cate the relationship between the dimensions and the fundamental attributes.
The directional cosines can be interpreted in the same way as the factor load¬
ings.
Relative to the fundamental attributes non-metric scaling has the same
advantages and disadvantages as factor analysis. The advantages of non-metric
scaling relative to factor analysis and discriminant analysis are that (1) it
does not assume the ratings scales are metric and (2) because scales are deter¬
mined independently of the attributes, they can uncover dimensions not repre¬
sented in the attributes. However, they do require additional, hard to collect
data on similarity judgments and the scaling procedures are difficult and ex¬
pensive to use. Finally, the number of dimensions which can be identified is
severely constrained by the number of stimuli.
6.2.4 Modeling Consumer Preferences
We have chosen to describe the consumer-response process as one of per¬
ception, preference and choice. It is often tempting to skip the intermediate
preference models and use the perception measures directly in a choice model.
The purpose of separating these steps is to avoid compounding performance and
attractiveness characteristics which influence both preference and choice, and
availability, awareness, accessibility, and related characteristics which in¬
fluence choice but not preference. The determination to model preference as
a distinct step is tested by comparing importance weights and predictive abil¬
ity of models including an intermediate preference step with similar models
which exclude the intermediate preference step.
While it is important to identify the underlying cognitive dimensions and
to know product or service positions along these dimensions, it is necessary
also to know the relative importance of these dimensions. The analysis of con¬
sumer preferences is directed toward finding a function which maps measures of
consumer perceptions into a preference-rating index which ranks the alternatives
consistently with consumer's preferences.
The preference models considered in this study determine the relative
importance of the fundamental attributes or cognitive dimensions by measuring
or estimating a linear compensatory* model of the form shown in equation (6.1).
* Non-linear models are not developed in this study. For a discussion
of these models, see Green and Wind [1972], Hauser and Urban [1976],
and Johnson [1974].
181
This model states that consumer i's preference index for product j, P.., is
' J
the weighted sum of his or her perceptions, of alternative j for attri¬
bute or dimension k. The specific models evaluated are the expectancy-value
model which uses stated attribute importances, preference regression, and
first-preference logit which statistically estimate importance weights using
preference data, revealed-preference logit which statistically estimates impor¬
tance weights based on choice behavior, and unit weights which assumes equal
importance for each fundamental attribute or cognitive dimension. The char¬
acteristics of these models are described below and summarized in Table 6-2.
Expectancy-Value Model - Each respondent is asked to rate the importance
of each fundamental attribute in reaching their choice decision. The expec¬
tancy-value model (Fishbein [1967], Rosenberg [1956]> Wilkie and Pessemier
[1973]) then represents the preference function as a linear sum of importances
times attribute ratings, as shown in equation (6.2).
Pij = I wik dijk (6,2)
where w.^ = individual i's importance rating for attribute k
d. = individual i's rating (also called belief) of shopping
J location j on attribute k
(In this model P.. is often called i's attribute toward j.)
' J
The appeal of the expectancy-val ue model is individual-specific weights.
Its drawbacks are the scaling problems inherent in using self-explicated impor¬
tances and the often questioned ability of consumers to accureately provide
these weights. Furthermore, because the self-explicated weights must be mea¬
sured in the original survey, expectancy value can only be used with funda¬
mental attributes.
Preference regression statistically estimates the importance weights
using rank order preference as the dependent variable and the consumers' per¬
ceptions as independent variables. The statistical techniques are either
monotonie regression (Johnson [1974]) or ordinary least-squares (OLS) regres¬
sion. Because recent simulation (Cattin and Wittink [1976], Carmone, Green, and Jain
[1976]) and empirical tests (Hauser and Urban 0976]) show that OLS performs
as well as the more complex and expensive monotonie regression, OLS is used
to estimate the preference-regression models in this study. In order to
obtain sufficient degrees of freedom to estimate the importance weights, the
analysis is made across individuals. This gives equation (6.3).
Pij = £k wk dijk (6.3)
where P^ and d^^ are defined as before.
The importance weights, w^, estimated are average importance weights for the
population. Preference regression uses full rank order information in the
estimation of importance weights. Its drawbacks are the metric assumption
and the inability to estimate individual importance weights.
182
PREFERENCE REGRESSION
weighted attributes
proportional to rank
order preference
statistically estimate
"average" weights
1ST PREFERENCE LOGIT
weighted attributes
proportional to prob¬
ability of 1st preference
maximum likelihood
estimate of "average"
weights
REVEALED PREFERENCE
preferences are
"revealed" by observed
choice behavior
maximum likelihood
estimate of "average"
weights
UNIT WEIGHTS
either cannot distinguish
differential weighting
or all attributes have
equal weight
EXPECTANCY VALUE
consumers can "self-
explicate" importance
weights
individual specific
importance weights
TABLE 6-2
Theoretical Constructs Behind Five Models
of Consumer Preferences
183
Preference logit assumes that the true preference, pl., is composed of
' J
an observable part, P.. as in equation (6.3), plus an error term, e.., as
I J • J
shown in equation (6.4).
pL = pij ♦ eij (6-4)
Assuming an appropriate probability distribution for the error term* makes it
possible to derive a functional form for the probability, L.., that consumer
' J
i ranks product j as his or her first preference. This probability is given
by equation (6.5).
Lij = exp(Pij)/Z exp(Pil) (6.5)
where the sum is over all products, 1.
This is called the preference-!ogit model. The average importance weights are
estimated by maximum-likelihood techniques (McFadden and Wills [1970]). The
appeal of the logit model is that it models stochastic behavior explicitly
(Bass [1974]), and it makes no metric assumptions about preference rankings.
Its drawbacks are that it uses only first-preference information and that it
estimates average importance weights to gain degrees of freedom.
Revealed preference assumes that the underlying preference weights can be
identified by observing choice behavior. It assumes that each individual
selects an alternative which has the greatest utility to him or her. When the
choice model directly incorporates the fundamental attributes or cognitive
dimensions, the importance weights, w.., are estimated jointly with the impor-
' J
tance of non-preference characteristics such as the time, effort, or cost of
obtaining a selected alternative. The advantage of the revealed-preference
model is that it does not rely on reported preference data but on observed
choice behavior- When repeated choice decisions are observed or reported, it
can use complete information in choice frequencies for all of the chosen
alternatives. The drawbacks of revealed preferences are that it estimates
average importance weights to gain degrees of freedom and that the estimates
of importance weights may be biased if the nonpreference choice elements are
not carefully specified.
Unit weights - All of the above models assume that we can distinguish the
relative importances that consumers place on different attributes. An alter¬
native hypothesis (Einhorn [1975], Sheth [1971]) is that the attributes are
important, but we cannot distinguish relative importance. Such a model assigns
unit or equal weights to each attribute in the perceptual structure, as shown
in equation (6.6).
PTj = l dijk <«-6>
The unit weights model does not require any estimation effort, but it also
does not provide the insights of the other models.
*The error terms are independent and identically distributed Wei bull
random variables (McFadden [1970]).
184
6.2.5 Modeling Consumer-Choice Behavior
The consumer-response process is designed to explain and predict indivi¬
dual based on a model of perceptions and preferences. The choice model
postulates that individual consumers associate a value or utility, v.., with
J
each available alternative and select that alternative which has the greatest
utility. Our estimate of the individual utility, v.., is a linear combination
■ J
of the preference index and situational variables influencing choice behavior,
as shown in equation (6.7).
A
V.• = Bn P. • + E B x • • (6.7)
ij 0 ij m mij v '
The true utility is equal to the estimated utility plus a random component
which represents unobserved influence and specification errors. Using the
same distributional assumption as for preference logit, we obtain the multi-
nomial-logit model (McFadden [1970]) which describes the probability of indi¬
vidual i choosing alternative j, C.., on a single occasion by equation (6.8).
' J
A
exp(v.•)
C, • = 4 (6.8)
U
z exp(v.,)
1 11
The parameters of the choice model are estimated by maximum-likelihood tech¬
niques McFadden and Wills [1970]). When the preference index has not been
estimated the estimated utility can be formulated in terms of the fundamental
attributes or cognitive dimensions, as shown in equation (6.9).
= Ek "k dijk + I \ Xijm t6'9)
The importance weights, w^, can be estimated simultaneously with the parameters
of the choice model. These are the importance weights of the revealed-prefer-
ence model described earlier.
The choice models considered in this study are all based on the multinom-
ial-logit formulation. The choice models will differ by inclusion of the pref¬
erence index based on different perception and preference models.
6.2.6 Linked Perception, Preference, and Choice Models
The objective of this study is to develop an integrated model of consumer
perception, preference, and choice. The individual models described in the
preceding sections are linked together to represent the consumer-response pro¬
cess. The linked models considered in this study include all feasible combin¬
ations of perception and preference models combined with the multinomial choice
model (see figure 6-1). The linked models exclude only the expectancy-value
model with cognitive dimensions because stated importances are available only
for fundamental attributes.
185
PERCEPTION MODELS
oo
en
PREFERENCE
MODELS
Fundamental
Attributes
Factor
Analysis
Discriminant
Analysis
Non-metric
Scaling
Expectancy
Val ue
y
Stated Importance Weights Available
for Fundamental Attributes Only
Preference
Regression
y
y
y
y
Preference
Logit
y
y
/
y
Revealed
Preference
y
y
y
y
Uni t
Weights
/
/
y
y
FIGURE 6-1
Linked Models of Choice Responses
6.3 Empirical Setting and Experimental Design
The empirical problem is to model consumers' choice of shopping locations
based on perceptions of and preferences for aspects of shopping location attrac¬
tiveness and accessibility. Historically, studies of destination-choice behav¬
ior and shopping-location market area emphasized the importance of accessibility
or distance of the shopping location from the shopper's residence location.
Some studies included measures of shopping-location size, usually retail floor
space or number of retail employees. Although measures of accessibility influ¬
ence consumers' choice of where to shop, the variety of large-scale shopping
areas within easy driving distance of most urban and suburban residences indi¬
cate that present and future behavior may be relatively more sensitive to char¬
acteristics of alternative shopping locations. Furthermore, from the perspec¬
tive of the managers of shopping centers or local officials concerned with the
success of central business districts where location is fixed, the sensitivity
of destination-choice behavior to the attractiveness of shopping locations pro¬
vides the only opportunity for developing strategies to attract shoppers.
Although size of shopping locations, which also represents the range of oppor¬
tunities available to the shopper, is surely a relevant measure of attractive¬
ness, it is unlikely to capture all the aspects of attractiveness which influ¬
ence shopping-location choice behavior. In order to understand the construct
of shopping location attractiveness from the perspective of consumers we must
determine the cognitive dimensions of shopping locations attractiveness, their
relative importances in forming preferences and their importance relative to
accessibility in influencing choice behavior.
This study develop models based on measures of seven shopping locations
including downtown Chicago and six suburban shopping centers of widely differ¬
ing characteristics, as described in chapter 3. The locations represent the
types of shopping opportunities available to residents in the suburbs north of
Chucago. The data were obtained by sampling individual shoppers at four of
these locations. The models estimated in this study use choice-based adjust¬
ments to eliminate estimation bias which would otherwise result from the use
of choice-based samples (Lerman and Manski [1975]). The effectiveness of
these adjustments was verified by performing two parallel analyses. The pri¬
mary analysis for the full set of seven shopping locations is reported in full.
The parallel analysis for the four shopping locations at which the sample was
collected is summarized and compared to the primary analysis. The comparison
confirms that the estimation methods used produce models which are unbiased.
This result is important for market research because it indicates that random
samples may be replaced by more efficient data collection designs.
The data used in this analysis include rank-order preference for each
shopping location, similarity judgments for all pairs of shopping locations,
direct ratings of each shopping location for sixteen attributes, self-expli¬
cated importances of those attributes and frequency of trips to each shopping
locations.
Five hundred of these respondents were randomly selected for analysis and
an additional 500 respondents were selected for saved data testing, from those
who provided fully completed surveys.
187
The data collected did riot include information on the costs (time, out-
of-pocket cost, physical effort, etc.) of travelling to each of the shopping
locations. Only the residential location of the shopper was obtained. For
this reason accessibility is represented by the distance between each shopping
location and the shopper's residence.
The experimental design is a full factorial for all feasible combin¬
ations of perception and preference models linked with the multinomial choice
model. These models are compared with each other and with selected base models
in terms of interpretabi1ity, predictive ability, and ease of use
and cost. These comparisons are described in the following sections.
6.4 Results of the Analysis
6.4.1 Perception Models
The most straightforward procedure for describing perceptions of shopping
locations is by the average ratings of each locations for the sixteen fundamen¬
tal attributes. These average ratings, illustrated in Figure 6-2, provide
useful information about each shopping location. They provide a basis for
assessing the relative strengths and weaknesses of each location and can be
used to identify gaps in the market. For example, Figure 6-2 indicates that
Chicago Loop is poorest in parking-related questions but does well in variety
of stores and merchandise. Woodfield does uniformly well but has top ratings
on only a few attributes. Korvette City is poor on most attributes but good
on price and specials. Careful examination of this figure reveals a number of
good insights into the existing pattern of perceptions. However, the complex¬
ity of the figure makes it difficult to focus on critical areas due to the
excessive amount of information displayed. It is appropriate to develop and
evaluate alternative, reduced representations of the public's perceptions of
shopping locations.
Such alternative perceptual maps are developed by use of factor analysis,
discriminant analysis and non-metric scaling, described earlier. Although the
methods of analysis differ, each of the perception models identifies cognitive
dimensions by structure matrices which relate them to the sixteen fundamental
attributes. The data were analyzed to develop three and more cognitive dimen¬
sions. The perceptual models developed by each model in three dimensions are
shown in Table 6-3. Although the models have superficial similarities, they
present striking differences in interpretation. The factor-analysis and non-
metric-scaling models have strong attribute loadings on a single dimension
indicating strong relationships within groups of attributes. The factor-anal¬
ysis model has strong loadings for fourteen of the sixteen attributes and mod¬
erate loadings for the remaining attributes (return and service, and availabil¬
ity of credit). The discriminant-analysis model has strong loadings for only
seven of the sixteen fundamental attributes. The remaining attributes do not
have strong loadings on any of the discriminant dimensions. The non-metric-
scaling and discriminant models include mixed signs for some of the major
loadings, that is, some attributes load positively and others negatively on
the same dimensions. These mixed loadings preclude identification of a natural
direction of goodness along the affected dimensions. The factor-analysis model
does not include mixed signs for any of the major loadings. The factor loadings
188
1. LAYOUT OF STORE
2. RETURN AND SERVICE L
1
3. PRESTIGE OF STORE I
1.0
4. VARIETY OF MERCH.
5. QUALITY OF MERCH.
6. AVAIL OF CREDIT
7. REASONABLE PRICE
8. "SPECIALS"
9. FREE PARKING
10. CENTER LAYOUT
IT. STORE ATMOS.
12. PARKING AVAIL.
13. CENTER ATMOS.
14. SALES ASSTS.
15. STORE AVAIL.
16. VARIETY OF STORES
•Figure 6-2
Map of Fundamental Attributes Ratings
for Seven Shopping Locations
O Chicago Loop
O Woodfield
A Golf Mill
A Edens
E3 Old Orchard
□ Plaza del Lago
X Korvette City
189
FUNDAMENTAL
ATTRIBUTES
FACTOR ANALYSIS
FACTOR LOADINGS
DISCRIMINANT ANALYSIS
DISCRIMINANT COEFFICIENTS
NON-METRIC
DIRECTIONAL
1
SCALING
COSINES
Variety.
Quality, and
Satisfaction
Value
Parking
Variety
Quaiity
vs.
Val ue
Parking and
Satisfaction
Variety
Quality
vs.
Val ue
Parking and
Satisfaction
1.
Layout of store
.619
.168
.256
.016
-.016
-.010
.217
.497
.840
2.
Return and service
.469
.290
.361
.047
-.019
.100
.318
.122
.940
3.
Prestige of store
.878
-.001
.042
.120
.493
-.138
.297
.804
.515
4.
Variety of merchandise
.614
.455
-.270
.372
-.290
.042
.929
.360
.084
5.
Quality of merchandise
.847
.026
.038
.032
.778
-.291
.295
.811
-.505
6.
Availability of credit
.342
.454
.121
.129
-.119
-.006
.880
-.085
-.468
7.
Reasonable price
-.057
.596
.121
.071
-.416
-.045
.485
-.853
.192
8.
"Specials"
.140
.747
.022
-.062
-.286
-.119
.786
-.594
.173
9.
Free parking
-.061
.028
.800
.063
-.032
1 .713
-.294
-.550
.782
10.
Center layout
.246
.071
.585
-.204
.057
.115
-.447
.035
.894
11.
Store atmosphere
.583
-.009
.505
-.149
.090
-.059
-.199
.452
.853
12.
Parking available
-.033
.088
.645
-.266
-.047
.165
-.463
-.473
.747
13.
Center atmosphere
.702
-.027
.460
-.070
.270
.322
-.099
.480
.872
14.
Sales assistants
.546
.137
.390
-.177
.116
-.077
-.052
.411
.910
15.
Store availability
■ 573
.350
-.064
.114
-.026
-.070
.372
.429
-.236
16.
Variety of stores
.652
.364
-.311
1 .180
.017
.181
.921
.385
-.054
TABLE 6-3
Structure Matrices for Three Dimensional Perception Models
in Table 6-3 represent (1) variety, quality, and satisfaction; (2) parking;
and (3) value. The discriminant dimensions represent (1) variety, (2) quality
versus value, and (3) parking. The non-metric-scaling dimensions represent
(1) variety, (2) quality versus value, and (3) parking and satisfaction. The
consistent grouping of fundamental attributes which describe variety, quality
satisfaction, value and parking indicate that these five aspects represent the
cognitive dimensions of shopping locations.
Factor-analysis and discriminant-analysis models of higher dimensions are
developed in an attempt to isolate these aspects (non-metric-scaling models of
higher dimensionality cannot be developed due to the small number of stimuli
-- shopping locations -- rated).
The four-dimensional factor-analysis model separates the variety, quality,
and satisfaction dimension into a variety dimension and a quality and satisfac¬
tion dimension (Table 6-4). The other dimensions are not affected. Increasing
the number of dimensions to five or six retained the same dimensions shown in
Table 6-4 and produced additional dimensions which were not strongly loaded by
any of the attributes.
The four-dimensional discriminant-analysis model separates the combined
dimension of quality versus value into separate dimensions (Table 6-4). It
also identifies the aspect of satisfaction which combines with value to pro1
duce a value versus satisfaction dimension . The four-dimensional model also
incorporates two additional fundamental attributes. Increasing the number of
dimensions to five or six retains the four dimensions shown in Table 6-4 and
forms additional dimensions which have mixed loadings and which cannot be
interpreted readily.
Based on these results, the perception models selected for use in the
balance of this study are (1) the four-dimensional factor-analysis model, (2)
the four-dimensional discriminant-analysis model, and (3) the three-dimensional
non-metric-scaling model. Each of these models identify cognitive dimensions
composed of the aspects of variety, quality, satisfaction, value and parking.
They provide a simpler intuitive interpretation than the fundamental attributes
and the cognitive dimensions identified are consistent with results of prior
studies (Singson [1975]). The factor-analysis model provides the clearest
interpretation.
These models can be used to develop perceptual maps of shopping locations
based on the underlying cognitive dimensions. These maps, as well as the per¬
ceptual map based on fundamental attributes, are shown in Figure 6-3. Two
things are immediately apparent from examination of these perception maps.
First, and most obvious, analysis using cognitive dimensions is simpler than
analysis of fundamental attributes. Second, there is general consistency among
the various perception maps. For example, note the low scores of Korvette City
and high scores of Old Orchard on quality (prestige, quality of merchandise,
sales assistants for fundamental attributes), the low scores for Chicago Loop
on parking (parking availability and free parking) and the high scores for
Chicago Loop and Woodfield on variety (variety of merchandise, availability of
a special store and variety of stores). These consistencies have strong face
validity which support the use of the reduced perceptual structures. However,
191
FUNDAMENTAL
ATTRIBUTES
FACTOR ATTRIBUTES
FACTOR LOADINGS
DISCRIMINANT ANALYSIS
DISCRIMINANT COEFFICIENTS
VARIETY
QUALITY AND
SATISFACTION
PARKING
VALUE
VARIETY
QUALITY
VALUE VS.
SATISFACTION
PARKING
1.
Layout of store
.267
.583
.200
.156
.067
-.110
-.086
-.023
2.
Return and service
.095
.528
.255
.343
-.094
.254
.287
.134
3.
Prestige of store
.388
.822
-.058
-.001
.156
.366
-.318
-.137
4.
Variety of merchandise
.665
.327
-.185
.309
.335
-.153
.295
.042
5.
Quality of merchandise
.307
.810
-.074
.037
-.196
1.114
.020
-.216
6.
Availability of credit
.159
.337
.049
.487
-.008
.170
.352
.025
7.
Reasonable price
.067
-.063
.113
.599
-.108
-.008
.586
-.009
8.
"Specials"
.223
.074
.008
.739
-.101
-.171
.225
-.115
9.
Free parking
-.150
.068
.811
.043
.066
-.066
-.007
1.714
10.
Center layout
.030
.308
.560
.074
-.066
-.233
-.335
.082
11 .
Store atmosphere
.080
.658
.400
.034
-.158
.087
-.056
-.053
12.
Parking available
.145
.105
.841
.108
oo
oo
CM
-.020
.018
.171
13.
Center atmosphere
.244
.694
.404
-.040
.109
-.123
-.510
.284
14.
Sales assistants
.173
.560
.319
.147
-.138
.015
-.168
-.084
15.
Store availability
.619
.320
.034
.204
.089
.035
.084
-.065
16.
Variety of stores
.829
.288
.173
.160
1.291
-.123
-.022
.145
TABLE 6-4
Structure Matrices for Four Dimensional Perception Models
VARIETY
Variety
vaux vs.
satisfaction
Duality vs.
Value
Variety
Duality and
Satisfaction
Parkins
pwwjhd
b) Discriminant Analysis: group centroids
Parkins and
Satisfaction
a) Non-metr1c Scaling: common space
positions
c) Factor Analysis: factor scores d) Fundamental Attributes: average ratings
STIMULI SET: ° Chtca9° L°°P ° P1a2â ^ U9°
• Woodfield B Old Orchard
A Edens X Korvette City
A Golf Mill
1. layout of si ore l
v.o
2. return ai® service l.
3. prestige of store i
4. variety of mercii. '
1.0
5. quality of merch. l
1.0
6. avail of credit
7. rca so! {able price ■—
1.0
8. "specials"
9. free parkinc
10. center layout
.11. store a twos.
12. parking avail.
1.0
13. center amos. l_
1.0
14. salfs ass7s. i—
1.0
10. store avail. i—
1.0
16. variety of siotxs i—
1.0
-ÙV,
j
Figure 6-3: Perceptual maps for the four models of consumer perceptions
193
there are important differences among the reduced perception maps. The non-
metric-scaling map is difficult to interpret because it combines quality and
value on a single dimension with opposite directionality. That is, the qual¬
ity versus value dimension implies that locations high in quality are also
low in value. Analysis of the fundamental attributes and factor scores indi¬
cates that there is a wide spread of quality judgments across locations while
all locations except Plaza del Lago (a small exclusive center with many spec¬
ialty stores) are rated closely for value. Similarly, the discriminant anal¬
ysis perception map has mixed loading for value and satisfaction.
Thus, although all the perception models provide useful insight and over¬
all consistency of interpretation, factor analysis is superior because of the
clearer loadings, absence of mixed loadings and the ability to identify four
important dimensions. All of the reduced perceptual maps are easier to work
with and understand than the fundamental attribute map which presents too much
data to synthesize readily. They are also more likely to identify the small
number of dimensions which people use in evaluating alternative shopping loca¬
tions. Final judgment on the importance of these differences include results
of the predictive ability tests.
6.4.2 Preference Models
The normalized importance weights for the direct preference models on the
three perception structures are shown in Table 6-5. The most important dimen¬
sion for each perception structure includes quality as a component. The impor¬
tance weights estimated by preference regression and preference logit are
similar for each of the perception structures. This robustness and the fact
that the estimated weights are significantly different from equal indicates
that the cognitive dimensions do have differential importances and that a unit-
weighting model probably neglects important information. The interpretation
of the importance weights for the discriminant and scaling dimensions is com¬
plicated by the mixed loadings described earlier and the negative importance
of the discriminant value-versus-satisfaction dimensions. These model struc¬
tures imply that an increase in value rating for a selected location will
reduce the preference for that location. This illogical result undermines
the usability of both of these model structures.
Normalized importance weights on fundamental attributes are shown in
Table 6-6. The estimated preference-regression weights include six negative
values. Furthermore, preference regression and first-preference logit models
produced very different importance weights. The negative importance weights
and the instability between estimated importance weights are due to the high
degree of multicol1inearity among the fundamental attributes.
The average reported importance weights shown in Table 6-6 provide useful
insight into the importances of fundamental attributes but provide no informa¬
tion on the importance of cognitive dimensions.
194
CONSUMER MODEL
Factor Analysis
Variety
Quality and
Satisfaction
Value
Parking
Preference Regression
.39
.57
.03*
.01*
Preference Logit
.30
.41
.23
.06*
Unit Weights
.25
.25
.25
.25
Discriminant Analysis
Variety
Quality
Value vs
Satisfaction
Parking
Preference Regression
.34
.43
-.12
.12
Preference Logit
.36
.45
-.03*
.16
Unit Weights
.25
.25
.25
.25
Non-Metric Scaling
Variety
Quality vs Value
Parking am
Satisfacti<
Preference Regression
.26
.43
.31
Preference Logit
.26
.49
.26
Unit Weights
.33
.33
.33
TABLE 6-5
Normalized Importance Weights
(*-Non-significant at 5%)
195
NORMALIZED IMPORTANCE WEIGHTS
ATTRIBUTE
FIRST
PREFERENCE
LOGIT
PREFERENCE
REGRESSION
EXPECTANCY
VALUE
EQUAL
WEIGHTS
1.
Layout of store
.12
.13
.05
.06
2.
Return and service
.03
.03
.07
.06
3.
Prestige of store
.14
.17
.05
.06
4.
Variety of merchandise
.08
.10
.07
.06
5.
Quality of merchandise
.12
.06
.07
.06
6.
Availability of credit
.02
.03
.06
.06
7.
Reasonable price
.13
.01
.07
.06
8.
"Specials"
.07
-.03
.06
.06
9.
Free parking
.01
-.01
.07
.06
10.
Center layout
.04
-.07
.06
.06
11.
Store atmosphere
.03
-.04
.06
.06
12.
Parking available
.01
-.07
.06
.06
13.
Center atmosphere
.03
.10
.06
.06
14.
Sales assistants
.04
-.04
.07
.06
15.
Store availability
.08
.07
.07
.06
16.
Variety of stores
.06
.05
.06
.06
TABLE 6-6
Normalized Importances for Fundamental Attributes
196
6.4.3 Choice Models
Choice-model estimation provides estimates of revealed preference impor¬
tances as well as the relative influence of attractiveness or preference mea¬
sures and accessibility measures in determining choice behavior. The results
of the various estimations are summarized in terms of the normalized impor¬
tance weights and modified 3 coefficients. The choice model (equation 6.7)
is formulated with distance as the only accessibility measure as in equation
(6.10).
v.j = 30 2 wk dijk + 3i Xx (6.10)
k
Normalization of the preference importance weights gives equations (6.11) and
(6.12).
= e« l \ ^ dijk+ 6° <6-1"
- So • Î wj; djjk + 8, X, (6.12)
Table 6-7 presents both the normalized importance weights, wk , and the ratio
Bo /(3o + 3i)> which indicates the relative weights of the attractiveness
index for each combined preference/perception model. The similarity of the
importance weights between preference regression and preference logit for each
perception model is partially retained in the revealed-preference model. The
dimension in each perception model which includes quality continues to obtain
the highest importance weights. However, the estimated importance weights and
the rank order of importances for the other dimensions are different from those
estimated with the direct preference models. For the factor-analysis and dis¬
criminant-analysis models, the importance of variety is lower while the impor¬
tance of both value and parking is higher. For the non-metric-scaling model,
the importance of parking is lower.
These results may be due to the confounding of the importance weights
obtained by the revealed preference models with aspects of accessibility. The
residential location of the sample is the North Shore suburbs of Chicago.
Thus, for this data set, the accessibility characteristics which influence
choice behavior are highly correlated with both variety, which is highest for
Woodfield and Chicago Loop -- the two remotest shopping locations -- and with
parking, which is lowest for Chicago which is the second most remote shopping
location. These effects may be exaggerated by the use of distance as the sole
accessibility measure in this analysis. We suggest, but are not able to test
with this data set, that the normalized importance weights obtained by reveal¬
ed preference analysis would be closer to those obtained by direct preference
analyses if a more complete specification of accessibility were used.
The differences in importance weights may also be due to differences in
preferences and choice behavior for different types of purchases. One plaus¬
ible hypothesis is that consumers make more frequent trips for smaller pur¬
chases to conveniently located centers and less frequent trips to more distant
197
NORMALIZED IMPORTANCE WEIGHTS
CONSUMER MODEL
Factor Scores
Variety
Quality and
Satisfaction
Value
Parking
Preference Regression
.38
.54
.07*
.01*
.88
Preference Logit
.30
.41
.23
.06*
.90
Revealed Preference
.05
.43
.29
.19
.91
Unit Weights
.25
.25
.25
.25
.90
Discriminant Analysis
Variety
Quality
Value vs.
Satisfaction
Parking
Preference Regression
.34
.43
-.12
.12
.90
Preference Logit
.36
.45
-.03*
.16
.90
Revealed Preference
.06
.56
.08
.30
.90
Unit Weights
.25
.25
.25
.25
.91
Non-Metric Scaling
Variety
Quality vs.
Value
Parking and
Satisfaction
Preference Regression
.26
.43
.31
.97
Preference Logit
.26
.49
.26
.97
Revealed Preference
.22
.76
.03
.97
Unit Weights
.33
.33
.33
.96
Choice Analysis Importance Weights
Table 6-7
198
high variety centers to shop for major purchases. Both the preference ratings
and reported frequencies merge these two types of shopping behavior. Based on
this study we can posit, but not test, the hypothesis that trips for major
purchases influence the formation of preference rankings to a greater degree
than represented by the proportion of trips actually made for these purposes.
Table 6-7 also reports the relative importance of attractiveness versus
accessibility for the different perception and preference models. The attrac¬
tiveness ratio is stable across models within a perceptual technique but sys¬
tematically varies across perceptual techniques. This difference between per¬
ceptual models means that the variance in the scales for attractiveness is
changing across perceptual technique. The apparent stability for each type of
perception model is encouraging to the interpretability of the choice models.
6.4.4 Linked Model Structure
The perception, preference and choice models are linked together in a
unified structure. The interpretability of the linked model structure is
determined by three characteristics. First, identification of the underlying
aspects or cognitive dimensions of attraction provides a basis for identifying
the dimensions which consumers use in discriminating among shopping locations.
This is best accomplished by the factor-analysis models, which identify four
cognitive dimensions that are readily interprétable in terms of the fundamen¬
tal attributes and provide an easy-to-understand representation of market
structure.
Second, the estimated importance weights for the cognitive dimensions
provide a basis to infer the relative impact of changes in shopping location
characteristics in preferences and choice behavior. Factor scores provide
the most useful set of importance weights. The estimated importance weights
for the non-metric-scaling and discriminant-analysis dimensions combined with
the mixed loadings of attributes in dimensions produce unrealistic interpreta¬
tions of the effect of improving selected attributes. Multicollinearity of
fundamental attributes produces some negative importance weights with corres¬
ponding unrealistic interpretations. Importance weights estimated for funda¬
mental attributes based on stated importances provide useful information about
the effect of changes in fundamental attributes but do not provide any insight
into the importance of cognitive dimensions. The range of estimated importance
weights obtained by direct and revealed-preference models lead to differences
in detailed interpretations which require further study. These differences
suggest the need to consider the implications of the different models in inter¬
preting destination-choice behavior.
Third, the relative importance of the attractiveness index versus the
accessibility measure indicates the influence of these different elements in
choice behavior. The sensitivity to the attractiveness and accessibility
effects is not easily compared due to the absence of a common measure for
attractiveness and accessibility. However, the stability of the relative
values provides confidence in the consistency of the models developed.
199
6.4.5 Summary
This analysis identifies factor analysis with either of two direct-preference
models (preference regression or first-preference logit) as the best structure
for representing perceptions of and preferences for shopping locations. In
the absence of preference-ranking data, factor analysis with the revealed-
preference model can be used to obtain importance weights. Fundamental attri¬
butes with expectancy-value importance weights provides an alternative and
complementary set of interpretations.
6.5 Predictive Ability
A good model provides useful insights into the behavioral process and
also predicts well. The interpretive ability of alternative structures was
evaluated in the previous section. This section evaluates the predictive
ability of each of the alternative models. Each model is evaluated for its
ability to predict individual and aggregate preference rankings and choice
behavior. These predictions are made for both the "estimation data" sample
and a "saved data" sample of equal size. The estimated models are compared
amongst themselves and against a set of reference models.
6.5.1 Prediction Formation
Individual predictions are made by applying the alternative model struc¬
tures to each individual's ratings on the fundamental attributes and distance
for each shopping location. The prediction process, described in Figure 6-4,
consists of the following sequence of steps.
First, perception measures are obtained by applying perception
models to the fundamental attribute ratings (non-metric-scaling
models are developed using individual similarity measures rather
than attribute ratings) to obtain perception scores for the cogni¬
tive dimensions, Fundamental-attributes models are formulated
directly on the individual ratings. Factor-analysis models obtain
factor scores by applying the factor-score coefficient matrix to the
fundamental-attribute ratings. Discriminant-analysis models obtain
discriminant measures by applying the discriminant-coefficients
matrix to the fundamental-attribute ratings. Non-metric-scaling
models use similarities measures to obtain individual weights for
scale coordinates.
Second, the perception scores formulated are combined with the
estimated importance weights, w^, to obtain individual preference
ratings for each shopping location.
Third, the preference ratings are rank ordered to obtain
individual preference ranks which are used in the analysis of
preference prediction.
Fourth, the preference ratings and accessibility measures are
applied to the choice models to predict overall ratings and choice
probabilities for each shopping location. These predictions are
used in the analysis of choice prediction.
200
PERCEPTION
MODEL
STRUCTURE
IMPORTANCE
WEIGHTS
FUNDAMENTAL
ATTRIBUTE RATINGS OR
SIMILARITIES DATA
SCORES ON
COGNITIVE
DIMENSIONS
ATTRACTIVENESS
RATINGS
CHOICE
MODEL
t
COEFFICIENTS
PREDICTED
PREFERENCE
RANKING
ACCESSIBILITY
MEASURES
OVERALL CHOICE
RATINGS FOR
SHOPPING LOCATIONS
PREDICTED
CHOICE
PROBABILITIES
FIGURE 6-4
Prediction Process
201
Predictions are made with each of the fifteen perception/preference model
combinations described earlier. These include fundamental attributes with
preference regression, expectancy value and unit weights and factor analysis,
discriminant analysis, and non-metric scaling each with preference regression,
first-preference logit, revealed preference, and unit weights.
6.5.2 Tests of Preference Prediction
The individual preference rankings are compared to reported preference
rankings. These data are used to compute four measures of predictive accuracy:
(1) the percent of times each model correctly predicts first preferences and
(2) the percent of times each model correctly predicts the seven preferences.
These measures emphasize individual preference recovery. The preference pre¬
dictions are compared to base models which assume (1) all shopping locations
are equally likely to be preferred most, second most, etc., and (2) all shop¬
ping locations are likely to be most preferred in proportion to their market
share of first preferences.
6.5.3 Tests of Choice Prediction
The individual choice probabilities are compared to the actual choice
frequencies. These data are used to compute four measures of predictive
accuracy: (1) the percent of times the chosen alternative is predicted to
have the greatest choice probability and (2) the information about choice
behavior provided by the model relative to a model which predicts probabilities
equal to individual choice frequencies. The choice predictions are compared
to base models which assume (1) all shopping locations are equally likely to
be chosen, (2) shopping locations are likely to be chosen in proportion to
their market share of choices, and (3) accessibility or distance influences
choice behavior but attractiveness ratings do not. The choice predictions
are also compared to a model based on stated preferences.
6.5.4 Preference Prediction Results
The preference prediction measures for each perception-preference model
are reported in Table 6-8. The results presented include predictions with
the estimation sample and the saved data sample. Factor analysis dominates
both non-metric scaling and discriminant analysis in preference prediction.
That is, all the factor-analysis models predict better than any of the dis¬
criminant or scaling models by both measures of predictive accuracy. The
differences in predictive ability between the discriminant analysis and scal¬
ing models are small. The predictive ability of the fundamental-attributes
model is similar to that of the factor-analysis models when used with unit
weights or expectancy value, but much lower when used with preference regres¬
sion. The poor performance of preference regression on fundamental attributes
is surprising. A priori, one would expect that fundamental attributes would
contain more information than the reduced perception models, but in this case
the data are colli near and it appears that the col linearity degrades prediction.
Such multicollinearily also produces unreliable importance weights and thereby
also degrades interpretability. All of the models provide better predictions
than the base models described earlier.
202
ESTIMATION SAMPLE
SAVED DATA SAMPLE
CONSUMER MODEL
Base Models
Equally likely
Market Share
Factor Scores
Preference Regression
First Preference Logit
Unit Weights
Discriminant Analysis
Preference Regression
First Preference Logit
Unit Weights
Non-metric Scaling
Preference Regression
First Preference Logit
Unit Weights
Fundamental Attributes
Preference Regression
First Preference Logit
Unit Weights
Expectancy Value
TABLE 6-8:
PREFERENCE RECOVERY PREFERENCE RECOVERY
First Rank First Rank
14.3
14.3
14.3
14.3
26.8
m
"""
50.6
32.9
47.3
55.0
37.0
50.8
36.' 6
48.7
33.0
44.0
31.4
35.5
27.5
38.1
29.2
35.3
28.8
40.3
30.1
36.8
27.8
38.9
28.7
36.6
25.1
23.1
20.5
34.8
24.4
22.7
19.6
32.4
24.8
24:1
20.6
39.6
30.6
41.4
30.9
55.6
37.9
51.6
36.2
50.7
36.3
47.6
34.2
51.4
36.6
47.6
34.7
Preference Prediction Tests
203
6.5.5 Choice Prediction Results
The results of choice prediction tests are presented in Table 6-9. The
predictive ability of these models with respect to choice behavior is less
differentiated than with respect to preference ranking. All of the models
predict with similar accuracy and all predict much better than models based
on the assumption of equally-likely choice or market-share likely choice.
In addition, the models with perception/preference-based models predict sig¬
nificantly better than models based on accessibility or distance measures
alone. Finally, the perception/preference-based models predict similarly to
models based on stated preference rankings.
6.5.6 Summary of Predictive Ability Analysis
Models based on factor analysis predict best among the cognitive models
for preference and as well as other models for choice. Within each cognitive
or perception model there is little differentiation in predictive ability.
The fundamental-attributes model predicts well when combined with expectancy
value or unit weights but less well when used with preference regression.
These results support the conclusions based on interpretability. Factor-
analysis models which rated best for interpretability also rate best for pre¬
dictive ability.
6.6 Ease of Use and Cost
The primary objective of this research is to develop models of choice
behavior which will enable planners and managers to identify improved product
or service strategies. Expenditures for development and application of dif¬
ferent model structures should properly be evaluated relative to the benefits
of the improved strategies. The problem of identifying and evaluating such
benefits is both beyond the scope of this research and related to each appli¬
cation situation.
Thus, we choose to focus on the lesser problem of comparing alternative
model structures in terms of their ease of use and development cost. The ele¬
ments of ease of use and cost are (1) the data required to perform the required
analysis, (2) the capability required to perform and interpret the analysis,
and (3) the cost of computer utilization.
6.6.1 Ease of Use and Cost of Perception Model Development
Factor-analysis, discriminant and fundamental-attribute models require
information on attribute ratings for each shopping location. Non-metric scal¬
ing requires this information and measures of similarity-dissimilarity between
pairs of shopping locations.
The programs required to develop factor-analysis and discriminant-analysis
models are readily available in many standard statistical packages, are well
documented, simple to access and easy to use, and provide easily interprétable
output. Both models are inexpensive to run with costs around $20* to develop
*Cost estimates based on use of CDC 6400 computer billed at $510 per
CPU hour.
204
58
CONSUMER MODEL
Base Models
Equally Likely
Market Share
Distance Only
Individual Model
Factor Analysis
Preference Regression
First Preference Logit
Revealed Preference
Unit Weights
Discriminant Analysis
Preference Regression
First Preference Logit
Revealed Preference
Unit Weights
Non-Metric Scaling
Preference Regression
First Preference Logit
Revealed Preference
Unit Weights
Fundamental Attributes
Preference Regression
First Preference Logit
Revealed Preference
Unit Weights
Expectancy Value
ESTIMATION SAMPLE
Per Cent
Correctly
Predicted Information
14.3
0.0
18.5
24.7
31.9
32.6
38.7
100.0
SAVED DATA SAMPLE
Per Cent
Correctly
Predicted Information
14.3
0.0
18.2
24.0
30.4
32.8
37.4
100.0
32.7
36.4
31.0
37.1
32.9
37.3
31.2
38.1
32.6
38.5
31.4
38.8
32.5
36.7
31.2
37.5
32.4
34.3
30.8
34.7
32.5
34.5
30.9
34.8
32.4
35.5
30.5
35.8
32.8
34.1
30.7
33.9
31.8
31.8
32.2
31.9
33.5
33.7
34.1
33.3
24.7
30.9
23.4
26.1
0.4
-8.5
-36.7
10.0
32.7
32.8
32.7
32.9
32.8
35.8
37.8
39.2
37.8
37.7
31.2
31.3
31.3
31.1
31.2
28.9
39.2
39.7
38.7
39.1
TABLE 6-9
CHOICE PREDICTION TESTS
perception structures at three or four different levels of dimensionality
after initial preparation of data files. These models are transferable to
new data sets by use of factor score or discriminant coefficients to compute
scores along cognitive dimensions based on fundamental-attribute ratings.
In contrast, non-metric scaling requires the use of specialized programs
which are not readily available. A series of runs required to develop common
perceptual spaces, estimated individual weights and compute directional cosines
for attribute vectors (needed for interpretation) for three different levels of
dimensionality costs about $200. Application of the non-metric-scaling model
to new data requires reestimation of individual weights based on reported
similarities data.
Fundamental attributes requires no analytic effort or cost of use. How¬
ever, the retention of a large number of attributes increases the cost of
estimating importance weights.
6.6.2 Ease of Use and Cost of Preference Model Development
Direct preference models (preference regression and first-preference logit)
require measures of fundamental attributes or cognitive dimensions and reported
preference rankings. In addition, these models must be augmented by accessi-
-bility and frequency-of-visits data in order to develop the final choice models.
The revealed-preference models require the same data with the exception of
preference-rankings data which are not required.
Regression programs needed for estimation of the preference-regression
model are readily available, widely documented and easy to use. Logit-esti-
mation programs needed for estimation of first-preference logit and revealed-
preference models are becoming more widely available, and are well documented
but somewhat more difficult to use.
Regression estimation is relatively inexpensive. A typical preference
regression run costs between five and ten dollars for 500 observations with
ten to twenty variables. Logit estimation is somewhat more expensive and
costs increase rapidly with an increase in the number of variables. Logit
estimations on cognitive dimensions which require 9 or 10 variables (3 or 4
cognitive dimensions and 6 sampling variables) cost about $30, while an esti¬
mation on fundamental attributes which requires 22 variables (16 attributes
and 6 sampling variables) costs about $120.
Unit weights and expectancy values have no estimation cost but expectancy
values requires the collection of self-reported importance for the fundamental
attributes.
6.6.3 Ease of Use and Cost of Choice Model Development
All of the alternative perception/preference models are integrated with a
logit-choice model. The choice models are estimated with cognitive scores or
preference ratings developed using the perception and preference models, acces¬
sibility measures and reported frequency of trips to different shopping loca¬
tions. Each logit estimation costs between $25 and $30.
206
6.6.4 Ease of Use and Cost of Model Sets
The overall ease of use and cost of sequences of perception/preference
models is determined by the interaction of these model structures. The least
expensive (in cost and effort) sequence of models is the combined model which
includes the least expensive perception models, fundamental attributes, and
the least expensive preference model, unit weights. However, other interactions
are somewhat more complicated. Table 6-10 summarizes this information in terms
of rankings for least cost and ease of use for perception models alone (part a),
for preference models alone (part b), and for combined perception/preference
models (part c). The combined model ranking shows that the high cost of non-
metric scaling dominates preference model costs so that the four most expensive
models are those using non-metric scaling. The fundamental attributes are
least expensive when used with low- or no-estimation cost preference models
such as preference regression or unit weights (and expectancy value not inclu¬
ded in the tables). However, the fundamental-attributes models are very ex¬
pensive when preference-importance weights are estimated by use of logit anal¬
ysis (first-preference logit and revealed-preference logit). The factor-anal¬
ysis and discriminant-analysis models have essentially identical costs for
each preference model. The ease of use and cost differences vary widely over
the range of rankings. In particular, there is a significant increase in cost
between the two models tied for second rankings. A cost index for developing
an integrated perception/preference model and using it to estimate a choice
model is presented in Table 6-11. These numbers provide an estimate of the
relative costs of estimation of alternative models rather than an estimate of
actual total costs. Total costs will exceed the index costs to the extent that
it is necessary to explore multiple specifications of the models developed.
Overall, we see that fundamental attributes is easiest to use and least
costly when combined with either preference regeression or unit weights. These
models are followed closely by factor-analysis or discriminant-analysis models
with all the preference models. The ratio between fundamental-attributes-with-
unit-weights model costs and factor analysis or discriminant analysis with
first-preference logit is three to one. This is a relatively small difference
considering the magnitude of the costs involved and the peripheral costs of
data preparation and manipulation. The inverse in cost to use fundamental
attributes with first-preference logit or revealed-preference logit is a fac¬
tor of 5 to 6 over the least cost model. Finally, the non-metric-scaling
models cost about 10 times the least cost model.
These results indicate a need to make some tradeoff between the increased
interpretability of factor-analysis models used with an estimation procedure
to obtain importance weights and the additional costs of the increased analytic
effort.
a. PERCEPTION MODELS RANKINGS
Model
Rank
Model
Rank
Factor
Analysis
Preference
Regression
Discriminant
Analysis
Non-Metric
Scaling
2 4
PREFERENCE MODELS RANKINGS
First
Preference
Logit
Revealed
Preference
Logit
Fundamental
Attributes
1
Unit
Weights
1
c. JOINT PERCEPTION/PREFERENCE MODELS RANKINGS
Preference
Regression
First
Preference
Logit
Revealed
Preference
Logi t
Unit
Weights
Factor
Analysis
5
9
7
3
Discriminant
Analysis
5
9
7
3
Non-Metric
Scaling
14
16
15
13
Fundamental
Attributes
2
12
11
1
TABLE 6-TO
Rankings of Models in Terms
of Least Cost and Ease of Use
208
First Revealed
Preference Preference Preference Unit
Regression Logit Logit Weights
Factor
Analysis
50
75
60
45
Discriminant
Analysis
50
75
60
45
Non-Metric
Scaling
230
255
240
225
Fundamental
Attributes
35
145
130
25
TABLE 6-11
Cost Index for Developing Perception/Preference
Structure and Choice Model Estimation
209
6.8 Sensitivity Tests
6.8.1 Description of the Tests
As a final test of the choice models developed, it was desirable to
undertake some sensitivity tests on the models and determine whether the
models exhibited differences in sensitivity on the attractiveness variables.
Sensitivity tests for this type of model are somewhat different from those
normally undertaken. In this case, the variables to be manipulated are the
distributions of ratings of individuals on the various fundamental attributes.
Without tying the perceptual measures to planning variables, the process be¬
comes one of guess work in terms of changes that would stem from specific
transportation or planning actions.
The generation of sensitivity tests was undertaken by setting new dis¬
tributions on selected attributes, where the distribution was drawn from
information on other shopping locations. It proved to be too expensive to
investigate more than one attribute change at a time. Five changes were
selected, each one affecting one shopping location. These are described in
Table 6-13. The selection of tests is somewhat limited, partly by the costs
of running such tests and partly by the desire to make changes relevant to
transportation or urban planning, rather than ones relevant to shopping-cen¬
ter managers.
The basis of the attribute-distribution changes made are now described.
For the first sensitivity test of Table 6-13, the distribution for the Loop
was taken as the average for Korvette City, Plaza del Lago, and Edens Plaza.
All three have free parking, but it is either insufficient or much is remote
from the stores in the center. In this respect, these centers were seen to
come closest to the situation anticipated in the Loop. The resulting distri¬
bution is shown in Table 6-14. For the second test, the Old Orchard distri¬
bution was adjusted to approximate the Chicago Loop, as shown in Table 6-15.
The third test was approximated by changing the Old Orchard distribution to
something close to that of Plaza del Lago, which is the one for which parking
is least adequate, apart from the Loop. The distribution is shown in Table
6-16. The fourth and fifth tests utilized different distributions on shopping
center atmosphere. Test four was based on an average of Old Orchard and Wood-
field, while test five was based on Korvette City, and the distributions are
shown in Table 6-17-
To generate the data for testing, the distributions of Tables 6-14
through 6-17 were used to generate random ratings for each of the 500 indivi¬
duals used to calibrate the choice models. Thus, for each test, each indivi¬
dual received one changed rating in his set, all others being left unaltered.
The ratings were then standardized as in the original analysis and new factor
relationships for the factor-analysis models.
Three models were subjected to tests in this analysis. These were the
factor-analysis, revealed-preference regression model, the factor-analysis,
first-preference logit model, and the fundamental-attributes model. The non-
metric-scaling model was not tested. The principal reason for this is the
expense of generating the new inputs. Unlike the factor-analysis models, it
would be necessary to undertake a new fitting of the three-dimensional space
for each changed attribute.
210
POLICY CHANGE
SHOPPING
LOCATION
ATTRIBUTE
AFFECTED
1.
Free Parking
Chicago Loop
Free Parking
2.
Parking Fees Imposed
Old Orchard
Free Parking
3.
Reduce Available Parking
Old Orchard
Ability to park
where desired
4.
Pedestrian Mall
Chicago Loop
Center Atmosphere
5.
Pedestrian Mall
Chicago Loop
Center Atmosphere
TABLE 6-13
Basis of Sensitivity Tests on Choice Models
211
Rating
Percentage
Good - 1
60.7
2
18.4
3
15.3
4
4.0
Poor - 5
1 .6
TABLE 6-14
Changed Distribution of Free Parking in the Loop
Rating
Percentage
Good - 1
1 .8
2
1.2
3
3.0
4
7.8
Poor - 5
86.3
TABLE 6-15
Parking Fees at Old Orchard
212
Rating
Percentage
Good - 1
30.0
2
24.0
3
24.0
4
14.0
Poor - 5
8.0
TABLE 6-16
Reduced Parking at Old Orchard
Rating
Percentage
Test 4
Test 5
Good - 1
60.0
15.0
2
23.0
18.0
3
10.0
28.0
4
4.0
20.0
Poor - 5
3.0
19.0
TABLE 6-17
Pedestrian Mall in the Loop
213
6.8.2 Results of the Sensitivity Tests
The results for the five tests are shown in Tables 6-18, 6-19, and 6-20.
A general comment may be made about the tests. It appears that the selected
changes have generally a rather small effect upon the market shares. At this
point, this should be taken more as an indication of problems in generating
significant changes in the input data than as an indication of a lack of sen¬
sitivity of the models. It is also worthwhile to point out that the most
important attributes were found to be attributes relating to the shopping
location, per se, rather than to attributes under the direct control of pub¬
lic policy makers.
The only model to produce counter-intuitive results is the factor-anal¬
ysis, first-preference logit model. In this case, the majority of the respon¬
ses on the two subject shopping locations are counter-intuitive. All three
improvements in the Chicago Loop (tests 1, 4 and 5) result in lower market
shares for the Loop than in the base estimate. Similarly, Old Orchard exhi¬
bits lower market shares than the base for all tests, but with larger declines
for changes in the Loop than for changes in itself. It appears, then, that
this model has undergone a change in its base through manipulating the ratings.
Apart from this, the model shows about the same range of sensitivity as the
factor-analysis, revealed-preference model. As expected, the fundamental-
attributes model shows the greatest sensitivity to parking changes than the
other models. The two factor-analysis models both show the pedestrian mall
(test 4) as having the greatest impact on improving the relative share of
trips to the Loop, while the fundamental-attributes model shows free parking
as having the greatest effect.
6.8.3 Conclusions
All of the models exhibit a relative insensitivity to small changes in
the distributions of the attribute ratings. The factor-analysis, first-pref¬
erence logit model shows a rather disquieting shift in all market shares on
the basis of the adjusted data, suggesting a need to examine more carefully
the process by which this model reacts to changes in the input variables.
Apart from this, the two factor-analysis models are equally sensitive (in
terms of range), while the fundamental-attributes model is the most sensitive.
In general, however, it must be concluded that the models are not satis¬
factory in their present form for predicting market shares for changed circum¬
stances. This conclusion is drawn from the need to obtain a revised set of
ratings on each attribute as the basis for prediction. Such values must either
be obtainable directly from a knowledge of the mapping between planning para¬
meters and attribute perceptions, or reliance must be placed on the ability of
individuals to react to a hypothetical situation to provide revised ratings.
This latter process is not likely to be satisfactory.
214
Shopping Center
Sensitivity Tests - Market Share (%)
Base
Estimate
Test 1 -
Free Parking
in Loop
Test 2 -
Parking Fees
at Old Orchard
Test 3 -
Reduced Parking
at Old Orchard
Test 4 -
Pedestrian
Mall in Loop
Test 5 -
Pedestrian
Mall in Loop
Loop
13.00
12.79
12.64
13.23
12.59
12.56
Edens
14.56
14.77
14.69
14.49
14.61
14.62
Golf Mill
20.10
20.39
20.28
20.01
20.16
20.18
Korvette City
6.71
6.81
6.77
6.65
6.73
6.74
Old Orchard
29.73
29.07
29.55
29.71
29.90
29.89
Plaza del Lago
6.38
6.52
6.47
6.40
6.45
6.45
Woodfield
9.52
9.66
9.60
9.51
9.56
9.56
TABLE 6-18
Results of Sensitivity Tests on the Factor-Analysis,
Revealed-Preference Model
Shopping Center
Sensitivity Tests - Market Share (%)
Base
Estimate
Test 1
Test 2
Test 3
Test 4
Test 5
Loop
12.53
12.55
12.58
13.12
12.60
13.28
Edens
14.62
14.60
14.63
14.51
14.61
13.96
Golf Mill
20.19
20.16
20.20
20.04
20.16
18.56
Korvette City
6.74
6.73
6.75
6.66
6.73
5.28
Old Orchard
29.91
29.96
29.82
29.74
29.90
31.34
Plaza del Lago
6.45
6.44
6.45
6.41
6.44
7.48
Woodfield
9.57
9.55
9.57
9.52
9.56
10.10
TABLE 6-19
Results of Sensitivity Tests on the Factor-Analysis,
First-Preference Logit Model
216
Shopping Center
Sensitivity Tests - Market Shares {%)
Base
Estimate
Test 1
Test 2
Test 3
Test 4
Test 5
Loop
14.30
13.28
12.59
12.93
12.58
12.59
Edens Plaza
14.33
15.29
14.58
14.55
14.60
14.60
Golf Mill
19.77
21.09
20.13
20.08
20.16
20.15
Korvette City
6.51
7.12
6.73
6.68
6.72
6.72
Old Orchard
29.35
26.46
29.94
29.78
29.89
29.88
Plaza del Lago
6.31
6.78
6.45
6.43
6.46
6.46
Woodfield
9.43
10.02
9.59
9.59
9.60
9.60
TABLE 6-20
Results of Sensitivity Tests on the Fundamental-Attributes Model
217
7. SUMMARY AND RECOMMENDATIONS FOR FUTURE RESEARCH
7.1 Summary
This research project has provided a number of valuable insights into
extensions of individual-choice models to decisions other than choice of
travel mode. These insights relate both to the structure of the models and
to the methods for calibration and use.
The most important conclusion of this research, from the viewpoint of
the planner and decision maker, is that people select nongrocery shopping
destinations on the basis of quality, variety, and value, rather than on
attributes for which size might serve as a useful proxy. Thus, the research
has shown that, for the population surveyed, conventional transportation
planning models of trip distribution capture relatively little of the behav¬
ioral content of the destination-choice decision. A second important con¬
clusion is that it appears to be feasible to build destination-choice models
using individual choice theories and attitudinal information. Moreover, by
making appropriate corrections for choice-based sampling, consistent estimates
of importance weights for the choice models can be obtained and these do not
vary when the choice set is reduced.
An important methodological conclusion is that it does not appear to
be necessary or even particularly desirable to use the more complex and
expensive psychological-scaling methods to analyze attitudinal data. Our
research has shown that factor analysis provides results that are generally
consistent with those from scaling techniques, but use less data, require
less complex measurement devices and are cheaper to use. Factor-analysis
results are also easier to use for prediction than are the scaling results.
For this data set, factor analysis produced a satisfactory perceptual
space, from which both preference regression and first-preference logit models
were found to produce consistent preference weights. The direct use of funda¬
mental attributes was also found to perform well on all model tests, but poses
a problem of providing too much information and, for certain modeling processes,
generates a far too costly calibration and estimation phase.
The prototype destination-choice model was a logit model of choice, using
either fundamental attributes or factor scores (for 4 factors) with importance
weights derived from either preference regression or first-preference logit.
The project was not successful in determining the feasibility of segmenting
the population on the basis of socioeconomic characteristics. In part, this may
be due to our inexperience in identifying and measuring appropriate socioeconomic
variables. However, it is also due to the relative lack of suitable techniques
for identifying market segments by either prior or posterior classification. In
the course of our research in this area, it was necessary to develop modifica¬
tions of various statistical procedures in order to obtain usable tests.
Finally, it proved to be feasible to survey shoppers, using a self-adminis¬
tered survey that requested extensive attitudinal information. No difference
was found between fully self-administered surveys and those filled out with
interviewer assistance. Although extensive procedures were not utilized to
achieve high response rates, the response was found to be much higher than is
usually associated with a hand-out, mail-back survey. It must be noted, however,
that complete returns were obtained for only about 6% of those given a question¬
naire, or about 23% of the returned questionnaires.
218
The models developed in this project are not suitable in their present
form for application to planning problems. First, the models were calibrated
on a very biased demographic sample, exhibiting a much higher than average
income distribution and a higher than average level of education. Thus, the
findings cannot be generalized across demographic groups. Second, the present
form of the models requires that attitudinal data be available both for cali¬
bration and prediction. As previously stated, it is not generally possible to
predict preferences and perceptions, hence the models are currently restricted
to providing behavioral insights into the destination-choice process.
7.2 Future Research Directions
On the basis of the research reported here, a number of future directions
for research can be identified. A number of these research tasks can be pur¬
sued with the data already collected for this project. These will be identi¬
fied first.
The first task is to extend the prototype models to include a better
specification of site accessibility. Data from the second destination-choice
survey and the mode-choice survey would permit such an extension. This speci¬
fication should replace the distance measure used in this research. Second,
further validation checks are needed on the prototype models. Limited checks
could be made by using the second destination-choice data set, which contains
four of the seven shopping locations used in the prototype models, together
with five new locations. The additional locations also include different
types of shopping sites, namely strip developments and another downtown shop¬
ping location. Thus, this data set would also permit testing of whether the
prototype model is valid for choices of shopping sites other than specific
shopping centers.
A third research direction is to seek a mapping between attitudinal
measures (i.e. perceptions, preferences, saliences) and physical planning
parameters. Since there are detailed records of a variety of physical char¬
acteristics in existence for the Chicago-area shopping locations, it is
possible to explore this issue fairly readily. Measures have been obtained
on sixteen perceptual variables in the two destination-choice surveys. The
attribute groupings determined in the factor-analysis models could be used
as a basis for seeking physical correlates. This is an essential step if the
prototype model is to find practical implementation as a planning tool. A
fourth research direction, that is enabled by the data already collected in
this project, is to pursue certain issues of efficacy of attitude measurement.
Between the two destination-choice surveys, a number of measurement changes
were introduced. First, certain questions were structured on seven-point
scales in the second survey, compared with five-point scales in the first
survey. It is not generally known what the effect is of such variations.
Second, perceptions were obtained on nine stimuli (shopping locations) against
seven in the first survey. This should make a 3-dimensional nonmetric-scaling
model more reliable and potentially more similar to the factor-analysis model.
This should be tested, since it could potentially weaken or strengthen the
conclusions of this research to recommend factor analysis as the reduction
methodology.
219
Fifth, improved data were obtained in the second survey pertaining to
socioeconomic variables. These data could be used to explore further the
issue of market segmentation. In addition, attempts should be made to invest¬
igate multidimensional segmentation strategies, i.e., those in which two or
more segmentation variables are used simultaneously. Additional research is
also required here on statistical techniques that can assist in indicating
whether a particular segmentation scheme is effective. A sixth issue relates
to the logit choice model. While extensive research has now been undertaken
on the effects of the Independence of Irrelevant Alternatives Axiom (11A) on
mode choice, using the logit model, no extensions have been made to the area
of destination choice. It would be desirable, particularly given the more
complex issues of choice sets, to explore 11A properties and problems in the
context of destination choice. At the very least, diagnostic tests proposed
in recent research (CRA [1976]) should be applied to the models developed in
this research.
A more critical issue has been purposively ignored by this research
project. This is the issue of choice set. Early im this research, the problem
of identifying choice sets was recognized and assumptions made to bypass it.
Specifically, the models developed are based on what is almost certainly a
contrived choice set, comprising those shopping locations listed by the project
personnel on the questionnaire that were indicated as being known to the
respondents. The issue of choice-set definition was referred to briefly in
chapter 2 (p. 21), but the hypotheses advanced there were not tested in this
research. Within the limitations of the project, it was not found to be
possible to collect data on the range of alternative shopping sites alluded
to in chapter 2. Hence, further pursuit of this issue is likely to require
the collection of additional data. The definition of choice sets may, in
fact, require the pursuit of a different choice structure, based upon the
definition of acceptable alternatives. Some preliminary notions in this
direction were advanced in one of the working papers from this project
(Peterson and de Bettencourt [1976]), together with some suggestions for
appropriate data collection.
Eighth, it has been stated in this report that the findings with
respect to the salient, reduced attributes of destination attractiveness
undermine the traditional use of measures of size (floor area, employment) in
conventional travel-forecasting models. However, this finding needs to be
verified in at least two directions. First, it is necessary to collect data
from other locations and other socioeconomic groups in order to confirm the
identification of salient, reduced attributes. Second, the models based on
these measures should be compared in both descriptive and predictive power
to models using more traditional measures, to test more directly the compar¬
ative advantages of the models developed in this research.
Finally, as in most areas of travel-forecasting research, attempts have
been made here to develop demand models from cross-sectional data, in which
the dynamics of decision making and decision changing cannot be discerned.
There is a need to test these models, and most others developed for forecasting
purposes, on dynamic behavior changes, traced out from time-series data. With¬
out such data, the forecasting ability of these models cannot be established.
Indeed, it is probable that the models are not effective at forecasting because
they are based on cross-sectional associations in the data and cannot respond
to dynamic changes and instabilities in behavior.
220
REFERENCES
Aaker, D.A. and J.M. Jones, "Modeling Store Choice Behavior," Journal of
Marketing Research, 8 (Feb. 1971), 38-42.
Appelbaum, W., "Methods for Determining Store Trade Areas, Market Penetration,
and Potential Sales," Journal of Marketing Research, 3 (May 1966), 127-41.
Aspinwall, !_., "The Characteristics of Goods and Parallel System^ Theories,"
Managerial Marketing, Perspectives and Viewpoints, edited by E.J. Kelley
and W. Lazer (Homewood, Illinois: Irwin, 1962), 633-652.
Bass, F.M., "The Theory of Stochastic Preference and Brand Switching," Journal
of Marketing Research, 11 (Feb. 1974), 1-20.
Baumol, W.J. and E.A. Ide, "Variety in Retailing," Management Science, 3 (1956),
93-101.
Ben-Akiva, M., Structure of Passenger Travel Demand Models, Ph.D. dissertation,
Department of Civil Engineering, Massachusetts Institute of Technology (1973).
Berkson, J., "Application of the Logistic Function to Bio-Assay," Journal of
the American Statistical Association, 39 (Sept. 1944), 357-365.
Bishop, W. and E. Brown, "An Analysis of Spatial Shopping Behavior," Journal of
Retailing, 45 (summer 1969), 23-30.
Boyle, B., The Decision To Grant Credit, Ph.D. dissertation, Department of
Electrical Engineering, Massachusetts Institute of Technology (1974).
Brand, D., "Travel Demand Forecasting: Some Foundations and a Review," Urban
Travel Demand Forecasting, D. Brand and M.L. Manheim (eds.). Highway Research
Board Special Report (1973), 143.
Brown, R.C., "The Use and Mis-Use of Distance Variables in Land-Use Analysis,"
The Professional Geographer, 20 (Sept. 1968), 337-341.
Brown, F.E., "Price Image Versus Price Reality," Journal of Marketing Research,
6 (May 1969), 185-195.
Bruner, J.S., J.J. Goodnow and G.R. Austin, A Study of Thinking (New York: Wiley,
1956).
Brunner, J. and J.L. Mason, "The Influence of Drive Time on Shopping Center
Preference," Journal of Marketing, 32 (April 1968), 57-61.
Bucklin, L.P., Shopping Patterns in an Urban Area, Institute of Business and
Economic Research, University of California (1967).
Bucklin, L.P., "Trade Area Boundaries: Some Issues in Theory and Methodology,"
Journal of Marketing Research, 8 (Feb. 1971), 30-37.
Burnett, K.P-, "The Dimensions of A1ternatives in Spatial Choice Processes,"
Geographical Analysis, 5(3) (1973), 181-204.
221
Burnett, K.P., "Two Disaggregated Models for Trip Distribution Over Time,"
unpublished mimeo, University of Texas at Austin (1974a).
Burnett, K.P., "A Three-State Markov Model of Choice Behavior Within Spatial
Structures," Geographical Analysis, 6(1) (1974b), 53-68.
Burnett, K.P., "A Bernoulli Model of Destination Choice," Transportation Research
Record, (527) (1974c), 33-44.
Carmone, F.J., G.E. Green and A.K. Jain, The Robustness of Conjoint Analysis:
Some Monte Carlo Results, Working Paper, University of Pennsylvania (1976).
Carroll, J.D., "Polynomial Factor Analysis," American Psychological Association
Proceedings, 4 (1969), 103-104.
Carroll, J.D., "Individual Differences and Multidimensional Scaling,"
Multidimensional Scaling: Theory and Application in the Behavioral Sciences.
R.N. Shepard, A.K. Romney and S. Nerlove (eds.), (New York: Academic Press)
(1972), 105-157.
Carroll, J.D. and J.J. Chang, "A 'Quasi-nonmetric' Version of INDSCAL, a
Procedure for Individual Differences Multidimensional Scaling," Psychometric
Society, Stanford, Cal. (March 1970).
Carrothers, G.A.P., "An Historical Review of Gravity and Potential Concepts of
Human Interaction," Journal of the American Institute of Planners, 22 (1956),
94-102.
Cattin, P. and D.R. Wittink, A Monte Carlo Study of Metric and Non-Metric
Estimation Methods for Multi-Attribute Models, Research Paper no.341, Graduate
School of Business, Stanford University (1976).
Chang, J.J. and J.D. Carroll, How To Use Profit, a Computer Program for Property
Fitting by Optimizing Nonlinear or Linear Correlation, Bell Telephone Laboratories
(1970).
Charles River Associates, A Disaggregate Behavioral Model of Urban Travel Demand,
(Washington, D.C.: Federal Highway Administration, U.S. Department of Transporta¬
tion, 1972).
Charles River Associates, Disaggregate Travel Demand Models: Phase I Report,
National Cooperative Highway Research Project 8-13 (1977).
Clark, W.A.V., "Consumer Travel Patterns and the Concept of Range," Annals of
the Association of American Geographers, 58 (1968), 386-96.
Cox, W.E., Jr. and E.F. Cooke, "Other Dimensions Involved in Shopping Center
Preference," Journal of Marketing, 34 (October 1970), 12-7.
de Bettencourt, J.S. and G.L. Peterson, A Threshold Theory for Choice Behavior
Studies: General Considerations and Initial Experimental Results, Working Papers,
Transportation Center, Northwestern University (1976).
de Leeuw, J., "The Positive Orthant Method for Nonmetric Multidimensional Scaling,"
Mimeographed Report, Department of Data Theory for the Social Sciences, University
of Leiden, Netherlands (1969).
222
Dobson, R. and J.F- Kehoe, "Disaggregated Behavioral Views of Transportation
Attributes," Transportation Research Record, (527, 1974), 1-15.
Dobson, R. and G.C. Nicolaidis, "Preferences for Transit Service by Homogeneous
Groups of Individuals," Transportation Research Forum, Proceedings (1974).
Domenich, T. and D. McFadden, Urban Travel Demand: A Behavioral Analysis
(North-Holland Pub. Co., 1975).
Einhorn, H.J., "Multi-Attribute Models - Why Weights Don't Matter," American
Marketing Association, Combined Proceedings, Ser. no.37 (1975), 95-98.
Fishbein, M., "Attitudes and the Prediction of Behavior," Attitude Theory and
Measurement, M. Fishbein (ed.), (New York: Wiley, 1967).
Fisk, G., "A Conceptual Model for Studying Customer Image," Journal of Retailing,
37 (winter 1961-62), 1-8, 54.
Forbes, J.D., "Consumer Patronage Behavior," Proceedings. Fall Conference,
American Marketing Association, 1968, 381-85.
Friedman, Milton, "The Use of Ranks to Avoid the Assumption of Normality
Implicit in the Analysis of Variance," Journal of the American Statistical
Association, 1 937, pp. 675-701.
Gallaqher, R., Information Theory and Reliable Communication (New York: Wiley,
1968).
Gleason, T.C., "A General Model for Nonmetric Multidimensional Scaling,"
Michigan Mathematical Psychology Program 67-3, University of Michigan (June 1969).
Golob, T.F. et al, "An Analysis of Consumer Preferences for a Public Transportation
System," Transportation Research, 6(1) (March 1972)s 81-102.
Green, P.E. and F.J. Carmone, Multidimensional Scaling and Related Techniques
in Marketing Analysis (Boston: Allyn & Bacon, 1970).
Green, P.E. and V. Rao, "Conjoint Measurement for Quantifying Judgmental Data,"
Journal of Marketing Research, 8 (August 1971), 355-363.
Green, P.E. and V. Rao, Applied Multidimensional Scaling (New York: Holt Rinehart
& Winston, 1972).
Green, P.E. and Y. Wind, Multiattribute Decisions in Marketing (Hinsdale, Illinois:
Dryden Press, 1973).
Green, P.E. and Y. Wind, "New Way to Measure Consumer's Judgments," Harvard
Business Review (July-August 1975).
Halley, R.I., "Benefit Segmentation: a Decision-Oriented Research Tool,"
Journal of Marketing, 32 (July 1968), 30-35.
Hanson, S., "On Assessing Individuals' Attitudes Towards Potential Travel Desti¬
nations: A Research Strategy," Transportation Research Forum Proceedings (1974),
363- 370.
Harman, H.H., Modern Factor Analysis (Chicago: University of Chicago Press, 1967).
223
Hauser, J.R., A Normative Methodology for Modeling Consumer Response to Innovation:
Issues, Models, Theory, and Use, Sc.D. thesis, Operations Research Center,
Massachusetts Institute of Technology (1975).
Hauser, J.R., Testing the Accuracy, Usefulness, and Significance of Probabilistic
Choice Models: an Information Theoretic Approach, Working Paper, Transportation
Center, Northwestern university (April 1976).
Hauser, J.R. and F.S. Koppelman, "Designing Transportation Services: a Marketing
Approach," Transportation Research Forum Proceedings (1977), 638-652.
Hauser, J.R. and G.L. Urban, Direct Assessment of Consumer Utility Functions:
von Neumann-Morgenstern Theory Applied to Marketing, Working Paper, Sloan School,
Massachusetts Institute of Technology (1976), 843-76.
Hauser, J.R. and G.L. Urban, "A Normative Methodology for Modeling Consumer
Response to Innovation." Operations Research 25(4) (July-Aug. 1977), 579-619.
Herrmann, R. and L. Beik, "Shopper Movements Outside Their Local Trade Area,"
Journal of Marketing, 32 (Oct. 1968), 45-51.
Huff, D.L., "A Topographical Model of Consumer Space Preferences," Regional
Science Association Proceedings, 6 (1960), 159-173.
Huff, D.L., "A Probabilistic Analysis of Shopping Center Trade Areas," Land
Economics, 39 (1963), 81-90.
Johnson, S.C., "Hierarchical Clustering Schemes," Psychometrika, 32 (1967),
241-254.
Johnson, R.M., "Tradeoff Analysis of Consumer Values," Journal of Marketing
Research, 2 (May 1974), 121-127.
Kelley, E.J., Shopping Centers: Locating Controlled Regional Centers. New
York: The Eno Foundation for Highway Traffic Control, Columbia University
Press (1956), chapter T6.
Klemmer, E.T. and N.W. Shrimpton, "Preference Scaling Via a Modification of
Shepard's Proximity Analysis Method," Human Factors, 5 (1963), 163-168.
Koppelman, F.S., "Prediction with Disaggregate Models: the Aggregation Issue,"
Transportation Research Record, 527 (1975).
Kruskal, J.B., "Multidimensional Scaling by Optimizing Goodness of Fit to a
Nonmetric Hypothesis," Psychometrika. 29 (March 1964), 1-27.
Kruskal, J.B. and F.J. Carmone, How To Use M-D-SCAL: a Program To Do Multi¬
dimensional Scaling and Multidimensional Unfolding, Bell Telephone Laboratories
(1969).
Lerman, S.R., C.F. Manski and T.J. Atherton, Non-Random Sampling in the
Calibration of Disaggregate Choice Models, Cambridge Systematics (1976).
224
Lingoes, J.C., "An IBM 7090 Program for Guttman-Lingoes Smallest Space Analysis,"
Behavorial Science, 10 (1965), SSA I, 183-184; SSA II, 487; 11 (1966), SSA III,
75-76; SSA IV, 407; SSAR I, 322; SSAR II, 323; 12 (1967), SSAR IV, 74-75.
Lingoes., J.C., "An IBM 7090 Program for Guttman-Lingoes Conjoint Measurement,"
Behavorial Science, 12 (1967), 501-502.
Lingoes, J.C. and L. Guttman, "Nonmetric Factor Analysis: A Rank Reducing
Alternative to Linear Factor Analysis," Multivariate Behavioral Research, 2
(1967), 485-505.
Liou, P.S., G.S. Cohen and D.T. Hartgen, A Systems-Level Planning Application of
the Disaggregate Modeling Technique. Report PRR6Z. (Albany, NY: Department of
Transportation, 1974).
Lisco, T.E., The Value of Commuters' Travel Time, a Study in Urban Transportation,
Ph.D. Thesis, Department of Economics, University of Chicago (1967).
Lovelock, C.H., "A Market Segmentation Approach to Transit Planning, Modelling
and Management," Transportation Research Forum Proceedings (1975).
Luce, R.D. and P- Suppes, "Preference Utility and Subjective Probability,"
Handbook of Mathematical Psychology, R.D. Luce, R.R. Bush and E. Galanter (eds.)
(New York: Wiley, 1965), 249-410.
MacKay, D.B., "A Microanalytic Approach to Store Location Analysis," Journal of
Marketing Research, 9 (May 1972), 134-40.
Marble, D.F. and S.R. Bowlby, "Shopping Alternatives and Recurrent Travel Patterns,"
Geographic Studies of Urban Transportation and Network Analysis, F.E. Horton (ed.),
Northwestern University Studies in Geography, no. 16 (1968), 42-75.
Mason, J.B. and C.T. Moore, "An Empirical Reappraisal of Behavioral Assumptions
in Trading Area Studies," Journal of Retailing, 6 (1970), 31-37.
McFadden, D., "Conditional Logit Analysis of Qualitative Choice Behavior," Frontiers
in Econometrics, Paul Zaremblea (ed.), (New York: Academic Press, 1970) 105-142.
McFadden, D., "The Theory and Measurement of Qualitative Consumer Behavior: The
Case of Urban Travel Demand," Working Paper 1, University of California (Sept.
1972).
McFadden, D., "Quantal Choice Analysis: A Survey," paper presented to the NSF-NBER
Conference on Decision Rules and Uncertainty, Berkeley (1974).
McFadden, D. and H. Wills, "XL0GIT, A Program for Multinomial Logit Analysis,"
Travel Demand Forecasting Project, Institute for Transportation and Traffic
Engineering, University of California (1970).
McGee, V., "Multidimensional Scaling of N Sets of Similarity Measures: A Non-Metric
Individual Differences Approach," Multivariate Behavioral Research, 3 (1968) 233-248.
225
Mood, A.M. and F.A. Graybill, Introduction to the Theory of Statistics (New York:
McGraw-Hill, 1963).
Nerlove, M. and S.J. Press, "Univariate and Multivariate Loglinear and Logistic
Models," Rand Corporation Report R-1306-EDA/NIH (1973).
Nicolaidis, G.C., "Quantification of the Comfort Variable," Transportation Research.
9(1) (1975), 55-66.
Pennell , R., "Additive Representations for Two Dimensional Tables," Research
Bulletin RB 7-29. Princeton, N.J., Educational Testing Service ( 19707*1
Pessemier, E.A., Market Structure Analysis of New Products, Market and Communication
Opportunities, Report no.76-106, Marketing Science Institute (1976).
Peterson, G.L., J. de Bettencourt, and P.K. Wang, "Markov-Based Linear Programming
Travel Models," Department of Civil Engineering, Northwestern University (1976),
Philpot, J.C. and R.C. Reizenstein and D.J. Sweeney, "Identifying Determinants of
Store Patronage Using Factor Analysis," Proceedings. Third Annual Conference of
Association for Consumer Research (1972), 201-12.
Quarmby, D.A., "Choice of Travel Mode for the Journey to Work: Some Findings,"
Journal of Transport Economics and Policy, 1(3) (1967).
Rei11 y, W.J., Methods for Study of Retail Relationships, Research Monograph #4.
Bureau of Business Research, University of Texas (1929).
Reynolds, F. and W. Darden, "Intermarket Patronage: A Psychographic Study of
Consumer Outshoppers," Journal of Marketing (October 1972), 50-54.
Rosenberg, M.J., "Cognitive Structure and Attitudinal Effect," Journal of
Abnormal and Social Psychology, 53 (1965), 367-372.
Roskam, E.E., Metric Analysis of Ordinal Data: Psychology (Voorschoten, Holland:
VAM Press, 1968^
Rummel, R.J., Applied Factor Analysis (Evanston, 111.: Northwestern University
Press, 1970).
Rushton, G., "The Scaling of Location Preferences," Behavioral Problems in
Geography, K.R. Cox and R.G. Golledge (eds.), Department of Geography,
Northwestern University (1969), 199-237.
Shepard, R.N., "The Analysis of Proximities: Multidimensional Scaling with an
Unknown Distance Function I," Psychometrika, 27 (1962), 125-140.
Shepard, R.N., "A Taxonomy of Some Principal Types of Data and of Multidimensional
Methods for Their Analysis," Multidimensional Scaling, volume I, R.N. Shepard,
A.K. Romney and S.B. Nerlove (eds.) (New York: Seminar Press, 1972).
Shepard, R.N. and J.B. Kruskal, "Nonmetric Methods for Scaling and for Factor
Analysis," American Psychologist, 19 (1964), 557-558.
226
Sheth, J.N., "Canonical Analysis of Attitude-Behavior Relationships," Institute
of Management Sciences, 18th international meeting (1971).
Simmons, J., The Changing Pattern of Retail Location, Research Paper no.92,
Department of Geography, University of Chicago (1964).
Singson, R., "Multidimensional Scaling Analysis of Store Image and Shopping
Behavior," Journal of Retailing, 51(2) (1975).
Skinner, R.W., "Hidden Consumer Motives in Supermarket Selection," American
Journal of Agricultural Economics, 51 (1969), 1154-68.
Slater, P., "The Analysis of Personal Preferences," British Journal of
Statistical Psychology, 13 (1960), 119-135.
Sneath, P.H.A., "The Application of Computers to Taxonomy," Journal of
General Microbiology, 17 (1957), 201-226.
Sorensen, T., "A Method of Establishing Groups of Equal Amplitude in Plant
Sociology Based on Similarity of Species Content and Its Application to
Analyses of the Vegetation on Danish Commons" Biological Skr., 5 (1948),
1-34.
Spear, B.D., The Development of a Generalized Convenience Variable for
Models of Mode Choice, Ph.D. dissertation, Department of Environmental
Engineering, Cornell University (1974).
Stopher, P.R., "A Probability Model of Travel Mode Choice for the Work
Journey," Highway Research Record, 283 (1969), 57-65.
Stopher, P.R. and T.E. Lisco, "Modelling Travel Demand: A Disaggregate
Behavioral Approach - Issues and Application," Transportation Research
Forum Proceedings (1970), 195-214.
Stopher, P.R. and A.H. Meyburg, "Travel Demand Estimation: A New Prescrip¬
tion," Traffic Engineering and Control, 15(19) (1974), 879-884.
Stopher, P.R. and A.H. Meyburg, Urban Transportation Modeling and Planning,
(Lexington, Mass.: Lexington Books, 1975).
Talvitie, A.P., "Aggregate Travel Demand Analysis With Disaggregate or
Aggregate Travel Demand Models," Transportation Research Forum Proceedings,
(1973) 583-603.
Thompson, B., "Intraurban Retail Structure: The Supermarket Sector," Journal
of Retailing, 45 (fall 1969), 69-80.
Thompson, D., "New Concept: Subjective Distance," Journal of Retailing, 39
(spring 1963),1-6.
Thurstone, L., "A Law of Comparative Judgment," Psychological Review, 34
(1927), 273-286.
227
Torgerson, W.S., "Multidimensional Scaling: I Theory and Method," Psychometrika,
17 (1952), 401-419.
Torgerson, W.S., Theory and Methods of Scaling (New York: Wiley, 1958).
Torgerson, W.S., "Quantitative Judgment Scales," Psychological Scaling:
Theory and Applications, Gulliksen and Messick (eds.) (New York: Wiley, 1960).
Tversky, A., "A General Theory of Polynomial Conjoint Measurement," Journal of
Mathematical Psychology, 4 (1967), 1-20.
Tversky, A. and A. Zivian, "An IBM 7090 Program for Additivity Analysis,"
Behavioral Science, 11 (1966), 78-79.
Urban, G.L., "Perceptor: A Model for Product Positioning," Management Science,
8 (April 1975), 858-871.
Ward, J.H., Jr., "Hierarchical Groupings To Optimize an Objective Function,"
Journal of the American Statistical Association, 58 (1963), 236-244.
Warner, S.L., Stochastic Choice of Mode in Urban Travel: A Study in Binary
Choi ce (Evanston: Northwestern University Press, 1962).
Watson, P.L., "System Service Requirements, Other Than Travel Times and Costs,
To Be Included in the Models," Transportation Center, Northwestern University
(1973).
Watson, P.L. and P.R. Stopher, "The Effects of Income on the Usage and Valuation
of Transport Modes," Transportation Research Forum Proceedings (1974).
Watson, P. and R. Westin, "Transferability of Binary Choice Models," unpublished
manuscript, Northwestern University (1973).
Westin, R.B., "Predictions from Binary Choice Models," Journal of Econometrics,
2 (May 1974), 1-16.
Westin, R.B. and P.L. Watson, "Reported and Revealed Preferences as Determinants
of Mode Choice Behavior," Journal of Marketing Research, 12 (August 1975), 282-289.
Wilkie, W.L. and E.A. Pessemier, "Issues in Marketing's Use of Multi-Attribute
Attitude Models," Journal of Marketing Research, 10 (Nov. 1973), 428-441.
Young, F.W., Polynominal Conjoint Analysis of Similarities: Definitions for a
Speci fi c A1gorithm. Research Paper 76. University of North Carolina Psychometric
Laboratory (1969).
Young, F.W. and W.S. Torgerson, "T0RSCA, A FORTRAN IV Program for Shepard-Kruskal
Multidimensional Scaling Analysis," Behavioral Science, 12 (1967), 498.
Yuill, R.S., "Spatial Behavior of Retail Customers: Some Empirical Measurements,"
Geografiska Annaler, 1 (1967) 49B, no.2.
228
APPENDIX A
PRETEST CROSSTABULATIONS
AGE
SEX
Under 21
22 - 29
30 - 39
40 - 49
50 - 59
60 and
Over
MALE
18.7
18.6
6.8
20.3
8.5
27.1
FEMALE
14.8
18.2
16.6
22.2
15.3
7.1
TABLE A-l
Crosstabulation of Age and Sex for the Pretest Survey
INCOME
SEX
Less Than
$10,000
$10,000-
$14,999
$15,000-
$19,999
$20,000-
$24,999
$25,000-
$49,999
$50,000
and over
MALE
5.1
20.4
22.0
16.9
20.3
10.2
FEMALE
11.0
14.0
13.2
13.2
25.1
15.0
TABLE A-2
Crosstabulation of Sex and Income for the Pretest Survey
AGE
INCOME
Less Than
$10,000
$10,000-
$14,999
$15,000-
$19,999
$20,000-
$24,999
$25,000-
$49,999
$50,000
and over
UNDER
22
13.3
17.3
13.3
10.7
13.3
17.3
22 - 29
28.8
17.5
13.7
18.8
11.2
7.5
30 - 39
1.5
14.9
23.9
13.4
34.3
9.0
40 - 49
0
11.5
10.4
15.6
34.4
25.0
50 - 59
4.8
11.1
9.5
15.9
38.1
19.0
60 AND
OVER
18.6
23.2
20.9
7.0
16.3
4.7
TABLE A-3
Crosstabulation of Age arid Income for the Pretest Survey
A-3
APPENDIX B
1974 QUESTIONNAIRE
THE TRANSPORTATION CENTER • NORTHWESTERN UNIVERSITY
LEVERONS HALL • 2001 SHERIDAN ROAD • EVANSTON. ILLINOIS 60201
July 24, 1974
Dear Shopper:
The Transportation Center of Northwestern University is conducting a project which is aimed at helping to improve
the planning of transportation for shopping trips. To assist in this work, we would like to ask for your help in
spending a little of your time to answer this questionnaire. The purpose of the questionnaire is to find out what
attracts people to different stores and shopping centers.
Your answers to this questionaire will be kept strictly confidential and will only be used for scientific inquiry.
Information about individuals will not be divulged and your response is anonymous, since we do not request your name.
However, your response is most important to us to ensure the success of our project.
If you have any questions about this project or the questionnaire, please feel free to ask one of the survey
interviewers handing out the questionnaires, or telephone either Dr. Peter Watson at 492-5017 or Dr. Peter R. Stopher
at 492-5183. In answering the questionnaire, if you have any comments or suggestions to offer, please feel free to
add these on your couriered questionnaire.
Thank you for your valuable help on this project.
Peter Rf?tôpher
We would like you to complete this questionnaire in connection with the shopping center visit you made when
you received the questionnaire.
We would like you to indicate hew
familiar you are with each of the shopping
centers in our list. For each center,
please check the box which best describes
your familiarity with that center.
a* °
* £
a H
# * '
-* <?.»
*" ï> *
c •
• s
-4? 0}
«
4*
q, *
<? 4? J9
f c
v s
■
■i
y>
Plaza del Logo
Sdens Plaza
Old Orchard Shopping Center
Talisman Shopping Center
Coif Mill Shopping Center
Iaorenoe Wood Shopping Center
Rorthpoint Shopping Center
Randhurst
Arlington Market Shopping Center
Mount Prospect Plaza
Market Place Shopping Center
Doer,field Commons Shopping Center
Deertrook Shopping Center
Morthbrook Meadows
llorthbrook Shopping Plaza
Rolling Meadows Shopping Center
loans ton Downtown
hod field
Chicago Loop
torvette City
Others (Please name)
[ ]
3 E 3 t
Si
B-l
If this shopping center were not available for the particular purchases that you came to make, which other shop¬
ping centers would you think about going to? Please check those that you would consider.
Plaza del Logo [ ]
Edens Plaza [ ]
Old Orchard Shopping Center [ ]
Talisman Shopping Center [ ]
Golf Mill Shopping Center [ ]
Laurence Wood Shopping Center [ ]
tlorthpoint Shopping Center [ ]
Randhuret
Arlington Market Shopping Center
Mount Prospect Plaza
Market Place■ Shopping Center
Deerfield Commons Shopping Center
Deerbrook Shopping Center
Northbrook Meadows
Northbrook Shopping Plaza
Rolling Meadows Shopping Center
Evans ton Downtown
Woodfield
Chicago Loop
Korvette City
Other (Please name)
If all the following shopping centers
were equally easy to get to, which of
them would you prefer to shop at for the
goods you came to buy?
Please indicate your order of prefer¬
ence by placing a number beside each
center. Start with number 1 for the
most preferred shopping center, number
2 for the second most preferred, and so
on down to the least preferred shopping
center.
Please rank aU the shopping centers.
Chicago Loop
Edens Plaza (Wilmette)
Coif Mill Shopping Center
Korvette City (Dempster S
Waukegan)
Plaza del Logo
Old Orchard
Woodfield
Again, if all the shopping centers were
equally easy to get to, how similar do you think
they are to each other? In answering this ques¬
tion, please think about your preference to shop
at them for the goods you came to buy. Check
the box which best describes how similar they
are. Please be sure to do this for all pairs
of shopping centers.
ÔT le
> J?
/
<9
Woodfield
and Chicago Loop
Edens Plaza
and
Coif Mill
Woodfield
and
Plaza del Lago
Cnicago Loop
and
Golf Mill
Old Orchard
and
Wood fie Id
Golf Mil I
and
Korvette City
Chicago Loop
and
Old Orchard
Plaza del Lago
and
Coif Mill
Korvette City
and
Old Orchard
Plaza del Lago
and
Edens Plaza
Chicago Loop
and
Korvette City
Woodfield
and
Coif Mill
Edens Plaza
and
Korvette City
Old Orchard
and
Golf Mill
Edens Plaza
and
Woodfield
B-2
0
ti
* i"
5^
//
✓
<?
Old Orchard
and
Plaza del Logo
Korvette City
and
Plaza del Lago
Chicago Loop
and
Edena Plaza
Voodfield
and
Korvette City
Plaza del Logo
apd
Chicago Loop
Edena Plaza
and
Old Orchard
For the store and shopping center where you
received this questionnaire, how important were
the following characteristics of the store and
the shopping center to you when you decided to
tome here? Please be sure to tell us how
important they are to you and not how available
they are.
layout of store
5restige of store
luality of merchandise
Reasonable price
Rase of returning or servicing mer&andise
.vailability of credit
vailability of sale items ("specials")
ree Parking
tores located in a compact area
tore atmosphere (heating, cooling,
noise, crowds, etc.)
hopping center atmosphere (pedestrian-
only area, flowers & shrubs,
covered walk-ways, etc.)
vailability of a specific store
bility to park where you want
ourteous S helpful sales assistants
mber and variety of stores
iriety or range of merchandise
u b
? #
«4
ts
3 S
ij 5»
I# sr a-
e
4 *
u S
'S
o
£L
*!
«* §
3
u
* s
2 a
S S
B-3
In this question, we would like you to rate each of the shopping centers on these characteristics. We have
provided a range from good to poor for each characteristic
center fits on this range.
We would like you to tell us where you feel each shopping
For example:
J
#
6
s
s ?
f
-i
«
s
ta
i
S
*4
3
* ?
1 è
i s
S S
i i
good ^
Eating
Facili¬
ties
poor
Please do this for all the shopping centers on all the characteristics,
center, please guess where you think it would fit.)
s
' 3 ? 3. _
§•
J
I
3
s
•*«
S
£
S
o
s
(If you are not familiar with a shopping
J S -
s
è?
I
o
i
*
£
- ?"
4* O
«
Ê .?
a? °
i
f
«
V
S
s i
good
good
Layout
of
store
poor
Ease of
returning
or
servicing
merchandise
poor
good
Prestige
Of
3 tore
Variety
or range
Of
merchandise
gOOi
poor
poor
good
duality
>f
nerchandise
Availability
of
credit
' good
poor
poor
good
Reasonable
>rice
Availability
of sale
items
("spéciale") '
"good
poor
poor
B-4
a
o
j
S
a
■o
•o
3
i
»?
A,
?
£
«*4
3
2?
3
60
3
•o
V
è
3
o
3
Free
parking
good
poor
store good
atmosphere
(heating,
cooling,
noi3e,
crouds,
eta-> poor
good
Shopping
center
atmosphere
poor
good
Availability
of a specific
store
poor
Stores
located in
a compact
area
good
poor
Ability to
park uhere
you want
good
poor
Courteous S
helpful sales
assistants
good
poor
Number and
variety of
stores
good
poor
So that we can study your answers in the context
of the type of shopping trip that you are making,
we would like you to complete this part of the ques¬
tionnaire in connection with the shopping center
visit you made when you received the questionnaire.
Please name the stores you came to visit:
1.
2.
3.
4.
5.
Hew did you travel to the shopping center?
'• Automobile (driver) [ ]
2- Automobile (passenger) [ ]
3. Taxi [ ]
Bus [ ]
5. Other public transport [ 1
Motorcycle f J
?• Bicycle
8. Walking [ J
Please check the major items you- came to purchase.
Clothing, including shoes S accessories
Home furnishings
Major appliances
Small appliances
Housewares
Toys i '
Gifts
Books, stationary
Lawn or garden equipment
Tard goods
Sporting goods
Window shopping
Eating out
Drugs and cosmetics
Tools and hardware, paints
Jewelry, clocks, watches
China, glass, flatware
Home entertainment
Automotive accessories
Groceries, specialty foods or candies
Liquor
Other (please specify)
-5
Did you travel to the shopping center directly from:
Home
Work
Another shopping center
(Pleaee name)
Other
(Please specify).
And where will you go to next?
Home
Work
Another shopping center
(Please name)
Other
(Please specify).
So that we can determine how well served you are
public transport, please add your address.
Street
I or block
City, town, or village
Zip code
In order to help with the classification of results
of this survey we would be very grateful if you would
provide the information in this last section. We
appreciate that you may feel hesitant about revealing
some of this information, which is of a personal nature,
but these factors are important to this study. We will
use the information only for research purposes and it
will not be divulged to anyone.
In each case please check the appropriate boxes.
emale
[ ]
[ ]
Age:
Under 16 years
16 - 21
22 - 29
30-39
Occupation:
[ ]
[ 3
[ ]
[ 3
40 - 49
SO - S9
60 years or over
t]
[ 3
[ 3
Military
Salesman/Buyer
Teacher/Professor
Professional/Technical/ManageriaI
Craftsman/Mechanic/Factory Worker
Clerical/Secretarial/Office Worker
Student
Housewife
Governmental
He tired
Other
What is your approximate family income before
taxes? (Your income plus that of your husband
or wife or the person on whom you are dependent.)
$10,000 and under [ ]
$10,00) - $15,000 [ ]
$15,001 - $20,000 [ ]
$20,001 - $25,000 t 3
$25,001 - $50,000 [ 3
Over $50,000 [ 3
How long have you lived in this area, that is,
north-west Chicago and/or the north and north¬
west suburbs? If not a resident, please check [ ].
Less than 1 year [ ]
1-3 years [ ]
4-6 yeare [ ]
7-10 years
[ 3
more than 10 years [ ]
THANK YOU VERY MUCH FOR YOUR COOPERATION
B-6
APPENDIX B
1975 QUESTIONNAIRE
THE TRANSPORTATION CENTER • NORTHWESTERN UNIVERSITY
LEVERONE HALL • 2001 SHERIDAN ROAD • EVANSTON, ILLINOIS 60201
July 1975
Dear Shopper:
The Transportation Center of Northwestern University is conducting
a project which is aimed at helping to improve the planning of transpor¬
tation for shopping trips. To assist in this work, we would like to ask
you to help by spending a little of your time to answer this question¬
naire. The purpose of the questionnaire is to find out what attracts
people to different stores and shopping centers.
Your answers to this questionnaire will be kept strictly confidential
and will only be used for scientific inquiry. Information about individuals
will not be divulged and your response is anonymous, since we do not re¬
quest your name. However, your response is most important to us to ensure
the success of our project.
If you have any questions about this project or the questionnaire,
please feel free to ask one of the survey interviewers handing out the
questionnaires, or telephone Dr. Peter R. Stopher at 492-5183. In answer¬
ing the questionnaire, if you have any comments or suggestions to offer,
please feel free to add these on your conpleted questionnaire.
Thank you for your valuable help on this project.
B-7
we would like you to complete this questionnaire in connection with the shopping center visit you made
when you received the questionnaire.
we would like you to indicate how frequently you have visited eûcu.of the skipping locations in our list.
below each. location, please circle the number which best describes how frequently you have visited that location.
m
u X £
i. (U • *r *•
OJ . X J
«* . i. e i. J- i/t m c — «/»
s„
d a
"O r—
■^o r— r—
■£. S
3
8
-- *» a.
o f o
u cn
a. -t- to
se o
55 S
î«
<8
I have heard of it but I
have never visited it 22222222222
i have been there but
not in the past year 33333333333
During the past year I have been there on the average:
less than once a month
1 or 2 times a month
3 or 4 times a month
'lore than 4 times a
month
2 2 2
3 3 3
Suppose that all of the following shopping locations were equally easy to get to. we would like you to
indicate your preference for shopping at each location for the goods you came to purchase when you received
'.his questionnaire.
we have provided a scale ranging from most preferred to least preferred. be1j0w the shopping location
which you most prefer circle the num5er 'j below the location milch you least prefer circle the number 7;
below fach of the remaining locations circle the nimber which indicates your preference for that location.
m •#-
8
o IE
tÉ ^
o 5
+J
VI
c
0>
3
C
at
S—
J. •
to -a
o ce
o
O!
to
u
<
' u
<0 CI
.* js
u
«
cn
jë c
O CI
^ (_)
in
a a.
g-
-»
x: o
o cn
L
to
«
c.
2E o
5
in
tii
o
•o
55
c
0»
to
3
o
8
m
8
3
Host Preferred
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
S
6
6
6
6
6
6
6
6
Least Preferred
7
7
7
7
7
7
7
7
Please continue on next
page
- i
I have never heard of it 11111111111 1 111 1 llllll
2 2 2 2 2 2
3 3 3 3 3 3
13-8
please circle the numbers next to the major items you came to purchase when you received this questionnaire
1 Clothing, Including shoes & accessories 13 Tools and hardware, paints
2 Furniture 14 Jewelry, clocks, watches
3 Home furnishings 15 China, glass, flatware
4 Major appliances 16 Home entertainment
5 Small appliances 17 Automotive accessories
6 Housewares 18 Groceries, specialty foods or candles
7 Toys 19 Liquor
8 Books, stationery 20 Window shopping
9 Lawn or garden equipment 21 Eating out
10 Yard goods 22 Other (please specify)
11 Sporting goods y
12 Drugs and cosmetics
in deciding where to purchase the items you circled in the previous question, how important to tqu.were
the following characteristics of shopping locations? we have provided a scale which ranges from extremely
important to of no importance. below each characteristic please circle the number which you feel best indi¬
cates the importance of the characteristic, please be sire to tell us uqjl important the characteristics are
to you and not how available they are.
M p
cu C
■O JC
CJ i/t
£ s a
° 1/1 i
** o *-
o «. °
2
07 07 C •*- •»-
I £ fO
Extremely Important
Of No Importance
U1
i
u
ra
07
en »
c «/>
T- "O
ï
•4-»
■M
E
O
T3
07
07
«
07 U
JC U
O
f-
o
M
«4-
o--.
c
07 07
k M
07 •»-
£
en
c
4->
<o
O. C
.s
m
S
^ U
.O 07
ra o.
r— tA
*»- s
ro —
5
k.
«o
O.
07
t
tores lo(
area
:ore atmt
cooling,
«t
<
U.
in
i/>
1
1
1
i
i
2
2
2
2
2
3
3
3
3
3
4
4
4
4
4
5
5
5
5
5
6
6
6
6
6
7
7
7
7
7
—- - >,
r— to I
3
O
>>
> P- T-
Û. U o •»-
£• S
T- *-»
ï I
a °
07 O)
t- o>
1 s
îiiiiii i
222222 2 2
333333 3 3
444444 4 4
555555 5 5
666666 6 6
7 7 7 7 7 7 7 7
Please continue on next page(Inside)
1 1
2 2
1 1
2 2
B-9
In this question we would like you to rate each Of-
we have provided a range frcm very good to very poor pc
the shipping locations fit on this range by circling ON
ar a
t-
at at g
f 4-»
CO CO
a ï c
g" -s
Ï
s
C H—
i 5
I/» f-*-* TJ
-C o U.
r— y Ol fl
S SES «
■r- J- Br
O p-
X
5
§ •
s
4J —
Very Good
Layout
of
Store
Very Poor
Very Good 1
Prestige
of
Store
Very Poor
Very Good
Quality
of
merchandise
Very Poor
Very Good
Reasonable
Price
Very Poor
to N Ul
«3
r- e
0> <M O. a
o
V\ -r- i/t 4->
Q. a C C
§to <u a
o
O LU Q
^ 55
^ C
o «y
-se O
c a»
X «-
o *—
*3 O
** O
Very Good
Ease of
returning
or
servicing
merchandise
Very Poor
Very Good
Variety
or range
of
merchandise
Very Poor
Very Good
Availability
of
credit
Very Poor
Very Good
Availability
of sale
itens
("specials")
Very Poor
Please continue on next page
B-1 0
a
ai
gf S.| S Si? R9 3 8s>gs g 1 2
Il§ a g- p- g i g !» aig
Êl ? ï if ? * -SUT » sis S§-
n p 3 3 2 ^ ^ '^gqQ a
<3 ? s • s» ? * s» ? &
O L
M»Q.
a < Z » 3* p
ni pf
g (l) {S _ __ W t?»V
<P ff s S&
§ Q «5 R g ë «< Q "3 tî* «<
&
je;
■5S-U&
SBS-P
Store
local
a coo
area
S * f»
3 vi
S
s=
I8-"
rt
8
Q
O
O
o
a.
a
T3
O
O
t
SS"
Dempster Street (at Skokle
Swift)
Edens Pie»
Downtown Evenston
Golf H111 Shopping Center
North Michigan Avenue
(Chicago)
Old Orchard
Downtown Skokle (Oakton St.
& Nil es Center Rd.)
State Street (Chicago)
I 5 s
9 r> 2
h> o
Dempster Street (at Skokle P J ^
S«1") op
Edens Plaza ï j o
Downtown Evanston
Golf Mill Shopping Center
c —
" 5
O X
g 9
North Hlchlgan Avenue
(Chicago) g Q
Old Orchard JS jj
P d
Downtown Skokle (Oakton St. £)
& Nil es Center Rd.)
State Street (Chicago)
Hoodfield
Woodfleld
Reasonable price
3 g <
to in in
| I 3
| % g
3d
i2!
layout of store |
Prestige of store q
g o
Quality of merchandise — 3
3 S>
O O) >
Ease of returning or -o Z!
servicing merchandise S 0
</> x
Availability of credit m S
o m
Availability of sale Items o S
("specials") S c
Free parking S 3
Z 30
Stores located In a compact S S
a rea 0 §
y ">
Store atmosphere (heating, x g
cooling, noise, crowds, etc.) g
3>
Shopping center atmosphere (pedes- § S
trlan-only area, flowers & shrubs, p; q
covered walk-ways, etc.) 5
Availability of a specific store
? 5
Ability to park where you want „ 0
K S
ss
Courteous S helpful sales assistants s;
"n
Number and variety of stores 0
Variety or range of merchandise |g
S
3
3
<
m rn 2.
> ' «
m 0
2 S
| ! f lië
Q !•. =»
">j <y* in ut ro —»
2 a
CD
II
~0
~o «
Dempster Street (at Skokle m 5
Swift)
o, w j» u ro — Edens Plaza
-4 » oi * w m -> Downtown Evans ton
-w m oi « u ro — Golf Mill Shopping Center
il
Fn Q
™ %
m
fin
5
-4 O.': w ' 4. C4 ro - North Michigan Avenue
(Chicago) i H
£ 5
-J O. W 4» U) N. -4 Old Orchard _ 5
§ 5
Downtown Skokle (Oakton St. 5 in
1 Ni les Center Rd.) jn g
30
State Street (Chicago) 2 5
£ ?
Woodfleld ~ g
TO
i 3
N Ol OI
m
p
6
c a
il
q s| q q 1 •?"
Please circle tie nlmber next to the MODE OF TRAVEL you used to reach the shopping center.
1 Automobile (driver) 3 Taxi 5 Other public transport 7 Bicycle
2 Automobile (Passenger) 4 Bus 6 Motorcycle 8 Walking
From what location did you travel directly to the shopping center? (Please circle)
1 Home 3 Another shopping center (Please name)
2 Work 4 Other (Please specify)
WAT TIME DID YOU LEAVE THAT LOCATION? :
WHAT TIME DID YOU ARRIVE AT TIE SHOPPING CENTER? :
HOW MUCH DO YOU THINK YOUR TRIP TO THE SHOPPING CENTER COST YOU?
To WAT LOCATION WILL YOU GO TO AFTER LEAVING THE SHOPPING CENTER? (Please circle.)
1 Home 3 Another shopping center (Please name)
2 Work 4 Other (Please specify)
WHAT IS YOUR HCME ADDRESS?
Block Number
Street
City, town, or village
Zip Code
10
16. Master 17. Doctorate 18. Other
Occupation:
1 Salesman
2 Teacher/Professor
3 Professional/Technical/Managerial
11 12 13. College studies, no degree
please specify
4 Craftsman/Mechanic/Factory Worker
5 Clerical/Secretarial/Office Worker
6 Student
7 Housewife
8 Governmental
9 RetT red
10 Other
23
. 1 1
27
28
1
32
1 1
33
36
1 .
in order to help with the classification of results of this survey we would be very grateful if you
would provide the information in this last section. we appreciate that you may feel hesitant about reveal¬
ing some of this information, which is of a personal nature, but these factors are important to this study.
we will use the information only for research purposes and it will not be divulged to anyone,
in each case please circle the number preceding the appropriate entry.
SEX: 1 Female 2 Male
Age:
1 Under 22 years 3 30-39 years 5 50-59 years 7 Over 65 years
2 22 - 29 years 4 40-49 years 6 60-65 years
Last year of school completed.
1 2 3 4 5 6 7 8
14. AA/AS 15. BA/BS
please specify
12 13 14 15 16 17 18 19 20 21
Are you presently married?
1 Yes 2 No
Please circle the AGES of all children living with you.
less than 1 1 2 3 4 5 6 7 8 9 10 11
22 and over
wat is your approximate family income before taxes? (tour income plus that of your husband or wife or the
person on wiom you are dependent.)
1 $10,000 and under 5 $25,001 - $30,000 8 $40,001 - $45,000
2 $10,001 - $15,000 6 $30,001 - $35,000 9 $45,001 - $50,000
3 $15,001 - $20,000 7 $35,001 - $40,000 10 Over $50,000
6 $20,001 - $25,000
How many years have YOU lived in this area, THAT is, NORTH-WEST CHICAGO ANCHOR THE FORTH AND NORTH-WEST
suburbs?
Not a resident Less than 1 year 123456789 10 or more years
thank you for your cooperation.
B-13
Office
Use Only
20
21 '22
39
37 38
I I I
42
43
TT
45 "46
4-7V
49
iii)
50
1 1
54
1 1
55
1 1
59
1 1
60
1 1
64
1 1
65
1
69
1
70
72
ÎÎ^Ti
4
80
APPENDIX C
ADDITIONAL TABULATIONS AND CROSS-TABULATIONS
FOR THE FIRST OLD ORCHARD
SURVEY
SHOPPING CENTER
VISITED
SEX
FEMALE
MALE
PLAZA DEL LAGO
90.4%
9.6%
EDENS PLAZA
78.7
21.3
OLD ORCHARD
82.1
17.9
GOLF MILL
75.5
24.5
TABLE B-l:
Sex by Shopping Center
SHOPPING CENTER
AGE GROUP
60 AND
VISITED
UNDER 16
16-21
22
29
30 - 39
40 - 49
50 - 59
OVER
PLAZA DEL LAGO
3.2%
14.4
22
9
22.9
14.1
14.1
8.5
EDENS PLAZA
4.0
19.3
19
8
18.0
18.7
14.2
6.1
OLD ORCHARD
3.8
22.5
22
8
15.3
16.0
12.4
7.2
GOLF MILL
7.3
23.6
24
3
16.4
15.1
10.3
3.0
TABLE B-2: Distribution by Age Groups by Shopping Center
Visited
C-l
c~>
I
PO
SHOPPING CENTER
VISITED
OCCUPATION
Mil i-
tary
Sales
Teacher
Profes¬
sional
Crafts
Cleri¬
cal
Student
House¬
wife
Govern¬
mental
Retired
Other
PLAZA DEL LAGO
0.7
3.5
12.0
6.7
0.4
4.2
18.7
43.3
0.7
3.5
6.3
EDENS PLAZA
0
7.2
8.1
19.0
1.3
9.2
21.9
28.0
0.7
0.9
3.8
OLD ORCHARD
0.1
4.4
10.4
16.4
0.8
10.5
23.9
26,2
0.8
2.6
3.9
GOLF MILL
0.2
3.7
6.8
17.8
2.7
12.8
25.0
25.2
0.8
1.2
3.6
TABLE B-3: Distribution of Occupations by Shopping Center Visited
SHOPPING CENTER
INCOME GROUP
VISITED
$10,000
AND LESS
$10,000-
$14,999
$15,000-
$19,999
$20,000-
$24,999
$25,000-
$50,000
OVER
$50,000
PLAZA DEL LAGO
9.4
17.6
10.5
17.6
29.7
15.2
EDENS PLAZA
11.8
13.8
15.7
17.4
29.7
11.6
OLD ORCHARD
12.7
16.5
17.1
16.8
26.7
10.2
GOLF MILL
14.4
21.5
23.9
20.2
16.4
3.7
TABLE B-4: Distribution of Income Groups by Shopping Center Visited
SHOPPING CENTER
VISITED
LENGTH OF RESIDENCE
LESS THAN
1 YEAR
1 - 3
YEARS
4-6
YEARS
7 - 10
YEARS
MORE THAN
10 YEARS
PLAZA DEL LAGO
3.3
10.4
13.8
10.4
62.1
EDENS PLAZA
3.0
9.7
9.0
13.3
65.1
OLD ORCHARD
3.1
9.1
9.5
10.6
67.7
GOLF MILL
3.1
11.4
9.8
11.9
63.8
TABLE B-5: Distribution of Length of Residence by
Shopping Center Visited
C-3
OCCUPATION
NUMBER
(PERCENT)
MILITARY
11 ( 0.2)
SALESMAN
305 ( 4.3)
TEACHER
641 ( 9.0)
PROFESSIONAL
1182 (16.7)
CRAFTSMAN
108 ( 1.5)
CLERICAL
777 (11.0)
STUDENT
1700 (24.0)
HOUSEWIFE
1889 (26.7)
GOVERNMENTAL
55 ( 0.8)
RETIRED
146 ( 2.1)
OTHER
274 ( 3.9)
TABLE B-6: Distribution of Occupations
for the Sample
C-4
APPENDIX D
DERIVATION OF PSYCHOLOGICAL DISTANCES FOR MULTIDIMENSIONAL SCALING
As mentioned 1n the body of the report, there are two sources of
Information for determining psychological distances between stimuli.
Figure F-l of the report shows a d1rect-s1m1lar1ty question, while Figure
F-2 shows an indirect-similarity question. These constitute the two sources
of information for psychological-distance estimation. The principal for
deriving the distances is rather simple in concept, being based on simple
notions of dominance and implied dominance.
Simply stated, the derivation of psychological distances from direct-
similarity measurements is based upon a simple count of the number of times
one stimulus is considered dissimilar from other stimuli. The larger the
number of times that a stimulus is rated as dissimilar to other stimuli, the
greater is its distance from any other stimulus. This is, however, almost
an oversimplification. The goal of the dissimilarity analysis is to determine
the distances between all pairs of points, i.e., to determine all interpoint
distances. To do this, it becomes necessary not just to examine dominance
of one stimulus over others, but rather the dominance of pairs of stimuli
over other pairs.
To illustrate the method by which these distances are computed, it is
necessary to assume that data have been obtained in which pairs of stimuli
have been rated in similarity to other pairs. It is important to note here
that measurement techniques will normally mandate that only those pairs with a
common stimulus can be compared. Thus, a respondent may be asked to say
which of the following two pairs of shopping centers are more similar in terms
of his or her preference to shop for a particular commodity:
Old Orchard and Golf Mill or Old Orchard and Woodfield.
D-l
This question is answerable, since it simplifies to a question of whether Golf
Mill is more similar to Old Orchard than is Woodfield. On the other hand, a
respondent should never be asked to compare the similarity of two pairs, such
as :
Old Orchard and Golf Mill or Woodfield and the Loop.
Although such a comparison obviously exists, it is nearly impossible to make
the judgment. It could be likened to a request to say whether an orange and
a plum are more similar to each other than an apple and a banana. This is
clearly a most difficult, if not impossible, judgment to make.
Assuming that there were 5 objects being judged, a matrix may be con¬
structed of the comparisons of each pair with other pairs. In the matrix, a
*
zero is entered if the row pair was considered to be more similar than the
column pair- Otherwise, a one is entered. A zero is also entered for all
cells where no comparison was made. The resulting matrix from a hypothetical
response, is shown in Figure D-l.
Next, the row sums are obtained, as shown in Figure D-l, and the
matrix is permuted in order of decreasing row sums, as shown in Figure D-2.
Where rows have the same sums, ordering is arbitrary. In the permuted
matrix, intransitivities are identifiable as ones below the diagonal.
However, some care is necessary here. Where such ones appear close to the
diagonal for tied rows, they may be removed by reordering the tied rows.
Thus, interchanging row (2,4) and row (4,5) in Figure D-2 will move one of
the ones back above the diagonal. Attending to this yields a new matrix,
Figure D-3. There now remains only one intransitivity, rather than the
D-2
(1,2)
(1,3)
(1,4)
(1,5)
(2,3)
(2,4)
(2,5)
(3,4)
(3,5)
(4,5)
(1,2)
0
0
1
0
0
0
0
0
0
(1,3)
1
0
1
0
0
0
0
0
0
(1,4)
1
1
1
0
0
0
1
0
0
(1,5)
0
0
0
0
0
1
0
0
0
(2,3)
1
1
0
0
0
0
0
0
0
(2,4)
1
1
1
0
1
1
1
0
0
(2,5)
1
1
0
0
1
0
0
0
0
(3,4)
0
1
0
0
1
0
0
0
0
(3,5)
0
1
0
1
1
0
1
1
0
(4,5)
0
0
1
1
0
1
1
1
1
Figure D-l
Matrix of Dissimilarities
D-3
(2,4)
(4,5)
(3,5)
(1,4)
(2,5)
(1,3)
(2,3)
(3,4)
(1,2)
(1,5)
(2,4)
0
0
1
1
1
1
1
1
0
(4,5)
1
1
1
1
0
0
1
0
1
(3,5)
0
0
0
1
1
1
1
0
1
(1,4)
0
0
0
0
1
0
1
1
1
(2,5)
0
0
0
0
1
1
0
1
0
(1,3)
0
0
0
0
0
0
0
1
1
(2,3)
0
0
0
0
0
1
0
1
0
(3,4)
0
0
0
0
0
1
1
0
0
(1,2)
0
0
0
0
0
0
0
0
1
(1,5)
0
0
0
0
1
0
0
0
0
Figure D-2
Permuted Dissimilarities Matrix
D-4
(4,5)
(2,4)
(3,5)
(1 ,4)
(2,5)
(3,4)
(2,3)
(1,3)
(1,2)
(1,5)
(4,5)
1
1
1
1
1
0
0
0
1
(2,4)
0
0
1
1
1
1
1
1
0
(3,5)
0
0
0
1
1
1
1
0
1
(1,4)
0
0
0
0
1
0
1
1
1
(2,5)
0
0
0
0
0
1
1
1
0
(3,4)
0
0
0
0
0
1
1
0
0
(2,3)
0
0
0
0
0
0
1
1
0
(1,3)
0
0
0
0
0
0
0
1
1
(1,2)
0
0
0
0
0
0
0
0
1
(1,5)
0
0
0
0
1
0
0
0
0
Figure D-3
Adjusted Permuted Dissimilarities Matrix
D-5
apparent five of the preceding matrix. A number of strategies are possible
with intransitivites. If there are few, they may be left alone, since their
presence provides potential information and will not cause undue problems in
the subsequent analysis steps. If many intransitivities are present, the
individual may be dropped from-futher analysis, or all below-diagonal ones
can be changed, arbitrarily to zeros (thus removing information on that
comparison), or the analyst may assume that a genuine mistake was made and
interchange the "incorrect" cells. (In this case, that would involve setting
row (1,5), column (2,5) to zero and row (2,5), column (1,5) to a one.) For
this example, the intransitivity will be left alone.
If it is desired, the matrix of Figure D-3 can be powered to fill in some
of the zeroes that result from no comparisons being made. This is equivalent
to the use of a round-robin tournament for inferring results of dominance.
Thus, a second-order matrix is made up of elements, c.-, which are the sum of
' J
the cross multiples of the row (d^) and column (d^) in which the element lies.
The powering of the matrix of Figure D-3 to the second order is shown in
Figure D-4. Because an intransitivity was retained in the matrix of Figure
D-3, some new intransitivities are generated in Figure D-4. These should be
ignored, since they contribute further intransitivities that cannot be resolved
by adjusting the permuted matrix. Apart from these values, a one is now entered
in the matrix of Figure D-3 wherever a non-zero value occurs in the powered
matrix, as shown in Figure D-5. Further powering may be undertaken until no
new information can be obtained. For this exposition, no further powering is
undertaken, although the remaining above-diagonal zeros could possibly be
removed by doing so.
D-6
(4,5)
(2,4)
(3,5)
(1.4)
(2,5)
(3,4)
(2,3)
(1,3)
(1,2)
(1,5.)
(4,5)
0
0
1
3
3
4
5
3
2
(2,4)
0 '
0
0
0
0
2
4
4
3
(3,5)
0
0
0
1
0
2
3
3
1
(1,4)
0
0
0
1
0
1
1
1
2
(2,5)
0
0
0
0
0
0
1
2
2
(3,4)
0
0
0 '
0
0
0
1
2
1
(2,3)
0
0
. 0
0
0
0
\
0
1
2
(1,3)
0
0
0
0
1
0
0
0
1
(1,2)
0
0
0
0
1
0
0
0
0
(1,5)
0
0
0
0
0
0
1
1
1
Figure d-4
Second-Order Matrix From Figure D-3
0-7
(4,5)
(2,4)
(3,5)
(1,4)
(2,5)
(3,4)
(2,3)
(1,3)
(1,2)
(1,5)
(4,5)
1
1
1
1
1
1
1
1
1
(2,4)
0
0
1
1
1
1
1
1
1
(3,5)
0
0
0
1
1
1
1
1
1
(1,4)
0
0
0
0
1
1
1
1
1
(2,5)
0
0
0
0
0
1
1
1
1
(3,4)
0
0
0
0
0
1
1
1
1
(2,3)
0
0
0
0
0
0
X
1
1
1
(1,3)
0
0
0
0
0
0
0
1
1
(1,2)
0
0
0
0
0
0
0
0
1
(1,5)
0
0
0
0
0
0
0
0
0
Figure D-5
Augmented Permuted Dissimilarities Matrix
9
7
6
5
4
4
3
2
1
0
D-8
It may also be noted that the powering has placed a one in row (2,5) and
column (1,5). As a result, it may now be assumed that this intransitivity was
a mistake in the judgments and the one below the diagonal may be changed to
a zero, as has been done in Figure D-5.
The row sums are next recomputed and the matrix is permuted again, if
necessary. In this case, no further permuting is needed. Rank scores are
next assigned, beginning with 1 for the row with a zero sum. If any rows are
ties, they are assigned a value equal to the combined raw ranks divided by the
number of ties. Thus, the two rows that are tied in Figure D-5 will be assign¬
ed a score of 5.5 each. These scores are now the estimated distances between
the stimuli, as shown in Figure D-6. This completes the derivation of distances.
Using indirect similarities data, the approach is basically very similar.
In this case, it is assumed that individuals have rated the stimuli on a set
of attribute scales, where the attributes are assumed to specify completely
the perceptual space to be measured. In general, a Euclidean model is then
applied to determine the set of distances between the stimuli. Thus, the dis¬
tance between stimulus j and stimulus k, d^, is given by equation (D.,1).
where a^, a^ = ratings of stimuli j and k, respectively, on attribute m.
A set of possible ratings of the same 5 stimuli on 6 attributes are shown
in Figure D-7. These values are used in equation D.l to compute the derived
distances as shown in Figure D-8. Thus, the interpoint distances are computed
in a single step from the scale ratings. It may be noted that the distances in
D-9
STIMULI
1
2
3
4
5
1
0
2.0
3.0
7.0
1.0
2
2.0
0
4.0
10.0
5.5
3
3.0
4.0
0
5.5
8.0
4
7.0
9.0
5.5
0
9.0
5
1.0
5.5
8.0
10.0
0
Figure D-6
Matrix of Direct Interpoint Distances
SHOPPING
CENTER
ATTR
[BUTE
1
2
3
4
5
6
1
2
1
3
2
1
2
2
3
4
2
4
2
3
3
1
2
1
2
1
2
4
1
3
1
3
7
6
5
5
2
5
2
1
4
Figure p-7
Ratings of Shopping Centers on 7-point Scales
of 6 Attributes
D-10
figure D-8 arè not monotonically related to those in Figure D-6. This would
suggest that erroneous judgments have been made or that the set of 6 attri¬
butes is insufficient to specify the perceptual space derived from the similarity
measures; However, such conslusions cannot be confirmed until the full MDS
analysis is completed.
0-11
STIMULI
1
2
3
4
5
1
0
4.1
2.5
7.9
4.2
2
4.1
0
3.9
6.4
4.8
3
2.5
3.9
0
7.4
6.0
4
7.9
6.4
7.4
0
8.6
5
4.2
4.8
6.0
8.6
0
Figure D-8
Matrix of Derived Interpoint Distances
D-12
APPENDIX E
CLASSIFICATION WITH RESPECT TO DIFFERENCES IN
PERCEPTION OF DISSIMILARITIES
An Important element of data collected in the shopping destination
choice study is the perceived dissimilarity between pairs of shopping centers.
This information is used to determine the multidimensional space which
describes the perception of shopping center attractiveness. The methods used
to determine the space assume some degree of commonality of perception with
respect to the attractiveness space.
The project staff has hypothesized that different population groups,
identified by their socio-economic characteristics, perceive attractiveness
differently. The purpose of this note is to propose a procedure to test this
hypothesis in addition to the within group hypothesis (that there is common¬
ality of agreement within an identifiable group) and the total population
hypothesis (that there is commonality of agreement across the entire popula¬
tion).
The proposed procedure is based on an extension of Friedman's two way
analysis of variance by ranks (Friedman, 1937). Friedman's test is briefly
described below. The application to the test of the population hypothesis and
the within group hypothesis is shown. The extension to the test of differ¬
ences between groups is described next. A statistic is proposed to undertake
statistical testing of the between group hypothesis.
Friedman's Test
The data consists of B sets of J stimuli (dissimilarity perceptions)
ranked in order of increasing dissimilarity from 1 to J. The ranking of
the j stimulus by the b individual is designated R^. Friedman's statistic
is:
E-l
T - b-IU\) 5 ^(Rjb -R»2
* B*J(J+l) 5 [B(Rj " R)]2
J
(1)
where R = -y-
That is, T measures the degree of agreement in ranking by the dispersion
of average rank for stimuli across individuals from the average ranking
across all dissimilarities.
The null hypothesis is that there is no consistent agreement in rankings
among individuals in the population. That is:
Under this hypothesis, Friedman shows that for large values of B, T is dis¬
tributed approximately as a chi-square random variable with J-l degrees
of freedom. The value of T can be compared to the chi-square distribution.
Larger values of T tend to reject the null hypothesis and indicate that there
is agreement in ranking structure in the population.
Friedman's test may be applied to groups within the population by com¬
putation of a group specific value for T, that is:
. = R. = R
J
(2)
(3)
E-2
where the subscript n indicates values for the n group.
In this case the null hypothesis is:
H0: Rln = R2n = • ' • RJn = R ^
Joint Group Test
The proposed extension is to first create a new statistic which can be
used to jointly test within group ranking structure for N groups. The pro¬
posed statistic for this purpose is simply the sum of group statistics taken
over all groups:
£Vônfrr £ <En£ (5)
IJ J
Since each Tn is a chi-square random variable with J-l degrees of freedom,
ET is distributed as a chi-square random variable with N(J-l) degrees of
n
freedom. In this case the null hypothesis is the joint hypothesis:
V «In = «2n »n (6>
Large values of ETn indicate that there is a systematic ranking structure
in some or all of the subgroups. However, the value of ETn can not be used
to determine if there are different ranking structures among groups. Large
values of ETn may be obtained when there is rank agreement in the total
population across all groups or when there is rank agreement solely within
groups (but rank differences between groups).
E-3
Disaggregation of ETn
This can be seen by decomposition of ETn into its component parts:
£T = tttttt £ b 2 (r- " R)2
n J(J+1) n . jn
i J
Jfilrrz Bn ^ [(Rjn - Rj>2 - (Ro -R'2]
' 1 Û
TOÎTT „E Bn J <Rjn - R/
+ J(0+1) I Bn f i " R)2
JT3TTT e b„ e (Rj„ " R>2 + TUTTT E (Rj " R)2
J 1 J J
12 E B E (R. - R.)2 + T. (7)
J(J+1) _ n , jn j
n J
That is ETn is equal to T plus a term which measures the departure of
average within group rankings, R\ , from average population rankings, R\.
J
Thus:
ETn > T (8)
The equality holds when agreement in rank structures is identical among
groups. The inequality holds otherwise. Thus, the difference, ETn - T-j ,
is an appropriate measure of difference in rank structure between groups.
Difference Test
From equation 7,
ET - T = (\2y, E B E (R. - R.)2,
n J(J+1) n n j jn j
E-4
When B 1s large 1s distributed normal with mean 1/2(J+1) and variance
(J+l)(J-1)/l2 B. When Bn 1s large R^n is distributed normal with mean
1/2(J+1) and variance (J+l)(J-1)/12 B . The covarlance between Ff. and R.
n J jn
1s (J+l)(J-l)/12 B. Thus, the difference, Rjn - Rj is normally distributed
with mean zero and variance [(J+l )(J-l )/12](^- - h. Rewriting equation 9
n
we obtain:
ZTn ' T = V" I (~BJ1) Z. C(J+l)(J-l) , 1 1 1 > (*jn'*j)2]
n J ^Bn " B '
where the term in braces 1s a squared unit normal variate. Adjusting equation
10 for structural dependencies we can show that ET - T is a chi square
variable with degrees of freedom (N-l)(J-l). The ratios (J-l)/J and (B-Bn)/B
take account of the dependencies in the squared unit normal variables. Thus,
the difference measure which indicates the degree of difference in ranking
among groups has statistical properties and can be used to test the null
hypothesis.
Hq: Rjn = R", V j,n (11)
Conclusion
The Index ETn - T defined in equation 9 provides an intuitive and
statistical measure for the existence of differences in the ranks between
a priori designated groups. The ease of computation of this index suggests
its potential usefulness as a tool for screening market segments subject to
tests which confirm its relationship to differences identified 1n model
estimation.
E-5
REQUEST FOR FEEDBACK TO
The DOT Program Of University Research
DOT/RSPA/DPB-50/79/39 - "A METHOD FOR UNDERSTANDING AND PREDICTING
DESTINATION CHOICES" - D0T-0S-40001
YES NO
CD CD Did you find the report useful for your particular needs?
If so, how?
CD CD Did you find the research to be of high quality?
D □ Were the results of the research communicated effectively
by this report?
CD CD Do you think this report will be valuable to workers in the
field of transportation represented by the subject area of
the research?
CD CD Are there one or more areas of the report which need
strengthening? Which areas?
□ CD Would you be interested in receiving further reports in this
area of research? If so, fill out form on other side.
Please furnish in the space below any comments you may have concerning the
report. We are particularly interested in further elaboration of the above
questions.
COMMENTS
Thank you for your cooperation. No postage necessary if mailed in the U.S.A.
FOLD ON TWO LINES, STAPLE AND MAIL.
RESEARCH FEEDBACK
Your comments, please ...
This booklet was published by the DOT Program of University Research and
is intended to serve as a reference source for transportation analysts, planners,
and operators. Your comments on the other side of this form will be reviewed
by the persons responsible for writing and publishing this material. Feedback
is extremely important in improving the quality of research results, the transfer
of research information, and the communication link between the researcher
and the user.
FOLD ON TWO LINES, STAPLE AND MAIL.
Fold
Fold
DEPARTMENT OF TRANSPORTATION
RESEARCH AND SPECIAL PROGRAMS AdfTll 111 Strati Of!
WASHINGTON D.C. 20590
PORTAOE AND PEER PAID
DEPARTMENT OF
TRANSPORTATION
Official Business
PENALTY FOR PRIVATE USE, »S00
DOT 513
OFFICE OF UNIVERSITY RESEARCH (DPB-50)
Research and Special Programs Admini Stration
U.S. Department of Transportation
400 Seventh Street, S.W.
Washington, D.C. 20590
Fold REQUEST FOR INFORMATION FROM THE UNIVERSITY RESEARCH PROGRAM Fold
□Check here if you would like to be placed on the mail Hit for the
University Research Program Solicitation Booklet (DT-68C)
IF YOU WISH TO BE ADDED TO THE MAIL LIST FOR FUTURE
REPORTS, PLEASE FILL OUT THIS FORM.
Name Title.
Use Block Letters or Type
Department/Office/Room
Organization
Street Address
City State Zip .
U.S. DEPARTMENT OF TRANSPORTATION
RESEARCH AND SPECIAL PROGRAMS ADMINISTRATION
WASHINGTON, D.C. 20590
OFFICIAL BUSINESS
PENALTY FOR PRIVATE USE, $300
POSTAGE AND FEES PAID
RESEARCH AND SPECIAL PROGRAMS
ADMINISTRATION
DOT 513