PLACE IN RETURN BOX to remove this chedtout from your record.
To AVOID FINE return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

1M clam“

HOW PEOPLE EVALUATE BUDGET PRIORITIES:
A CONJOINT ANALYSIS APPROACH

By

Chia—Nan Yeh

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Political Science

1998

ABSTRACT

HOW PEOPLE EVALUATE BUDGET PRIORITIES:
A CONJOINT ANALYSIS APPROACH

By

Chia-Nan Yeh

The purpose of this study is to examine how people evaluate budget priorities by
applying a method called conjoint analysis. Specifically, we evaluate how people make
tradeoffs among four spending areas: education, welfare and health, crime and public
safety, and natural resources and environmental control while holding other programs and
tax rate constant. Traditional survey questions usually are addressed in a “wish list” type
of format with unrealistic choices among incomplete sets of budget options, thus they
actually measure people’s “desires” rather than their “true preferences.” In contrast,
conjoint analysis asks respondents to make more “real world” decisions by forcing them
to trade off competing interests and hence could offer researchers a way to collect more
accurate information about voters’ true preferences.

Conjoint analysis is a tradeoffs exercise in which respondents are forced to make
preference decisions based on the relative importance of one service or policy over
another. Conjoint analysis is defined as a method which identifies the joint effect of two
or more independent variables on the ordering of a dependent variable.

Our findings reveal that respondents’ policy preferences from traditional surveys
and conjoint estimates are quite different; sometimes respondents’ preference orders for a
particular area even go in completely opposite direction. In the area of overall spending,

for instance, traditional surveys show that voters most prefer a decrease in overall

spending but conjoint estimates indicate that the status quo is virtually voters’ first
choice. The implication from our findings is that if policymakers make policy decisions
based on inaccurate and incomplete information from traditional surveys, these policies
cannot reflect peoples’ true preferences and hence the quality of democracy might be
discounted. Furthermore, in order to illustrate the applicability of conjoint analysis,
simulations have been performed to predict the gubernatorial election and evaluate
current welfare reform in Michigan. Lastly, reliability tests demonstrate that our conjoint

estimates yield answers at least as reliable with previous studies on conjoint reliability.

To the memory of my brother, Chia—Lin

ACKNOWLEDGMENTS

I would like to thank those who made intellectual, emotional, and financial
contributions to me. First, I wish to express my gratitude to my father, Lien-Fa Yeh, my
mother, Pao-Lien Lu, my sister, Chia-Mei Yeh, and my parents in—law Ching-Tai Chen
and Chiu—Fang Chao for their financial and emotional support.

I thanks to the members of my committee - Larry Heimann, Thomas Hammond,
Paul Abramson, Paula Kearns, and Brian Silver. I sincerely appreciate Larry Heimann,
who wore the hats of committee chair, employer, and friend. This dissertation would not
been completed without his advice, assistance, and encouragement. I would also like to
thank Thomas Hammond, who became my chair in the final stage, for his patience and
guidance to correct my lousy writing and providing me insightful suggestions. In
addition, I want to thank Roger Calantone of the business school for his suggestions on
designing conjoint plan. Moreover, I gratefully acknowledge the assistance of Seon—Jou
Kang, Brandon Prins, John Davis and Mark Hurwitz for helping me collect student data.
Among my Taiwanese friends, I want to express my gratitude to Lu-Huei Chen, Jong-
Tian Wang, Ching-Lung Chang, Min-Chieh Lin, Chien-Chin Chen for their
encouragement and insightful suggestions.

Lastly, special thanks go to my wife, Nancy Chen, who shared everything with
me during the process of writing my dissertation and now have to prepare for delivering
our first baby. I appreciate your support more than I can say. I think my graduation will

be the best gift for our first baby in this coming New Year.

TABLE OF CONTENTS

LIST OF TABLES

LIST OF FIGURES

INTRODUCTION

The Importance of Studying Budget Priorities
A New Survey Tool: Conjoint Analysis
Objectives of This Study

Organization of This Study

LITERATURE REVIEW

Public Opinion and Policy
Public Attitudes Toward Government Spending

Explanations of the Inconsistencies Between Policymakers and the Public
Toward Government Spending

Existing Approaches to Evaluate Budget Priorities

Summary

THE LOGIC OF CONJOINT ANALYSIS

The Concept of Conjoint Analysis
Compositional vs. Decompositional Approaches

The Theory of Conjoint Analysis:
An Information Integration Theory Approach

Design and Analysis of Conjoint Analysis
Why Conjoint Analysis Works

The Advantages and Limits of Conjoint Analysis

vi

ix

xi

11

12

14

24

26

26

28

3O

35

49

52

Classification of Conjoint Analysis
Current Applications of Conjoint analysis
Summary
RESEARCH DESIGN
Step by Step Procedure of Conducting a Conjoint Analysis
Summary
RESULTS AND DISCUSSIONS
Interpreting Conjoint Results from Voter Sample
Interpreting Conjoint Results from Student Sample
Comparison Between Traditional Survey Questions and Conjoint Results
Summary
GROUP DIFFERENCES IN CONJOINT ANALYSIS
Gender Differences
Generational Differences
Religious Differences
Party Differences
Children
Home Ownership
Summary

CONJOINT APPLICATIONS: SIMULATION OF GUBERNATORIAL
ELECTION AND POLICY EVALUATION

Simulation of Gubernatorial Election in Michigan

Welfare Policy Evaluation

vii

54

56

56

58

58

74

76

76

90

98

103

105

106

112

118

123

128

132

136

137

I37

146

Summary
RELIABILITY AND VALIDITY TESTS
Why Reliability and Validity Check are so Important
Definitions of Reliability and Validity
Methods and Designs
Results and Discussions
Summary
SUMMARY AND CONCLUSIONS
The Need of New Survey Tool
Empirical Findings by Using Conjoint Analysis
Potential Applicability of the Conjoint Analysis
APPENDIX: CONJOINT SURVEY FORMS FOR STUDENT AND VOTER

BIBLIOGRAPHY

viii

149

150

150

153

157

159

173

175

175

177

181

183

193

Table 3.1:

Table 3.2:

Table 3.3:

Table 3.4:

Table 4.1:

Table 5.1:

Table 5.2:

Table 5.3:

Table 5.4:

Table 5.5:

Table 5.6:

Table 5.7:

Table 5.8:

Table 5.9 :

LIST OF TABLES

The Relationship among a Profile, Attributes, and Levels
The 2x2x2 Design and the Two One-Half Fractions

A Graeco-Latin Square for the 2x3 Design

A Full Profile with 2x3 Design in LINMAP Estimation
Alternative Data Collection Methods

Orthogonal Array for Voters (First Round of Survey)
Part-worth Utilities from OLS and LINMAP (Voters)
Overall Utilities of 10 Profiles for Voters (OLS)

Goodness of Fit Measures for OLS and LINMAP
Part-worth Utilities from OLS and LINMAP (Students)
Overall Utilities of 10 Profiles for Students (OLS)
Comparison of Overall Utilities Between Voters and Students
Goodness of Fit Measures for OLS and LINMAP (Students)

Comparison of Budget Priorities Between Traditional Surveys and
Conjoint Analysis (Voters)

Table 5.10 : Comparison of Budget Priorities Between Traditional Surveys and

Table 6.1:

Table 6.2:

Table 6.3:

Table 6.4:

Conjoint Analysis (Student)
Gender Differences in 4 Major Policy Areas
Generational Differences in 4 Major Policy Categories
Religious Differences in 4 Major Policy Categories

Party Differences in 4 Major Policy Categories

ix

27

43

45

47

65

78

80

88

90

91

95

96

98

100

103

109

115

120

125

Table 6.5:

Table 6.6:

Table 7.1:

Table 7.2:

Table 7.3:

Table 7.4:

Table 8.1:

Table 8.2:

Table 8.3:

Table 8.4:

Table 8.5:

Table 8.6:

Children Differences in 4 Major Policy Categories

Differences Between Home Owners and Renters in 4 Major Policy
Categories

Simulation Results for Gubernatorial Election (All Voters)
Simulation Results for Gubernatorial Election (Independent Voters)
Simulation Results for Welfare Policy Evaluation

Simulation Results for Maximizing Voters’ Utilities

Reliability of Conjoint Estimates for Ranking Data (Voters)
Reliability of Conjoint Estimates for Rating Data (Voters)

Reliability of Conjoint Estimates for Ranking Data (Students)
Reliability of Traditional Survey Questions for Ranking Data (Voters)
Reliability of Traditional Survey Questions for Rating Data (Voters)

Reliability of Traditional Survey Questions for Ranking Data (Students)

129

133

143

145

147

148

161

163

164

169

171

173

Figure 3.1:
Figure 3.2:
Figure 5.1:
Figure 5.2:
Figure 5.3:
Figure 5.4:
Figure 5.5:
Figure 5.6:
Figure 5.7:
Figure 5.8:
Figure 6.1:
Figure 6.2:
Figure 6.3:
Figure 6.4:
Figure 6.5:
Figure 6.6:
Figure 6.7:
Figure 6.8:

Figure 6.9:

LIST OF FIGURES

Preference Models

Classification of Conjoint Analysis

Part-Worth Utilities for Voters (OLS)

Part-Worth Utilities for Voters (LINMAP)

Relative Importance of Attributes for Voters (OLS)
Relative Importance of Attributes for Voters (LINMAP)
Part-Worth Utilities for Students (OLS)

Part-Worth Utilities for Students (LINMAP)

Relative Importance of Attributes for Students (OLS)
Relative Importance of Attributes for Students (LIN MAP)
Gender Differences in Education Spending

Gender Differences in Welfare Spending

Gender Differences in Crime Spending

Gender Differences in Natural Resources Spending
Generational Differences in Education Spending
Generational Differences in Welfare Spending
Generational Differences in Crime Spending
Generational Differences in Natural Resources Spending

Religious Differences in Education Spending

Figure 6.10: Religious Differences in Welfare Spending

Figure 6.11: Religious Differences in Crime Spending

xi

41

55

81

81

85

85

92

92

93

94

110

110

111

111

116

116

117

117

121

121

122

Figure 6.12:
Figure 6.13:
Figure 6.14:
Figure 6.15:
Figure 6.16:
Figure 6.17:
Figure 6.18:
Figure 6.19:
Figure 6.20:
Figure 6.21:
Figure 6.22:
Figure 6.23:

Figure 6.24:

Religious Differences in Natural Resources Spending
Party Differences in Education Spending

Party Differences in Welfare Spending

Party Differences in Crime Spending

Party Differences in Natural Resources Spending
Children Differences in Education Spending
Children Differences in Welfare Spending

Children Differences in Crime Spending

Children Differences in Natural Resources Spending
Home Differences in Education Spending

Home Differences in Welfare Spending

Home Differences in Crime Spending

Home Differences in Natural Resources Spending

xii

122

126

126

127

127

130

130

131

131

134

134

135

135

CHAPTER 1

INTRODUCTION
The Importance of Studying Budget Priorities

The purpose of this study is to examine how people evaluate budget priorities by
applying a technique called conjoint analysis. This topic is important for several reasons.
First, the budget is definitely a significant policy issue. Since budgeting allocates scarce
resources among rival groups, it appears regularly on the agenda of United States
electoral politics, and candidates for public office routinely raise this concern in their
campaign rhetoric. Indeed, one of the most distinctive features of government spending
is its long-standing presence as an issue in US. politics. Other social programs and
political controversies come and go over time, nevertheless, disagreements over the
proper level of governmental spending for all programs remain fixed on the public
agenda.

Second, one of the major characteristics of government budgeting is that decisions
about how money will be spent are decided not by taxpayers but by their representatives.
The distinction between payers and the deciders leads to two important features of public
budgeting: public accountability and political acceptability. Accountability means to
guarantee that all the money is spent as approved, and to report accurately to the public
on how money was spent. Acceptability means that bureaucrats who make budget
decisions are constrained by what the public wants (Rubin 1993: 133-34). However, a
puzzle arises when we try to figure out what the public wants. Do policymakers and the

public always have the same preferences for government spending? Unfortunately,

evidence from survey research has shown that there appear to be significant
inconsistencies between politicians and public opinion with respect to budget priorities.
During the past decade, most political candidates and incumbent officeholders have
requested reductions in public spending. Even so, public opinion on this issue seems to
head in the opposite direction: while evidence from public opinion research has shown
that the public also support cuts in overall spending for both federal and state
governments (Sears and Citrin 1982; Peterson 1985; Modigliani and Modigliani 1987),
when asked about cutting any particular program (especially social programs), most of
the citizens oppose such actions (Ladd et a1. 1979; Citrin 1979; Welch 1985).

Third, we know very little about the topic of public preferences toward
government spending. While there have been some previous studies on this topic, the
results and conclusions have varied widely. For instance, one explanation of the
preference inconsistencies between politicians and the public is that people just do not
understand the complexities of government budgeting (Croker 1981). An alternative
explanation is that people want to eliminate the waste and inefficiency in government
rather than a specific program (Ladd et a1. 1979; Eismeier 1982). These reasons may
contribute to the preference inconsistencies between politicians and the public; yet we
believe that there is a deeper problem in the way we typically assess the voter’s policy
preference. Traditionally, researchers have depended on surveys for evaluating the
“desires” of their target populations. Typical questions on a survey would usually be
addressed in a “wish list” type of format with disaggregated questions and unbounded
financial constraints. For example, two typical questions to evaluate government

spending are: (1) “Government should provide many services, increase spending a lot” or

“Government should provide many fewer services, reduce spending a lot,”l and (2) Are
you in favor of increasing government spending on program, reducing it, or
keeping it about the same?”2 Facing such questions, respondents would usually want
more of, perhaps, every service. One problem of this format is that the respondents will
usually overstate their demand because they will not be charged for the increased level of
service provision, and expressions of preference for “more” of a service will be expected
to have only a small marginal impact upon themselves (Wilson 1983: 335-3 6).
Therefore, what traditional survey measure is the people’s “desire” rather than their “true
preference.”

Another problem of this format is that it provides respondents unrealistic choices
among incomplete set of options. That is, single dimensional questions addressed by
traditional surveys only ask people to evaluate the importance of a single attribute (factor
or dimension) in isolation from the multitude of attributes that necessarily consist of a
policy package or budget proposal. In the real decision process, however, the people’s
choice of a particular policy option is made on the basis of evaluating and combining
multiple attributes. Some policy areas, such as environmental protection, are difficult to
oppose when considered alone, but tend to lose support when compared with competing
interests such as lost employment. Thus, traditional surveys cannot tell how people make
tradeoffs among different attributes and hence may produce misleading information about

people’s preferences about government spending.

 

' NES survey question.
2 CBS survey questions in 1980.

Because in the political world politicians always make tradeoffs among
competing interests, it is extremely important to introduce a new survey tool which would
allow people to choose from more realistic policy options. The new survey tool should
define not only people’s “desires,” but more importantly, their actual “preferences.”
While the former can be addressed with a standard “wish list” type of question on the
survey, the latter can only be addressed with precisely designed tradeoff type questions.
For all these reasons, there is ample justification for a more detailed analysis of public
attitudes toward budget priorities by using a new survey tool.

A New Survey Tool: Conjoint Analysis

One way to improve the limits of traditional survey is through the use of a
technique called conjoint analysis. It provides respondents with more realistic choices
and hence collects more accurate information about people’s budget priorities. Currently,
conjoint analysis and related techniques represent the most widely applied methodologies
for measuring and analyzing consumer preferences (Carroll and Green 1995). However,
this important technique has received very little attention in the policy analysis and public
opinion literature.

Conjoint analysis is an exercise in making tradeoffs in which respondents are
forced to make preference decisions based on the relative importance of one service or
policy over another. Conjoint analysis is defined as a method which identifies the joint
effect of two or more independent variables (attributes) on the ordering of a dependent
variable (profile). A typical conjoint survey asks respondents to state their relative
preference of various alternatives (profiles) by rank ordering the alternatives from most to

least preferred. Researched products or services are sets of preferences called “profiles.”

Each profile is defined by a specific set of attributes (factors). The primary outputs of a
conjoint analysis are a set of intervally scaled “utilities” or “part-worths” for each level
of each defined attribute. The part-worth utilities from conjoint analysis are then used to
calculate the relative importance of attributes and overall utilities of profiles, and to
simulate all kinds of what-if scenarios.
Objectives of This Study

The objectives in this study are as follows: First, we intend to provide an
overview of conjoint analysis: the concept, theory, and design of conjoint analysis, the
procedure to conduct conjoint analysis, why conjoint analysis works, the advantages and
limits of conjoint analysis and classification of conjoint analysis and related techniques.

Second, to demonstrate the applicability of conjoint analysis, we will apply it to
the topic of the public preferences toward government spending. Specifically, a conjoint
study was conducted to evaluate people’s budget priorities in Michigan. 1n the early
19905, Michigan suffered a fiscal crisis due to economic slowdown, income tax change,
tax base erosion, and spending pressure (Gold 1995: 301-305). Even though the fiscal
crisis has largely been relieved recently, state government has had to fight spending
pressures because the constitution of Michigan requires a balanced budget. A survey of
public attitudes toward budget priorities in Michigan can serve two purposes: First, it
may be used as a type of performance indicator to measure the public satisfaction with
current services. Second, it can help policymakers to decide how to set the budget

priorities across different policy areas. In this study, I will examine samples of Michigan

voters3 and undergraduate students4 and assess their budget priorities for the state
government spending in four general categories: education, health and welfare, public
safety and corrections, and natural resources and environmental control, while holding
other programs and tax rates constant.5 The four categories explain 80.2 percent of total
state expenditures in the 1996 fiscal year. Meanwhile, the four categories also represent
the largest increase in spending items from 1990-96 fiscal years.

Third, group attitudes provide an underlying structure that organizes a broad range
of policy opinions and candidates evaluations. In this study, I will discuss spending
preference differences among different groups such as gender, age, parties, generations,
and religions, etc.

Fourth, one of major advantages of conjoint analysis is that it provides the
opportunity to conduct computer choice simulations to answer various “what-if”
questions. In this study, I will show how the public preferences fi'orn conjoint analysis
can be used in a simulation of the gubernatorial election and policy evaluation on welfare
reform.

In addition to examining how people evaluate budget priorities, my last objective
is to examine whether conjoint analysis can provide stable (reliabile) answers. Studies of
public opinion have shown that people tend to have unstable responses to survey
questions. If we ask the same respondents the same question in repeated interviews,
probably only about half give the same answers. Furthermore, the literature has also

shown that survey questions not only measure public opinion but also channel and shape

 

3 A random sample from the Meridian township voters (N=128).
4 A convenient sample from the undergraduate students of political science, Michigan State university (N=77).

it by framing issues, ordering alternatives, and setting the environment of the question
(Zaller 1992). Hence, we will evaluate whether conjoint analysis can reduce two major
problems in public opinion studies: response instability and response effects. Put another
way, we will examine whether conjoint analysis can provide stable and valid results.
Organization of This Study

The design for this study on budget priorities is as following: Chapter 2 reviews
public opinion toward government spending and the existing approaches and major
findings on the study of public attitudes toward government spending. Chapter 3
introduces the methodology of conjoint analysis and the procedure to conduct it. Chapter
4 discusses the research design: the sample, questionnaire, and procedures. Chapters 5
and 6 examine the results from conjoint and traditional survey questions and analyze the
preference differences among different groups. Chapter 7 demonstrates the applicability
of conjoint analysis by simulating the gubernatorial election and policy evaluation on
welfare reform. Chapter 8 checks the reliability of both conjoint and traditional survey
questions and validity from conjoint results. Chapter 9 summarizes this study and discuss

applicability of conjoint analysis for future research.

 

5 For comparison purpose, five NES alike questions will be asked before conjoint questions.

CHAPTER 2

LITERATURE REVIEW

Democratic accountability requires that the public be reasonably well-informed
about policy, and hence the public adjusts its preferences for “more” or “less” policy in
response to what policymakers actually do. Therefore, to examine how people evaluate
budget priorities, it is necessary to find out: Does public opinion really influence the
policymaking? Do policymakers and the public always have the same preferences for
government spending? If not, what cause the inconsistencies? Do current survey tools
provide satisfactory answers about public preferences toward government spending? In
this chapter, we review the relationship between public opinion and policy, public
attitudes toward government spending, explanations of the inconsistencies between
policymakers and the public toward government spending, and existing approaches to
evaluate government spending.

Public Opinion and Policy

That public policy should be controlled by the public is a major requirement of
democratic theory. In reality, we often evaluate the quality of democratic government by
the responsiveness of policymakers to the preferences of the public and by formal
opportunities for, and the practice of, citizen participation in political life. An important
question is then to ask to what extent public opinion affects public policy? Some
economics-style theories suggest that public opinion is a major influence on policy-
making. Downs (1957) in his influential An Economic Theory of Democracy assumes

that voter preferences can be arrayed on a single ideological continuum, from the political

left to the political right, with citizens voting for the candidate closest to their own
position on this ideological spectrum. In addition, Downs’ model also assumes
competition between two political parties or candidates for majority support. These
assumptions drive the outcome: Parties and candidates converge toward the middle of the
spectrum, with the winner being the candidate closest to the median voter. Since the
winning program converges toward voter preference at the mid-point of the ideological
continuum, the policy result renders an accurate representation of the composite position
of the electorate as a whole.

Evidence from survey research does support the above idea. Monroe (1979,1983)
shows that public policy in America is associated with majority preferences about two-
thirds of the time. Page and Shapiro (1983) argue that major changes in people’s
preferences are followed by coherent changes in policy about two-thirds of the time.
Farkas, Shapiro, and Page (1990) compare changes in public opinion in national surveys
from 1973 to 1978 with changes in spending. Their findings suggest that public opinion
affects spending on the environment, space, education, and drugs. Page and Shapiro
(1992) characterize the public as composed of individuals whose preferences are fairly
stable over time, particularly in the short-run. They also suggest that opinion change is
effectively parallel across various subcategories of the American public. Stimson’s
(1991) study of a wide range of public opinion data further reveals that public preferences
for policy in various domains are closely connected over time. Simply, preferences
across a range of issues tend to move in the same liberal or conservative direction,
producing what Stimson has called “policy sentiment” or “policy mood.” That

preferences for policy generally move in the same liberal or conservative direction

denotes an underlying structure to those preferences. In addition, evidence from survey
research also shows that the liberalism or conservatism of the public in the different states
strongly influences whether states make liberal or conservative policies (Wright,
Erickson, and McIver 1987; Erickson, Wright and McIver 1989 1992). Furthermore,
some research suggests that aggregate voter turnout in elections is a predictor of
respective welfare-benefit levels in the US and western democratic nations (Peterson and
Rom 1989; Hicks and Swank 1992; Hill, Leighley, and Anderson 1995). Hill and
Anderson (1995) provide a causal model which suggests that a reciprocal causal
relationship between mass and elite preferences shapes state policy liberalism. In
addition, party competition and lower-class mobilization enhance elite and policy
liberalism.

However, not everyone agrees with the argument that the public’s policy
preferences have a substantial impact on what government does. Crigler (1990) claims
that public opinion, the media, and Congress seldom attend to the same agenda. And
other research on the agenda of Congress indicates that public opinion surveys seldom
reach specific conclusions which offer guidance on the usually complicated policy
matters before Congress. Kingdon (1995) argues that policy emerges from a “primeval
soup” of ideas, speeches, bills, and proposals, where the contents of the soup come from a
variety of sources, including public opinion. According to Kingdon, the mating of a
policy solution and a policy problem may be due to nothing besides the efforts of policy
advocates who attach their policy to an opportune problem. If policy proposals come
from only a small segment of public opinion, then the other segments have fewer

opportunities to redress their needs.

10

In sum, most evidence from public opinion research does Show a close
relationship between public opinion and policy. The public does adjust its preferences for
“more” or “less” policy activity in response to what policymakers actually do. In
contrast, although many studies have confirmed the influence of public opinion on
policies, whether policymakers follow public opinion to make policies is not clear.

Public Attitudes Toward Government Spending

Evidence from public opinion has repeatedly shown that there are two remarkable
and contradictory features of peoples’ preferences toward government spending. First, it
is known that most of citizens favor both tax reduction and increased levels of
governmental services. For example, V.O. Key (1961) found in Public Opinion and
American Democracy that a sizable fraction of the electorate wanted stronger social welfare
protections while also favoring tax cuts “even if it means putting off some important things
that need to be done.” Key pointed out that both attitudes, though appearing contradictory,
were real. As he wrote, “a simple calculus of self-interest makes simultaneous support of
tax reduction and expansion of social welfare activities entirely consistent For the system
as a whole, however, this type of opinion combination is irrational and creates problems in
program making The balance of forces drives policymakers back toward concealed and
indirect taxation, which may be regressive in its incidence” (168).

Sears and Citrin (1982) show that well over 70 percent of Californians before and
after Proposition 13 (1978) thought their taxes were too high, yet majorities favored
increases in spending for a wide variety services - mental health, police, fire, prisons and

corrections, school, and transportation. Only for welfare did a majority of the public

11

favor cutbacks. They conclude that people do, indeed, want “something for nothing” in
public policies.

At the same time, however, there appears to be a inconsistency in the views of the
public about reducing government spending. One the one hand, a majority of the public
strongly favors cuts in government spending in general, when no specific programs are
mentioned. One the other hand, when asked whether or not cuts should be made in
specific programs, with only a few exceptions, a majority either favors maintaining
current spending levels or prefers even more spending (Ladd et al. 1979; Bennett and
Bennett 1990). Free and Cantril (1967), for example, argue that there appears to be an
“ideological schizophrenia” within the US. public: people favor increasing spending on
social programs, even while they repeatedly express demands for a smaller government.

Explanations of the Inconsistencies Between Policymakers and the Public
Toward Government Spending

Several explanations have been addressed for the apparent distinction between
taxing and spending preferences. The simplest explanation is that just as taxing and
spending decisions have been institutionally divorced so they have been divorced in the
minds of the public (Axelrod 1967; Brittan 1975). Another explanation, suggested by
Downs (1963), is that because government forcibly provides complex bundles of different
programs, taxpayers never reach equilibrium in their dealings with governments. That is,
for each taxpayer there is some preferred reallocation of resources either within the public
sector or between the public and private sectors, thus there is a tendency for every

taxpayer to believe that taxes are too high. Finally, some argue that for some taxpayers

12

dissatisfaction with tax levels may be rooted not in a belief that the overall tax burden is
too great but rather that it is unfairly distributed (Kuttner 1980).

For the contradiction about reducing government spending, four explanations
have been suggested by Crocker (1981). The first explanation involves the public’s
views about the relationship between inflation and government spending. Inflation
consistently has been chosen as one of the most important problem facing the country
since the end of Vietnam War. Moreover, the public seems to blame government
spending for causing inflation. For instance, a series of NBC News/Associated Press
surveys conducted in October, November, and December 1978, asked which of nine
approaches was best to fight inflation. The most frequently named approaches were to
cut government spending - 31 percent in October, 32 percent in November, and 25
percent in December.

The second explanation is that the public has little knowledge of the complexities
of economics and government budgeting. Therefore, it is argued, the public holds
contradictory views because most people lack an understanding of economic problems.
For example, according to the April CBS News/New York Times Poll, 51 percent of
those responding said that they did not know enough to have an opinion on Reagan’s
budget cutting proposals, and 52 percent said that they did not know enough about
Reagan’s tax cut proposal to have an opinion.

The third explanation, suggested by many polling experts, is that the public views
the solution to be the elimination of waste and inefficiency in government. There is
substantial evidence that, to many people, waste and inefficiency is the major problem

with the government. In 1978 a Harris survey found that 94 percent of those surveyed

agreed “that too much tax money is wasted by inefficient bureaucracy.” A 1981 Gallup
Poll found that the public believes that an average of 48 cents out of every dollar the
federal government collects in taxes is wasted.

The last explanation is that the questions in polls simplify the issues and,
consequently, lead to contradictory results. For instance, one might get different results if
the questions on specific program reductions were asked in a way that indicated the
tradeoff between increased spending and balancing the budget. Alternatively, one could
argue that the questions on overall cuts in government spending do not inform
respondents about what cuts are to be made.

One difficulty with the first two explanations is that neither give much credit to
the public. If the public favors overall spending reductions and also favors maintenance
of various social programs without increased taxes, how does it view the means of
achieving both these ends? Similarly, the explanation of inefficient government also
cannot solve the problem for policymakers. Students of public administration have
learned that the motive of bureaucrats are driven by constraints, not profits, and public
administration is different from business. Therefore, achieving the goal of efficient
government is easier said than done. What we really should try is to use a better survey
instrument to measure people’s true preferences rather than their desires on different
spending programs.

Existing Approaches to Evaluate Government Spending
Several studies have tried to explain the complexities of public preferences toward

government spending by using all kinds of statistical techniques and other survey tools.

Eismeier (1982) examines the nature and determinants of individual budget
priorities by using multiple discriminant analysis. Specifically, he uses multiple
discriminant analysis to determine how respondents with different budget priorities
differed according to several political, socioeconomic, and attitudinal variables, including
tax levels. Eismeier finds that the growing dissatisfaction with taxes is rooted in the
belief that government is not only too big but also inefficient and unresponsive. In other
words, dissatisfaction with spending levels is due to people’s lack of confidence in the
performance of government. His second finding is that patterns of partisan and
socioeconomic differences vary substantially across policy areas. However, he also
points out that the inconsistent attitudes of budget priorities may not be as much at the
individual level as it is at the aggregate level. That is, while the individual taxpayers’
policy preferences may be internally consistent, they may be incompatible with the policy
preferences of other taxpayers.

Wilson (1983) examines people’s “sincere preferences”' on government spending
through experimental surveys.2 In particular, he estimates the frequency of the expression
of “sincere preferences” for additional public spending - preferences which are
accompanied by a willingness to pay additional taxes in order to obtain additional public
goods or services. The format of the surveys is a modified n-prisoners’ dilemma game
which includes a series of questions about people’s preferences toward specific policy

areas. The survey design is best explained by the following example: “Do you think that

 

1 Individual expressing “true” or “real” preferences for increased, constant, or reduced government spending are
considered to be expressing “sincere preferences.”
2 The surveys were conducted on Eugene, Oregon (N=1 115) and Tempe, Arizona (N=509) in 1977 and 1981

respectively.

the city should spend more, less, or the same amount on crime prevention?” There are
three options for respondents.

1. If the answer is “spend more,” the following question is asked: “ Would you be willing
to pay more in taxes in order to spend more on crime prevention?” If the answer is “yes,”
no further question. If the answer is “no,” then proceed with question “Would you
suggest that there be no change in service level if it means increased taxes or would you
reallocate money from some other program? The answer will be either “no change” or “
reallocate.”

2. If the answer is “spend the same,” no further question.

3. If the answer is “spend less,” the following question is asked: “If the city were to
reduce its spending in this area, what should it do with the money saved?” The answer
will be either “reduce taxes” or “ reallocate.”

The structure of the survey was designed to permit individual respondents to
overstate their preferences for the government spending (“would you like to see the city
spend more on the provision of a service”) and then attempt to hold them accountable for
their preferences (are you willing to pay more in taxes in order to have more of the
service” .

The results from Wilson’s study illustrate that approximately two-thirds or more
of the demands for additional service expressed on the experimental survey constitute
“sincere” preferences. The proportion of “sincere” preferences ranges from a low of 62
percent to a high of 84 percent in Tempe and from a low of 66 to a high of 78 percent in
Eugene. The range of “insincere” preferences (those who decide they do not want the city

to spend more on the provision of a good or service if it means as increase in taxes) varies

16

from a high of 12 percent (Tempe) and 14 percent (Eugene) to a low of 6 percent
(Tempe) and 5 percent (Eugene).

Welch (1985) explores the possibility of increasing spending level by means other
than taxes, such as reallocation or nontax revenue. She creates a fourfold categorization
called “willingness to pay” scale to demonstrate the willingness of individuals to raise
revenue to support increased spending. The four categories are:

1. Taxers - those willing to increase taxes.

2. Probably realistic revenue raisers - those who were not willing to raise taxes but
willing to adopt new user fees, increase revenue through intergovernmental aid, or
reallocate funds to pay for increased service in priority areas.

3. Probably not realistic revenue raisers - those who prefer reallocation but wanted to
spend more in more than half of the programs.

4. Free lunchers.

The findings from her study suggest that support for increasing government
spending is high. Nevertheless, relatively few favor a “free lunch,” increase spending
without any reasonable means to pay for it. Most people are willing to raise additional
revenue to pay for these services, or at least to reallocate from less preferred to more
preferred services. From 72 percent to 88 percent of the respondents are found in one of
the first two categories (taxers and probably realistic revenue raisers), where support for
raising revenue in realistic ways is apparent. Yet, very few people think we can depend
solely on greater efficiencies in government to pay for increased services. Therefore, she

concluded: “The taxing-spending paradox may not be as great as previously thought”

(p.316).

Sanders (1988) assesses the two opposite explanations of public attitudes on
government spending: the spending attitudes of the public have been described as either
too irrational, wanting something for nothing, or too rational, only supporting programs
they benefited. He runs 16 logistic regressions that dependent variables are composed of
each of the individual spending questions (programs) while independent variables include
a composite index3 and 6 demographic variables. From his findings, Sanders concludes
that “some of the traditional views we have of the public’s attitudes on public spending
issues may be more myth than reality. The public may not simply be asking for
something for nothing, nor do they seem to be saying ‘help me.’ The attitudes people
have are more complex than that” (p.321). Specifically, he suggests that most citizens
desire increasing spending on some of the programs and less on other programs. In fact,
everyone desires to cut spending in at least one domain, and on average the number of
programs people prefer to increase tends to balance out with those they want to cut. In
addition, the role of self-interest is limited; most programs cannot rely solely on the
support of those who benefit from them. Sanders points out that the two domains where
self-interest plays the strongest role - food stamps and welfare — are also two of the three
domains where most people want to cut. Therefore, he suggests that building support for
a program needs to establish the legitimacy of the beneficiaries among the broad public.4

Jacob (1994) examines the nature, sources, and consequences of peoples’ attitudes

toward government spending by using a variation of Guttman scaling technique. He

 

3 The index is created as follows: First, respondents were asked if they received benefits from 7 of the programs
(16 programs total in the survey). Second, a composite index was formed adding up the number of those seven
areas in which an individual received support.

argues that most people have very limited ideas of political phenomena, so it is
inappropriate to view all of the program-specific responses as equally valid
representations of people’s spending preferences. Therefore, he evaluates patterns of
responses across programs by fitting a cumulative scaling model. According to this
model, there is an underlying trait or dimension that ranges from a position “reduce
spending in all programs (or a specific program),” at one extreme, to “increase spending
in all programs (or a specific program),” at the opposite extreme. He uses 1984 NES
postelection survey as his data in research (which is slightly different from the traditional
NES survey questions). The survey includes the following questions: “If you had a say in
making up the federal budget this year, which programs would you like to see increased
and which reduced?” After that, the respondents were asked the following question about
10 program areas: “Should federal spending on __ be increased, decreased, or kept
about the same?”

Jacoby’s analysis suggests that the public has coherently structured orientations
toward spending on welfare-related programs. Yet, opinions on nonwelfare spending are
highly varied and largely independent of each other. In addition, the general term
“government spending” triggers individuals to think about welfare spending, even if the
particular programs are not explicitly called to public attention. That is, most people
seem to translate the phrase “government spending” into government spending on

programs that will benefit the poor, blacks, and other disadvantaged people.

 

4 Establishing legitimacy is easier with programs which offer benefits for most people (such as environment) or
that most people might potentially need (such as Social Security or health). For a detailed discussion, see
Sanders (1988, p.322).

19

Furthermore, individuals’ responses to social welfare programs are generally quite
consistent with their political predispositions. However, in the area of nonwelfare
programs, people’s preferences are very difficult to predict on a systematic basis. Put
another way, Jacoby argues, “nonwelfare spending preferences are not political attitudes
at all” (p.355).

Wlezien (1995) employs a thermostatic model to analyze public preferences for
spending on defense and a set of five social programs. The thermostatic model refers to
the public behaving like a thermostat: when the actual policy “temperature” differs from
the favored policy temperature, the public will send a signal to adjust policy accordingly,
and once sufficiently adjusted, the signal would stop. The public’s preferences for
spending in various policy areas are captured by the measures of net support for
spending.’ Time-series regression is used to estimate defense spending preference, five
categories of social spending, and aggregate social spending.

From his study, Wlezien’ points out that public preferences for spending in the
specific policy areas do not generally reflect appropriations in those areas. Nonetheless,
the public does adjust its preferences for more or less spending in the particular areas in
response to what policymakers appropriate for social programs taken together. The
reasOn for this difference, Wlezien suggests, is that specific information about spending
for those programs is not regularly and widely available to the public. Thus, it may be
that the public obtains only general information about social programs. The public still

responds to spending for social programs, but in a general way. In general, he argues that

 

5 The net support for spending is calculated by subtracting the percentage of people who think we are spending
“too much” from the percentage of people who think we are spending “too little.”

20

his findings are consistent with Eastonian model: there is a negative feedback of spending
decisions on the public’s preferences, whereby the public adjusts its preferences for more
spending downward when appropriations increase, and vice versa, at least in the defense
spending area and across a set of social spending areas.

Hansen (1998) examines public preferences over deficits, taxes, and spending by
using tradeoff type of survey questions. He uses the data of 1995 NES Pilot Study which
asked respondents 12 questions with a complete budget tradeoffs design. The budget
choices in the survey were divided into two revenue categories: taxes and deficits and two
spending categories: domestic and defense. Respondents then were asked to accept or
reject a budget proposal that made them better off on one dimension (by raising domestic
spending, increasing defense spending, lowering taxes, or reducing the deficits) but left
them worse off on another (by raising taxes, enlarging the deficits, reducing domestic
spending, or cutting back on defense). From his study, Hansen discovered that
respondents overwhelmingly rejected the tradeoffs and revealed their preference for the
policy status quo. He also evaluated the characteristics of individual preferences, such as
their completeness, consistency, and coherence and found that peoples’ preferences over
the public budget were remarkably well structured: most people are able to determine
their views, identify tradeoffs, and maintain consistent and coherent preferences.

Some overall conclusions can be drawn from these findings about public
preferences toward government spending. First, the idea that the public wants
“something for nothing” is not true if we can hold them accountable for their preferences.
If the people favor more services, we could hold them accountable by asking them either

to increase taxes or reallocate spending on various programs. These tradeoff procedures

eliminate the problem with the traditional survey questions that respondents usually
overstate their demand because they will not be charged for the increased level of service
provision. Wilson’s study reveals that more than two-thirds of the demand for additional
service is “sincere” preferences - a willingness to pay more taxes to obtain more services.
Welch discovered that most people are willing to raise taxes to pay for more of services,
or reallocate funding from other different programs. Sanders’ findings show that on
average the number of programs people prefer with increases tends to balance out with
those they want to cut. Hansen found that public opinion is remarkably well structured
and overwhelmingly partial to the policy status quo. Thus, previous studies have
demonstrated the limits of traditional surveys: they virtually measure people’s desires
rather than their true preferences. To obtain people’s true preferences, we have to use
tradeoff format of survey design and hence hold respondents accountable for their
preferences.

Second, all these studies have tried either different statistical techniques or
different survey tools to unveil the complexities of public attitudes toward budget
priorities (Eismeier’s discriminant analysis, Wilson’s prisoners’ dilemma format of
survey, Welch’s “willingness to pay” scale, Sanders’s logistic regressions, Jacoby’s
scaling approach, and Wlezien’s time-series analysis). Although these efforts are not
small tasks, the effects of these applications are limited due to the problem of the data
they used. Specifically, for those data which came from traditional surveys, the problems
are caused by the fact that these questions only asked people to evaluate the importance
of a single attribute rather than multiple attributes that necessarily consist of a policy

package or budget proposal. For example, from traditional surveys we could obtain a

22

respondent’s preferences that he or she favors a budget proposal “increases in education,
crime, and natural resources spending and cuts in welfare spending.” However, we do
not know how this respondent would evaluate a budget proposal “increases in education,
crime, and natural resources spending and cuts in welfare spending” versus another
budget proposal “increases in crime and natural resources spending and cuts in education
and welfare spending.” For those data which came from different survey tools, the limits
lie in the fact that respondents only evaluate tradeoffs of partial (e. g., two-factors)
attributes rather than all attributes which necessarily consist of a policy package or budget
proposal. For instance, from Hansen’s study we know that a respondent favors tax
increases in order to cut the deficit and opposes cuts in domestic spending in order to
increase defense spending. Even so, it is still very difficult to find out how a respondent
chooses budget proposals with tax increases, lowering deficit, increases in domestic
spending, and cuts in defense spending versus tax decreases, lowering deficit, decreases
in domestic spending, and cuts in defense spending. In other words, we do not know how
a respondent chooses his or her most preferred policy proposal from different policy
packages.

In fact, in the real decisionmaking process, people’s choices are made on the basis
of evaluating and combining multiple attributes simultaneously. Lacy (1998) examines
two models of individual-level responses to issue questions when respondents have
nonseparable preferences. His results reveal that question-order effects occur on issues
for which people have nonseparable preferences. Hence, the results from previous
studies are not satisfactory because we obtained only fragmented and incomplete

information about public opinion toward budget priorities. In contrast. conjoint analysis

23

provides more accurate and complete information about people’s preferences by
producing part-worth utilities for each level of each attribute (i.e., utilities of cuts in
domestic spending and increases in domestic spending). Part-worth utilities are then used
to calculate the relative importances of attributes (how important an attribute is) and
overall utilities of profiles (i.e., a profile with tax increases, lowering deficit, increases in
domestic spending, and cuts in defense spending). In addition, we also could obtain
people’s preferences on different policy packages by simulating all kinds of policy
combinations. This accurate information is very difficult to acquire from the traditional
survey or from other survey tools previously discussed.

Finally, some survey formats (such as Hansen’s tradeoff design) address questions
in a direct and obtrusive way (such as raising taxes in order to increase domestic
spending) which may cause respondents to underestimate their preference for a particular
policy package because respondents were held accountable for the cost of increased
levels. On the contrary, since conjoint analysis ask respondents rank ordering alternatives
from most preferred to least preferred rather than ask respondents making tradeoffs
between two competing answers (i.e. increasing in tax in order to increasing domestic
spending), it is less obtrusive and hence respondents are more willing to reveal their true
preferences.6

Summary
In general, the literature on public opinion has demonstrated that peoples’ policy

preferences have a substantial impact on what government do. In addition, previous

24

studies have tried various statistical techniques and survey instruments to unveil the
complexities of people’s budget priorities; however, the effects of these applications are
limited due to the constraints of traditional survey questions and different survey designs
(i.e., Wilson’s prisoners’ dilemma format, Welch’s “willingness to pay” scale, and
Hansen’s tradeoff format). Conjoint analysis forces respondents to make tradeoffs among
various programs so that respondents evaluate all the attributes simultaneously (which
necessarily consist of a budget proposal) rather than evaluate attributes one by one or
pairs by pairs. Even though it is not perfect,7 conjoint analysis does provide more
accurate and complete information about people’s preferences toward public spending

when compared with the traditional survey.

 

6 For the full profile data collection method only. Collecting data by a two-factor-at-time method is also
obtrusive to respondents. See chapter 4 for a detailed discussion of the differences between full profile and
two-factor-at-time methods.

7 The limits of conjoint analysis will be discussed in chapter 3.

25

CHAPTER 3

THE LOGIC OF CONJOINT ANALYSIS

Studying budget priorities has both substantive and methodological significance.
The purpose of this chapter is to discuss the concept, theory, design, and analysis of
conjoint analysis, why conjoint analysis works, advantages and limits of conjoint
analysis, classification of conjoint analysis, and current applications of conjoint analysis.

The Concept of Conjoint Analysis

Conjoint analysis is both a method of measurement and a method of analysis. It is
a multivariate technique used specifically to understand how consumers decide
preferences for products and services based on the simple premise: consumers evaluate
the utility of a service or policy idea (real or hypothetical) by combining the separate
amounts of utility provided by each attribute. The word “conjoint” has to do with the
notion that the relative values of attributes considered jointly can be measured when they
might not be measurable if taken one at a time (Barron 1977). Quite often respondents
are asked to express the relative value of various alternatives to them by ordering the
alternatives from most desirable to least desirable. The attempt in a conjoint analysis
solution, then, is to assign values to the levels of each of the attributes so that the
resulting values or utilities are as monotonic as possible with the input rank-order
judgments.

In conjoint analysis, services and products are described by “profiles.” Each
profile is a combination of one arbitrarily selected level for each of the attributes.

Attributes are the key dimensions (features) of services or policies, while levels are those

26

specific points along the key dimensions (Hu 1996). The major objective of conjoint
analysis is to identify the single profile which includes the most favored level for each of
the attributes. Table 3.1 illustrates the relationship among a profile, attributes, and levels.

Table 3.1: The Relationship among a Profile, Attributes, and Levels

Profile X Attribute L_ey_e_l
A2 A -- Al, Al, A3
B1 B -- _Ifl, B2, B3
C3 C -- C1, C2, _C_3_
D2 D -- D1, 2;, D3

This table can be explained as follows: Profile X is composed of the second level
of attribute A (A2), the first level of attribute B (Bl), the third level of attribute C (C3),
and the second level of attribute D (D2). Hair et a1. (1995) point out that there are two
objectives of conjoint analysis: (1) to evaluate the contributions of determinant attributes
and the values of these separate levels on the determination of consumer preferences, and
(2) to build a valid model of consumer judgments that is useful in predicting consumer
acceptance of any combination of attributes, even those not originally evaluated by
consumers (556-559). In order to achieve these objectives, coefficients called “utilities"
(or “part-worths”) among different levels of attributes are first estimated, and then
“relative importance” among attributes and “profile utilities” are developed to
quantitatively measure preferences in consumer decisions. The meanings of the terms is

as follows:

27

0 Level utility (part-worth): a numerical expression of the value that a respondent places
on each level of each attribute.

o Attribute relative importance: the computation of the relative importance for each
attribute depends on the relative range between maximum and minimum level utilities
within an attribute. This is based on the assumption that the larger the difference between
maximum and minimum level utilities within attribute, the more determinative and
salient the attribute is in the overall evaluation of profiles. The relative importance is
described in a percentage term to reflect its weighted importance across all involved
attributes. The

relative importance is defined as follows:

Max(X,-,)- Min(Xu')
R1 = , x 100% (3.1)
.I Z[Max(){lj) -— MIN(X”)I

 

Where
R1]- = the relative importance of attribute j.
Max (V ij) = the maximum level (i) utility in attribute j.
Min (VD) = the minimum level (i) utility in attribute j.
0 Profile utility: the overall utility of a profile calculated by summing all utilities of
attribute levels defined in that profile.
Compositional vs. Decompositional approaches

The compositional or build-up model, such as regression or discriminant analysis,
involves collecting respondent’s judgments about each level of each attribute (e.g., a
respondent’s spending preference toward education, welfare, crime, and natural

resources), and then combining these ratings to compose an overall preference.

28

In contrast, conjoint analysis uses a decompositional model. Knowing
respondents’ choice preferences between pairs of objects, and knowing the characteristics
of the objects presented to respondents (as manipulated through an experimental design),
respondents react to a set of “total” profile descriptions. These preferences are
decomposed to determine how much utility is associated with each level of each attribute.
Therefore, it is the job of the analyst to find a set of part-worths for the individual
attributes , which given some type of composition rule (e.g., an additive one), are most
consistent with the respondent’s overall preferences. This avoids many of the subject
artifacts and demand effects of compositional approaches.

The major advantages of the compositional approach include a strong and
thoroughly tested conceptual foundation and ease of operationalization. However, these
models have been criticized because they cannot accurately simulate decision processes
in the real world. For example, they do not include trade-offs between attributes in the
decision process (Johnson 1974), they tend to overweight less important attributes
(Beckwith and Lehmann 1975), and they describe a more complicated decision process
than probably exists in the real world (Olshavsky and Granbois 1979).

The decompositional models (such as conjoint analysis) tend to be strong where
compositional methods are weak. Conjoint analysis provides a more realistic description
of the decision process. Because people are evaluating a complete service or policy, they
are able to incorporate the same attribute-discounting and trade-off processes they use
when making actual decisions. The limit of conjoint methods is that they are
cumbersome to administer due to the tedious procedures necessary to parameterize the

models.

29

The Theory of Conjoint Analysis: An Information Integration Theory Approach

The seminal theoretical contribution to conjoint analysis was made by Luce, a
mathematical psychologist, and Tukey, a statistician. Thereafter, a number of theoretical
contributions (Krantz 1964; Tversky 1967) and algorithmic developments (Kruskal 1965;
Carroll 1969, 1973; Young 1972; Shocker and Srinivasan 1977) appeared.
Information Integration Theory

Louviere (1988) provides a theoretical basis for conjoint analysis based on the
information integration theory (IIT) developed by Anderson (1981,1982). Information
integration theory allows us to study information processing as revealed by people’s
response to multiattribute options. Compared to other theories, Louviere argues that IIT
is a better theory because it possesses a theory of errors; error theory is important because
it permits one to falsify models.l

The basic assumptions of integration information theory are as follows:
1. The unknown and unobservable overall utility that a consumer possesses in his or her
mind regarding the p-profile is linearly'related to a consumer’s response on a category-
rating or ranking scale. That is,

Up=a+bRp+ep (3.2)

where
Up is the overall utility to measure of the p-th profile,

Rp is the observed response on a category-rating or ranking scale,

 

' [IT is similar in conceptual orientation to Social Judgment Theory (Adehnan et al., 1974) and bears a
strong resemblance in experimental format to axiomatic conjoint measurement such as Krantz (1964) and
Tversky (1967).

30

ep is a normally distributed error term with zero mean and constant variance 02, which
satisfies assumptions of OLS (ordinary least squares).

2. The category-ranking scale used by a subject under approximate experimental
instructions and task conditions approximates an interval measurement level.

3. A subject’s response strategy reveals his or her decision model. The response strategy
can be approximated by algebraic conjoint models amenable to experimental design and
statistical parameterization.

Assumption 2 may be relaxed through the use of alternative response scales, such
as binary scale (yes/no) or multiple choice scales (from most preferred to lest preferred,
etc.).

Algebraic and Statistical Formalities

IIT assumes that simple algebraic conjoint models are valid approximations to the
unobservable evaluation and decision process of consumers. However, the decision
model that consumers “really use” is unknowable. A simple additive model is a good
start to illustrate the conjoint method. The additive model can be expressed as follows
with three determinant attributes:

Up = C + W(X,p) + W(X,p) + W(X3p) (3.3)
where Up is the overall utility to measure of the p-th profile, C is an additive constant, and
W(X,p), W(X2p), and W(X3p) are part-worth utilities for the three attributes of the p-th
profile.

Each level of each attribute may have a different part-worth utility. Different

profiles can be described by combinations of different attributes with different levels. An

31

additive model implies that consumers add the separate part-worth utilities to evaluate
each profile. Assuming the levels of attributes 1, 2, and 3 as X”. Xzb, and Km and the
corresponding part-worth utilities, as W(X.,), W(X2b), and W(X3c). Each a-th, b-th, c-th
combination of levels thus represents one profile.

Hence, instead of using Up, we will use Uabc with no loss of generality, but it
means Up. The total number of profiles that one can produce is the product of the a, b,
and c levels. The total number of profiles is therefore a combinatorial problem defined
by the set of possible combinations of the levels a, b, and c. Such problems are called
factorial problems (which will be discussed later).

Now we rewrite the additive model for a single subject as follows:

Uabc = C + W(X|.) + W(X2b) + W(X3c) (34)

where all terms were defined previously.

Recall that we cannot observe Uabc directly. Instead, we observe a subject’s
evaluation, Raw on a rating or ranking scale. Substituting Equation 3.2 and transposing,
we get:

R

abc = C + W(X,a) + W(X,_b) + W(X3c) + eabc (3.5)

where all terms were defined previously. Equation 3.5 suggests that a subject’s rating or
ranking are an additive function of the true but unknown part-worth utilities. The error
assumptions of ITT imply that ANOVA or OLS can be used to test the additive model
hypothesis of equation 3.4.

If a subject evaluates each of the axbxc combinations more than once, there are

sufficient observations to test equation 3.4. If equation 3.5 is true, there will be no

32

nonadditivities in a subject’s response data. That is, only the purely additive terms will
be nonzero in a statistical analysis - there will be no significant interaction effects in the
analysis of variance or multiple regression if equation 3.4 is true. This conclusion can be
proved if we subtract the response to level 2 of attribute 1 from the response to level 1 of
attribute l, holding other attributes constant:
Rlbc - Rm = [C + W(Xn) + W(X2b) + W(X3c) + elbc] (3-6)
- [C + W(X.2) + W091.) + W(X3c) + 62b. 1,
= W(Xn) - W(X12) + (6le - eZbc)'

Equation 3.6 shows that the difference in the part-worth utilities of level I and 2
of attribute 1 is a constant; therefore, curves of response data for each level of attribute 1
graphed against levels of attribute 2 (or 3) must be parallel. Responses to levels of one
attribute are independent of responses to levels of other attributes, a property that
generalizes to any number of attributes.

Therefore, if a consumer’s decision process is additive, we expect to maintain the
null hypothesis that no interactions in an analysis of variance or multiple regression.
Thus, the additive model hypothesis can be falsified by using ANOVA or OLS to test the
main and interaction effects of subject’s evaluation. However, subjects must rate or rank
each profile at least twice in order to have enough degrees of freedom to perform the
necessary statistical tests. If an experimental (factorial) design is small (e.g. 2*2*2),
multiple replications are doable. However, as the size of an experiment grows, by
increasing the numbers of attribute numbers, by increasing the numbers of attribute

levels, or by both, it becomes impractical to ask subjects to rank or rate multiple profiles.

33

Therefore, one must find other ways to solve this problem, such as fractional factorial
design (which will be discussed later).
Now let us turn our attention to estimating the part-worth utilities - the W(X,a),
W(X2b), and W(X3c). Repeating equation 3.5 for convenience, we get:
Rabc = C + W(X,a) + W(X2b) + W(X3c) + eabc (3.7)
Considering the result of averaging 3.7 over attribute 2 and 3 (or subscripts b and c):
R... = C + W(Xla) + W(X2.) + W(X3.) + C... (3.8)
= [C + “’09) + W(X3.)l + W09.) + C...
where dots (.) refer to the effect of averaging. Equation 3.8 constitutes a theorem that
expresses that the marginal means (R..-) of a subject’s response data are equal to the
unknown part-worth utilities (W(Xk)’s) up to a linear transformation. That is, the
marginal response means are interval scale measures of the unknown part-worth utilities
because a linear transformation is the only permissible transformation for interval scales.
This conclusion is true only if an additive model is “correct.” Hence, if an
additive model passes the statistical tests described earlier, one can use the marginal
response means to represent the consumer’s part-worth utilities for each attribute. In
addition, the (W(Xk)’s) must be a function of the Xk’s. (The different composition rules
and preference models will be discussed in the next section). Therefore, one can test
additive conjoint models and, if they cannot be rejected, one can infer relationships

between consumer’s positions on different profiles and part-worth utilities.

34

Design and Analysis of Conjoint Analysis

In conjoint analysis, the most widely used model is an additive main-effects (no
interactions) part-worth model. This statement includes the use of different composition
rules and preference models for consumers in conjoint analysis.

Choice of Composition Rules: Compensatory vs. Noncompensatory Models

The first step in conjoint analysis is to consider the principle ways in which
attribute information of each respondent can be processed. Two major approaches,
compensatory model and noncompensatory model, can be used in conjoint analysis.

1. Compensatory models are those cases in which all alternatives can be ultimately
described in terms of single utility numbers that are commensurate with each other. The
models are compensatory in the sense that a low value on some attribute can be
compensated for a high value on some other attribute. In these models a single holistic
value or utility is attached to each multidimensional profile. Tradeoffs between attributes
are generated by different profiles that have equal utilities. The simplest compensatory
model is the main-effects, additive utility model. The composition rule for this model is
an additive one over all the attributes selected by the respondent; if the respondent
selected attributes A, B, and C, then an additive composition rule would be expressed as
A + B+ C (as seen in the Equation 3.3).

In conjoint analysis, the additive main-effects model is the most widely used.
Except using additive model in conjoint analysis, there are numerous models available.
For instance, the additive model belongs to the class of algebraic conjoint models known
as simple polynomials (Krantz and Tversky 1971). The class of simple polynomials

includes additive, multiplicative, distributive, and dual-distributive models. The

35

multilinear model is the general form for these simple polynomials, and it can be
expressed as follows for three determinant attributes:
Uabc = C0 + C,W(X,a) + C2W(X2b) + C3W(X3c) (3.9)

+ C4W(X,a) W(X2b) + C5W(X2,)W(X3C)

+ C6W(X2b)W(X3c) + C,W(X,a) W(X2b)W(X3c),
where the Co is the origin of the utility scale and CI - C7 are scaling constants. Equation
3.9 implies that attributes can be complements, substitutes, or independent. The
multilinear form allows a general approach to study decision processes, because it
contains completely multiplicative process (all attributes are complements), distributive
processes (one attribute complements the other two, which are jointly independent); and
dual-distributive processes (one attribute is independent of two others that are

complements). These models can be written as follows:

Uabc = C0 + C,[W(X,QW(X2b)W(X3c)], [multiplicative] (3.10)
Um = C0 + C,W(X,a)[C2W(X2b) + C3W(X3c)], [distributive] (3.1 1)
Uabc = C0 + C,W(X,a) + C2W(X2b)W(X3c), [dual-distributive] (3.12)

where all the terms are as defined previously.

Equation 3.10 - 3.12 are subsets of the multilinear form in which one or more C
terms are equal to zero. Thus the multilinear form can be used as the general form for
estimation and testing: all results that apply to this form also apply to equations 3.10 -
3.12. Although there are other composition rules available, but they are not often used in

conjoint study.2

 

2 For a discussion of these composition rules, see Green and Wind (1973), and Barron (I977).

36

2. Noncompensatory models do not allow tradeoffs between attributes. Comparisons are
made on an attribute-by-attribute basis, and in general the multidimensional profiles are
not collapsed into single utility numbers. For example, one attribute could be expressed
in terms of pounds and another in terms of rating points. Since component values are
applied to each attribute separately, the dimensions do not have to be commensurate.3
Preference Models

Utility preference models are the mathematical formulations that define the utility
levels for each of the attributes. In practice, the attributes are modeled as either a linear
(vector model), quadratic (ideal-point model), or piecewise mart-worth model) function
(Green and Srinivasan 1978).

The Vector Model

The vector model is represented by a single linear function that assumes
preference will increase as the quantity of attribute j increases (preference decreases if the

function is negative). Preference for the j-th attribute is defined as:

S=2M%» on)

i=1
where:

Si = Preference for the stimulus object at level i,

j = 1, 2, ..., t denote the set of attributes that have been chosen,

W]. = the individual’s weight assigned to each of the j attributes. One weight is derived
for each attribute,

Xi,- = Level of the j-th attribute for the i-th level.

 

3 For a discussion of noncompensatory models, see Yoon and Hwang (1995).

37

The vector model for the attribute with three levels would appear as a straight
line, with the levels on the line (see Figure 3.1). The vector model requires that a single
parameter be estimated for the variable treated as a vector. In contrast to the part-worth
model (will be discussed later), the vector model defines the attribute variable not as a
series of dummy variables but as a single linear variable where values are the measured
values or levels associated with the attribute.

The Ideal Point Model

The ideal point model is operationalized as a curvilinear function that defines an
optimum or ideal amount of an attribute. The ideal point model is appropriate for many
qualitative attributes, such as those associated with taste or smell. For instance, too much
sweetness may be less than optimal, while just the right amount is highly preferred.

The ideal point model establishes an inverse relationship between preferences and
the weighted distance df between the location of the i-th level and the individual’s ideal

point, 1].. The ideal point model is expressed as :
, 2
d} =2"ng —1',) (3.14)
[=1

where:

j = 1, 2, ..., t denote the set of attributes that have been chosen,

W). = the individual’s weight assigned to each of the j attributes. One weight is derived
for each attribute,

Xi,- = Level of the j-th attribute for the i-th level,

I]. = the individual’s ideal point, j.

38

The ideal point model for the attribute with three levels would appear as a curve
with the center of the curve higher than either end (see Figure 3.1). The highest point is
the ideal quantity of the attribute. Mathematically, the implications of specifying each of
the models ultimately extends to the number of parameters that must be estimated. For
the ideal point model, 2t parameters must be estimated (WJ- and I]. ), and for the part-worth
model, (i-l) parameters must be estimated, where i is specified to be the number of levels
for each of j attributes.

The Part-Worth Model

The part-worth model is the simplest of the utility estimation models. This model
represents attributes by a piecewise linear curve (see Figure 3.1). This curve is formed by
a set of straight lines that connect the point estimates of the utilities for the attribute
levels.

The part-worth model posits that
s, = Zfi(Xii) (3.15)
.I=|

where
Si = Preference for the stimulus object at level i,
j = 1, 2, t denote the set of attributes that have been chosen,
Xij denote the level of the j-th attribute for the i-th stimulus (level),
f]- is the function denoting the part-worth utility of different levels of the stimulus object,
Xij for the j-th attribute.
The part worth model reflects a utility function that defines a different utility

(part-worth) value for each of the i levels of a given attribute. Because of design

39

considerations, most conjoint studies constrain the number of levels to be less than 6,
though in reality, the number of levels varies from 2 to 9.

The implications of specifying a given preference model (linear, ideal point, or
- part-worth) extend beyond the actual shape of the preference curve being modeled. Each
model requires that a different number of parameters be estimated. The part-worth model
requires that each level of an attribute be defined by a dummy variable distinct column
within the design matrix. As would be expected, a total of H dummy variables are

required to estimate i levels.

40

Figure 3.1: Preference Models

1. Vector Model

Preference

4}

 

 

 

it

Levels of Attributej

2. Ideal Point Model

Preference
A

 

Ideal Point lJ

j X
Level of Attribute j

 

V

3. Part-Worth Model

Preference
A

   

 

 

V

Chosen Levels of Attributej

41

Fractional Factorial Design

Factorial design involves combinations of levels of determinant attributes. When
the size of an experiment grows, by increasing the numbers of attribute numbers, by
increasing the numbers of attribute levels, or by both, it becomes impractical to ask
respondents to rank or rate all the profiles. The usual solution for this factorial problem is
to use fractional factorial design. A fractional factorial design is a sample of treatments
selected from a complete factorial design. We should be cautious of factorial designs
because, without all possible combinations of attribute levels (complete factorials),
information is lost. Not only is information lost in that tests of some effects cannot be
conducted, but information lost is confounded with information obtained.4

In using fractional factorial design, one is willing to trade off the measurement of
all possible interactions effects in order to obtain a smaller number of replicates in which,
for example, all single-factor (main) effects and two-factor interactions can still be
estimated without confounding. By using this class of designs one assumes that all
higher-order interactions (as three-factor and beyond) are negligible. The most common
type of fractional factorial involves the case in which all main effects and all two factors
interactions can be estimated on an uncorrelated (orthogonal) basis. However, one
assumes the higher interactions (involving three or more factors) are negligible and can
be ignored. For instance, let us consider a two-attribute case in which both attributes

have two levels. Table 3.2 illustrates both the complete factorial and two one-half

 

‘ Confounded means that some effects are correlated (often highly or perfectly correlated) with other
effects. This is the multicollinearity problem in econometrics.

42

fractions of a 2 x 2 x 2 design. Levels of the three two-level attributes are represented as
-1 and +1.

Table 3.2: The 2 x 2 x 2 Design and the Two One-Half Fractions

 

 

 

 

Main Effects Two-Way Interactions Three-Way Interactions
X, X2 X3 X, X2 X, X3 X2 X3 X, X2X3
-1 -1 -1 1 1 1 -1
-1 -1 1 1 -1 -1 1
-1 1 -1 -1 1 -1 1
-1 1 1 -1 -1 1 -1

1 -1 -1 -1 -1 1 1
1 -1 1 -1 1 -1 -1
1 1 -1 1 -1 -1 -1
1 1 1 1 1 1 1
Fraction A
-1 -1 1 1 -1 -1 1
-1 1 -1 -1 1 1 1
1 -1 -1 -1 -1 -1 1
1 1 1 1 1 1 1
Fraction B
-1 -1 -1 1 1 1 -1
-1 1 1 -1 -1 1 -1
1 -1 1 -1 1 -1 -1
1 1 -1 1 -1 -1 -1

 

Table 3.2 shows that both one-half fractions contain columns that are duplicates
of one another. Specifically, X, is exactly negatively related with X2 x X3 in Fraction A,
and exactly positively correlated in Fraction B. Therefore X, is perfectly correlated with
the X2 x X3, and the effect of X, can be estimated only if X2 x X3 is equal to zero.
Otherwise, the effect represented by the column X, includes both the effects of X, and X2
x X3, and these effects cannot be separated. Likewise, X2 is perfectly confounded with

the X, x X3 interaction. Like X,, neither the X2 nor the X3 estimates can be interpreted

43

unless both X, x X2 and X, x X3 are exactly zero. This is the same logic in regression
analysis; when interpreting a single coefficient, we will say “holding other variables
constant.” Lastly, the three-way interactions column labeled X, X2 X3 is perfectly
correlated with the intercept or grand mean.

It would not be difficult to ask respondents to rate or rank 8 combinations.
However, with the increase of attributes, levels, or both in a design, it will be impractical
to execute it. For instance, let us consider a design with five attributes and each of
attribute with three levels: it is a 3x3 x3x3x3 experiment with total 243 combinations.
Obviously, it would be impossible to carry out an evaluation study of this magnitude and
it would cause confusion and fatigue among the respondents. In such cases, one might
wish as few combinations ( profiles) as possible. With fractional factorial design, the 3’
design can be reduced to 16 combinations.’

Orthogonal Arrays

In using fractional factorial design, the researcher is willing to trade off
information about higher types of interactions for convenience and lower cost in
constructing and testing fewer experimental combinations. Among fractional factorial
designs, orthogonal arrays are most widely used in conjoint analysis. Before we discuss
orthogonal arrays, let us turn our attention to the latin square design. Latin square design
is a type of fractional factorial in which high parsimony in number of combinations is

achieved by neglecting all interaction effects. In many evaluation-type of experiments

 

’ Some useful fractional factorial design catalogs include Hahn and Shapiro (1966), Conner and Zelen
(1959), and Anderson (1984).

this may be sufficiently accurate, particularly if one is permitted to transform the original
response data monotonically before estimating the model’s parameter values.

Let us consider a simple example that combining two (orthogonal) latin squares in
order to obtain a graeco-latin square, as illustrated below for two factors, each at three

levels:

Table 3.3: A Graeco-Latin Square for the 2x3 Design

 

Factor B
1 2 3
1 A,B, A283 A382
Factor A 2 A232 A38, A, a,
3 A383 A,B2 A28,

 

Note that each pair A,, B, appears exactly once in the Table 3.3 and that each Ai and Bj
separately appears once in each row and column.

Orthogonal arrays build on the preceding notion (illustrated by the graeco-latin
square) by developing even more highly fractionated designs in which, nevertheless, all
main effects can be estimated on an unconfounded basis, assuming that all interaction
effects can be neglected. Orthogonal arrays represent the most parsimonious (in the sense
of lowest number of combinations) set of designs available for main-effect parameter
estimation (Green 1974).

In sum, to use fractional factorial design one must assume that the unobserved
effects, which are confounded with effects of interest, are equal to zero. If such

assumptions are untenable or if one is uncertain about their satisfaction in a particular

45

design being considered, one should consider alternative designs with different
confoundment properties.6
Estimation Methods

To analyze input data, there are four major estimation methods available:
MONANOVA (Kruskal 1965), PREF MAP (Carroll 1973), LINMAP (Shocker and
Srinivasan 1977), and OLS regression. Because LINMAP and OLS are two most popular
estimation methods for nonmetric and metric data respectively, the discussion will be
focused on LINMAP and OLS methods.

LINMAP Estimation

LINMAP is one of the most popular nonmetric conjoint estimation procedures.
The use of linear programming enables LINMAP to obtain global optimum parameter
estimates, whereas there is no guarantee that other procedures would achieve global
optimums. In addition, in LIN MAP, attribute weights can be considered to be non-
negative, and part worth functions can be constrained to be monotone or of the ideal point
type; such constraints cannot be readily imposed for other approaches.

To show an intuitive explanation of the LINMAP procedure, consider two
attributes X, and X2 with three levels, respectively. Let a0, a, and a2 be the part-worths for
X, and b0, b, and b2 be the part-worths for X2. The full profile design with all 9 possible

cards is presented below.

 

° For a discussion of fractional factorial design, see Green, Carroll, and Carmone (1978), and Louviere
(1988), Chapter 2.

46

Table 3.4: A Full Profile with 2x3 Design in LINMAP Estimation

 

Card # Attribute X, Attribute X2 Utility Rank
1 0 0 a0 +bo 9
2 O 1 a0 +b, 6
3 O 2 a0 +b2 3
4 1 O a, +b0 8
5 1 1 a, +b, 5
6 1 2 a, +b2 2
7 2 0 a2 +bo 7
8 2 1 a2 +b, 4
9 2 2 a2 +b2 1

 

The utility of the profile (card) is equal to the sum of the part-worths of its
attributes. The last column displays a respondent’s rank order for the profiles, with Card
9 being the most preferred and Card 1 the least preferred.

Since Card 9 is the most preferred, its utility should be greater than that of all
other cards, i.e.

a2+b2>a0+b0,a0+b,,a0+b2,a, +b0,a, +b,‘a, +b2.a2+b0,anda2+b,,
Likewise, the fact that Card 6 is preferred to Card 1, 2, 3, 4, 5, 7, and 8 implies that
a, +b2>a0+bo,ao+b,_a0+b2.a, +b,,‘aI +b,,a2+b0,anda2+b,,

Similar constraints result from comparing, in turn, the third, fourth, fifth, sixth,
seventh, eighth, and ninth ranked cards to other cards of lower preference. Note that
LINMAP uses ordinal information about preferences, even if the input data consists of
ratings, which are (nominally) measured on an interval scale.

LINMAP uses the mathematical technique of linear programming to determine
the values for a0, a, a2. b0, b, and b2 in such a way that the paired comparisons are satisfied
as best as possible, taking into account the extent to which any inequalities are violated

(Srinivasan and Shocker 1973).

47

OLS Estimation

The OLS regression approach to conjoint analysis offers a simple, yet robust
method of deriving alternative forms of respondent utilities. It is commonly available,
inexpensive to use, and easy to interpret estimation method. The attractiveness of the
OLS model is in part a result of the ability to scale respondent choices using rating scales,
rather than rankings. The OLS procedure has the important advantage of providing
standard errors for the estimated parameters (Green and Srinivasan 1978). The ability to
implement designs having larger numbers of attributes and levels (through fractional
factorial designs) has made this methodology the de facto standard for conjoint analysis.
The objective of OLS conjoint analysis is to produce a set of additive part-worth utilities
that identify each respondent’s preference for each level of a set of product attributes. In
application, the OLS model solves for utilities by using a dummy matrix of independent
variables. Each independent variable indicates the presence or absence of a particular
level. The dependent variable is the respondent’s evaluation of one of the profiles

described by the independent variables. This model is expressed:

U, =C+ZZW,,X”+ep (3.16)
i .I

Where:

Up = the overall utility to measure of the p-th profile,

C = constant or intercept,

Wij = the part-worth utility of i-th level j-th attribute estimated in the regression,
X ij = dummy variable representing the presence or absence of attribute j, level i in

alternative X, and

48

ep = error term for profile P.
Why Conjoint Analysis Works

Conjoint analysis came from the psychometric tradition which is rigorous and
idealistic, whereas its adoption by market research community has been approximate and
pragmatic. Although the professional success of conjoint practitioners attests to how well
it works, we do not have a clear account of why it works. Huber (1987) provides the
following reasons for the success of conjoint analysis:

I. Conjoint asks respondents make tradeoflfs which are similar to those in the market.

A conjoint task is valuable because it forces the respondent to evaluate conflicting
attributes, as between the type of TV and its price. People usually try to avoiding making
such decisions by searching for unambiguous solutions involving dominance or following
a relatively clear heuristic. However, the market also requires such decisions and people
make them when they must in the marketplace or the conjoint task.

2. The simplification in conjoint analysis may mirror that in the market.

It would be misleading to think that conjoint analysis uses a small number of
attributes to simplify decision process if people use a very different kind of simplification
(e. g. evaluating all factors) in the marketplace. However, there is evidence that people’s
decisionmaking is based on remarkably few dimensions or factors (Olshavsky and
Granbois 1979). If this is true, then conjoint analysis may indicate those few attributes on
which people base their decisions.

In addition, to the extent that conjoint analysis is used to predict aggregate shares,
it does not matter whether an individual’s choice of attributes is unstable over time. As

long as conjoint analysis reflects an unbiased choice of attributes at the time, the

49

aggregate shares will also be unbiased. The criteria for conjoint analysis to work at the
aggregate level are considerably less strict than for individuals.
3. Conjoint uses orthogonal arrays.

An orthogonal design is that the levels of different attributes across profiles are
uncorrelated. Such design makes sure that an estimate of one attribute is unaffected by
the estimate of the other attributes. Most econometric models could suffer moderate
levels of multicollinearity without much harm. However, respondents often simplify the
conjoint task by focusing on one or two variables; this leads to substantial difficulties (i.e.
wrong sign) if any attributes in the design are correlated. This explains an inherent
advantage in an orthogonal array. For an orthogonal array, the main-effect estimate for
each attribute is independent of the others, whereas in the correlated case, the
independence does not hold. If attributes are correlated, misspecification leads to biased
estimates. The particular misspecification that so often occurs in conjoint is
simplification, where a number of attributes are effectively ignored. With orthogonal
arrays, the estimated coefficients for the attributes remain unbiased. In the correlated
case, misspecification results in distortions in the coefficients for both the attributes
focused on and those ignored. Therefore, orthogonal arrays play an important role of
increasing the robustness of conjoint analysis by making it less likely that coefficients
have counter-intuitive signs.

4. Conjoint simulations can explain heterogeneous preferences in a market.

Typically, part-worth functions are estimated at the individual level, and then

these are aggregated to produce estimates of market share under various conditions or

scenarios. These simulations implicitly reject the notion that one homogeneous customer

50

can explain choices in the marketplace. Instead, one is forced to deal with each customer
having an idiosyncratic preference function, or at least with an explicit clustering in
which strongly differing tastes are represented in different clusters. This practice of
preserving the heterogeneity of individuals in simulations facilitates the representation of
two important properties of markets that are difficult to achieve with other market
research techniques. These are the properties of differential substitution and dominance.

Differential substitution refers to the notion that a new competitor in a market
tends to take share differentially from those brands with which it is most similar. For
example, Diet Pepsi took most share from Classic Pepsi and Coke, and had relatively
little impact on the lemon-lime soft drink (i.e., 7-Up) category. Dominance refers to the
idea that a brand that is equal on most attributes but slightly worse than its competitor on
others gets a very low share.

Researchers understand these two phenomena and expect simulations to reflect
them in positioning and new-product studies. Unfortunately, most models of market
research have had difficulty explaining differential substitution and dominance. In
contrast, both phenomena arise naturally out of a conjoint simulation. For instance,
differential substitution occurs because individuals who like Pepsi tend to like Coke.
Generally, changes in any brand will have a greater impact on similar brands than
dissimilar ones. Dominance is represented in a conjoint simulation since a brand that is
dominated consistently loses out to that competitor and achieves almost no share.

In summary, conjoint analysis works because it is derived from a task that forces
respondents to trade off attributes in ways that may mirror actual buying behavior. The

orthogonal arrays provide not only efficacy but also strong degree of robustness against

51

misspecification. Finally, the preservation of the utility function at the level of individual
or segment permits us to simulate a market that behaves in ways we expect.
The Advantages and Limits of Conjoint Analysis

Advantages of Conjoint Analysis

Several advantages of conjoint analysis are noteworthy:
1. Under conjoint analysis, respondents perform a very realistic task - making trade-offs
between different products and services. This avoids many of the subject artifacts and
demand effect of traditional compositional approaches in that respondents’ separate
evaluations of each attribute level are aggregated to form estimates of respondent
preferences (Conjoint experiment 1997).
2. Conjoint analysis employs a decompositional model which differs from almost all
other multivariate methods in which it can be cafried out at the level of the individual
respondent. This implies that the researcher can develop a separate model for predicting
the preferences of each respondent. Most other methods of analysis take measures from
each respondent but can only perform an analysis using all respondents combined.
3. Most multivariate methods require the assumption of linear relationships between the
predictor and the dependent variables (preference). Conjoint analysis is not hampered by
this assumption. It can make separate predictions for the effects of each level of the
dependent variables and does not assume that the dependent variable increases (or
decreases) in equal amounts for each unit change in the predictor variables. Thus
conjoint analysis can handle any type of non-linear relationship, even the complex
curvilinear relationship in which one value is positive, the next negative and the third

positive again (Hair et al. 1995).

52

4. Most importantly, individual level models can be used to simulate a market, that is, to
predict how a group of respondents would respond to a product or service with a
particular set of attributes and values.

5. In a study of the residential choice behavior, Lerman and Louviere (1978) compared
the specification of conjoint analysis with 27 different utility specifications. The results
showed that each of the 27 specifications was statistically inferior to the specification of
conjoint study. They suggested that conjoint analysis can provide information that not
only upgrades the predictive power of an econometric choice model, but this information
probably could not have been obtained in any other way.

Limits of Conjoint Analysis
Some disadvantages of conjoint analysis are:

1. Some profiles may seem artificial or unlikely to respondents. This is because the use
of orthogonal arrays. For instance, in the case of orthogonal arrays a necessary and
sufficient condition that the main effects of any two factors be uncorrelated is that each
level of one factor occurs with each level of another factor with proportional frequencies.
In order to make factors with proportional frequencies, therefore, one must create some
unrealistic and unlikely profiles.

2. Compared to traditional surveys, filling out a conjoint survey is demanding for
respondents, especially when attributes, levels, or both become larger and larger. The
conjoint results acquired under such conditions may not be representative of the real life
behavior of the individual where one may have more time and motivation to deliberate on

the choice from among a small set of options.

53

3. Respondents may not consider all attributes when evaluating hypothetical situations.
Usually respondents will simplify their decisionmaking by focusing on a couple of the
most important attributes.
4. Attributes are often limited to only a few levels to keep the number of treatment
combinations from becoming too large. Generally one will reduce attributes to six or less
than six to avoid the problem of information overload.
5. Conjoint analysis is costly. Since the personal interview is most widely used approach,
it is very expensive to collect a large sample size.
Classification of Conjoint Analysis

Figure 3.2 (adapted from Carroll and Green 1995) provides a classification of the
various types of conjoint analysis. Conjoint analysis can be conducted in two major
methods: regular or traditional conjoint analysis, and hybrid conjoint analysis. In
addition, a related approach - self-explicated model - often is used in similar studies.
Regular conjoint analysis is based on a decompositional method to derive a set of utilities
for all individual attribute levels defined in profiles chosen by a special design. In the
self-explicated model, the respondent first evaluates the levels of each attribute on a 0-10
(say) desirability scale. Then, the respondent then is asked to allocate, say, 100 points
across the attributes so as to reflect their relative importance. Part-worths are obtained by
multiplying the importance weights with attribute-level desirability ratings. The hybrid
conjoint contains not only those terms in the regular conjoint but also a compositional
term derived from respondents’ self-explicated reactions to a set of attributes and their

levels (Green and Srinivasan 1990; Carroll and Green 1995).

54

 

 

 

Classification
of Conjoint
Analysis

 

 

Figure 3.2: Classification of Conjoint Analysis

 

ull Profiles
Only

 

 

 

 

Self-Explicated
and Profiles

 

 

 

1 ———>

Full Profiles

 

Partial Profiles
(Subset)

 

Traditional Conjoint
(Individual Analysis)

 

 

Continuous Variables

 

 

 

 

Monanova: Kruskal (I965);
Prefinap: Carroll (1973);
Linmap: Shocker &
Srinivasan (1977);

OLS Regression

 

 

 

Pekelrnan & Sen (1979);
Bretton-Clark: Herman
(1988);

Krishnamurthi & Wittink
(1989)

 

 

 

Constrained Attribute
Levels

 

 

 

Partially Aggregated
Model

 

 

 

Complete Set of
Full Profiles

 

 

Subset of
Full Profiles

 

 

 

 

 

 

 

Data Only

Self-Explicated I

 

 

55

Order Constraints:
Srinivasan, Jain, &
Malhortra (I983)

 

 

Componential
Segmentation: Green &
DeSarbo (1979);
Optimal Scaling:
Hagerty (I985);
Cluster Analysis:
Karnakura (1988)

 

 

 

Bayesian: Cattin, Gelland,
& Danes (1983);
Monotonic Constraints:
Van Der Lans & Heiser
(I990)

 

 

 

Hybrid Models: Green,
Goldberg, & Montemayor
(1981);

Green (1984)

 

 

J Adaptive Conjoint
rlAnalysis: Johnson (1987)

 

 

Casemap: Srinivasan
(1988);
Srinivasan & Wyner
(1989)

 

 

Current Applications on Conjoint Analysis

Conjoint analysis has been extensively applied to marketing research for the
following purposes: new product/concept development, pricing, market segmentation,
competitive analysis, repositioning, advertising, and distribution (Wittink and Cattin
1989; Wittink, Vriens, and Burhenne 1996). A few researchers have noticed that conjoint
analysis could be applied to the studies of policy analysis and public opinion (Green and
Srinivasan 1978; Louviere 1988; Shukla and Bruno 1991; Johnson 1995), but the
literature has paid very little attention to this important method. There have been
reported applications to health care system and services (McClain and Rao 1974; Parker
and Srinivasan 1976; Wind and Lawrence 1976; Chakraborty, Ettenson, and Gaeth 1994),
water quality-cost alternatives (Whitmore and Cacadias 1974), education policy and
school administration (Sparling 1977; Hooley 1981; Ramsing 1982; Zufryden 1983;
Shukla 1990), public bus service preferences, residential choice behavior (Louviere 1988:
63-67), competing values in public opinion (Shamir and Shamir 1995), and evaluation
utilization (Johnson 1995).

Since it is versatile and applicable to a wide variety of problems and there is
computer software available, conjoint analysis is a very promising method which could
be applied to the study of policy analysis and public opinion.

Summary
In a nutshell, conjoint analysis is a powerful alternative to traditional surveys. It
requires tradeoffs that are similar to those in the real decision process. It employs the
best elements of experimental design: it is decompositional rather than compositional; it

uses orthogonal arrays to reduce combination, it can simulate preferences under various

56

scenarios. The result from conjoint analysis is that we obtain more accurate information
about people’s preferences. Thus, conjoint analysis has great potential to be applied to

the study of policy analysis and public opinion.

57

CHAPTER 4

RESEARCH DESIGN

The design for this study can be divided into two parts. In the surveys we first ask
respondents the traditional survey questions and then conjoint questions. The data for
this study were collected from a convenient sample of 77 undergraduate students in
Michigan State university and a random sample of 128 Meridian Township voters in
Michigan. (A detailed description will be discussed in Step 8 of the procedure).

Step by Step Procedure of Conducting a Conjoint Analysis

Green and Srinivasan (1978), Churchill (1995), and Hu (1996) all provide some
suggestions or guidelines for conducting a conjoint survey. Based on the literature
review and the purpose of this study, we provide a step by step procedure as follows: (1).
Select Attributes and Determine Attribute Levels, (2). Select Preference Model and
Composition Rule, (3). Determine Data Collection Method, (4). Construct fractional
factorial design, (5). Select Stimulus Presentation Method, (6). Select Measurement Scale
for the Dependent Variable, (7). Collect Data, (8). Select the Estimation Method, (9).
Interpret the Results and Decide on Aggregation of Judgments, and (10). Check
Reliability and Validity.
1. Select Attributes and Determine Attribute Levels

The first step in the process involves deciding on the attributes to be used and

specifying the actual levels for each attribute when constructing the profiles.

58

The Principles of Selecting Attributes

When choosing attributes, researchers should be guided by the principles that the
attributes used should be both feasible and important to individuals. Feasible attributes
are those that the organization can do something about, in that the organization has the
technology or other resources to make the changes that might be indicated by consumer
preferences. Important attributes are those that actually affect consumer choice. The
attributes actually used can be determined by managerial judgment or by any of the other
techniques which are productive at the exploratory research stage of an investigation,
including depth interviews with individual subjects, focus groups, analysis of insight-
stimulating examples, and so on. In any single conjoint analysis study, only a handful of
the attributes that could be used will be used, so it is important that they are carefully
selected. When the number of attributes that need to be varied exceed reasonable limits
with respect to data collection problems, a series of pilot studies can be conducted. In a
typical conjoint analysis, the number of attributes actually used is less than six.

The Principles of Deciding Attribute Levels

The number of levels for each attribute has a direct influence on the number of
attributes respondents will be asked to evaluate and consequently on the burden placed on
each subject. In general, we wish to minimize that burden; however, at the same time, we
wish to end up with good estimates of the utility of each attribute level. Our ability to
generate good estimates requires that the number of attributes be relatively large versus
the number of parameters that need to be estimated, and the number of parameters in turn
depends on the preference model being used. Thus, it is an essential question to ask: Is

the model linear in the sense that more or less of the attribute can be expected to be most

59

desired, or is it nonlinear in a systematic way, or could there be a very nonsystematic
relationship between preference and attribute levels?

Another factor that affects the choice of attribute levels is their effect on consumer
choice. Using levels which are similar to those in existence increases respondents’ belief
in the task and the validity of their preference judgments. In addition to that, using
attribute levels which are outside the range normally encountered decreases the
believability of the task for respondents but can increase the accuracy by which the
parameters can be estimated statistically. The general recommendation is to make the
ranges for the various attributes somewhat larger than what is normally found but not so
large as to make the options unbelievable.

Attributes and Levels Used in This Study

Based on the feasibility and importance considerations, in the current study four
determinant attributes are used. They are four largest spending categories in the state
government of Michigan: education, health and welfare, public safety and corrections,
and natural resources and environmental control. Each of these attributes has three
spending levels: increase spending 5 percent, decrease spending 5 percent, or maintain
current spending levels.1 The four categories explain 80.2% of total state expenditures
(actual spending) in the 1996 fiscal year. Meanwhile, the four categories also represent

the largest increase in spending items from 1990-96 fiscal years.2 Based on Governor

 

' We consider the most widely discussed issue “taxes vs. government spending.” However, adding one
more attribute “taxes” will lead to a increase in minimum profiles from 9 to 16; not even mention double
the minimum profiles (32). Such large profiles will definitely cause the problem of information overload.
Since this a pilot study, we think it is better to keep the format as simple as possible. Therefore, we only
use 4 determinant attributes in this study.

2 Calculated based on the report of the Department of Management and Budget in Michigan.

60

Engler’s platform, 3 the four categories should be an appropriate indicator of his budget
priorities. Furthermore, the four categories include some of the most salient issues of
today such as education, Medicaid, health care, crime, and environmental quality.
Studies have showed that respondents usually decide and answer issues most salient to
them (Sniderman 1993: 221). Hence, focusing on these salient issues can help
respondents answer questions more easily and reduce the information overload problem.

In addition, a pretest of 10 cases was conducted to decide whether the design is
reasonable and appr0priate and whether respondents can correctly interpret the attributes
and levels. For instance, we attempted to determine whether the breakdown of the
expenditures was reasonable and whether the levels of each factor were appropriate.

2. Select Preference Model and Composition Rule

Mentioned earlier, generating good estimates requires that the number of
attributes be relatively large versus the number of parameters that need to be estimated,
and the number of parameters depends on the preference model being used. In other
words, we should decide which preference models will be applied: vector, ideal point, or
part-worth models?

However, choosing a preference model is not so simple as it appears to be (linear,
quadratic, or piecewise). First, Green and Srinivasan (1978) point out that the complexity
(and number of parameters required) increases from vector to part-worth. This results in
the precision of estimated parameters decreasing from vector to part-worth because more

parameters must be estimated from the same number of observations. On the other hand,

 

’ According to Governor John Engler’s State of the State Address (1998), his next steps to make Michigan
first in the 21 st century includes cutting taxes, protecting the environment, education, strengthen families,

61

ability to fit an arbitrary pattern of true preferences increases from vector to part-worth
(i.e., the vector model can fit only a linear pattern of true preferences, whereas the part-
worth model can fit linear, quadratic, or any other pattern). Since these factors tend to go
in opposite directions, no a priori recommendations can be made on which model to use.
Later studies have proved that simple models predict individual preferences better than
complex models (Pekelman and Sen 1979; Cattin and Punj 1984).

In addition to selecting a preference model, whether interactions should be
included in the preference model also affect the model’s ability to accurately predict
people’s preferences. This issue involves deciding the composition rule, that is, whether
the simple additive model is enough to accurately predict preferences or we should
consider using multiplicative, distributive, or dual-distributive model. Once again,
according to previous studies, the conclusions are that the simpler models perform better
than the more complex models (Akaah and Korgaonakar 1983; Neslin 1981, and
Goldberg 1967).

Even though these findings seem to favor simple models, Hagerty (1986) showed
that simple models did predict better for individuals ’ preferences but complex models
predict better for market share. Now it seems that the choice among preference models
depends on the trade-off between the more complex interaction models and their costs.
Specifically, the major obstacle for using a complex model is the increased burden placed
on a respondent by the need for a large design matrix to obtain the degrees of freedom

necessary to estimate a more complex model. In order to obtain more accurate

 

and fighting crime.

62

prediction, for instance, Hagerty’s analysis used up to 32 profiles. Such design definitely
will cause respondents’ tendencies to become careless when rating or ranking profiles.

Based on above considerations, the additive part-worth model is used in this
study for several reasons. First, so far there is no theory can indicate that people’s
preferences for the four spending categories (education, welfare, crime, and natural
resources) are linear, curvilinear, or piecewise. Nevertheless, the part-worth model
provides the greatest flexibility in allowing different shapes for the preference function
along each of the attributes. Second, since the design in this study only has three levels
for each attribute, the part-worth model is indistinguishable from the quadratic (ideal
point) form (Hagerty 1986). Third, since no theory shows that there are interactions
among the four spending categories, an additive model should be enough to accurately
predict preferences. In practice, conjoint studies rarely apply models beyond the additive
form because the simple model is easier to apply and the simpler model (additive)
performs better than the more complex models such as the multiplicative or distributive
models (Akaah and Korgaonakar 1983). Hence, the additive part-worth model should
work well in this study.
3. Determine Data Collection Method

Two major data collection procedures are used in conjoint analysis: full profile
approach and two-factor-at-a-time procedure.

The fitll-profile approach (or concept evaluation) uses the complete set of factors
(see Table 4.1 for a comparison of full profile and two-factor-at-a-time procedures).
Respondents provide rank orders (or rating) of preference for product concepts which

differ simultaneously with respect to all attributes been studied. The full profile approach

63

has the advantages of greater “realism,” since respondents are choosing among concepts
which are more elaborately specified, and at least theoretically, of being able to quantify
interactions among attribute. The major drawback of this procedure is the possibility of
information overload and the resulting temptation on the part of the respondent to
simplify the experimental task by ignoring variations in the less important factors or by
simplifying the factor levels themselves. The conjoint results acquired under such
conditions may not be representative of the real life behavior of the individual where one
may have more time and motivation to deliberate on the choice from among a small set of
options. Because of the problem of information overload, the full profile procedure is
usually limited to six factors in any specific sort.4

The two-factor-at-a-time procedure (or paired comparison) considers factors on a
two-at-a-time basis. The respondent is asked to rank the various combinations of each
pair of attribute levels from most preferred to least preferred. The two-factor-at-a-time
approach is simple to apply and it can reduce information overload on the part of the
respondent. However, two-factor-at-a-time approach sacrifices realism; people are
usually unclear as to what should be assumed about the t-2 attributes that are not being
considered in a specific evaluation task. In addition, there is a tendency for respondents
either to forget where they are in the table or to adopt patternized types of responses, such
as always attending to variations in one attribute before considering the other.

Furthermore, studies have shown that the full profile approach usually overstates

the importance of the most preferred attribute while the two-factor-at-a-time approach

 

‘ To overcome the information overload problem, one could use “bridge-type” factors design. See Green
and Srinivasan (1978) for a discussion.

usually overstates the importance of the least preferred attribute (Green and Srinivasan
1978).

In this study, the full profile approach is used to collect data because it provides
greater realism and entails fewer judgments by the respondent (when implemented by
various kinds fractional factorial design). In addition, our design only has four factors
with three levels in each attribute, hence the problem of information overload should be
reduced to a minimal or negligible level.

Table 4.1: Alternative Data Collection Methods

 

 

 

 

Full Profile Approach Two-Factor-at-a-Time
Profile A Education
Education 5% Health 8 Welfare -_5_°/_o % 5%
-5% 7 3 1
Health 8. Welfare -5%
0% 8 5 2

Crime 8. Public Safety 5%
5% 9 6 4
Natural Resources 0%

 

 

Note: 1 denotes the most preferred combination and
9 denotes the least preferred combination.

4. Construct Fractional Factorial Design

As previous discussed, it would be too costly to carry out a conjoint survey if we
use the full set of profiles. For example, there are four attributes (factors) in this study:
education, welfare and health, public safety and corrections, and natural resources. Every
attribute has three levels: decrease 5% spending, status quo, and increase 5% spending.
If we array all possible combinations of the four attributes, a total of 3x3x3x3 = 81
alternatives would have to be tested. Obviously, it would be too costly to carry out an

evaluation study of this magnitude and it would cause confusion and fatigue among the

65

respondents. An appropriate way is to use various types of fractional factorial design to
reduce the number of combinations to a manageable size while at the same time
maintaining orthogonality. The most common type of fractional factorial design involves
the case in which all main effects (single factor) and all two factor interactions can be
estimated on an uncorrelated (orthogonal) basis. However, one assumes the higher
interactions (involving three or more factors) are negligible and can be ignored.’ In this
study, the main-eflects plan is adopted which permits one to estimate only the main
effects, assuming negligible interactions (Green 1974).

Orthogonal main-effect plans can be built for symmetrical factorial experiments
involving (5" - 1)/(s -1) factors, each having 3 levels, with s“ combinations, where s is a
prime or the power of a prime number (Addelman 1962). In this study, s=3, n = 2 and (3"
- 1)/(s -l) = 4. So, we only need sn (3’) = 9 combinations for this study. If we used any
fewer than this number of profiles, we would not be able to derive a set of utilities. And
if we used only the minimum number, we could not check the respondent’s internal
consistency in evaluating the profiles (Curry 1997). Therefore, it is generally
recommended that we include 1.5 to 2 times the minimum number of profiles in the
ranking or rating task. A necessary and sufficient condition that the main effects of any
two factors be uncorrelated is that each level of one factor occurs with each level of
another factor with proportional frequencies (Addelman, 1962).

For comparison purposes, 10 profiles will be used in a random sample while 18

profiles will be used in the convenient sample. Such design is based on the consideration

 

5 For a discussion of fractional factorial design, see Green, Carroll, and Carmone (1978) and Louviere
(1988), Chapter 2.

66

that 10 profiles are more manageable in a mail survey (voter sample). On the other hand,
it is not too difficult to ask students rank ordering 18 profiles in the classrooms (see
Appendix for both voter and student survey forms). Therefore, in the voter sample we
cannot check respondent’s internal consistency while in the student sample we could
check respondent’s internal consistency.
5. Select Stimulus Presentation Method

Step 5 in the process involves selecting. the form of presentation of the stimuli and
the nature of the judgments to be secured from subjects. The presentation of the
hypothetical stimuli in the full profile approach has involved variations and combinations
of three basic approaches: verbal description, paragraph description, and pictorial
representation. The paragraph description is used in both voter and student samples.
6. Select Measurement Scale for the Dependent Variable

Related to the issue of the form of presentation of the stimuli is the nature of the
judgments that will be secured from respondents; that is, the measurement scale for the
dependent variable. The two most common approaches measured respondents’
preferences for each alternative are: nonmetric (rank order, paired comparison) or metric
(rating scales assuming approximately interval scale or ratio scales). Some of the main
reasons advanced by those using rank-order judgments are their case of use by subjects,
ease of administration, and a desire to keep the judgment task as close as possible to a
consumer’s behavior while actually making decision. In addition, ranking data are likely
to be more reliable, since it is easier for a respondent to say which he or she prefers

something more as compared to expressing the magnitude of his or her preference. Those

67

using rating scales believe that they are less time-consuming, are more convenient for
respondents to use, and are easier to analyze (through the ordinary least squares).

For comparison purposes, we used both ranking and rating approaches to collect
data. In all the first round of surveys, ranking is used. In the second round of voter
surveys, half of the respondents were asked to rate 12 profiles while another half of the
respondents were asked to rank order 10 profiles.

7. Select the Estimation Method

Step 7 in the execution of a conjoint analysis involves selecting the technique by
which the input data will be analyzed. Four estimation methods are available for this
study: MONANOVA (Kruskal 1965), PREFMAP (Carroll 1973), LINMAP (Shocker and
Srinivasan 1977), and OLS regression. For the comparison purpose, both LINMAP and
OLS were used to estimate all input data. LINMAP is one of the most popular nonmetric
conjoint estimation procedures. The use of linear programming enables LINMAP to
obtain global optimum parameter estimates, whereas there is no guarantee that other
procedures would achieve global optimums. The OLS regression approach to conjoint
analysis offers a simple, yet robust method of deriving alternative forms of respondent
utilities. It is commonly available, inexpensive to use, and easy to interpret estimation
method. The attractiveness of the OLS model is a result of the ability to scale respondent

choices using rating scales.

68

8. Collect Data

Samples

The data for this study were collected from a convenient sample of 77
undergraduate students in Michigan State university and a random sample of 128
Meridian Township voters in Michigan.

Purposive (nonrepresentative) sample: A purposive sample of undergraduate
students were collected from the Michigan State University. In July 1997, we collected
77 completed surveys from the students enrolled in the second summer school session. In
August 1997, we obtained 22 completed follow-up surveys out of the 79 students.“ Even
though a nonrepresentative sample cannot be generalized to make inferences about the
distribution of attitudes in the total population, several advantages make it attractive: (1)
it is less expensive, (2) it can collect a large number of cases at relatively small cost, (3)
researchers can use questions that reflect their theoretical interests, and (4) small
subcultural groups or other special populations can be studied (Abramson 1983: 23-26).

Random Sample: A random sample of 128 registered voters were collected from
the Meridian Township in Michigan by using random numbers created by uniform
distribution through the software called Crystal Ball. 7 To maximize the response rate,
the subjects were first contacted by telephone and asked to participate. Several days later,
the respondents who agreed to participate received a survey package by mail.8 The

package included a cover letter, the questionnaires (both traditional and conjoint

 

" Originally, we planed to collect the follow-up sample in a two month interval. However, the summer
session only lasts one and a half months, hence the data were collected in August (one month interval).

7 Crystal Ball is software for business managers. 11 provides all kinds of simulations for decision making.
8 It is extremely difficult to carry out conjoint survey by telephone. The best way to conduct conjoint
survey will be the personal interview. Due to the budget constraints, we could only afford the mail survey.

69

questions), a postage-paid return envelope, and a lottery ticket as an incentive. The cover
letter thanked them for participating and asked that they returned the completed survey
within 10 days. In the late November 1997, the telephone solicitation netted 421
completed contacts, of which 25] agreed to participate. Later, we obtained the total
number 128 completed surveys, a response rate of 50 percent of those who agreed to
participate or 30 percent of the original completed contacts.

Approximately two months later, we sent questionnaires to those who previously
agreed to participate in our follow-up surveys. For comparison purposes, we divided 88
agreed subjects into two groups (44 subjects in each groups). In the first group, we asked
respondents to rank order 10 profiles but with a different orthogonal array. In the second
group, we asked respondents to rate 12 profiles in a 0-100 points scale. We obtained 34
completed surveys from ranking data while 27 completed surveys from rating data.

Format

In both student and voter surveys, we first asked 5 traditional questions before
conjoint questions. In student surveys, traditional questions were presented with the
format of unbounded financial resources, such as “Generally speaking, do you think the
State of Michigan should: increase overall spending, decrease overall spending, or keep
overall spending at current levels?” and “In the area of (i.e., education), you
would prefer to see the State of Michigan: increase spending in the area of education,
decrease spending in the area of education, or maintain education spending at current
levels?” In voter surveys, traditional questions were presented by asking respondents
expressing their own ideal spending levels (except first general question), such as “In the

area of (i.e., education), you would prefer to see the State of Michigan:

70

increase spending in the area of education by ___%, decrease spending in the area of
education by ___°/o, or maintain education spending at current levels?”

In the conjoint questions, we asked respondents to rank order 18 profiles (student
sample) or 10 profiles (voter sample) from the most preferred to the least preferred. In
the rating survey (second round for voter), we asked respondents to rate 12 profiles in a 0-
100 points scale. This rating approach will allow respondents filling in tied scores. In
the end of the surveys, respondents were asked to fill in the demographic information in
order to analyze group differences.

9. Interpret the Results and Decide on Aggregation of Judgments

Step 9 involves interpreting the results and deciding if the responses from
individual subjects will be aggregated and, if so, how?

The Principles of Making Aggregation Judgments

Although it is possible to derive the utilities for each level of each attribute at the
individual level, individual results are very difficult for researchers to use for developing
marketing strategy. The other extreme is to pool the results across all respondents and
then to estimate one overall utility function. The middle ground is to form segments from
groups of respondents in such a way that the models for the groups will have predictive
power close to that found for the individual-level models while having some clear
marketing strategy implications for researchers. This raises the question of how these
groups should be formed. Most typically it is done by forming segments that are
homogeneous with respect to the benefits that the respondents want from the product or

service. Operationally, this often translates into estimating utilities for the individual-

71

level models and then clustering respondents into groups that are homogeneous with
respect to the utilities assigned to the various levels of the individual attributes.

An attractive feature of conjoint analysis is that it allows market share predictions
for selected product alternatives. For example, a common choice rule is the first choice
rule, which assumes that each respondent chooses the object with the highest predicted
preference. Given the estimated utilities for each level of each attribute, the researcher
then can investigate which of several product options being considered is likely to appeal
most to respondents and what may also be the share of preference for each of other
options. In this study, we ran several simulations for gubernatorial election in Michigan
and welfare policy evaluation. In addition, the researcher can also attempt to link
individual part-worth utilities with selected consumer characteristics, such as whether
Democrats have higher utilities for an increase in welfare expenditures than Republicans
or not.

Interpreting Results

We analyzed the results by comparing traditional surveys results with conjoint
result: how different are respondents’ preferences in traditional and conjoint surveys?
We also examined the differences among different groups such as gender, religion, party
identification, etc. under traditional and conjoint results. In addition, we simulated the
gubernatorial election in Michigan to see which candidates (John Engler or Jeffrey
Fieger) would win the election based on their positions on the four policy areas. We also
evaluated respondents views of current welfare reform in Michigan: Do voters in

Meridian Township satisfy the current welfare spending level?

72

10. Check Reliability and Validity

Test of Reliability

Conjoint analysis can provide more accurate information on people’s preference.
But can it also provide reliable results? Public opinion studies have shown that people
tend to have unstable responses to survey questions. The major explanations for the
response instability are: (1) people do not have meaningful beliefs (Converse 1964), or
(2) people have true attitudes but survey tools cannot detect it due to measurement errors
(Achen 1975). Thus, it is interesting to know if a different survey tool (conjoint analysis)
can provide better reliability.

In this study, all the surveys asked respondents traditional survey questions and
conjoint questions. In conjoint analysis, several different tests were used to check
reliability. Reibstein, Bateson, and Boulding (1988) provide four types of reliability for
conjoint analysis:

(1) Reliability Over Time asks: “Would the answers be the same at different time?

(2) Reliability Over Stimulus Set asks: “Would the answers be the same if a different set
of profiles had been used?”

(3) Reliability Over Attributes Set ask: “Would the utilities for a given set of attributes
have been the same if these attributes had been included in a study with other attributes?”
(4) Reliability Over Data Collection Procedure ask: “Would the answers obtained have
been the same if a different data collection procedure had been used?”

In our study, we evaluate the following three criteria: reliability over time,
reliability over stimulus set, and reliability over data collection method. First, we test

reliability by using alternate forms method with spaced testing: after finishing a conjoint

73

analysis task, a subset of the respondents will be approached after a period of the time
(one or two months) and asked for preference judgments on a second set of profiles. The
second profiles (“alternate form”) would also have n descriptions from the same factor
levels as the first set but would avoid duplication of profiles from the original one.
Product moment correlation and Spearman rank order correlation are used to test
reliability.

Tests of Validity

Except for checking reliability, we also test validity (only in student sample) by
the following ways:
I. Cross-validity: the ability of the model to predict the ranking (or first choice) of a
holdout set of profiles (only for the student sample). In addition, data used for the
reliability tests also provide a method of cross-validation. The parameters of the
preference function estimated from the first set of preference data can be used to predict
the preferences for the second set. The predicted preferences can be correlated to the
actual to obtain a measure of cross-validity.
2. Predictive validity: predictive validity is measured by holdout profiles (student sample
only). Holdout profiles are additional profiles which the respondent evaluates beyond
those generated by actual conjoint design. In the student sample, we include three
holdout profiles in our design. In particular, we use first 15 profiles (Profile 1-15) to
predict last three profiles (Profile 16-18).

Summary
In this study, we provided a 10 step procedure to conduct conjoint analysis: (1)

Four determinant attributes are were chosen: education, health and welfare, public safety

74

and corrections, and natural resources - and each of these attributes with three spending
levels: increase spending 5 percent, decrease spending 5 percent, or maintain current
spending levels, (2) An additive part-worth model was selected, (3) A full profile
approach was utilized to collect data, (4) 10 profiles were evaluated in voter surveys
while 18 profiles were evaluated in student surveys, (5) A paragraph description was used
to present stimulus, (6) Rank order and rating were chosen to measure dependent
variables, (7) Data were collected from a convenient sample of 77 undergraduate students
in Michigan State university and a random sample of 128 Meridian Township voters in
Michigan, (8) LINMAP and OLS were utilized to estimate part-worth utilities, (9)
Traditional and conjoint results were compared, part-worth utilities differences among
groups were analyzed, and simulations for gubernatorial election and policy evaluation
were executed, and (10) Product moment correlation and Spearman rank order correlation

were adopted to test reliability while holdout choices were used to test validity.

75

CHAPTER 5

RESULTS AND DISCUSSIONS

In the student sample, respondents were asked to rank order 18 profiles in both
first and second round of surveys but with a different orthogonal array in the second
round of survey. In the voter sample, respondents were asked to rank order 10 profiles in
the first round of survey. In the second round of the survey, half (N = 44) of voters were
asked to rank order 10 profiles while another half (N = 44) were asked to rate 12 profiles
in a 0 - 100 points scale. Both ranking and rating data were estimated by ordinary least
squared regression (by SPSS 8.0) and LINMAP (by CONJOINT LIN MAP). In this
chapter we first discussed the part-worth utilities, relative importances of attributes,
overall utilities of profiles, and goodness-of-fit measures obtained from Meridian
Township voters and undergraduate students in Michigan State University. The logic of
LINMAP and OLS has been discussed in Chapter 3. Since OLS is widely used procedure
in political science, the interpretations of conjoint analysis will be emphasized on OLS
results. Then, the results from both traditional survey and conjoint analysis were
compared by using different simulation models.

Interpreting Conjoint Results from Voter Sample

Interpreting the Part-worth Utilities from Voter Sample

The additive part-worths model was used in both voter and student samples. For
each individual, the part-worths (weights or regression coefficients) were estimated by
OLS. The purposes of part-worths were used to determine (1) the importance of levels of

attributes, (2) the relative importances of attributes, and (3) the overall utilities of profiles.

76

The OLS regressions were estimated at the individual level, for example, resulting in 128
equations for voters (first round of survey). Furthermore, these 128 “individual
equations” were averaged, resulting in an “aggregate equation” for voters. Then, the
aggregate equation is used to calculate the relative importances of attributes and estimate
the overall utilities for 10 profiles.

Recalled Equation 3.16 (in Chapter 3), the OLS model can be expressed:
Up = C+ZZW”X” +ep
’ ,I

Where:
U p = the overall utility to measure of the p-th profile,
C = constant or intercept for profile P,
Wij = the part-worth utility of i-th level j-th attribute estimated in the regression,
X ij = the dummy variable representing the presence or absence of attribute ,, level, in
alternative X, and
ep = error term for profile P.

For illustration, let us consider the example of voters in the first round of survey
with an orthogonal array of 10 profiles (see Table 5.1). Now for Wij and X ij» the i= 0,
1, and 2 ( O = -5%,1= 0%, and 2 = +5%) whilej= 1, 2, 3, and 4 (l = education, 2 =

welfare/health, 3 = crime/public safety, and 4 = natural resources/environmental control).

77

Table 5.1: Orthogonal Array for Voters (First Round of Survey)

 

Profile Education Welfare Crime Natural
1 0 0 O 0
2 O 1 1 2
3 O 2 2 1
4 1 O 1 1
5 1 1 2 0
6 1 2 O 2
7 2 O 2 2
8 2 1 0 1
9 2 2 1 0
1 O 1 0 2 0

 

From Table 5.1, for instance, the Profile 1, 4, and 7 can be written as: 1

 

U1 = C + W01X01 + W02 X02 + W03 X03 + W04X04 + 81 (5.1)
U4 = C + W11 X11 + W02X02 + W13 X13 + W14X14 + e4 (52)
U7 = C + W21X21 + W02 X02 + W23 X23 + W24 X24 + e7 (53)

In these cases, different profiles represent different spending preferences and can
be expressed as follows:

Profile 1 indicates that a voter prefers cutting spending across-the-board, Profile 4
shows that a voter favors a decrease in welfare spending and maintaining spending
education, crime and natural resources at current levels, and Profile 7 reveals that a voter
prefers increasing spending in education, crime, and natural resources while cuts in

welfare.

78

Part-worth Utilities of Attributes

Table 5.2 reveals the voter’s part-worth utilities from OLS and LINMAP
estimations. Although the scaling on each set of part-worth utilities differs due to
computational differences, the utility estimates from each technique are quite similar.l
This can be seen by comparing the %AUtility columns in Table 5.2, which shows the
percent change in part-worths that come from either spending increases or decreases
when compared with the current level of spending.

In the area of education (see Figure 5.1 and 5.2) , for example, both OLS and
LINMAP show that voters suffer a significant utility loss from decreases in education
spending (utility drop from 0.357 to -l .302 in OLS and from 3.091 to -lO.833 in
LINMAP) and a mild support in utility from increasing education spending (utility
increase from 0.357 to 0.945 in OLS and from 3.091 to 7.741 in LINMAP). In the areas
of welfare, crime, and natural resources, voters are better off when spending on status
quo while worse off when either spending cuts or increased. The results indicate that in
general Meridian Township voters are satisfied with most of current spending levels in

Michigan.

 

' Cattin and Wittink (1976) report that the difference between OLS and LINMAP is about 0.03 units in
terms of Pearson’s rho, with OLS being the better approach when the attribute weights are normally
distributed (compensatory model) and LINMAP being the preferred method when the attribute weights
show a lexicographic structure (noncompensatory model). Jain et al. (1978) report that MONANOVA,
LINMAP, LOGIT, and OLS produce roughly the same level of cross-validity, with LOGIT and LINMAP
being slightly preferred procedures.

79

Table 5.2: Part-worth Utilities from OLS and LINMAP (Voters)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

OLS LINMAP
Factor Level Utility %A Utility Utility %AUtllIty
-5% -1.302 -464.5% -10.833 -450.5%

Education 0% 0.357 3.091
5% 0.945 164.5% 7.741 150.4%

Relative Importance 52.44% 50.44%
-5% -0.052 -115.0% -0.672 -121.6%

Welfare/Health 0% 0.344 3.105
5% -O.292 -185.0% -2.433 -178.4%

Relative Importance 14.83% 15.04%
-5% -0.250 -178.2% -2.680 -200.4%

Crime/Public Safety 0% 0.320 2.668
5% -0.070 -121.8% 0.012 -99.6%

Relative Importance 13.31% 14.52%
-5% -0.447 -216.2% -4.587 -265.2%

Natural Resources 0% 0.385 2.776
5% 0.062 -83.8% 1.810 -34.8%

Relative Importance 19.42% 20.00%

Constant 5.521 N/A

N= 128 128

 

80

Part-Worth Utility

Part-Worth Utility

Figure 5.1: Part-Worth Utilities for Voters (OLS)

 

1.5

1.04

-.5 I

I
_a
o

 

-1.5

 

 

 

EDU

WELFARE

CRIME

NATURAL

-5% 03/0 +5%

Spending Level

Figure 5.2: Part-Worth Utilities for Voters (LIN MAP)

 

 

 

 

 

10
0 I”
-10 . EDU
WELFARE
CRIME
-20 NATURAL
-5% 0% +5%

Spending Level

81

 

General Notes on Interpreting Part-worth Utility

The part-worth model indicates that the attribute levels are categorical and that no
assumption is made about the relationship between the attribute and the rank and scores.
The input data are scaled so that for each attribute, the utilities of the levels sum to zero.
In other words, the average utility for the levels of the attribute is zero (i.e., the average
utility for education is: -1.302 + 0.3 57 + 0.945 = 0 in OLS), and the actual part-worths
indicate how much better or worse than average that level is. It should be noted that these
utilities are all relative to the other levels of the attribute. That is, while we can state that
level A is preferred to level B, and we can also state the amount that it is preferred by, we
cannot say in any absolute way that level B is disliked. For instance, if a level of a
attribute has a negative average utility, this only indicates that the level is less preferred
than the average level for that attribute, and not that it is disliked in any absolute sense.

For example, consider the area of education with the following spending levels:
+100%, +50%, and +5%. In this case, it is safe to bet that +5% would have a negative
utilities for most respondents. However, this does not imply that they dislike +5%. If we
use different levels: +5%, 0%, and -5%, it is possible that the same respondents would
have highly positive utilities for +5% - even though their absolute “liking” for +5% did
not change across the two studies.

It should also be mentioned that interpretation of conjoint utilities is true of all
types of conjoint analysis, no matter how they are scaled (i.e., how you define levels).
For example, one method of scaling sets the utility for the “worst” level of a attribute to

zero. However, this does not mean that the lowest level has no utility to the respondent,

82

just that it is X utiles worse than the next lowest level. Furthermore, it does not imply
that the worst level of each attribute has the same utility, even though they all receive a
utility of zero.

Relative Importance of Attributes

The second purpose of part-worth utility is to determine the relative importance of
attributes. Relative importance is based on the assumption that the larger the difference
between maximum and minimum level utilities within attribute, the more determinative
and salient the attribute is in the overall evaluation of profiles. The relative importance of

an attribute can be calculated from Equation 3.1:

_ Max(X,.,. ) - Min(X,,. )
»’ T Z[Max(X,,) - Min(X..)]

ll

 

R] x 100%

Where
RIj = relative importance of attribute j.
Max (V ij) = maximum level (i) utility in attribute j.
Min (V,,) = minimum level (i) utility in attribute j.
For example, the relative importance for education in OLS is:

0.945 - (-l.302)

R1, = x 100%
0.945 — (—l.302) + 0.344 — (—0.292) + 0.320 — (—0.250) + 0.3 85 — (—o.447)

 

= 52.44%
As seen in Table 5.2, Figure 5.3 and Figure 5.4, for the attributes and levels
tested, Meridian Township voters view education as the most important attribute in their

evaluation of budget priorities with a relative importance rating of 52.44 percent in OLS

83

and 50.44 percent in LINMAP. For the attributes and levels tested, although natural
resources is the second most important attribute (19.42 percent in OLS and 20 percent in
LINMAP), its rating is only slightly higher than welfare and health (14.83 percent in OLS
and 15.04 in LINMAP) and crime and public safety (13.31 percent in OLS and 14.52 in
LINMAP). One thing should be mentioned: Because the relative importance is
determined by the range rather than direction of levels, it should be noted that the highest E
relative importance rating could go either way: the highest attribute’s relative importance

rating could be the most desirable or least desirable attribute.

 

84

Figure 5.3: Relative Importance of Attributes for Voters (OLS)

 

60

50!

40-1

304

201

101

 

 

 

Relative Importance

education crime and public saf
welfare and health natural resources

Attribute

Figure 5.4: Relative Importance of Attributes for Voters (LINMAP)

60

 

50¢

40!

301

20'

 

 

 

Relative Importance

 

10 .
EDU WELFARE CRIME NATURAL

General Notes on Interpreting Relative Importance of Attributes

Researchers usually use average utilities to compute the importance of each
attribute, as seen in equation 3.1. Although this form of analysis appears valuable and
informative, there are several problems inherent in it. Using the example in Table 5.2,
importance is typically interpreted as follows: in OLS, education has the greatest impact
on voter’s preference (52.44%), followed by natural resources (19.42%), welfare and
health (14.83%), and crime (13.31%). This statement may be incomplete and misleading:
If we changed the range of levels of an attribute, we would expect its range of utilities to
change and therefore its relative importance. For instance, if we tested spending levels
of natural resources from 0% to +5% (omitting -5%), the utility range for natural
resources would have been just 0.323 utiles, making it the least important among the
attributes. Therefore, it is more precise to qualify the earlier statement of relative
importance by saying that for the attributes and levels tested, education has the greatest
impact on voter’s preference, followed by natural resources, welfare and health, and
crime. The correct way to compute the importances is to compute them for each
respondent and then average across respondents. Since all attributes had identical levels,
the relative importances in this study did not suffer the problems of misleading and
incomplete information.

Overall Utilities of Profiles

The third purpose of part-worth utilities is to decide the overall utilities of

profiles. To illustrate how to derive the overall utility, recall Equation 5.1 to 5.3 (Profile

86

l, 4, and 7). We could obtain the overall utilities by replacing Wij and X”. with estimates
from the Table 5.2.
U, = 5.521 - 1.302*Edu - 0.052*Welf- 0.250*Crime - 0.447*Natura1
= 3.470
U4 = 5.521 + 0.357*Edu - 0.052*Welf+ 0.320*Crime + 0.385*Natural
= 6.531
U7 = 5.521 + 0.945*Edu - 0.052*Welf - 0.070*Crime + 0.062*Natural
= 6.406
Note: Edu, Welf, Crime, and Natural are equal to 1 because they are dummy variables (
l= presence, 0 = absence).

Table 5.3 reveals the overall utilities for 10 profiles and the simulations from
three models in OLS2 (there is more detailed discussion in Chapter 7). The maximum
utility model is simply the probability of choosing a profile as the most preferred. The
BTL (Bradley-Terry-Luce) model computes the probability of choosing a profile as the
most preferred by dividing the profile’s utility by the sum of all the simulation total
utilities. The logit model is similar to BTL but uses the natural log of the utilities instead
of the utilities. The overall utilities show the profile 8 has the highest overall utility with.
a score 6.9 while profile 1 has the lowest utility with a score 3.5. This means that among
the 10 profiles, voters most prefer an increase in education spending, cuts in crime
spending, and maintaining welfare and natural resources at current levels while voters

least prefer cuts in across-the-board. Likewise, from the three simulation models we

 

2 We only report the OLS results because LIN MAP produce the similar results in maximum utility model
but cannot use BTL and logit models.

87

know that the profile 8 is the most preferred option with a probability 100 percent in

maximum utility model, 12.63 percent in the BTL model, and 27.89 percent in the logit

 

model.
Table 5.3: Overall Utilities of 10 Profiles for Voters (OLS)
Profile Edu Welf Crime Natur Overall Utility llllax Utility (%) BTL (%) Loglt (%)
1 -5% -5% -5% -5% 3.5 0.00 6.31 0.86
2 -5% 0% 0% 5% 4.9 0.00 8.99 3.78
3 -5% 5% 5% 0% 4.2 0.00 7.71 1.87
4 0% -5% 0% 0% 6.5 0.00 11.88 18.47
5 0% 0% 5% -5% 5.7 0.00 10.37 8.08
6 0% 5% -5% 5% 5.4 0.00 9.82 5.94
7 5% -5% 5% 5% 6.4 0.00 11.65 16.30
8 5% 0% -5% 0% 6.9 100.00 12.63 27.89
9 5% 5% 0% -5% 6.0 0.00 10.99 11.37
10 0% -5% 5% -5% 0.00 9.65 5.44

5.3

 

Measures of Goodness of Fit

SPSS (OLS) provides Pearson’s R and Kendall’s tau statistics to indicate how
well the model fits the data. Pearson’s R and Kendall’s tau statistics are correlations
between the observed and estimated preferences based on different assumptions and
formulae.3 In many conjoint analyses, the number of parameters is close to the number of
profiles rated, which will artificially inflate the correlation between observed and
estimated scores. Therefore, these coefficients may be very high in the voter sample. As
seen in Table 5.4, both Pearson’s R and Kendall’s tau are 1.000, a perfect correlation
between observed and estimated scores. The perfect correlations demonstrate that the

conjoint model fits the data well. With the caveats we just mentioned, it is better to look

 

3 Pearson’s R (product moment correlation) is the standard correlation statistics reported in all OLS results.
Kendall’s tau is the difference between proportions of the pairs in “right order” and “wrong order.” A tau
of 1.0 indicates a perfect rank order, a tan of -1 .0 indicates a perfect negative relationship, and a value of
zero indicates an unrelated ordering.

88

at other ways to measure goodness of fit and this will be discussed later in the chapter of
reliability and validity tests.

In a goodness of fit measure, LINMAP checks how the respondent evaluated each
pair of stimuli (profiles).4 Then it uses the respondent’s utility function to predict how he
or she will evaluate each of these pairs, and compares it to the actual results. Finally, it
calculates the percentage of all the pairs that were violated, for example, not consistent
with the predictions. In Table 5.4, we see 96.9 percent of respondents violated 0-5
percent pairs and there are only 4 (3.1 percent) respondents violated 5 percent or more
pairs comparisons (default threshold is 15 percent). Likewise, the result indicates that the
conjoint model fits the data well. It should be noted that the percentage of pairs violated
will typically be much larger when rating scales are used as opposed to a ranking
procedure. Ratings typically contain a numbers of ties. Paired comparisons involving

these tied levels lead to violations, since their predicted utilities are rarely equal.

 

‘ LINMAP used to measure goodness of fit by producing a Index of fit (C'). Index of fit is a poomess of fit
measure and can be interpreted as the ratio of the average poomess of fit (unexplained variation) to the
average goodness of fit (explained variation). However, the latest version CONJOINT LINMAP only
reports the percentage of paired comparisons that were violated.

89

Table 5.4: Goodness of Fit Measures for OLS and LINMAP

 

OLS
Pearson's R 1.000 Sig.= 0.000
Kendall‘s tau 1.000 Sig.= 0.000
LINMAP

 

Distribution of Violations (%)

 

 

Pairs Violated (%) N Percent
0 to 5 124 96.9%
6 to 10 3 2.3%
11 to 15 1 0.8%
N: 128

 

Interpreting Conjoint Results from Student Sample

Part-worth Utilities of Attributes

The results from students are quite similar to those from voters. The part-worth
utilities and the relative importance of attributes from student sample are captured in
Table 5.5. A very sharp linear function in education spending shows that students
strongly favor an increase in spending while strongly against cuts in spending (see Figure
5.5 and 5.6). The results from students are similar to voters but have a steeper linear
utility function. For students, this result is reasonable; however, in all other programs the
students also show quite similar preferences with the voters: all other three policy areas 1
have very flat utility functions. In the areas of crime and natural resources, students
prefer spending levels at the status quo and are against either spending increases or cuts.
The spending increases in crime and natural resources are almost indistinguishable from
status quo; in OLS the utility decrease from 0.434 to 0.410 in the crime spending and

from 0.477 to 0.470 in natural resources spending. One difference from voters is welfare

90

spending; students have slightly higher utility when welfare spending increases (from

0.176 to 0.548 in OLS and from 1.489 to 2.657 in LINMAP) while voters strongly

oppose increases in welfare spending (see Figure 5.1 and 5.2).

Table 5.5: Part-worth Utilities from OLS and LINMAP (Students)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

OLS LINMAP
Factor Level Utility %A Utility Utility %A Utility
-5% -3.324 -858.9% -21.169 -759.7%

Education 0% 0.438 --- 3.209
5% 2.886 558.9% 17.960 459.7%

Relative Importance 60.98% 60.42%
-5% -0.724 -511.6°/o -4.146 -378.4%

Welfare/Health 0% 0.176 1 .489
5% 0.548 211.5% 2.657 78.4%

Relative Importance 12.50% 10.50%
-5% -0.843 -294.5% -5.889 -260.9%

Crime/Public Safety 0% 0.434 3.661
5% 0.410 -5.5% 2.228 -39.1%

Relative Importance 12.54% 14.74%
-5% -0.947 -298.6% -6.040 -286.2%

Natural Resources 0% 0.477 3.244
5% 0.470 -1.4% 2.796 -13.8%

Relative Importance 13.99% 14.33%

Constant 8.1 10 N/A

N= 77 77

 

91

 

Part-worth Utility

Part-Worth Utility

Figure 5.5: Part-Worth Utilities for Students (OLS)

 

4

WELF

CRIME

NATURAL

 

 

.4
-5°/o 0% +5%

 

 

Spending Level
Figure 5.6: Part-Worth Utilities for Students (LINMAP)

 

 

 

 

 

30

20 I

10 1

0 I

-10 . EDU
WELFARE

-20 ' I C C I
CRIME

-30 NATURAL

-5% %0 +5%

Spending Level

92

Relative Importance of Attributes

From Table 5.5 and Figure 5.7 and 5.8, we see that the relative importances from
students are similar to those from voters. The relative importance scores (for the
attributes and levels tested) show that education has the greatest impact on students’
preferences (60.98%), followed by natural resources (13.99%), crime (12.54%), and
welfare and health (12.50%) in OLS (same preference sequence in LINMAP). Without a
doubt, education is the most important issue for students; the relative importance score is
8.54 percent higher than voters. In the spending areas of natural resources, crime, and
welfare have slightly different ratings, but the utility gaps between students and voters are
indistinguishable.

Figure 5.7: Relative Importance of Attributes for Students (OLS)

 

70

60-

50-

40!

30:

Ni

10¢

 

 

 

Relative Importance

education crime and public saf
welfare and health natural resource

Attribute

93

Figure 5.8: Relative Importance of Attributes for Students (LINMAP)

 

70

60-

50-

40¢

30-

20-

1Ol

 

 

 

Relative Importance

 

 

EDU WELFARE CRIME NATURAL

Overall Utilities of Profiles

The overall utilities and the three simulation results from student sample are
captured in Table 5.6. Among the 18 profiles, profile 17 has the highest overall utility
with score 12.50 while profile 1 has the lowest overall utility with score 2.3 in OLS.
Thus, among the 18 profiles, students most prefer increases in education and welfare
while maintaining crime and natural resources at the current levels. All three simulation
models also show that profile 17 has the highest probability among the 18 choices: 100%

in maximum utility, 8.53 percent in BTL and 39.68 percent in Logit models.

94

Table 5.6: Overall Utilities of 10 Profiles for Students (OLS)

 

Profile Edu Welf Crime Natur Overall Utility Max Utility (%) BTL(%) Logit(%)
1 -5% -5% -5% -5% 2.3 0.00 1 .56 0.00
2 -5% 0% 0% 5% 5.9 0.00 4.02 0.05
3 -5% 5% 5% 0% 6.2 0.00 4.26 0.08
4 0% -5% 0% 0% 8.7 0.00 5.98 0.96
5 0% 0% 5% -5% 8.2 0.00 5.61 0.56
6 0% 5% -5% 5% 8.7 0.00 5.98 0.95
7 5% -5% 5% 5% 11.2 0.00 7.64 10.78
8 5% 0% -5% 0% 10.8 0.00 7.40 7.63
9 5% 5% 0% -5% 11.0 0.00 7.56 9.55
10 5% -5% -5% -5% 8.5 0.00 5.81 0.75
11 -5% -5% 5% 0% 4.9 0.00 3.39 0.02
12 0% -5% 0% 5% 8.7 0.00 5.98 0.95
13 -5% 0% 0% -5% 4.4 0.00 3.05 0.01
14 0% 0% -5% 0% 8.4 0.00 5.73 0.66
15 5% 0% 5% 5% 12.1 0.00 8.26 26.53
16 0% 5% 5% -5% 8.6 0.00 5.86 0.81
17 5% 5% 0% 0% 12.5 100.00 8.53 39.68
18 -5% 5% -5% 5% 5.0 0.00 3.40 0.02

 

In addition, it is interesting to know what the first choice would be if we provide

students or voters the same 10 or 18 profiles. To compare the preferences of students and

voters on the same basis, we ran a simulation on the student sample by simulating 10

profiles used in the first round of voter sample.’ As seen in Table 5.7, voters’ first choice

is Profile 8 with overall utility 6.9 while students’ first choice is Profile 7 with overall

utility 13.2. We also see that the three highest overall utilities of profiles (Profile 7, 8,

and 9) in the student sample are almost twice as those in the voter sample (Profile 8, 4,

and 7). Looking at the Profile 7, 8, 9 in the voter sample and Profile 8, 9, 10 in the

student sample, we noticed all these profiles have an increase in education spending.

These results reveal the students’ decision pattern of choosing the profiles with increases

 

’ Of course, it could be done the other way, by running a simulation on voter sample using 18 profiles in

the student sample.

95

in education spending first, then evaluating other programs. Basically, the simulation can

be used in various “what-if” situations to achieve a particular research interest.

Table 5.7: Comparison of Overall Utilities Between Voters and Students

 

 

 

 

 

Voters
Profile Edu Welf Crime Natur Overall Utility Max Utility (%) BTL (%) Logit (%)
1 -5% -5% -5% -5% 3.5 0.00 6.31 9 0.86
2 -5% 0% 0% 5% 4.9 0.00 8.99 3.78
3 -5% 5% 5% 0% 4.2 0.00 7.71 1.87
4 0% -5% 0% 0% 6.5 0.00 11.88 18.47
5 0% 0% 5% -5% 5.7 0.00 10.37 8.08
6 0% 5% -5% 5% 5.4 0.00 9.82 5.94
7 5% —5% 5% 5% 6.4 0.00 11.65 16.30
8 5% 0% -5% 0% 6.9 100.00 12.63 27.89
9 5% 5% 0% -5% 6.0 0.00 10.99 11.37
10 0% -5% 5% -5% 5.3 0.00 9.65 5.44
Students
Profile Edu Welf Crime Overall Utility Max Utility (%) BTL (%) Logit (%)
Natur

1 -5% -5% -5% -5% 2.7 0.00 2.92 0.00
2 -5% 0% 0% 5% 7.3 0.00 7.78 0.10
3 -5% 5% 5% 0% 7.2 0.00 7.67 0.09
4 0% -5% 0% 0% 10.0 0.00 10.74 1.67
5 0% 0% 5% -5% 9.0 0.00 9.65 0.60
6 0% 5% -5% 5% 10.4 0.00 11.13 2.38
7 5% -5% 5% 5% 13.2 100.00 14.11 38.73
8 5% 0% -5% 0% 13.0 0.00 13.89 31.47
9 5% 5% 0% -5% 12.7 0.00 13.63 24.75
10 0% -5% 5% -5% 7.9 0.00 8.47 0.20

 

Measures of Goodness of Fit

In Table 5.8, we see that both Pearson’s R (0.996) and Kendall’s tau (0.976) are
slightly lower than those in the voter sample. This is because we use 18 profiles rather
than 10 profiles in the survey and hence increase the chance of a respondent making
mistakes and reduce the chance of a respondent making consistent judgments. The model
for students, however, still fits the data well. In addition to the Pearson’s R and Kendall’s

tau, in this sample we use 3 holdout profiles. Holdout profiles are additional profiles

96

which the respondent evaluates beyond those generated by actual conjoint design.
Specifically, we use first 15 profiles (Profile 1-15) to predict last three profiles (Profile 16
- 18). Holdout profiles have a numbers of uses. They can help assess the predictive
accuracy of the conjoint study; ensure that data are collected for particular purposes; or
test specific hypotheses. As seen in Table 5.8, the Kendall’s tau for 3 holdouts is 1.00
with confidence level 94%. The result exhibits that this model performs well and has a
very good ability to predict choices.

The results from LINMAP tell a little different story. There are 50.6 percent of
respondents (adding pairs violated from 6 to 25) violated more than 5 percent pairs and
5.2 percent respondents (adding pairs violated from 16 to 25) violated more than 15
percent pairs (default threshold). The results are reasonable because in the student
sample we have 18 rather than 10 profiles which increase the chance of violating pairs of
profiles.

In order to check holdout profiles, LINAMP provides two measures: average
correlation and average percentage of paired violation. The average correlation is an
index which measures the degree of agreement between the predicted rankings based on
the respondent’s utility function, and the respondent’s actual ranking. The measure
“percentage of pairs violated” also uses ranked results of the predicted and actual holdout
data. As seen in Table 5.8, for holdout profiles, we have an average correlation 0.78 and
average percentage of paired violation 13 percent. Even though these are not perfect, the

results from the student sample are satisfactory.

97

 

Table 5.8: Goodness of Fit Measures for OLS and LINMAP (Students)

 

OLS
Pearson's R 0.996 Sig.= 0.000
Kendall's tau 0.976 Sig.= 0.000
Kendall's tau for 3 Holdouts 1.000 Sig.= 0.059
LINMAP

 

Distribution of Violations (%)

 

 

Pairs Violated N Percent

0 to 5 38 49.4%

6 to 10 27 35.1%

11 to 15 8 10.4%

16 to 20 3 3.9%

21 to 25 1 1.3%

Average Correlation 0.78
Average % Paired Violated 13%
N= 77

 

Comparison Between Traditional Survey Questions and Conjoint Results

In this study, one of the major purposes is to compare results between traditional
surveys and conjoint analyses. In particular, we like to know what the “response effect”
of conjoint study is. In order to compare conjoint results with traditional surveys, we ran
several simulations to make conjoint results comparable to those in traditional surveys.
To compare conjoint results with the first question, “Generally speaking, do you think the
State of Michigan should: increase overall spending, decrease overall spending, or keep
overall spending at current levels,” we ran three simulations with increasing spending
across-the-board (+5%, +5%, +5%, +5%), maintaining spending at current levels across-
the-board (0%, 0%, 0%, 0%), and decreasing spending across-the-board (-5%, -5%, -5%,
-5%). In order to compare conjoint results with Question 2, 3, 4, and 5: “In the area of

(education, welfare, crime, or natural resources), you would prefer to see the

98

State of Michigan increase spending, decrease spending. or maintain spending at current
levels,” we ran simulations with the specified attribute level (+5%, 0%, or -5%) while
holding other attribute constant (by keeping other attribute levels as 0%). For instance,
we simulated three profiles with (+5%, 0%, 0%, 0%), (0%, 0%, 0%, 0%), and (-5%, 0%,
0%, 0%) to represent respondents who favor an increase in education spending, keeping
education spending at the current level, and cuts in education spending in traditional
survey questions. In traditional surveys, we report percentages that respondents
answered in every spending level (increase, status quo, and decrease) from question 1 to
5. In the conjoint study, we report results from two simulation models (BTL and LOGIT)
with probability in each profile. In this fashion, we could make conjoint results
comparable to those obtained from traditional surveys.
Voter Sample

Table 5.9 reveals the differences between traditional surveys and conjoint
analysis. The-columns “General” represent the actual percentage of respondents who
favor an increase, the status quo, or a decrease in overall spending or particular policy
areas. The columns “Conjoint” show the simulation results from two probability models
(BTL and LOGIT). The “General” columns can be viewed as the “direct response” from
respondents (respondents only evaluate single dimensional questions) while “Conjoint”
columns can be viewed as the “deliberate response” (respondents make trade-offs among
four attributes). In rows, we report traditional questions and profiles for overall spending
and four policy areas, followed by percentage (actual answers) or probability (simulation

results), and choice order made by respondents (1,2, and 3).

99

Table 5.9 : Comparison of Budget Priorities Between Traditional Surveys and

Conjoint Analysis (Voters)

General Conjoint

General Conjoint

General Conjoint

 

Overall
%lProb.
Choice

Education
%lProb.
Choice

Welfare
%th
Choice

Crime
%lProb.
Choice

Natural
%th
Choice

BTL LOGIT

BTL LOGIT

BTL LOGIT

 

Increase (5%.5%.5%.5%)

Status Quo (0%.0%,0°/o,0°/o)

Decrease (-5%.-5%.-5%,-5%)

 

28% 37.23% 31.19%
2 2 2

Increase (5%.0%,0%.0%)

22% 41.82% 66.77%
3 1 1

Status Quo (0%.0%,0%,0%)

50% 20.95% 2.10%
1 3 3

Decrease (-5%.0%,0%.0%)

 

47% 38.13% 60.19%
1 1 1

Increase (0%.5%,0°/o,0°/o)

41% 35.15% 33.44%
2 2 2

Status Quo (0%,0%,0%,0%)

12% 26.73% 6.36%
3 3 3

Decrease (0%,-5%,0%,0%)

 

22% 31.86% 24.04%
3 3 3

Increase (0%.0%.5%.0%)

42% 35.07% 45.38%
1 1 1

Status Quo (0%.0%.0%,0%)

36% 33.07% 30.58%
2 2 2

Decrease (0%.0%,-5%,0%)

 

34% 32.98% 30.20%
2 2 2

Increase (0%,0%,0%,5%)

47% 34.95% 44.59%
1 1 1

Status Quo (0%,0%,0%,0%)

20% 32.07% 25.21%
3 3 3

Decrease (0%,0%,0%,-5%)

 

44% 33.65% 33.54%
1 2 2

43% 35.30% 46.32%
2 1 1

13% 31.05% 20.15%
3 3 3

 

As seen in the Table 5.9, in the question of overall spending, voters’ first choice

is cuts in overall spending (50%), followed by an increase in overall spending (28%), and

maintaining overall spending at the current level (22%). However, based on tradeoffs

among four policy areas, the differences between traditional surveys and conjoint analysis

are quite dramatic (see the “choice” row). In conjoint analysis, the simulation models

could be interpreted as: the profile with the highest probability voters will choose is

maintaining overall spending across-the-board (41.82% in BTL and 66.77% in Logit),

followed by increasing overall spending across-the-board (37.23% in BTL and 31.29% in

Logit), and decreasing overall spending across-the-board (20.95% in BTL and 2.10% in

100

Logit). Therefore, in conjoint simulation, voters’ first choice is maintaining overall
spending at the status quo rather than a decrease in overall spending (in traditional
question).

In the areas of education, welfare, and crime, the simulation results are in
accordance with traditional surveys: they share the same preference orders, even though
the simulation probabilities are different from the percentage that respondents actually
chose. In the area of education, for instance, voters most prefer an increase in spending
(47% in actual choice and 38.13% in BTL), followed by maintaining spending at the
current level (41% in actual choice and 35.15% in BTL), and a decrease in spending
(12% in actual choice and 26.73% in BTL); same preference orders can be found in
welfare and crime spending. In the area of natural resources, however, the results from
traditional survey and conjoint study are quite different. Voters’ first choice on budget
priorities is an increase in natural resources spending (44%), followed by the status quo
(43%), and a decrease in spending (13%), while simulation models reveal that a profile
with highest probability voters will choose is the status quo on spending (35.30% in
BTL), followed by an increase in spending (33.65 in BTL), and a decrease in spending
(31.05% in BTL). Again, we see the different results between conjoint simulations and
traditional survey questions. These preference differences are due to the fact that conjoint
analysis forces respondents to make tradeoffs among different attributes while the
traditional survey lets respondents evaluate single but incomplete budget options. Hence,

the results from conjoint analysis are more accurate and complete than traditional survey.

101

In a nutshell, the results from conjoint study and traditional survey are consistent
in the areas education, welfare, and crime while inconsistent in the areas of overall
spending and natural resources.

Student Sample

The results fiom the student sample demonstrate more contradictory preference
differences between traditional surveys and conjoint study. In Table 5.10, we see that the
results from traditional survey and conjoint study are inconsistent in all five questions.
For instance, the students’ spending preferences in the traditional survey and conjoint
analysis go in opposite directions in the overall spending question: students’ first priority
is cuts in overall spending (51%), followed by an increase in overall spending (40%), and
the status quo on overall spending (9%). On the contrary, the profile with highest
probability from student sample is an increase in overall spending (51.33% in BTL),
followed by the status quo (39.09% in BTL), and cuts in overall spending (9.58% in
BTL). In the areas of education, welfare, crime, and natural resources, we also see
inconsistencies between traditional survey questions and conjoint simulations.

In sum, while respondents only evaluate single dimensional questions in
traditional surveys, they are forced to make tradeoffs among various attributes in the
conjoint exercises. The inforrnaiton we obtain from conjoint analysis is therefore more

accurate and complete than traditional surveys (as we seen in the Table 5.9 and 5.10).

102

Table 5.10 : Comparison of Budget Priorities Between Traditional Surveys and

Conjoint Analysis (Students)

General Conjoint General Conjoint General Conjoint

 

BTL LOGIT BTL LOGIT BTL LOGIT

 

Overall Increase (5%,5%.5%.5%) Status Quo (0°/o.0°/o,0°/o,0°/o) Decrease (-5°/o.-5%.-5%,-5%)
%lProb. 40% 51.33% 97.04% 9% 39.09% 2.96% 51% 9.58% 0.00%
Choice 2 1 1 3 2 2 1 3 3

 

Education Increase (5%,0%,0%,0%) Status Quo (0%,0%,0%,0%) Decrease (-5%.0%,0%.0%)
%lProb. 75% 44.00% 95.81% 0% 34.31% 4.12% 25% 21.70% 0.07%
Choice 1 1 1 3 2 2 2 3 3

 

Welfare Increase (0%,5%.0°/o.0%) Status Quo (0%,0%.0%.0%) Decrease (0%.-5%,0%,0%)
%lProb. 31% 35.16% 51.44% 22% 34.11% 36.46% 47% 30.73% 12.09%
Choice 2 1 1 3 2 2 1 3 3

 

Crime Increase (0%,0%,5%.0%) Status Quo (0°/o.0°/o.0%.0°/o) Decrease (0%,0%,-5%,0%)
%lProb. 36% 34.24% 38.62% 12% 34.94% 48.73% 52% 30.82% 13.00%
Choice 2 2 2 3 1 1 1 3 3

 

Natural Increase (0%.0°/o,0%,5°/o) Status Quo (0%.0%.0%.0%) Decrease (0%.0%,0%.-5%)

 

 

%lProb. 38% 35.79% 52.08% 6% 35.09% 41 .67% 56% 29.1 1% 6.24%
Choice 2 1 1 3 2 2 1 3 3
Summary

In this chapter we analyzed the part-worth utilities, relative importances of
attributes, and overall utilities of profiles. In general, voters and students are satisfied
with the current spending levels in most areas. The only exception is in the education
area; both voters and students reveal that they favor an increase in education spending.
Such results may reflect the demographic characteristics in our samples. From the
Meridian Township voter sample, over 95 percent voters have at least a college degree,
and from the student sample, education undoubtedly is the most important issue for
students. Thus, we were not surprised that both voters and students support an increase in

education spending. For welfare, crime, and natural resources spending, both voters and

103

 

students exhibit that they are satisfied with the current spending levels (although students
have slightly higher utility for welfare spending increase).

In comparisons between traditional surveys and conjoint analyses, the results
reveal significant differences between traditional surveys and simulations from conjoint
results: sometimes the preference orders for a particular area even go in completely
opposite direction. These results are striking because most studies use results obtained
from traditional surveys to interpret, analyze, evaluate, and predict all kinds of political
issues in fields of public opinion and policy analysis. If the information from traditional
surveys are inaccurate or incomplete, all research results based on these surveys will be
less persuasive and misleading. In addition, if policymakers make policy decisions based
on the traditional survey results, these policies will not reflect people’s true preferences
and hence we discount the quality of democracy. Therefore, it is important to introduce
new survey tools in order to collect more accurate and complete information about public
opinion. Conjoint analysis produces more accurate and complete information about
people’s budget priorities. The “more accurate” means that we obtain respondents’ true
preferences when holding them accountable. In Chapter 2 we have shown that
respondents will not overstate their preferences when holding them accountable. The
“more complete” means that respondents are forced to evaluate all factors (which
necessarily consists of a budget proposal) rather than single dimensional consideration.
Although conjoint analysis is not the only way to conduct a survey, it is indeed a better

way to collect more accurate and complete data and information.

104

CHAPTER 6

GROUP DIFFERENCES IN CONJOINT ANALYSIS

Group attitudes provide an underlying structure that forms a broad range of policy
opinions and candidates’ evaluations. Groups help people structure the political world in
two ways. First, the political environment itself is structured in terms of groups; in the
realm of politics the relevance of groups derives from their centrality to governmental
institutions, the policy process, and election campaigns. Thus, the distribution of the vote
across different social groups has been a regular characteristic in the discussion and
interpretation of election results. From the early Columbia University studies (Berelson,
Lazarsfeld, and McPhee 1954; Berelson, Lazarsfeld, and Gaudet 1944) to the analyses of
the most recent elections (Abramson, Aldrich, and Rhode 1998; Miller and Shanks 1996),
basically all studies provide at least some information about how voters’
sociodemographic characteristics relate to their political attitudes and candidate
preferences.

Second, while groups obviously play an important role in the aggregate patterns
of electoral politics, they are also a central part of individual political cognition.
Individuals have a predisposition to rely on cues and shortcuts when making choices and
forming preferences, and groups are ideal for this purpose in the realm of politics.

Groups help orient individuals to their social world; they provide standards, supply
information, define friend and foe. In numerous situations groups consists of frames of

reference that assist individuals in interpreting their social experiences. Such guidance

105

can be particularly powerful in the political world (Brady and Sniderman 1985; Dalton
1988:151-52)

The literature on public opinion has shown that at a given point in time, different
social groups may differ markedly from each other in the level of their policy
preferences; that is, in the proportions by which they favor or oppose various policies.
Group preferences can vary because of different interests, different social, economic,
political , and cultural environments, diverse experiences, or differing positions in
society, any one of which can lead to different perceptions, wants, and values (Page and
Benjamin 1992: 285-88).

The purpose of this chapter is to analyze group differences among gender,
generations, religion, party identification, children, and home owner by evaluating their
part-worth utilities, relative importance of attributes, and overall utilities of profiles.
Since our data solely came from the Meridian Township with a very limited sample size
(N =128), our results cannot be generalized to make inferences about the distribution of
attitudes in the US. To provide some references, we first report previous studies on
group attitudes. Then, we analyze group differences in terms of part-worth utilities and
relative importances of attributes among Meridian Township voters.

Gender Differences

Studies on public opinion have shown that there were persistent gender
differences on policy preferences, differences which remain after variables other than
gender are controlled for (Smith 1984; Shapiro and Mahajan 1984; Page and Benjamin

1992).

106

 

The most important gender differences are concerned with force and violence.
Smith (1984) reported that men are more likely than women to support violent options.
In addition, gender differences also found in “compassion” issues concerned with
policies which aid the poor, the unemployed, the sick, and others in need, or which
protect the well-being of people in general. Women were more supportive than men of
such policies; Shapiro and Mahajan (1984) indicated that gender difference for these
policies is over 3 percent, less than half that of the force issues.

Various researchers have explained these differences both by the way women

 

have been socialized as caretakers and nurturers and by the rise of a women’s political
consciousness (Stoper and Johnson; Gilligan 1982). Women may have extended the
scope of their more traditional roles to encompass the problems of society as a whole.
Therefore, women might be concerned about the issues that include policies to help those
in need and also a somewhat distinct set of regulatory policies (e.g., environmental
regulation, consumer protection, other safety issues) which are intended to protect the
public. In regulation and protection issues, Shapiro and Mahajan (1984) stated that there
is no gender difference for spending on the environment, although the overall average
may conceal a change which occurred: by the 19805 perhaps 5 percent more women than
men supported spending on these programs. In addition, there were striking differences
for particular types of regulations and safeguards. Women have shown more support for
the 55 mile per hour speed limit, jail terms for drunk drivers, and so on.

For Meridian Township voters, the policy preferences of gender are captured in
Table 6.1 and Figures 6.1 to 6.4. In the area of education (see Figure 6.1), the utility

functions for both women and men are linear but women are much better off with an

107

increase in education spending: for an increase 5 percent in education spending, women
feel better off by 0.787 utility than men while for a 5 percent decrease in education
spending, women feel worse off by 0.611 utility than men. The relative importance
ratings show that both women and men have identical preference orders: education is the
most important issue for both women and men, followed by welfare, natural resources,
and crime. Nevertheless, the relative importance also indicates that women are more
concerned with education than men, with an 18.46 percent relative importance rating
difference.

In the area of welfare and health (see Figure 6.2), women favor welfare spending
at the status quo but suffer a significant utility loss with cuts in welfare spending (utility
drops from 0.431 to -0.578). On the contrary, men have the highest utility gain with cuts
in welfare spending (utility gains from 0.280 to 0.333), yet have a dramatic utility drop
with a spending increase (utility drop from 0.280 to -0.612).

For crime and natural resources issues (see Figure 6.3 and 6.4), both women and
men reveal similar patterns of utility functions: they prefer the status quo while dislike
either increase or decrease in spending. In addition, in the area of natural resources, both
women and men suffer a dramatically utility loss with a decrease in spending (utilities

drop from 0.320 to -0.430 in women and from 0.433 to -0.460 in men).

108

 

Table 6.1: Gender Differences in 4 Major Policy Areas

 

Female Male
Attribute Level Utility Utility Difference
-5% -1.655 -1.045 0611
Education 0% 0.255 0.432 -0.177
5% 1.400 0.613 0.787
Relative Importance 58.24% 39.78% 18.46%
-5% -0.578 0.333 -0.91 1
Welfare/Health 0% 0.431 0.280 0.152
5% 0.147 -0.612 0.760
Relative Importance 19.24% 22.68% -3.44%
-5% -0.217 -0.275 0.057
Crime/Public Safety 0% 0.215 0.397 -0.182
5% 0.002 -0.122 0.124
Relative Importance 8.24% 16.1 1 % -7.87%
-5% -0.430 -0.460 0.030
Natural Resources 0% 0.320 0.433 -0.1 13
5% 0.110 0.027 0.083
Relative Importance 14.29% 21.42% -7.13%
Constant 5.575 5.482
N: 54 74

 

109

 

Part-Worth Utility

Part-Worth Utility

Figure 6.1: Gender Differences in Education Spending

2.0

 

1.5-1

 

-2.0

 

 

 

Female

Male

-5% 03/0 +5%

Spending Level

Figure 6.2: Gender Differences in Welfare Spending

 

0.0 I

-.2 u

-.4 u

-.6 n

-.8

 

 

 

 

Female

Male

-5% 03/0 +5%

Spending Level

110

 

Part-Worth Utility

Part-Worth Utility

Figure 6.3: Gender Differences in Crime Spending

 

 

 

 

 

0.0 .
-'2 0 —
Female
-.4 Male
-5% 0% +5%
Spending Level

Figure 6.4: Gender Differences in Natural Resources Spending

 

-.0 I

‘l
N
l

.0
h

Female

-.6 Male
-5% 0% +5%

 

 

 

 

Spending Level

111

Generational Differences

Generational change is a continuing process which transforms all societies
gradually. There are at least two possible reasons for the generational change. One
reason involves “life cycle”: age brings with it different economic and social
circumstances and physical conditions, so that peoples’ attitudes and opinions tend to
change as they grow older. The other reason involves “cohorts”: people in different age
groups have lived through different experiences - the Great Depression, World War II,
Vietnam War, periods of economic affluence or stringency - which led them, when they
were young and flexible, to form attitudes that they tend to carry through the remainder of
their lives. On the other hand, during a given period all age groups may be affected by
the same events and social and economic conditions, and by the same interpretations
offered through the media. When there are no strong generational or life cycle effects, all
age groups move about the same way in response to such “period effects” (Braungart and
Braungart 1989; Abrarnson 1989: 236-242).

Page and Benjamin (1992) argued that both life cycle and cohort differences have
sometimes appeared between the opinions of younger and older Americans over the last
fifty years. The older people have tended to be sensitive about the threats to their Social
Security pensions, though in more general terms the youngest age group has actually been
just as supportive of government spending on Social Security and Medicare, and the
elderly have not sought income transfers to themselves at the expense of transfers to the
young. Likewise, since the older people feel more vulnerable to street crime, they tended
to favor various types of protection. Page and Benjamin (1992) showed that the elderly

are more supportive of gun control than the middle-aged (p.303).

112

In addition, Page and Benjamin pointed out that new adult cohorts have typically
been the most supportive of the social welfare policies and benefits that were already in
place in the society in which they grew up. This has occurred in affluent European
welfare states as well as in the US. Moreover, the younger people have tended to be
more supportive of environmental protection (p. 303).

Although there is no consensus on the definition of political generations, Miller
proposed an appropriate definition and classification of political generations (Miller,
1992).l According to Miller, a generation is composed of a temporally bounded group of
people who have in common formative experiences that may have occurred over a
substantial time interval. The cohorts are defined by the election years in which
individuals were first eligible to vote for president. According to Miller, there are six sets
of cohorts or generations:

A. Pro-New Deal: (1) Pre-1920; (2) Republican Norrnalcy (1920-1928),

B. New Deal: (3) Roosevelt Era (1932-1944); (4) New Deal Consolidated (1948-1964),
C. Post-New Deal: (5) Years of Turmoil (1968-1976); (6) Reagan Era (1980-1988).

The cohorts within each generation seemed more like each other and less like those other
generations.

The oldest generations, the Pre-New Deal generation, are made up of a
population whose first votes, and then whose presumed periods of early adult or preadult

politicization, preceded the first Roosevelt election in 1932. The New Deal generation

 

' Alternatively, we obtained the similar conjoint results by using generational definitions based on William
Strauss and Neil Howe’s book, Generations: The History of Americgn’s Future, 1584 to 2069. Their
definitions of generations are as follows: 0.1.: birth years 1901 to 1924, Silent: birth years 1925 to 1942,
Boomer: birth years 1943 to 1960, and Ben birth years 1961 to 198].

113

was socialized predominantly under Democratic presidential leadership. It consists of
those whose first years of voting eligibility may have included the first Roosevelt election
in 1932 or any of the subsequent elections up through 1964. The Post-New Deal
generation is composed of the still-going set of population whose early adult experiences
included the turmoil of the late 19605, the Vietnam War, the counterculture, and the
Watergate scandals of the Nixon Administration. This Post-New Deal generation is
defined as open-ended and still-growing; it therefore includes as well those socialized by

the political dramas associated with the presidencies of Carter, Reagan, Bush, and

 

Clinton. Since there is no Pre-New Deal generation in our sample, we examined New _in
Deal and Post-New Deal generations with four generation groups.

Table 6.2 reveals the differences in policy priorities for different generations. For
the education issue (see Figure 6.5), most cohorts have a linear utility function favoring
an increase in education spending. The only exception is the New Deal 2 (New Deal
Consolidated) cohort which prefers spending at the status quo. For the relative
importance, education is the most important issue for all cohorts. In the areas of welfare
and health (see Figure 6.6), all cohorts prefer maintaining spending at the current level.
While most cohorts suffer significant utility loss from increase in welfare spending, New
Deal 2 suffers serious utility loss from cuts in welfare spending. For the spending of
crime and public safety (see Figure 6.7), most cohorts have flat utility functions favoring
Spending at the status quo except the New Deal 1 (Roosevelt Era). New Deal 1 (the
oldest cohort in this study) has linear utility function favoring increase in crime spending
With slightly utility increase (from 0.491 to 0.602) but suffering a great deal of utility loss

(from 0.491 to -1.093) with cuts in crime spending. The crime issue is also the second

114

most important issue for New Deal 1 (education is the first one). In the area of natural

resources (see Figure 6.8), most cohorts (New Deal 1, New Deal 2, and Post-New Deal I)

favor spending at the status quo but have dramatic utility loss with cuts in natural

resources spending. Post-New Deal 2 (the youngest cohort in this study) is the only

exception in that they favor either an increase in natural resources spending or at the

status quo (identical utility score 0.305).

Table 6.2: Generational Differences in 4 Major Policy Categories

 

Attribute Level New Deal 1 New Deal 2 Poet-New Deal 1 Post-New Deal 2
-5% -1.551 -0.780 -1.602 -1.019
Education 0% 0.227 0.589 0.313 0.321
5% 1.324 0.190 1.289 0.697
Relative Importance 52.01% 35.75% 58.28% 46.21%
-5% -0.023 -0.487 -0.041 0.204
Welfare/Health 0% 0.324 0.387 0.357 0.295
5% -0.301 0.099 -0.315 0499
Relative Importance 1 1 .31 % 22.82% 13.55% 21 .39%
-5% -1.093 -0.219 -0.243 -0.087
Crime/Public Safety 0% 0.491 0.433 0.33 0.187
5% 0.602 -0.214 -0.088 01
Relative Importance 30.65% 17.02% 1 1.55% 7.75%
-5% -0.065 -0.562 -0.369 061
Natural Resources 0% 0.199 0.372 0.455 0.305
5% -0.134 0.19 -0.086 0.305
Relative Importance 6.03% 24.41% 16.63% 24.65%
Constant 5.426 5.567 5.519 5.519
N: 8 22 64 34

 

New Deal 1: Roosevelt Era (1932-1944)

New Deal 2: New Deal Consolidated (1948-1964)
Post-New Deal 1: Years of Turmoil (1968-1976)
Post-New Deal 2: Reagan Era (‘ 1980-1988).

115

 

Figure 6.5: Generational Differences in Education Spending

1.5

 

1.01

.51

 

 

 

 

2. N01
5:”; -1.0 I - .-
3 N02
:5: I I I I
o -1.5 .
I PND1
g. - I
1:
E .2.0 PND2
-5% 0% +5%
Spending Level

Figure 6.6: Generational Differences in Welfare Spending

 

.2!

 

 

 

b “21 ND1
5 N02
“g - 4 ' I \ u a I I

O / ‘

g I \ PND1
é — a

a“: -.6 PNDZ

 

0% +5%

('2
a-

Spending Level

116

Figure 6.7: Generational Differences in Crime Spending

 

1.0

 

.5!

0.0 I

 

 

 

 

'.5 I —
34 N01
5
g '1 .0 I I I I IND2
; PND1
11: _ .
a“: -1.5 PND2

-5% 0% +5%

Spending Level

Figure 6.8: Generational Differences in Natural Resources Spending

 

 

 

 

 

o I
2 .3 ’o/
- I .e /
.0 ’0/ _
.' I ND1
g - 4 l’ {/ -' '-
2 // N02
t 7' I I I I
g ' 5' PND1
é _ I
a“: -.8 PND2
-5% 0% +5%

Spending Level

117

Religious Differences

It is well known that religion plays an important role in American Politics. After
reviewing numerous studies of voting behavior, Converse (1974) concludes that
“religious differentiation intrudes on partisan political alignment in an unexpectedly
powerful degree wherever it conceivably can” (p. 734). American religions differ in their
positions on a number of policy questions. In some cases these differences are related to
differing party loyalties due to the fact that religion has been one of the enduring bases of
party alignments. But many differences follow from distinctive religious and cultural
beliefs and experiences. According to Wald (1997), there are six religious categories:
Roman Catholic, Jewish, Mainline Protestant, Evangelical Protestant, Black Protestant,
and Secular. Previous studies have identified some basic political tendencies of these
religious groups. American Jews have long been identified with left-wing political
causes. Seculars tend to be young, mobile, well-educated, affluent, and to live in urban or
metropolitan areas. Because such characteristics are usually associated with
unconventional drinking, it has been thought that the seculars tend to share the liberal
political outlook of Jews. Historically associated with the Republican party and
conservative positions on many issues, the mainline Protestants have been faced with
growing liberal sentiment from clergy and denominational leaders. In contrast, the
evangelicals have been encouraged to move in a conservative direction by some of their
most vocal pastors.

Like Jews, Catholics coalesced around the Democratic party in the 19308, and the
Catholics’ commitment to the party was reinforced in 1960 when the Democrats

conferred their presidential nomination on a Catholic candidate. Some observers have

118

 

doubted the depth of this commitment, arguing that attachment to the more liberal
political party was a transient stage, reflecting the immigrant and working class
background of Catholics. Under conditions of economic parity with Protestants, it has
been suggested, Catholics will embrace the political conservatism that appears to be a
more natural outgrowth of Church doctrine. Although many Catholics have moved into
the economic mainstream and the Republicans have courted them aggressively (using the
abortion issue among others), it remains open to debate whether the Catholic masses have
moved to the right side of the political spectrum (Wald 1997: 172-76).

For religious differences, we could find some clues from the 1992 NES surveys.
For the question of maintaining or increasing government service, religious preferences
from the more government services to less government services are: Black Protestant,
Secular, Jewish, Roman Catholic, Evangelical Protestant, and Mainline Protestant. About
the welfare spending issue, religious preferences from spending more to spending less
are: Black Protestant, Secular, Jewish, Roman Catholic, Mainline Protestant, and
Evangelical Protestant.

Table 6.3 exhibits the religious differences from Meridian Township voters. In
the area of education (see Figure 6.9), all subgroups have almost identical preferences -
linear utility functions favoring an increase in education spending. Education is also the
most important issue for Seculars, Evangelical Protestants, and Mainline Protestants
(second most important issue for Catholics). For welfare spending (see Figure 6.10),
most subgroups favor spending at the status quo except the Catholics, which have a
negative utility function favoring cuts in welfare spending. The welfare issue is also the

most important issue for Catholics. In the area of crime (see Figure 6.11), the Mainline

119

 

Protestants slightly prefer an increase in crime spending (utility increase from 0.171 to
0.192) while other subgroups favor spending at the current level. For the natural
resources issue (see Figure 6.12), all subgroups favor spending at the status quo but
Seculars have similar utilities between the status quo (0.521) and an increase in spending

(0.453). In sum, we did not find any significant religious differences in four spending

 

areas.
Table 6.3: Religious Differences in 4 Major Policy Categories
Attribute Level Secular Catholic Mainline Evangelical Other
-5% -1.505 -1.005 -1.422 -1.342 -1.299
Education 0% 0.403 0.366 0.310 0.532 0.341
5% 1.102 0.638 1.112 0.810 0.958
Relative Importance 49.55% 31 .80% 61 . 16% 48.24% 51.95%
-5% -0.426 0.628 -0.268 -0.104 -0.047
Welfare/Health 0% 0.230 0.460 0.363 0.386 0.340
5% 0.196 -1.088 -0.096 -0.281 -0.293
Relative Importance 12.48% 33.22% 15.23% 14.95% 14.56%
-5% -0.009 -0.362 -0.363 -0.099 -0.265
Crime/Public Safety 0% 0.256 0.603 0.171 0.537 0.300
5% -0.247 -0.241 0.192 -0.438 -0.035
Relative Importance 9.56% 18.66% 13.38% 21 .86% 12.99%
-5% -0.973 -0.443 -0.193 0.077 -0.497
Natural Resources 0% 0.521 0.400 0.230 0.295 0.394
5% 0.453 0.043 -0.037 -0.372 0.103
Relative Importance 28.41% 16.32% 10.22% 14.95% 20.50%
Constant 5.624 5.469 5.496 5.493 5.524
N= 40 28 45 11 4

 

120

Part-Worth utility

Part-Wonh Utility

Figure 6.9: Religious Differences in Education Spending

 

1.5

Secular

Catholic

 

Mainline

 

 

Evangelical

 

—5'% 03/0 +5%

Spending Level

Figure 6.10: Religious Differences in Welfare Spending

 

1.0

 

0.0 I

 

 

 

'.5 I —
Secular
-1.0 I \\ I. . . Catholic
Mainline
-1 .5 Evangelical

 

-5% 03/0 +5%

Spending Level

121

Part-Worth Utility

Figure 6.11: Religious Differences in Crime Spending

 

.8
.6I A\
(7‘ \
\\
.4‘ / Q \
/, \\
0’ I \
.2-

-.4 l

 

-.6

  

\ ...t
,5:.\......

  

 

 

-5%

Spending Level

03/0

Secular

Catholic

I I I I
Mainline
_ I

Evangelical

5%

Figure 6.12: Religious Differences in Natural resources Spending

Part-Worth Utility

 

1.0

.51

0.0 I

-.5 I

'1.0I

-1.5

 

 

 

 

 

-5°/o

Spending Level

0%

Secular

Catholic

Mainline

Evangelical

+5%

122

Party Differences

The policy preferences of Democrats and Republicans tend to differ in ways
related to various economic, ethnic, and religious bases of the two parties as well as to
their different receptivity to messages from party leaders.

Since the New Deal the Democrats have enjoyed more support among working-
class and low-income people, and among Catholics, and Jews, blacks, and various white
ethnic groups, whereas the Republican have prevailed among the white and relatively
wealthy Anglo-Saxon Protestants. Over this period, white Southerners and some
Catholics and ethnics have gradually drifted away from their Democratic loyalty, and the
large ideological differences of 19305 and 19405 have diminished. Still, the parties in the
19905 continued to differ from each other in similar ways.

In accord with their working-class status and relatively low income, Democrats
have tended much more often than Republicans to favor a broad range of government
social welfare programs, such as education, for example, and assistance with medical
care, jobs, income, and housing.

In addition, Page and Benjamin (1992) argue that party groups are actually the
only ones for which they regularly found cases of divergent movements of opinion, with
the opinions of party groups opinions becoming more different from each other over time.
Such divergent partisan trends are related to party leadership of public opinion, especially
in the area of foreign policy. For instance, two years after the Republican President
Eisenhower took office in 1952, Republicans had tended to move from their isolationist

posture toward favoring active U.S. involvement abroad, while Democrats’ opinions

123

stayed much the same. Similar situations occurred in Watergate, Vietnam War, etc.
(Pp.307-31 1).

Table 6.4 demonstrates the utility differences of four policy areas among party
groups. For education spending (see Figure 6.13), Democrats have a linear utility
function favoring an increase in education spending while Republicans are satisfied with
the current spending on education. The utility function for Democrats looks like a
straight line that favoring an increase in education spending with its highest utility score
2.017 while opposing cuts in education spending with lowest utility score -2.098. The
utility for Independents also indicates a linear function favoring an increase in education
spending. Democrats and Independents view education as the most important issues with
relative importance scores 54.37 percent and 54.07 percent respectively, Republicans
only view education as the third important issue in the relative importance score.

In the area of welfare and health, Democrats are the only group favoring increased
spending: they are better off when spending increases (utility increase from 0.104 to
0.639) while worse off when spending decreases (utility decrease from 0.104 to -0.744).
In contrast, both Independents and Republicans favor spending at the status quo but
Republicans suffer a serious utility loss when spending increases; the utility decreases
from 0.534 to -1.054. To what extent that Republicans oppose an increase in welfare
spending? If we look at the relative importance score, we will find out that Republicans
strongly object to increase spending: welfare is the number one issue for Republicans
with a relative importance score 33.75 percent.

For the crime and public safety issues, (see Figure 6.15), Democrats are the only

group favoring an increase in crime spending. Both Republicans and Independents are

124

satisfied with the current spending and suffer a utility loss when either spending increases
or decreases. However, before we say that Democrats prefer increasing crime spending,
we should also notice that the crime issue is the least important area for Democrats with a
relative importance score 5.02 percent only.

For the natural resources spending (see Figure 6.16), Democrats prefer an increase
in natural resources spending while both Republicans and Independents favor the current

spending level, although Independents’ utility is almost indistinguishable between the

status quo and an increase in spending.

Table 6.4: Party Differences in 4 Major Policy Categories

 

Attribute Level Democrat Independent Republican
-5% -2.098 -1 .630 0459
Education 0% 0.081 0.354 0.566
5% 2.017 1.276 0106
Relative Importance 54.37% 54.07% 21.78%
-5% -0.744 -0.124 0.520
Welfare/Health 0% 0.104 0.327 0.534
5% 0.639 -0.203 -1.054
Relative Importance 18.28% 9.86% 33.75%
-5% -0.212 -0.442 -0.133
Crime/Public Safety 0% 0.043 0.378 0.482
5% 0.169 0.064 0349
Relative Importance 5.02% 15.27% 17.66%
-5% -0.919 -0.740 0.128
Natural Resources 0% 0.148 0.378 0.567
5% 0.771 0.361 0695
Relative Importance 22.33% 20.81% 26.81%
Constant 5.641 5.545 5.414
N: 38 39 51

 

125

Part-Worth Utility

Part-Worth Utility

Figure 6.13: Party Differences in Education Spending

 

 

 

 

 

-1 I
Democrat
.2 ' - -
Independent
-3 Republican
-5% 0% +5%

Spending Level

Figure 6.14: Party Differences in Welfare Spending

 

1.0

5‘IIIIIIIIIIII....-'...

0.0-I

 

O
«I‘Q.

 
  
   

O

  

 

 

-.5 I .
.0
.9 —
so Democrat
-1.0 I ‘. _ _-
Independent
-1 .5 Republican
-5% 0% +5%

Spending Level

126

Figure 6.15: Party Differences in Crime Spending

 

 

 

 

g. —

3 Democrat
E - -

3' Independent
t I I I I

6L“ '-6 Republican

 

-5% 01% +5%

Spending Level

Figure 6.16: Party Differences in Natural Resources Spending
1.0

 

.5.

0.0 .

 

 

 

 

-1.0 . _ _
Independent
13613-1 Republican
5% 0% +. A

Spending Level

127

Children
From the Table 6.5 (also see Figure 6.17), we see whether having children at

home or not only makes a minor difference in the area of education: both groups have
linear utility functions favoring spending increases while people with children at home
have slightly higher utility gain when an increase in education spending. Similarly, both
groups view education as the most important issue and the group with children at home
has relative higher relative importance score (55.51 percent in “with children at home” vs.
44.65 percent in “without children at home”). However, in the areas of welfare, crime,
and natural resources (see Figure 6.18, 6.19 and 6.20), there is no significant difference
between the two groups; both groups prefer the status quo than either increases or cuts in

spending.

128

Table 6.5: Children Differences in 4 Major Policy Categories

 

 

Utility Difference
Attribute Level WI Children WIO Children
-5% -1.523 -1.060 -0.462
Education 0% 0.384 0.329 0.055
5% 1.139 0.732 0.407
Relative Importance 55.51% 44.65% 10.86%
-5% 0.185 -0.311 0.496
Welfare/Health 0% 0.353 0.333 0.02
5% -0.538 -0.022 -0.515
Relative Importance 18.57% 16.04% 2.53%
-5% -0.229 -0.274 0.044
Crime/Public Safety 0% 0.308 0.333 -0.025
5% -0.079 -0.06 -0.02
Relative Importance 1 1 .21 % 1 5.1 1% -3.90%
-5% -0.268 -0.644 0.376
Natural Resources 0% 0.438 0.328 0.1 1
5% -0.169 0.317 -0.486
Relative Importance 14.72% 24.21% -9.49%
Constant 5.478 5.569 -0.091
N: 67 61

 

129

Part-Worth Utility

Part-Worth Utility

Figure 6.17 : Children Differences in Education Spending

1.5

 

1.0-
.5I
0.0I

-.5 I

With Children
-2.0 Without Children
-5% 0% +5%

 

 

 

 

Spending Level

Figure 6.18: Children Differences in Welfare Spending

 

-.0 I

 

 

 

-.2 I
‘4 ‘ —

With Children
-.6 Without Children

 

-5% 03/0 +5%

Spending Level

130

Part-Worth Utility

Figure 6.19: Children Differences in Crime Spending

.4

 

.3I

.2I

 

 

 

 

-.2 i
-.3 I With Children
-.4 Without Children
-5% 0% +5%
Spending Level

Figure 6.20: Children Differences in Natural resources Spending

Part-Worth Utility

.6

 

-.8

 

 

 

 

\Nlth Children

Without Children

-5% 03/. +5%

Spending Level

131

 

Home Ownership

Home ownership represents a traditional indicator of social economic status in the
United States. Therefore, we could view home owners as middle or high income families
while viewing renters as low income families. People of low income are considerably
more likely to favor social welfare spending while people of middle or high income are
more likely to be against social welfare spending. In other issues, class cleavages are
usually weak indicators for various spending preferences. Table 6.6 exhibits the utility
differences between home owners and renters. First we notice that there is no difference
on education spending; both groups have a linear utility function favoring spending
increases (see Figure 6.21). For welfare and crime spending (see Figure 6.21 and 6.22),
both home owners and renters are satisfied with current spending levels. The only
difference between home owners and renters occurs in the area of natural resources:
renters prefer spending increases (utility increase from 0.142 to 0.198) while home
owners favor spending at the current level. Since there are only 15 cases in the renter
category (while 113 cases in home owner category), we are not surprised by the results

that no significant utility difference is found between home owners and renters.

132

Table 6.6: Differences Between Home Owners and Renters in 4 Major Policy

 

 

Categories
Utility Difference

Attribute Level Home Owner Renter

-5% -1.276 -1.503 0.227
Education 0% 0.369 0.272 0.097

5% 0.907 1.231 -0.324
Relative Importance 50.67% 54.79% -4.12%

-5% -0.003 -0.417 0.414
Welfare/Health 0% 0.329 0.453 -0.124

5% -0.326 -0.036 0290
Relative Importance 15.20% 17.45% -2.25%

-5% -0.285 0.009 -0.293
Crime/Public Safety 0% 0.347 0.120 0.227

5% -0.062 -0.128 0.067
Relative Importance 14.65% 4.97% 9.68%

-5% -0.422 -0.640 0.218
Natural Resources 0% 0.417 0.142 0.275

5% 0.004 0.498 -0.493
Relative Importance 19.48% 22.97% -3.49%
Constant 5.512 5.591 -0.080
N: 1 13 15

 

133

Part-Worth Utility

Part-Worth Utility

 

Figure 6.21: Home Differences in Education Spending

 

1.5

1.0I ’

 

 

-2.0

 

Home Owner

Renter

-5% 65/. +5%

Spending Level

Figure 6.22: Home Differences in Welfare Spending

 

.6

 

 

 

 

Home Owner
-.6 Renter
-5% 0% +5%

Spending Level

134

Figure 6.21: Home Differences in Crime Spending

 

 

 

 

 

Z:
“3 -.2 .
.r:
5 '—
3 -.3 i Home Owner
t. - -
£1.“ -.4 Renter
-5% 0% +5%

Spending Level

Figure 6.24: Home Differences in Natural Resources Spending

 

 

 

 

 

0.0 I

-.2 q
.3:
E -.4 '1
D
g —

I
g ‘6 '/ Home Owner
5
a -.8 Renter
-5% 0% 5%

Spending Level

135

Summary

In this chapter we examined the group differences in terms of part-worth utilities
and relative importances of attributes. Among Meridian Township voters, we found
some similar group differences with previous studies: women are more concerned with
compassion issues such as education and welfare than men; the New Deal 1 generation
(the oldest cohort in this study) favors an increase in crime; Post-New Deal 2 (the
youngest cohort in this study) favors either an increase in natural resources spending or
the status quo; Democrats have a linear utility function favoring education spending
increased while Republicans are satisfied with current spending level on education;
Democrats also prefer an increase in welfare spending while Republicans favor spending
at the status quo but suffer a dramatic utility loss with an increase in welfare; people with
children at home have a slightly higher utility increase when education spending
increases.

As we noted before, our data came from the local area with a limited sample size
(N=128), thus our conjoint results cannot be generalized to inference about group
attitudes in the US. Despite these limitations, in this study we have shown Meridian
Township voters’ policy preferences are based on their trade-off j udgments rather than on

single dimensional consideration of incomplete budget options.

136

CHAPTER 7

CONJOINT APPLICATIONS: SIMULATIONS OF GUBERNATORIAL
ELECTION AND POLICY EVALUATION
One of major advantages of conjoint analysis is that it provides an opportunity to
conduct a computer simulation of choices by voters. A computer simulation can answer
various “what-if” scenarios and determine the likely results of attribute/attribute level
tradeoffs for the conceptualized services or products (e.g., different budget proposals in
our study). Choice simulations reveal respondent preferences for the specific services
defined by the researcher. In this study, first we ran simulations to identify which
candidate (John Engler/Republican or Jeffrey Fieger/Democrat) would win the
gubernatorial election in Michigan and then show how the underdog candidate could
surpass his opponent by changing his policy positions. Second, we show how candidates
can target a segment of voters (i.e., Independents) in order to win the gubernatorial
election. Third, we illustrate a policy evaluation in welfare reform by simulating several
policy alternatives against the current welfare spending level. Lastly, we demonstrate
how a governor can distribute revenue to a single spending area in order to maximize
voters’ utilities.
Simulation of Gubernatorial Election in Michigan
Simulating the gubernatorial election is related to the topic of issue voting. Issue

voting includes two major evaluations: retrospective voting and prospective voting.
Retrospective voting examines the incumbent’s performance on policy, and how likely is

it that his opponent would outperform him? Prospective voting examines the policy

137

platforms addressed by the candidates and evaluates which candidate’s policy promises
are most similar to what the voter believes the government should be doing (N iemi and
Weisberg, 1993: 98-99).

An important question is to what extent the voters’ voting behavior is based on
policy issues rather than on their social backgrounds or party loyalties. The authors of
The American Voter argue that the voters are often ill informed about public policy and
may not be able to vote based on issues. They reported that in 1956 on 16 different issues
only 18 to 36 percent of the electorate satisfied what the authors considered three
prerequisites for issue voting: (1) being familiar enough with the issues to have an
opinion, (2) knowing what the government is doing on the issue, and (3) seeing the
parties differ on the issue. Converse (1964) concluded that some issues were marked by
“nonattitudes,” with respondents answering questions basically randomly simply to
satisfy the overinsistent interviewer; this further emphasized the relative unimportance of
issues in voting. However, Abramson et al. (1995) found that there is a reasonably high
relationship (from 41% to 80%, depending on how many prerequisites for issue voting
are satisfied) between the voter’s issue position and the candidate he or she supported on
various issues from 1972 to 1992 presidential election (pp. 179-189). They argue that
failure to meet the prerequisites for issue voting does not mean that the voter has ill-
formed preferences. On the contrary, the voter’s ability to see differences between
candidates varies because the political environment changes from election to election (p.
182)

In our gubernatorial election simulation, voters actually evaluate candidates based

on both the prospective voting and retrospective voting. Specifically, voters will evaluate

138

both governor John Engler’s performance and his platform. On the other hand, voters
will tend to focus more on Jeffrey F ieger’s platform because he is the challenger
(although voters will still possibly evaluate the performances of the past Democrat
governors). Since it is not necessary to meet the prerequisites for issue voting, we can
run gubernatorial election simulations without knowing how many criteria the Meridian
Township voters can meet.
Three Choice Simulation Models

Researchers generally use three types of models to simulate market behavior
based on conjoint data: maximum utility (first choice) model, Bradley- Terry Luce/BTL
(average choice) model, and logit (share of preference) model. Each model estimates the
respondent’s preferences by combining their utilities with a set of competing profiles on
the study attributes. None of these models can support absolute statements such as:
“Proposal A will get 20% votes” or “Proposal C will beat Proposal B in 2-to-l votes.”
Conjoint models produce only relative results and are best used for ranking alternative
courses of action. For example, they can tell us that Pr0posal A is expected to be least
preferred or that voters will preferred Proposal C rather than Proposal B (Curry 1997a).

Maximum utility (first choice) model identifies the profile with the highest utility
as the profile of choice. This profile is selected and receives a value of l; ties receive a
0.5 value. After the process is repeated for each respondent’s utility set, the cumulative
“votes” for each profile are evaluated as a proportion of the votes or respondents in the
sample. Although it looks conceptually and computationally simple, the first choice
model has one major drawback. That is, it tends to produce more extreme results than are

observed in actual buying situations; it overestimates preferences for the most preferred

139

profiles and underestimates preferences for those that are least preferred. This occurs
because first choice models make no allowance for respondent error (Curry 1997b).

The Bradley-Terry-Luce (average choice) model estimates choice probability in a
different fashion. It computes the probability of choosing a profile as the most preferred
by dividing the profile’s utility by the sum of all the simulation total utilities. The BTL

model can be expressed as follows:

 

UP
P(P) = (7.1)

2 Up
where:
P(p) = probability for choosing Profile P; p = 1,..., p, and
Up = overall utility for Profile P.

Unlike the Maximum utility model, which assigns a respondent’s entire
preference to the profile with the highest utility, the Share of Preference (logit) model
splits the probability of purchase among all competing products or according to their
utilities. It is critical to ask: why should the logit model be better? First it recognizes that
consumers do not always purchase the product for which they have the highest utility.
Second, many products and services represent low-involvement purchases for consumers;
for these products consumers often have no clear first choice.

Share of preference (logit) models use an assigned choice probability that is
proportional to an increasing monotonic function of the alternative utility. The choice
probabilities are computed by dividing the logit value for one product by the sum for all

other products in the simulation. The logit transformation for converting a subject’s

140

utility for profile A, (U A), into a preference share, P(A), for profile A in a market made up

of two products, A and B. is:

“’4

p(A)=_le_e;__

fill”

‘ . (7.2)
e ' +6

The preference share for Profile B is:

Ill/B

P(B) : I211?

w,

8+8

(7.3)

where:

P(A) = probability for choosing Profile A; P(B) = probability for choosing Profile B,
e = 2.7183,

b = a measure of the individual’s involvement in the product category, and

U A = overall utility for Profile A; UB = overall utility for Profile B.

This expression can be generalized to any number of profiles by adding a term in
the denominator for each additional profile.

Researchers use the logit model because it has three intuitively desirable
properties. First, the higher a profile’s utility, the greater is its preference share. Second,
the individual’s preferences across all profiles sum to 1. And third, the greater the
individual’s involvement in the category (i.e. the larger b becomes), the more distinct are
the preference shares.

Although the logit model generally produces better estimates than the first choice
model, it overestimates preference for similar products. If we add identical (or very

similar) products to our model, the total preference share for those products is artificially

141

inflated (the “red bus/blue bus” problem). This is because the independence from
irrelevant alternatives (IIA) assumptionl implied by the logit rule.

In addition, the use of BTL and logit models usually involves arbitrary scaling
assumptions; that is, the results are not invariant over linear transformations on the
utilities estimated by conjoint analysis (Green and Srinivasan 1990).

In summary, while the literature shows that the first choice model can provide the
best overall validation, choice behavior models (BTL and logit) have a strong
probabilistic component which will be very useful because (1) consumers do not always
choose the alternative with the highest utility, and (2) consumers sometimes have a low
involvement for some services or products.

Simulation Results for Gubernatorial Election

In this simulation, first we identify two candidates’ positions on 4 major policy
categories. Since Engler is the incumbent governor, we assume that his policy positions
are maintaining all spending areas at current levels. On the other hand, based on the
Democrats’ policy preferences in Meridian Township, we could assume that Democrat
candidate Jeffrey F ieger’s policy positions are increase in spending in education, welfare,
and natural resources while maintaining crime spending at current level. A simulation is
then executed to identify which candidate will win the gubernatorial election in Meridian

Township. After we obtain the results, we run another simulation to see how the

 

' The independence of irrelevant alternatives problem arises from the assumption that the relative
probability of choosing two existing altematives is unaffected by the presence of additional alternatives.
However, this is not the case in reality; adding an irrelevant alternative changes the probabilities assigned
to all categories. For an example, see Kennedy (1992), p244. Empirical evidence presented by Kamakura
and Srivastava (1984) suggests that the [IA assumption is reasonable at the individual level.

142

underdog candidate can surpass his opponent by changing his policy positions.
Simulation results include profile utilities and results from three choice models.

Simulations results for two candidates John Engler and Jeffrey Fieger are captured
in Table 7.1. In the Simulation 1, John Engler’s policy positions are maintaining all
spending areas at the status quo while Jeffrey F ieger proposes a platform with increasing
spending on education, welfare, and natural resources but maintaining crime spending at
the status quo. Simulation results show that profile utilities for Engler and F ieger are 6.9
and 6.6 respectively. The maximum utility model shows that Engler will win the election
with probability 100 percent. The BTL and logit models also reveal that Engler has a
higher probability to win: 51.37% (Engler) vs. 48.63% (F ieger) in the BTL model and
59.16% (Engler) vs. 40.84% (Fieger) in the logit model.

Table 7.1: Simulation Results for Gubernatorial Election (All Voters)

 

 

Simulation 1 Simulation 2 Simulation 3
Profile Profile Profile
Attribute Engler Fieger Engler Fieger Engler Fieger
Education 0% 5% 0% 5% 5% 5%
Welfare/Health 0% 5% 0% 0% 0% 0%
Crime/Public Safety 0% 0% 0% 0% 0% 0%
Natural Resources 0% 5% 0% 5% 0% 5%

Simulation Results Simulation Results Simulation Results

 

Candidate Engler Fieger Engler Fieger Engler Fieger
Profile Utility 6.9 6.6 6.9 7.2 7.5 7.2

Max Utility 100% 0.00% 0.00% 100% 100% 0.00%
BTL 51.37% 48.63% 49.06% 50.94% 51.10% 48.90%
Logit 59.16% 40.84% 43.42% 56.58% 58.00% 42.00%

 

143

Can F ieger surpass his opponent Engler by changing his policy positions?
Simulation 2 provides the answer. In the Simulation 2, Fieger changes his policy
positions by prOposing another platform with increasing spending in education and
natural resources but maintaining welfare and crime spending at current levels while
Engler has identical policy positions with Simulation 1. Now the results from Simulation
2 indicate that the profile utilities for Engler and F ieger are 6.9 and 7.2 individually. The
maximum utility model reveals that Fieger will win the election in this simulation.
Similarly, the BTL and logit models also indicate that Fieger has a higher probability to
win: 50.94 (F ieger) vs. 49.06% (Engler) in the BTL model and 56.58% (Fieger) vs.
43.42% (Engler) in the logit model.

Suppose that Engler knew Fieger’s campaign strategies. How can Engler regain
his advantage position? In Simulation 3, Engler proposes another platform with an
increase in education while maintaining welfare, crime, and natural resources at current
levels while Fieger’s policy positions are same as Simulation 2. Now we see Engler has
profile utility 7.5 while Fieger has profile utility 7.2 - Engler beats F ieger again by
changing his policy positions. All three simulation models also show that Engler will
win the election or has higher probability to win. If we continue the simulation exercise,
this process simply generates a majority rule cycle. However, our purpose here is to
demonstrate how we can apply conjoint analysis to the gubernatorial election rather than
examine candidates’ campaign strategies.

Targeting A Particular Group
In addition to simulating in the aggregate level (all voters), conjoint analysis can

target a specific group for the purpose of segmentation. For instance, we could target

144

different parties, religious groups, gender, generations, etc., to find out what their policy
preferences are and which candidate they will support. In order to demonstrate this
segmentation, let us consider a situation that assuming two parties will support only their
own candidates. Then, in this case gaining the support from independent voters may be
the major concern for both candidates. One way to attract independent voters is to
propose a platform that meets the voters’ needs. In order to do 50, candidates can change
their policy positions on four major spending categories. Table 7.2 reveals the simulation

results for independent voters.

Table 7.2: Simulation Results for Gubernatorial Election (Independent Voters)

 

 

 

Simulation 4 Simulation 5
Profile Profile
Attribute Engler Fieger Engler Fieger
Education 0% 5% 5% 5%
Welfare/Health 0% 5% 0% 5%
Crime/Public Safety 0% 0% 0% 0%
Natural Resources 0% 5% 0% 5%
Simulation Results Simulation Results
Candidate Engler Fieger Engler F ieger
Profile Utility 7.0 7.4 7.9 7.4
Max Utility 0.00% 100% 100% 0.00%
BTL 48.69% 51.31% 51.79% 48.21%
Logit 40.74% 59.26% 63.34% 36.66%

 

Note: N=35

In Simulation 4, both candidates have their original policy positions. F ieger
obviously has the advantage: all three simulation models show that Fieger will win the
election. The profile utilities for Fieger and Engler are 7.4 and 7.0 respectively. Fieger

has the advantage that 100 percent probability in maximum utility model and 51.31

145

percent in BTL and 59.26 percent in logit models. Under this condition, Engler becomes
the underdog and has to prOpose another platform. After changing his policy positions by
increasing education spending while maintaining welfare, crime, and natural resources at
current levels (Simulation 5), now Engler surpasses Fieger with overall utility 7.9
(Engler) vs. 7.4 (Fieger). In this fashion, when candidates intend to gain support from a
particular group, they can target this group by simulating different scenarios - changing
their policy positions in this case (if the position changes do not violate their basic
political beliefs or their party’s ideological positions).
Welfare Policy Evaluation

One of the major achievements of John Engler is his welfare reform. In his
website, John Engler’s achievement in welfare reform are: “statewide, nearly half of all
welfare recipients who can work are working and earning a paycheck, and more than
160,000 families have achieved independence from cash assistance. The state welfare
caseload has declined for four years in a row, and since 1991, the number of recipients
has fallen by more than 325,000 -- more than any other state.”2 One way to evaluate
whether welfare reform is a success or not is to examine voters’ satisfaction with the
current welfare spending level. In order to so, we could compare three welfare spending
levels with an increase in spending, a decrease in spending , and maintaining spending at
the current level. As shown in Chapter 5, we could operationalize three welfare spending
levels by using three profiles with an increase in welfare spending while maintaining

other spending areas at status quo (0%, +5%, 0%, 0%), a decrease in welfare spending

 

2 Michigan John Engler 1991-1997 Accomplishments.
http://www.migov.state.mi.us/issues/engleraccomp9197.html

146

while maintaining other spending areas at status quo (0%, -5%, 0%, 0%), and maintaining
all spending areas across-the-board ((0%, 0%, 0%, 0%).

Looking at the Table 7.3, we found that Meridian Township voters are very
satisfied with the current welfare spending level - with a profile utility 6.9 (Profile B)
while opposing either an increase in welfare spending (Profile C’s utility 6.3) or a
decrease in welfare spending (Profile A’s utility 6.5). The BTL and logit models also
indicate the current welfare spending level (Profile B) has a higher probability to gain
support. Based on the simulation, we can conclude that Engler’s welfare reform is very

successful, at least within Meridian Township.

Table 7.3: Simulation Results for Welfare Policy Evaluation

 

Attribute Profile A Profile B Profile C
Education 0% 0% 0%
Welfare/Health -5% 0% 5%
Crime/Public Safety 0% 0% 0%
Natural Resources 0% 0% 0%

Simulation Results

Profile A Profile 8 Profile C

 

Profile Utility 6.5 6.9 6.3

Max Utility 0.00% 100% 0.00%
BTL 33.07% 35.07% 31 .86%
Logit 30.58% 45.38% 24.04%

 

Now let us consider another situation. Due to the low unemployment rate and the
economic resurgence in Michigan, Department of Budget and Management Office
estimates that there is extra revenue at hand this year. Assuming Governor John Engler
decides to distribute extra revenue to a single spending area in order to maximize voters’

utilities. What is the possible way to distribute the money? How can governor John

147

Engler find out the voters’ true preferences rather than their desires? Conjoint analysis
can serve this purpose of finding the best combination of different spending levels by
running several simulations. Since we know voters are satisfied with the current welfare
spending level, we could increase the spending level either on education, crime, or
natural resources. Specifically, we run three simulations with profiles that increasing
education spending while keeping other spending areas at the status quo (+5%, 0%, 0%,
0%), increasing crime and public safety spending while keeping other spending areas at
the status quo (0%, 0%, +5%, 0%), or increasing in natural resources spending while
maintaining other spending areas at the status quo (0%, 0%, 0%, +5%).

The simulation results are captured in Table 7.4. Obviously, an increase in
education spending (7.5 utility for Profile D) is the most favorite option for the Meridian
Township voters, followed by increasing the crime spending (6.5 utility for Profile E),
and an increase in natural resources spending (6.0 utility for Profile F). Thus, in order to
maximize the voters’ utilities, Governor Engler should increase education spending when
there is an extra revenue at hand.

Table 7.4: Simulation Results for Maximizing Voters’ Utilities

 

Attribute Profile D Profile E Profile F
Education 5% 0% 0%
Welfare/Health 0% 0% 0%
Crime/Public Safety 0% 5% 0%
Natural Resources 0% 0% 5%

Simulation Results

Profile D Profile E Profile F

 

Profile Utility 7.5 6.5 6.0

Max Utility 100% 0.00% 0.00%
BTL 37.53% 32.65% 29.81%
Logit 62.92% 23.67% 13.41%

 

148

Summary

In this chapter, we demonstrate the potential applicability of conjoint analysis by
simulating the gubernatorial election in Michigan and evaluating the political support of
welfare policy. We acknowledge that the voters do not vote for a particular candidate just
according to their spending preferences on the four policy areas nor will candidates’
campaign strategies solely rely on changing their policy positions. However, the
simulation exercise in the current study at least can provide candidates more accurate
information about the voters’ true preferences rather than their wishes or desires.
Similarly, evaluating spending levels may not be the only way to conduct a policy
evaluation. Nevertheless, conjoint analysis provides a new way to evaluate voters’
satisfaction with all kinds of policy issues. In conclusion, conjoint analysis has great

potential applicability to the field of public policy and public opinion.

149

CHAPTER 8

RELIABILITY AND VALIDITY TESTS

In this chapter, we examine the reliability and validity of conjoint analysis and
traditional survey questions. First, we point out the importance of reliability and validity
checks and two major explanations of unstable public opinion. Focus is then given to
definitions of reliability and validity by adopting a generalizability theory approach.
After evaluating four types of reliabilities, attention is given to the measures we use:
product moment correlation and Spearman rank order correlation. Lastly, we report the
reliability results from conjoint analysis and traditional survey questions.

Why Reliability and Validity Checks Are so Important

Research on public opinion usually proceeds on the assumption that people have
well organized attitudes on major political issues. Thus, surveys are passive instruments
to measure these attitudes. The typical view is that when people say they favor X, they
just recall a preexisting state of feeling favorably toward X. However, accumulated
studies on public opinion has shown that people tend to have unstable answers to survey
questions.
Random Response Variance

One explanation of the response instability is that people simply do not have
meaningful attitudes. Converse (1964) finds significant intertemporal instability in
response to a panel study that shows a sample of Americans the same set of issue
questions in 1956, 195 8, and 1960. His findings show that the correlation between

individuals’ positions on important policy issues in 1956 and 1960 is quite low. The

150

highest correlation on all issues was slightly less than 0.5. The correlation on most issues
fell in the range of 0.3 to 0.4. While the public is seemingly fickle in its issue positions,
partisanship remains fairly stable, exhibiting an intertemporal correlation slightly above
0.7. Converse also shows that the correlation between an individual’s responses in 1956
and 1958 and 1958 and 1960 is roughly equal to the correlation between 1956 and 1960,
suggesting that fluctuations in individual responses are not the result of steady changes
over time; rather, “the only change that occurred was random change” (p.243). Thus, he
argues that “large portions of an electorate simply do not have meaningful beliefs, even
on issues that have formed the basis for intense political controversy among the elite for
substantial periods of time” (p. 245). However, facing the interviewers, respondents still
politely choose an answer but in an almost apparently random fashion.

Achen (1975, 1983) challenges Converse’s original findings by showing that
measurement error in survey instruments can account for the over-time fluctuations in
individual responses (see also Erikson 1979; Judd and Milbrun 1980). Achen argues that
survey questions are stochastic or probabilistic measures of true attitudes. Each question
contains a stochastic component, and temporal changes in individual responses are likely
the product of random fluctuations in the error. Respondents are not to be indicted for
displaying unstable attitudes or preferences: “measurement error is primarily a fault of the
instruments, not of the respondents” (Achen 1975: 1229). Therefore, he concluded that
even when people’s “survey responses” fluctuate greatly, the members of the public have
underlying “true attitudes” that are overwhelmingly stable.

Zaller (1992) challenges both approaches and suggests that people have

conflicting opinions on most issues and these conflicting opinions might lead them to

151

support the issue either way. He argues that most people simply do not possess
preformed attitudes at the level of specificity demanded in surveys. Instead, they carry
around in their heads a mix of only partially consistent ideas and considerations
(ambivalence axiom). When questioned, they average across the opinions that happened
to be salient at the moment of response (response axiom) and that have been recently
thought about (accessibility axiom), and use them to decide among the options offered.
Therefore, Zaller argues that responses from individuals do not reflect anything that can
be described as either true attitudes or random guessing, rather, they simply reflect the
thoughts that are most accessible in their memories at the moment of response.
Response Effects

Aside from the random response variance that has been attributed to
measurement error, there exists systematic variance from artificial response effects. The
literature on response effects has shown that survey questions not only measure public
opinion but also shape and channel it by the way in which they frame issues, order the
options, change the tone of wording, add more options (i.e., the “don’t know” option),
and other contexts of the question.I For instance, a 1976 NORC experiment reported by
Turner and Krauss (1978) asked on one form about whether spending in each of 11
categories (e.g., space exploration, environment, defense) should be increased, followed
by a single question on whether the federal income tax paid by the respondent was too
high. On the other form the tax question preceded the spending items. In the latter case,

14% more respondents said taxes were too high than did so when the spending questions

 

' For a detailed discussion of response effects, see Schuman and Presser (1996), Questions Jami Answers in
Attitude Surveys.

152

had been posed first, a difference significant beyond the .001 level of confidence. These
findings have led researchers to a conclusion that respondents do not simply reveal
preexisting attitudes on surveys; instead, to some extent, people are using the
questionnaire to decide what their attitudes are.

If we know that people change their opinions over time and that their answers
depend on seemingly trivial changes in the order of questions, then the notions of
responsible electorate (Key 1966), reasoning voters (Popkin 1991), and rational publics
(Page and Shapiro 1992) become difficult to support. On the other hand, if people are
using the questionnaires to decide what their attitudes are, then can a different survey tool
(conjoint analysis in this study) produce a more accurate information or valid answer?
No matter which view we accept, examining the reliability and validity of survey
instruments has become an important and necessary procedure in the study of public
opinion.

Definitions of Reliability and Validity

Validity refers to the degree to which instruments truly measure the constructs
which they are intended to measure. A necessary but not sufficient condition for validity
of measures is that they are reliable. Reliability can be defined generally as the degree to
which measures are free from error and therefore yield consistent results. The
relationship between reliability and validity can be distinguished as follows:

m=x+x+x
Where:
X0 = an observed score,

X, = the true score,

153

X5 = systematic error, and

Xr = random error.

A measure is said to be valid when X0 = X,. On the contrary, reliability refers to
the consistency and reproducibility of the observed results. A measure is said to be
reliable when there is no variation in X0 due to chance or random errors, that is, Xr = 0.
Therefore, reliability is a necessary, but not sufficient, condition for validity (Bateson et
a1. 1987).

To apply this equation to our additive, main-effects part-worth model, it can be

written as follows (recall Equation 3.16):

Up =C+ZZWUXU +ep (3.16)
I" 1

Where:
Up = the overall utility to measure of the p-th profile,
C = constant or intercept,
Wij = the part-worth utility of i-th level j-th attribute estimated in the regression,
X ij = dummy variable representing the presence or absence of attribute j, level i in
alternative X, and
ep = error term for profile P.
Theoretically, a measure is valid if the part-worths Wij represents a true measure
of the respondent’s underlying part-worth for i-th level of the j-th attribute. Likewise, it
would be reliable if the Wij included no variation due to random error. However, this

would not preclude variations due to systematic error.

154

Operationally, the distinction between validity and reliability blurs because it is
impossible to measure X,, the true measurement. Instead of using surrogates or
approximations, the most widely used approach to this problem is the multitrait
multimethod matrix suggested by Campbell and Fiske (1959). Traditionally, reliability
check has been done on multi-item scales. There are three basic methods: test-retest,
internal consistency, and alternative forms. The basic difference among the three
methods lies in what the scale is to be correlated with to compute the reliability
coefficient. In test-retest, the identical set of measures is applied to the same subjects at
two different times. The two sets of obtained scores are then correlated. In internal
consistency, a measurement scale is applied to subjects at one point in time; subsets of
items within the scale are then correlated. In alternative forms, two similar sets of items
are applied to the same subjects at two different times. Scale items on one form are
designed to be similar (but not identical) to scale items on the other form. The resulting
scores from the two administrations of the alternative forms are then correlated. The
clear assumptions here are that there is a single construct called “reliability” and that
there are various ways of measuring it (Peter 1979).

Reliability and Validity Checks: A Generalizability Approach

One major problem not explicitly addressed by traditional approaches to
reliability is that measurement error can come from many sources, and each definition of
error changes the meaning of the “reliability coefficient.” For example, the components
of variance which are “true” and “error” in computing a test-retest correlation are

different from those for an internal consistency estimate. One approach to

155

simultaneously analyzing multiple sources of variance in a measurement procedure has
been addressed - generalizability theory.

Generalizability theory is based in part on the concept of sampling. However, the
primary focus is not on the sampling of the people from populations of people, but rather
on sampling conditions of measurement from universes of possible measurement
conditions. Generalizability theorists argue that the traditional approaches to reliability
do not address explicitly the sources of the measurement error being assessed. They
argue there is no such thing as a reliability score; rather the score must specify the
conditions of measurement over which reliability has been measured (Rentz 1987).

By adopting a generalizability perspective, a number of different reliability
measures can be computed from conjoint analysis (Reibstein et al. 1988):

Reliability Over Time

Reliability over time is assessing by taking one measure, then repeating it with the
same individuals and the same measurement devices at a subsequent time. The only
aspect varied is the time of administration; everything else is held constant. The question
is whether the Wij at time t is the same as that at time t + x where t is some time lag.

Reliability Over Attribute Set

This approach examines whether the part-worths for a given attribute level for an
individual depends on the other attributes or levels in the stimuli. Operationally, the tests
involve assessing the stability of part-worths computed for attributes that are common

when other attributes in the stimuli are varied.

156

Reliability Over Stimulus Set

The use of fractional factorial design in conjoint analysis reduces the problem of
information overload for respondents. However, it also raises a new kind of reliability
question: “reliability over stimulus set” addresses whether the derived part-worths are
sensitive to the fractional factorial used for estimation. If the part-worths are influenced
by the fractional factorial chosen, then using such designs would have to be reconsidered.

Reliability Over Data Collection Procedure

Conjoint analysis includes a large range of data collection procedures. In general,
these variants can be categorized along three dimensions: type of data, data-gathering
procedure, and dependent variable. Reliability over the data collection method procedure
addresses whether the part-worths would be different if a different methodological variant
were used.

Methods and Designs

This section looks at the reliability of conjoint estimates and traditional questions.
In the student sample, we collected the follow-up survey in a one month period.2 The
traditional questions were identical in the first and second round of surveys and were
presented to students with the format of unbounded resources, such as “Generally
speaking, do you think the State of Michigan should: increase overall spending, decrease
overall spending, or keep overall spending at current levels?” and “In the area of

(i.e., education), would you prefer to see the State of Michigan: increase

spending in the area of education, decrease spending in the area of education, or maintain

 

2 Because the summer session only last one and a half month, we could not collect the data in a two month
period like we did for the voter sample.

157

education spending at current levels?” In conjoint questions, we asked students to rank
order 18 profiles in both surveys but with a different fractional factorial design in the
second round of survey.

In the voter sample, the data for reliability tests were collected in a two month
period. In both first and second rounds of the surveys, traditional questions were

presented by asking respondents expressing their own ideal spending levels (except first

general question), such as “In the area of (i.e., education), would you prefer
to see the State of Michigan: increase spending in the area of education by %,
decrease spending in the area of education by %, or maintain education spending at

current levels?” In the conjoint questions, we asked respondents to rank order 10 profiles
from the most preferred to the least preferred in the first round of survey. Later, in the
follow-up surveys, we divided 88 subjects into two groups (44 subjects in each groups).
In the first group, we asked respondents to rank order 10 profiles but with a different
fractional factorial design while in the second group, we asked respondents to rate 12
profiles in a 0-100 points scale.

In conjoint analysis, we measure 3 types of reliability from above design:
reliability over time,3 reliability over stimulus set, and reliability over data collection
method. The use of test-retest with different fractional factorial design is to measure
reliability over time and reliability over stimulus set while the use of two data collection

methods (ranking and rating) is to measure reliability over data collection method.

 

3 Strictly speaking, we do not test reliability overtime directly if we adopt the definition above. Since we
test reliability in a one or two month period with different fractional factorial designs, this measure actually
provides more rigorous reliability test and hence can be viewed as an indirect way to measure reliability
over time.

158

Both OLS and LINMAP were used to estimate the part—worths from ranking and
rating data. Pearson product moment correlation of the part-worth utilities from the two
tasks provided a measure of reliability. In addition, the part-worths WU- estimated from
the first set of stimuli were used to predict utilities to the stimuli in the second setting Uzp;
the Spearman rank order correlation was used to check the reliability from the two tasks.
This procedure was then repeated with the part-worths estimated from the second set of
stimuli, predicting utilities to the stimuli in the first setting Ulp. Parker and Srinivasan
(1976) refers this method as cross-validation.

It should be noted that measures examining the utilities of the stimuli, Up, are
actually assessing the reliability of the model. Measures based on the Wij are assessing
the reliability of the part-worths. The Wij’s have the utmost importance in policy
decisions and are used as the basis of individual-level simulations. Reproducibility of the
output of the model is a less stringent test because, as a result of the compensating nature
of the model, it is possible to have stable Up’s with less stable Wij’s (Bateson et al. 1987).

Results and Discussions
Reliability of Conjoint Analysis

Voter Sample: Ranking Data

Table 8.1 illustrates the part-worths correlations for ranking data in the voter
sample between 2 time administrations. In the first column, we compare the product

moment correlations between first and second round utilities for both OLS and LINMAP.

The mean product-moment correlation for 32 subjects is 0.638 in OLS (median = 0.630)

159

and 0.596 in LINMAP (median = 0.601).4 Since LINMAP uses a different algorithm
(linear programming) to calculate part-worths, the minor difference is under our
expectation. In addition, using OLS (SPSS) we were able to use the part-worth utilities
derived from a respondent’s first round survey to predict his or her rankings of 10 profiles
in the second round (cross-validation). Comparing the predicted values in the first round
with actual responses in the second round produces a Spearman rank order correlation of
0.663 (Corrl). The procedure then was reversed (using second round utilities to predict
first round results), the Spearrnan’s rank order correlation was 0.594 (Corr2). The results
from rank order correlations (grand mean = 0.629) are in accord with those from product
moment correlations (grand mean = 0.617).

Compared to other conjoint reliability research, is conjoint reliability in this study
good enough? Bateson et al. (1987) summarize previous reliability studies in conjoint
analysis: they report that product moment correlations range from 0.49 to 0.92 (mean =
0.747) with time lag from 1 day to 2 month and sample size from 8 to 116, while
Spearman rank order correlations are from 0.51 to 0.90 (mean = 0.717) with time lag
from 1 day to 1 month and sample size from 41 to 68. However, all these studies
evaluate the reliability in the areas of consumer behavior and marketing research rather
than peoples’ political attitudes. Therefore, both results in our study demonstrate a

reasonably high reliability for a survey involving political issues.

 

" We also report the median because most studies on conjoint reliability report only the median.

160

Table 8.1: Reliability of Conjoint Estimates for Ranking Data (Voters)

Pearson Product Moment Correlation

Spearman Rank Order Correlation

 

ID OLS LINMAP Corr1 Corr2
13 0.503 0.415 0.624 0.314
16 0.477 0.531 0.612 0.273
22 0.933 0.935 0.939 0.967
23 0.464 0.590 0.733 0.382
24 0.323 0.281 0.370 0.248
26 0.331 0.366 0.292 0.273
32 0.526 0.433 0.503 0.407
36 0.946 0.935 0.879 0.927
38 0.939 0.867 0.918 0.939
43 0.640 0.416 0.527 0.673
45 0.719 0.683 0.794 0.733
46 0.887 0.891 0.909 0.891
50 0.436 0.513 0.304 0.176
53 0.684 0.749 0.758 0.406
56 0.274 0.110 0.427 0.212
61 0.621 0.580 0.709 0.544
65 0.848 0.815 0.770 0.823
66 0.506 0.543 0.612 0.515
71 0.669 0.651 0.736 0.673
72 0.854 0.680 0.855 0.839
77 0.227 0.092 0.418 0.243
88 0.692 0.708 0.721 0.624
90 0.967 0.955 0.960 0.964
95 0.749 0.603 0.758 0.713
102 0.908 0.825 0.906 0.918
106 0.454 0.317 0.539 0.370
119 0.616 0.388 0.620 0.693
122 0.601 0.382 0.552 0.624
133 0.626 0.716 0.770 0.648
139 0.824 0.847 0.802 0.823
164 0.634 0.666 0.485 0.564
166 0.522 0.599 0.426 0.602
N: 32 32 32 32
Mean= 0.638 0.596 0.663 0.594
Median= 0.630 0.601 0.715 0.624
Grand Mean= 0.617 0.629

 

Note: Corr1 measures the correlation of responses using their first round utilities to

predict a subject’s responses in the second round. C orr2 reverses this process, using the

second round utilities to predict a subject’s responses in the first round.

161

Voter Sample: Rating Data

Table 8.2 reveals the reliability tests of conjoint analysis for rating data in the
voter sample. The mean product moment correlation for 27 subjects is 0.716 in OLS
(median = 0.762) and 0.657 (median = 0.708) in LINMAP. Furthermore, the cross-
validation from Spearman rank order correlation is 0.684 (Corr1) and 0.689 (Corr2)
respectively. Both reliability scores show that rating data have even more reliability than
ranking data. Since ratings usually produce tied scores and hence should yield lower
reliability scores, the higher correlations in rating sample could be caused by specific
sample characteristics. If we look at the reliability scores in Table 8.1, we see that there
are a couple of outliers with very low reliability scores. Respondents 56 and 77, for
instance, only had the product moment correlation 0.274 and 0.227 in OLS respectively.
These low reliability scores indicate that both respondents answered the surveys
carelessly or have very low involvement in the survey. In contrast, the rating sample
shows a more homogenous product moment correlation and rank order correlation.
Hence, rating sample exhibits higher reliability scores than ranking sample in this study.

As we have shown, the reliability test we used here also provides a way to identify
careless respondents and hence we could remove them from the sample or give less
weight in the simulation in order to obtain more accurate information or predict market

share.5

 

5 Another way to identify careless respondents is to use holdout cases. We will discuss this later.

162

Table 8.2: Reliability of Conjoint Estimates for Rating Data (Voters)

Eggrson Product Moment Correlation

Spearman Rank Order Correlation

 

ID OLS LINMAP Corr1 Corr2
6 0.777 0.776 0.673 0.637
34 0.813 0.787 0.818 0.827
111 0.767 0.739 0.818 0.809
138 0.822 0.612 0.772 0.779
141 0.737 0.663 0.782 0.685
159 0.630 0.767 0.745 0.699
171 0.779 0.764 0.818 0.839
173 0.747 0.541 0.648 0.602
175 0.761 0.775 0.758 0.790
176 0.926 0.805 0.915 0.916
178 0.762 0.754 0.729 0.774
179 0.846 0.890 0.879 0.895
180 0.729 0.628 0.564 0.594
181 0.320 0.390 0.358 0.378
185 0.823 0.800 0.782 0.737
188 0.550 0.550 0.430 0.382
192 0.547 0.392 0.382 0.559
219 0.635 0.429 0.697 0.708
224 0.829 0.857 0.867 0.860
225 0.973 0.898 0.967 0.947
227 0.834 0.583 0.636 0.674
230 0.401 0.433 0.321 0.322
232 0.597 0.462 0.564 0.611
239 0.892 0.774 0.891 0.895
242 0.666 0.708 0.503 0.559
243 0.398 0.318 0.345 0.322
244 0.781 0.657 0.806 0.799
N: 27 27 27 27
Mean: 0.716 0.657 0.684 0.689
Median= 0.762 0.708 0.745 0.708
Grand Mean= 0.687 0.686

 

Student Sample: Ranking

The reliabilities of conjoint analysis for the student sample are captured in the
Table 8.3. The reliability correlations from student sample indicate higher reliability
scores than those from both voters’ sample: the product moment correlation is 0.733 in

OLS and 0.741 in LINMAP individually while the Spearman rank order correlation is

163

0.707 (Corr1) and 0.709 (Corr2). The higher reliability scores in the student sample are
expected because we used 18 rather than 10 profiles and hence respondents have more
chances to correct mistakes and make consistent responses. Nevertheless, we should
notice that this is not to say that the more profiles, the better reliability; too many profiles,
i.e., over 25 profiles, will cause the fatigue effects and the problem of information
overload for respondents.

Table 8.3: Reliability of Conjoint Estimates for Ranking Data (Students)

Pearson Product Momefnt Correlation Sgarman Rank Order Correlation

 

ID OLS LINMAP Corr1 Corr2
5 0.947 0.931 0.938 0.956
8 0.713 0.707 0.744 0.693
12 0.620 0.639 0.508 0.564
13 0.886 0.881 0.884 0.817
22 0.850 0.854 0.776 0.849
24 0.699 0.689 0.666 0.693
25 0.972 0.962 0.979 0.979
26 0.904 0.922 0.886 0.884
33 0.648 0.610 0.658 0.638
38 0.650 0.743 0.499 0.444
39 0.418 0.498 0.220 0.322
40 0.596 0.501 0.655 0.633
43 0.844 0.922 0.791 0.806
50 0.864 0.836 0.826 0.855
52 0.851 0.891 0.872 0.815
63 0.498 0.551 0.494 0.482
65 0.933 0.973 0.893 0.901
71 0.776 0.680 0.814 0.823
73 0.811 0.839 0.764 0.764
74 0.707 0.746 0.686 0.696
76 0.799 0.784 0.808 0.818
77 0.149 0.139 0.190 0.165
N: 22 22 22 22
Mean= 0.733 0.741 0.707 0.709
Median= 0.788 0.765 0.770 0.785
Grand Mean= 0.737 0.708

 

164

Validity of Conjoint Analysis

Cross-Validation

Data used for testing reliability also provide a method of cross-validation. The
overall profile utilities estimated from the first round are used to predict the preferences
for the second round. The predicted overall utilities in the first round are then correlated
with the actual overall utilities in the second round to obtain a measure of cross-
validation (Corr1). The procedure is then reversed by predicting from round 2 to round 1,
thus completing a double cross-validation (Corr2). We found that the cross-validity
scores are satisfactory in both voter and student samples: in Table 8.1 and 8.2, we
obtained the mean cross- validation scores (rank order correlation) 0.629 for ranking data
and 0.686 for rating data; in Table 8.3, the mean cross-validation score (rank order
correlation) is 0.708 for the student sample.

Holdout Profiles

Except for the measure of cross validation, the most widely used measure to test
validity is to include holdout profiles in the design. Holdout profiles are additional
profiles which the respondent evaluates beyond those generated by actual conjoint design.
The value of using holdout profiles is twofold. First, in the field holdout profiles are
useful in identifying inconsistent respondents whose responses are unlikely to correspond
to their behavior. These respondents can then be removed from the sample or given less
weight in the simulation. Second, it permits an immediate assessment of the relevance of
conjoint analysis to choice. Where a conjoint model appears not to correspond to choice,
it can be changed or improved. This permits the testing of different forms of conjoint

analysis and leads to continuous improvement in its predictive validity (Johnson 1997;

165

Huber 1987). In the student sample, we include three holdout profiles in our design. In
particular, we use the first 15 profiles (Profile 1-15) to predict last three profiles (Profile
16-18). As shown in Table 5.8 (Chapter 5), the Kendall’s tau for 3 holdouts in the
student sample is 1.00 with a confidence level 94%. The validity score from holdout
profiles indicates that the model’s predictive validity is satisfactory.

Besides, holdout profiles can be used to check validity for the specific group. For
example, one of the major party differences for Democrats and Republicans is their
attitude toward welfare spending; Democrats prefer an increase in welfare spending while
Republicans prefer cuts in welfare spending. In order to check the validity of the two
parties’ responses on budget priorities, we could choose two profiles which represent the
two parties’ preferences and see how well the prediction is. In order to do so, we choose
two profiles as holdouts from Profile 10 -18:

Democrat (Profile 17)

0 Education (+5%), Welfare (+5%), Crime (0%), Natural Resources (0%)
Republican (Profile 10)

0 Education (+5%), Welfare (-5%), Crime (-5%), Natural Resources (-5%)

We should point out that the holdout profiles may not completely reflect the two
parties’ policy positions because we need to choose profiles beyond the minimum
combinations (9 in this case).° Therefore, we have to choose holdout profiles between

Profile 10 to 18. In this case, we use the Profile 10 and 17 as holdout profiles. The

 

° In addition, due to the necessary and sufficient condition for orthogonal array, some profiles may be
unrealistic. A necessary and sufficient condition (for orthogonal array) that the main effects of any two
factors be uncorrelated is that each level of one factor occurs with each level of another factor with
proportional frequencies

166

results show that both party groups predict holdout profiles well: the Kendall’s tau for
two holdouts are 1.0 in both Democrat and Republican sample.

Although holdouts require less additional data to test validity, it is a relatively
weak measure of validity: the validity score in holdout profiles is obtained from the same
data in a single survey design rather than from a second set of profiles or from
respondent’s actual behavior (e.g., consumer’s buying behavior). Hence, in our study
cross-validation is a better and more appropriate measure of validity because it uses
different complete sets (fractional factorial design) to test validity.

Reliability of Traditional Survey Questions

Voter Sample: Ranking Data

Table 8.4 illustrates the reliability of traditional questions for ranking data in the
voter sample. The first row “identical answer” indicates the situation when a respondent
provides identical answers in the both time administrations. The second row “sign
violation” shows the situation where a respondent provides answers with different
directions in one of six following answers: from positive to zero, from positive to
negative, from zero to positive, from zero to negative, from negative to zero, and from
negative to positive. The row of “same direction” indicates when a respondent provide
answers with same direction ( i.e., from positive to positive) but with different intensity.
The row “combined” shows the combined percentage between “identical answer” and
“same direction” (we count the row “combined” as reliable responses). Then we
calculate the grand mean by averaging the combined percentages for four spending areas

and “identical answer” in overall spending.

167

Table 8.4: Reliability of Traditional Survey Questions for Ranking Data (Voters)

 

 

 

 

Overall Education Welfare Crime Natural Mean

Identical Answer 65.63% 43.75% 43.75% 43.75% 37.50% 46.88%
N: 21 14 14 14 12
Sign Violation 34.38% 28.13% 28.13% 34.38% 25.00% 30.00%
N: 11 9 9 11 8
Same Direction NIA 28.13% 28.13% 21.88% 37.50% 28.91%
N= NIA 9 9 7 12
Combined NIA 71.88% 71 .88% 65.63% 75.00% 71.09%
N: N/A 23 23 21 24
Grand Mean 70.00%
Total N= 32 32 32 32 32 32

Sign Violation: Differences blt lst and 2nd Surveys
Mean 5.01% 6.37% 7.07% 2.21%
Range 8.00% 24.00% 8.70% 4.30%

Same Direction: Differences bIt 1st and 2nd Surveys
Mean 14.36% 5.93% 10.43% 54.48%
Range 94.00% 8.40% 36.00% 468.00%

 

In the overall spending, the result reveals that respondents have the highest

consistency in the category “identical answers”: 65.63 percent (21 out of 32 cases) voters

provide identical answers in ranking sample. In other spending areas, the category

“identical answer” drops about 10 percent or more (i.e., identical answers in education,

welfare, and crime are 43.75 percent and natural resources is 37.59 percent). On average,

in ranking data 46.88 percent voters give identical answers. In addition, in all four areas

(education, welfare, crime, and natural resources), on average 28.91 percent voters offer

answers with the same directions. If we count the category “identical answer” and “same

direction” as “reliable responses,” we obtain a score (grand mean) 70 percent of

consistent responses. However, we should notice that in general 30 percent of the

168

respondents have sign violations. Among them, respondents gave crime issue most
inconsistent answers with 34.88 percent sign violation.

Since we asked respondents to fill in their preferred spending levels in traditional
questions, we could also examine the “intensity” of their policy preferences by looking at
the preferred spending level differences between first and second round of surveys. For
instance, a respondent’s preferred spending levels in education may be +100% in the first
round but only +5% in the second round. The preferred spending level difference is 95%
in this case. In the row “mean difference,” we averaged the preferred spending level
differences between the first round and second round of surveys in the four policy areas
for all respondents. In the education area, i.e., there are 9 cases in the “same direction”
and their preferred spending level differences are: 3.2%, 8%, 95%, 5%, 5%, 1%, 3%, 1%,
and 8% respectively. Therefore, the “mean difference” for education spending is 14.36%.
In the “range difference,” we calculated the range of preferred spending levels in each
spending area. For instance, the “range difference” for education spending is 94%

(95% - 1%). We repeat the procedure and then obtain the “mean difference” and “range
difference” from both categories “sign violation” and “same direction.”

In the category “sign violation,” the inconsistency between the first and second
round is reasonable: mean percentage differences are from a low 2.21 percent to a high
7.07 percent while the range differences are from a low 4.3 percent to a high 24 percent.
On the contrary, in the category “same direction,” we see significant preferred spending
level differences between the first and second round of surveys: the mean differences are
from a low 5.93 percent to a high 54.48 percent while range differences are from a low

8.4 percent to a high 468 percent.

169

Voter Sample: Rating Data

The reliability of traditional survey questions for rating data is captured in Table
8.5. First we notice that respondents provide highly consistent responses in the area of
overall spending: 85.19 percent of voters offer identical answers. Nevertheless, we also
see the percentages of identical answer drop approximately 40 percent in the individual
spending areas. Among them, education issue has the lowest percentage of identical
answer (25.93 percent) but also has lowest sign violation (25.93 percent) because 48.15
percent respondents provide answers in the same directions. Welfare and crime issues
have highest inconsistent responses with 40.74 percent sign violation in both areas. The
reliability scores in rating data are very close to those in the ranking data. On average,
69.63 percent respondents gave consistent responses in rating data (combining categories

identical answer and same direction).

170

Table 8.5: Reliability of Traditional Survey Questions for Rating Data (Voters)

 

 

Overall Education Welfare Crime Natural Mean
Identical Answer 85.19% 25.93% 40.74% 44.44% 44.44% 48.15%
N= 23 7 1 1 12 12
Sign Violation 14.81% 25.93% 40.74% 40.74% 29.63% 30.37%
N= 4 7 1 1 1 1 8
Same Direction NIA 48.15% 18.52% 14.81% 25.93% 26.85%
N: N/A 13 5 4 7
Combined NIA 74.07% 59.26% 59.26% 70.37% 65.74%
N= NIA 20 16 16 19
Grand Mean 69.63%
Total N: 27 27 27 27 27 27

Mean Difference
Range Difference

Sign Violation: Differences bIt 1st and 2nd Surveys

 

5.40%
6.20%

4.69%
8.00%

7.09%
29.00%

9.31%
49.00%

Same Direction: Differences bIt 1st and 2nd Surveys
6.29% 5.50% 5.25% 23.29%
23.00% 14.00% 7.00% 129.00%

 

Mean Difference
Range Difference

 

Looking at the preferred spending levels differences, we find the similar responses
with ranking data: preferred spending level differences in the sign violation category are
relatively lower than those from the same direction category with mean difference from
4.69 to 9.31 percent and range difference from 8 to 49 percent. In contrast, in the
category “same direction,” we see significant preferred spending level differences
between the first and second round of surveys: the mean differences are from a low 5.25
percent to a high 23.29 percent while range differences are from a low 7 percent to a high
129 percent.

Why do people show small preferred spending level differences when their

answers are inconsistent while fairly significant preferred spending level differences

171

appear when their answers are consistent but with different intensity? One possible
reason for this may be just due to chance because of the small case numbers in each
category. However, we see the significant differences between the category “sign
violation” and “same direction” in both ranking and rating data. Maybe there are some
psychological explanations for this phenomena; however, we do not know what is the
“correct” answer in the current study. No matter what reasons might be, further study is
necessary to unveil this question that people’s responses tend to show small differences
when they provide inconsistent answers while significant differences when they provide
consistent answers but with different intensity.

Student Sample: Ranking Data

Table 8.6 reveals the reliability of traditional survey questions for the student
sample. Since we only asked students to choose spending levels from increase, decrease,
or stay about the same in all spending areas, there are two categories in Table 8.5:
identical answer and sign violation. Similar to voter’s reliability in traditional survey
questions, students offer highly consistent answers in the area of overall spending with
identical answer 72.73 percent. Moreover, students offer most consistent responses in the
area of education: 90.91 percent students (20 out of 22 cases) provide identical answers in
the second round of survey. Unlike voters, students offer inconsistent answers in the
areas of natural resources and welfare with sign violations of 45.45 percent and 31.82
percent respectively. In general, 72.73 percent of the students provide reliable responses

with identical answers in five traditional questions.

172

Table 8.6: Reliability of Traditional Survey Questions for Ranking Data (Students)

 

 

 

Overall Education Welfare Crime Natural Mean
Identical Answer 72.73% 90.91% 68.18% 77.27% 54.55% 72.73%
N: 16 20 15 17 12
Sign Violation 27.27% 9.09% 31.82% 22.73% 45.45% 27.27%
N= 6 2 7 5 10
Total N: 22 22 22 22 22 22
Summary

Compared to previous conjoint reliability studies (with a mean product moment
correlation = 0.747 and a rank order correlation = 0.717), our conjoint estimates have
demonstrated satisfactory reliability scores in both voter and student samples. In ranking
data of voter sample, we obtained 0.617 of mean product moment correlation and 0.629
of mean rank order correlation. In rating data of voter sample, conjoint estimates
provided reliability scores 0.737 of mean product moment correlation and 0.708 mean
rank order correlation. In the student sample, conjoint results showed 0.737 of product
moment correlation and 0.708 of rank order correlation.

Since we used different measures to test reliability,7 we could not compare the
reliability between conjoint analysis and traditional survey. However, reliability of
traditional surveys at least could be served as an reference of sample characteristics.
From our traditional surveys questions, voters provided 70 percent “reliable” answers in

ranking data while 69.93 percent “reliable” answers in rating data. Similarly, students

 

7 Conjoint reliability is measured by product moment correlation and rank order correlation while
reliability in traditional survey is measured by percentage of identical answers and same directions that
respondents provided.

173

gave 72.73 percent “reliable” answers in ranking data. These results revealed a higher
than average reliability.8

In sum, before we can be fully confident of conjoint methods, further study needs
to be done to completely evaluate the reliability of conjoint estimates over time, over
different stimulus sets, and over different data collection methods - especially for those
political issues that tend to suffer greater numbers of stochastic shocks. Since this is the
first reliability and validity tests in conjoint analysis focused on political attitude, the

results we obtained here are encouraging and satisfactory.

 

3 For example, Zaller (1992) showed that respondents gave 45% to 55% reliable answers both times on
issues “American relations with Russia” and “Level of government services” (pp. 580-581).

174

CHAPTER 9

SUMMARY AND CONCLUSIONS

The contributions of this study are both methodological and substantial. First, we
have examined people’s budget priorities by asking them to make tradeoffs among four
spending areas: education, welfare and health, crime and public safety, and natural
resources and environmental control. Second, we have introduced a new survey tool -
conjoint analysis, which yields more accurate information about people’s preferences, to
the studies of public opinion and policy analysis. Now let us consider three broad
questions about this study. Why do we need a new survey tool? What do we learn from
the conjoint study? What is the potential applicability of conjoint analysis?

The Need of New Survey Tool

The quality of democratic government depends on the responsiveness of
policymakers to the preferences of the public. In order to do so, policymakers need to
collect accurate and complete information about the public’s true preferences. Studies
from survey research have shown that there are significant inconsistencies between
policymakers and public opinion with respect to public policy. Free and Cantril (1967),
for example, found that the public exhibited a combination of philosophical opposition to
welfare state policies and operational support for virtually the full range of welfare
programs. They argued that this “schizoid” belief pattern made it extremely difficult for
the government to engage in rational policy-making.

One of the reasons for this inconsistency is due to the limits of traditional surveys:

when a traditional survey asks voters to answer a set of disaggregated questions

175

 

unbounded by resource constraints, what it gets is the public’s “wish list” rather than their
true preferences. The “wish list” of result is one of the major reasons why we often see
the “schizoid” belief pattern from normal surveys. In addition, some policy areas, such as
environmental protection, when considered alone, it is difficult to oppose, but it tends to
lose support rather quickly when compared with competing interests such as lost
employment. Because in the political world policymakers always have to make tradeoffs
among competing interests, introducing a new survey tool is extremely important. The
new survey tool should define not only people’s “desires” but, more importantly, their
true “preferences.” While the former can be addressed with a traditional wish list type of
question on the survey, the latter can only be addressed with precisely designed tradeoff
type questions.

We believe that conjoint analysis is such a tool and can offer us a way to collect
more accurate information about voters’ true preferences. Conjoint analysis can yield
more accurate and complete information because it is different from traditional surveys in
two aspects. First, conjoint analysis is a trade-off exercise in which respondents are
forced to make preference decisions based on the relative importance of one policy or
service over another. It forces the respondents to make more “real world” decisions by
looking at complete descriptions of an array of policies or services and select the one(s)
they would choose in an actual “buying” environment. Specifically, a conjoint task is
valuable because it forces the respondent to evaluate competing interests, as between
more government services and tax increases. People usually try to avoiding making such

decisions by searching for unambiguous solutions involving dominance or following a

176

relatively clear heuristic. However, people make tradeoffs when they must “buy a
product” in the market or when they participated in the conjoint task.

Second, conjoint analysis employs a decompositional model which differs from
almost all other multivariate methods in that it can be carried out at the level of the
individual respondent. This means that the researcher can develop a separate model for
predicting the preference of each respondent. As shown in Chapter 7, we could simulate
all kinds of “what-if” scenarios to achieve a particular research purpose. In contrast, most
other methods of analysis take measures from each respondent but can only perform an
analysis using all respondents combined.

It is important to note that we are not saying that conjoint analysis is the only way
to conduct survey or even desirable in every situation. But when dealing with policy
issues that require voters to make tradeoffs among competing interests, conjoint analysis
is definitely one of best candidates. Of course, since conjoint analysis is a more difficult
and time-consuming task than traditional surveys, this additional information comes at
price. Nonetheless, when evaluating voters’ preferences on competing policies, usually
you get what you pay for and thus we believe that the additional investment in conjoint
analysis is worthwhile.

Empirical Findings by Using Conjoint Analysis

We have mentioned many advantages of conjoint analysis, but what do we learn
from our conjoint study? First, let us look at what we learned from both the traditional
survey and the conjoint study. In traditional surveys, we only know respondents favoring
an increase or a decrease in overall spending, or respondents’ preference orders among

three spending choices - increase, decrease or stay about the same. For instance, we

177

could acquire the individual preference for a particular area, such as a respondent
preferring an increase in education spending. However, we do not know how respondents
make tradeoffs among these spending areas and the weights for a particular spending
level. In contrast, in conjoint analysis we obtained respondents’ true preferences through
the part-worth utilities for each attribute level, the relative importances of attributes, the
profile overall utilities, and the possibility to simulate all kinds of “what-if” scenarios.
We have argued that conjoint analysis produces more accurate and complete information
about people’s budget priorities. The “more accurate” is possible because we hold
respondents accountable when they express their preferences (as discussed in Chapter 2).
The “more complete” is possible because respondents are forced to evaluate all factors
(which necessarily consists of a budget proposal) rather than the single dimensional
consideration.

From our samples, we pointed out that both voters and students are satisfied with
most of the current spending levels except education - they favored an increase in
education spending. However, the most striking part was that we obtained inconsistent
spending preferences between traditional survey and conjoint estimates. In traditional
survey questions, for instance, we found that voters most preferred a decrease in overall
spending, followed by an increase in overall spending, and maintaining overall spending
at the current level. In contrast, in conjoint estimates (voters making tradeoffs among
attributes), we showed that voters most favored maintaining overall spending at the
current level, followed by an increase in overall spending, and a decrease in overall
spending. These results revealed that there appeared significant inconsistencies between

voters’ answers based on single dimensional considerations and tradeoff exercises.

178

Similarly, in the area of natural resources, we found that in traditional survey questions
voters most preferred an increase in natural resources spending, followed by maintaining
natural resources spending at the current level, and a decrease in natural resources
spending. In contrast, in conjoint estimates, voters most favored maintaining natural
resources spending at the current level, followed by an increase in natural resources
spending, and a decrease in natural resources spending. Similar inconsistent results also
can be found in the student sample.

There is one important implication from above results. Most studies in public
opinion and policy analysis use data obtained from traditional surveys to interpret,
analyze, and evaluate all kinds of political issues and phenomena. If information
obtained from traditional surveys are inaccurate or incomplete, all research results based
on these surveys will be less persuasive and misleading. Therefore, if policymakers make
policy decisions based on the traditional survey results, these policies cannot reflect
people’s true preferences and hence the quality of democracy is discounted. Indeed,
some political analysts have used this contradictory survey data to blame the voters for
their apparent inconsistency while others accused politicians who told voters that was
possible to “have their cake and eat it too” (Free and Cantril 1967; Kuttner 1980; Sears
and Citrin 1982).

Second, our analysis of group differences in conjoint analysis also supported
some previous findings in the studies of public opinion. For example, we found that
women were more concerned with compassion issues such as education and welfare than
men; the generation New Deal 1 ( the oldest cohort in this study) favored an increase in

crime spending because of their vulnerability to street crime; Post-New Deal 2 (the

179

youngest cohort in this study) preferred either natural resources spending increased or the
status quo; Democrats favored an increase in education spending while Republicans were
satisfied with current spending on education; Democrats preferred increasing spending on
welfare spending while Republicans favored spending at the status quo but suffered a
serious utility loss when welfare spending increases; people with children at home had a
slightly higher utility increase with education spending increased.

We acknowledge that our sample size (N =128) is limited and hence these results
cannot be generalized to make inferences about the distribution of attitudes in the US.
Nevertheless, these findings are insightful because the results we obtained are based on
the respondents’ tradeoff exercise rather than on their single dimensional evaluation
(unrealistic choices among incomplete sets of options). In other words, these group
preferences in conjoint estimates can be viewed as their true preferences rather than their
wishes or desires.

Third, we conducted the first reliability and validity tests in conjoint analysis
which focused on political attitudes, and the results are encouraging. Although from our
voter and student samples we cannot provide a definite answer about which survey tool is
better, we know the conjoint methods yield answers at least as reliable with traditional
survey questions. One thing should be mentioned: before we can fully confident of
conjoint methods, further study needs to be done to completely evaluate the reliability of
conjoint estimates over time, over different stimulus sets, and over different data
collection methods - especially for those political issues that tend to suffer greater

numbers of stochastic shocks.

180

 

Lastly, we have demonstrated the applicability of conjoint analysis by various
simulations. Specifically, we showed that conjoint simulations can be served as
predicting the gubernatorial election in Michigan, targeting a particular group
(independent voters in this study), evaluate welfare policy, and distributing extra revenue
to maximize voter’s utility. These useful “what-if” simulations cannot be done through
the use of traditional surveys. Hence, in the fields of public opinion and policy analysis,
the potential for conjoint application is promising.

Potential Applicability of the Conjoint Analysis

As shown in Chapter 7, we have demonstrated the applicability of conjoint
analysis in different ways. Beyond these examples, conjoint analysis has great potential
to be applied in the areas of public opinion and policy analysis. One possible topic in the
area of public opinion is the concepts of materialism and postrnaterialism. Inglehart
(1971) operationalized the concepts of materialist and postmaterialist by asking
respondents to rank two most important goals from the following four goals faced in the
nation: (1) maintaining order in the nation, (2) giving the people more say in important
political decision, (3) fighting rising prices, and (4) protecting freedom of speech. Those
respondents who choose (1) and (3) are considered materialists and those who choose (2)
and (4) are considered postrnaterialists. Respondents who select one materialist value
and one postrnaterialist value are classified as having mixed value priorities.

In order to apply conjoint methods to this topic, we could dichotomize each of the
four questions to be either present or not. Since there are only 4 attributes with 2 levels in

each attribute, we could use a full set of profiles (2x2x2x2 = 16) to estimate the part-

181

worth utilities without causing the fatigue effect. What we can get from this conjoint
design are the responses in which respondents making tradeoff among these values. In
addition, the conjoint method is better than ranking because it is less obtrusive and can
avoid the problem of ranking.I

In addition to the application in public opinion, conjoint analysis provides a
tremendous potential for conducting cost-benefit analysis for public policies, measuring
people’s willingness to pay for public goods, engaging in program evaluation, and
exploring the decision-making process. Since policy-making and political decisions
always involves trade-offs among competing interests, the potential applicability for

conjoint analysis is encouraging and promising.

 

' Ranking forces respondents who may appreciate values equally to choose among them in their rank
ordering. But even if people have such meaningful rankings, when standing before actual decisions they
often balance and trade off values. For instance, while people may prefer freedom over equality, they may
end up opting for a solution that trades off some freedom over equality. The ranking task does not catch
this but conjoint analysis can. For a detailed discussion of the ranking problem, see Shamir and Shamir

(1995).

182

APPENDIX

CONJOINT SURVEY FORMS FOR STUDENT AND VOTER

Subject Consent Information

We appreciate your willingness to test a new survey instrument designed to measure
voter preferences. In order to ensure that you are fully informed and consent to

participate in this effort, please read through the following statements:

. Please understand that you are in no way obligated to participate in this
survey.

. You may discontinue your participation at any time without penalty.

. If you indicate your agreement at the end of this survey, you may be
reconnected by a researcher for a one-time, follow-up interview. This follow-
up interview is strictly optional and you may decline to participate at any time
without penalty. If you are not selected for a follow-up interview or choose
not to participate at that time, your involvement will be limited to this one
survey session.

. All responses are held in strict confidence. Subjects will remain anonymous
in any report of research findings. On request and within these restrictions
results may be made available to subjects.

. If you have any questions or concerns about this survey, please contact:
Professor Larry Heimann
Department of Political Science
Michigan State University
Ph: 517-353-3281 Fax: 517-432-1091
e-mail: heimann@pilot.msu.edu

You indicate your voluntary agreement to participate in this research by completing and
returning this questionnaire.

* Note: Subject consent form was presented to respondents in all surveys.

183

Overview on the State of Michigan Budget

Education expenditures covers all state contributions for elementary, secondary,
and higher education.

Health & Welfare expenditures covers state spending on social services, mental
health, Medicaid, and other health expenses.

Crime & Public Safety expenditures covers state spending on police services,
correctional facilities like prisons and jails, and other public safety expenses.

Natural Resources expenditures includes costs of natural resource management,
environmental protection, and environmental clean-up costs.

Currently, Michigan expenditures are as follows':

Education $11.26 billion 38.2%
Health 8. Welfare $9.83 billion 33.4%
Crime & Public Safety $1.86 billion 6.3%
Natural Resources $0.65 billion 2.2%
All other sgnding $5.87 billion 19.9%
Total Spending $29.47 billion 100.0%

Assume that "all other spending" is fixed at its current level. Any net cuts in
spending will be either placed in a "rainy-day" account or used to cut some form
of taxes in Michigan. Any net increase in spending will be funded by an increase
in some form of taxation in Michigan. The exact form such cuts or increases
would take have not yet been determined.

* Note: overview on the State of Michigan budget was presented to respondents in all surveys.

 

' Based on the actual spending in the 1996 fiscal year. Data came from the Department of Management and
Budget.

184

Survey Form (Student)

General Questions

1.

Generally speaking, do you think the State of Michigan should:
increase overall spending,

1:1 decrease overall spending,

C1 or keep overall spending at current levels?

In the area of education, you would prefer to see the State of Michigan:
1:1 increase spending in the area of education,

1:1 decrease spending in the area of education,

[:1 or maintain education spending at current levels?

. In the area of welfare and health, you would prefer to see the State of

Michigan:
13 increase spending in the area of welfare and health,
1:1 decrease spending in the area of welfare and health,
13 or maintain welfare and health spending at current levels?

In the area of crime and public safety, would you prefer to see the State of
Michigan:

Cl increase spending in the area of crime and public safety,

[:1 decrease spending in the area of crime and public safety,

1:1 or maintain crime and public safety spending at current levels?

. Inthe area of natural resources, would you prefer to see the State of Michigan:

1:1 increase spending in the area of natural resources,
1:1 decrease spending in the area of natural resources,
1:1 or maintain natural resources spending at current levels?

* Note: general questions were presented to both student surveys.

185

Survey Instructions

Below is a list of 18 gubernatorial candidates, A — J, and their positions on state
budget expenditures in four policy areas. In order to evaluate these candidates
and their positions, please follow these steps:

1. In the space on the left hand side of each candidate, please put a "V"
down for each candidate you would consider voting for.

2. After reviewing all the candidates you find acceptable, place a " 1" on the
right-hand side of the candidate you most prefer. After that, rank order
the remaining acceptable candidates.

3. After reviewing all the candidates you would not vote for, place an " l 8"
on the right-hand side of the candidate you least prefer. After that, place
a "17" next to the candidate who represents your second-least-preferred
position. Continue this process until all 18 candidates are rank-ordered.

After you have finished this, please turn to the next page for a few brief
demographic questions. All individual responses will be held in strict
confidence. Thank you very much for taking the time to fill this survey out.

* Note: survey instructions were presented to respondents in all surveys, except the profiles
for voter have only 10 rather than 18 and rating survey has different instructions.

186

Budgetary Proposals for Gubernatorial Candidates

(First Round for Student: Rank Order)

Education
Health & Welfare
Crime 8 Public Safety

Natural Resources

Education
Health 8. Welfare
Crime & Public Safety

Natural Resources

Education
Health & Welfare
Crime & Public Safety

Natural Resources

Education
Health & Welfare
Crime & Public Safety

Natural Resources

-5%

-5%

-5%

-5%

0%

5%

-5%

5%

0%

5%

5%

-5%

-5%

0%

0%

5%

5%

-5%

5%

5%

0%

-5%

0%

5%

5%

5%

0%

0%

Q—

-5%
5%
5%

0%

5%
0%
-5%

0%

-5%
0%
0%
-5%
_R__
-5%
5%
-5%

5%

187

0%

-5%

0%

0%

5%

5%

0%

-5%

0%

0%

-5%

0%

0%

0%

5%

-5%

5%

-5%

-5%

-5%

5%

0%

5%

5%

Background Information (for statistical analysis purposes only)

 

Sex: 1:1 Male 1:1 Female
Year in which you were born: 19 1:1 D
Education:
1:1 8th grade or less Cl Some high school
1:1 High school graduate or GED 1:1 Junior college/Technical school graduate
D Some college (1-3 yr.) Cl College graduate (4 yr.)
D Some post-graduate study [:1 Graduate degree
Race: Cl African-American or Black 1:1 Asian or Pacific Islander
U Native American 1:1 White or Caucasian
1:1 Hispanic or Latin-American 1:1 Other
Religion:
1:1 No religious group 1:1 Roman or Orthodox Catholic
1:1 Islamic 1:1 Jewish
1:] Mainline Protestant 1:1 Evangelical Protestant
1:1 Buddhist D Other non-christian

Political: Please place on an "X" on the location which best represents your party
affiliation

 

x x x x x x x
Strong Moderate Weak Independent Weak Moderate Strong
Democrat Democrat Democrat Republican Republican Republican

Are you currently married? D Yes 1:1 No
Do you have children living with you? 1:1 Yes 1:1 No

If so, what is the age of the oldest child? Youngest child?
Do you own your current home? 1:1 Yes D No

In a couple of months we'd like to recontact some of the people we've surveyed for a
short interview. If you would be willing to participate, please indicate in the space below.

I would be willing to be recontacted: Cl Yes 1:1 No
Please send me a Michigan Lottery ticket: D Yes 1:1 No

"‘ Note: background information was presented to respondents in all first round of surveys.

188

 

Budgetary Proposals for Gubernatorial Candidates

(Second Round for Student: Rank Order)

Education
Health & Welfare
Crime & Public Safety

Natural Resources

Education
Health & Welfare
Crime 8. Public Safety

Natural Resources

Educafion
Health & Welfare
Crime & Public Safety

Natural Resources

Education
Health & Welfare
Crime 8. Public Safety

Natural Resources

-5%

-5%

-5%

-5%

0%

0%

0%

0%

0%

-5%

5%

5%

5%

0%

-5%

0%

-5%

-5%

0%

-5%

5%

5%

L

0%

-5%

0%

0%

5%

0%

-5%

5%

Q_

-5%

5%

5%

0%

0%

5%

-5%

-5%

5%

5%

5%

0%

R

-5%

5%

0%

5%

189

-5%

0%

5%

-5%

5%

5%

0%

-5%

0%

5%

-5%

5%

0%

0%

5%

-5%

-5%

0%

0%

5%

-5%

0%

-5%

Survey Form (Voter)

General Questions

1.

2.

Generally speaking, do you think the State of Michigan should:
increase overall spending,

1:1 decrease overall spending,

U or keep overall spending at current levels?

In the area of education, you would prefer to see the State of Michigan:
1:1 increase spending by %,

U decrease spending by %,

1:1 or maintain education spending at current levels?

. In the area of welfare and health, you would prefer to see the State of

Michigan:
1:1 increase spending by %,
1:1 decrease spending by %,

1:1 or maintain welfare and health spending at current levels?

In the area of crime and public safety, would you prefer to see the State of

Michigan:
D increase spending by %,
1:1 decrease spending by %,

U or maintain crime and public safety spending at current levels?

In the area of natural resources, would you prefer to see the State of Michigan:
1:1 increase spending by %,

1:1 decrease spending by %,

1:1 or maintain natural resources spending at current levels?

"' Note: general questions were presented to both voter surveys.

190

 

Budgetary Proposals for Gubernatorial Candidates
(First Round for Voter: Rank Order)

_A_ _B_ _C_ _0_
Education -5% -5% -5% 0%
Health & Welfare -5% 0% 5% -5%
Crime & Public Safety -5% 0% 5% 0%
Natural Resources -5% 5% 0% 0%

_F_ _G_ _H_ _I_
Education 0% 5% 5% 5%
Health 8. Welfare 5% -5% 0% 5%
Crime & Public Safety -5% 5% -5% 0%
Natural Resources 5% 5% 0% -5%

Budgetary Proposals for Gubernatorial Candidates
(Second Round for Voter: Rank Order)

_A_ _B_ _G_ _D_
Education -5% -5% -5% 0%
Health 8. Welfare -5% 0% 5% -5%
Crime & Public Safety -5% 0% 5% 0%
Natural Resources -5% 5% 0% 0%

_F_ _G_ _H_ _I_
Education 0% 5% 5% 5%
Health & Welfare 5% -5% 0% 5%
Crime 8. Public Safety -5% 5% -5% 0%
Natural Resources 5% 5% 0% -5%

191

E

0%

0%

5%

-5%

K

0%

-5%

5%

-5%

0%

0%

5%

-5%

0%

-5%

5%

-5%

 

Survey Instructions (Second Round for Voter: Rating)

In this section, we are going to ask you to evaluate each of the 12 candidates using a 100
points scale, where 100 points means that this is your ideal candidate (on these four
issues) and 0 points means that this candidate and the policy position he/she represents is
a complete anathema to you. Please put your rating in the appropriate box. At the end of
the

survey are two brief yes or no questions comparing this format with the earlier format in
which you rank ordered the candidates.

Thank you very much for your time in complete this survey. Your participation is very
important to the success of this project.

Budgetary Proposals for Gubernatorial Candidates
(Second Round for Voter: Rating)

 

 

 

_A_ _B_ _c_ _c_ _E_ _F_
Education ' -5% 5% 5% 0% 5% 0%
Health 8: Welfare 0% -5% 0% 5% 0% -5%
Crime & Public Safety 0% 5% -5% -5% 5% 0%
Natural Resources 5% 5% 0% 5% 0% 0%
Candidate Rating [ pts| pts] pts] ptsT ptsj pts]

_G_ _H_ _I_ _J_ _K_ _L_
Education 0% -5% -5% 0% 5% 0%
Health & Welfare -5% -5% 5% 0% 5% 0%
Crime & Public Safety 5% -5°/o 5% 5% 0% 0%
Natural Resources -5% -5% 0% -5% -5% 0%
Candidate Rating [ ptsl pts[ pts] pts] pts] pts|

 

Format Questions:

1. Compared with the first survey where you rank ordered 10 candidates, did you find this
100 point rating method easier to complete? Yes No

2. Compared with the first survey where you rank ordered 10 candidates, did you find this
100 point rating method took more or less time to complete? Yes No

192

BIBLIOGRAPHY

193

BIBLIOGRAPHY

Abramson, Paul R. 1983. Political Attitudes in America: Formation and Change. W.H.
Freeman and Company.

----- . 1989. Generations and Political Change in the United States.” Research in Political
Sociology 4: 235-280.

----- , and Charles W. Ostrom, Jr. 1991. “Macropartisanship: An Empirical
Reassessment.” American political Science Review 85: 181-92.

----- , John Aldrich, and David Rohde. 1995. Change and Continuity in the I 992
Elections. Revised Edition, CQ Press.

Achen, Christopher. 1975. “Mass Political Attitudes and the Survey Response.”
American Political Science Review, 69: 1218-31.

----- . 1983. “Toward Theories of Data: The State of Political Methodology.” In Political
Science: The State of the Discipline, ed. Ada Finifter. Washington, DC: The
American Political Science Association. pp. 69-94.

Addelman, Sidney. 1962. “Orthogonal Main-Effect Plans for Asymmetrical Factorial
Experiments.” Technometrics 4: 21-46.

Akaah, Ishmael and Pradeep K. Korgaonkar. 1983. “An Empirical Comparison of the
Predictive Validity of Self-Explicated, Huber-Hybrid, Traditional Conjoint, and
Hybrid Conjoint Models.” Journal of Marketing Research 20: 187.

Anderson, Norman H. 1981. Foundation Measurement of Information Integration
Theory. New York: Academic Press.

----- . 1982. Methods of Information Integration Theory. New York: Academic Press.

Axelrod, Robert. 1967. “The Structure of Public Opinion on Policy Issues.” Public
Opinion Quarterly 31: 51-60.

----- . 1972. “Where the Votes Come From: an Analysis of Election Coalitions.”
American political Science Review 66: 11-20.

Barron, Hutton F. 1977. “Axiomatic Conjoint Measurement.” Decision Science 8: 548-
559.

194

Bateson, John E, David Reibstein, and William Boulding. 1987. “Conjoint Analysis
Reliability and Validity: A Framework for Future Research.” Review of
Marketing.

Beck, Paul Allen. 1974. “A Socialization Theory of Partisan Realignment.” In The
Politics of Future Citizens, ed. Richard Niemi, et al. San Francisco: Jossey-Bass.

Beckwith, N., and Lehamnn, D. 1975. “The Importance of Halo Effects in Multiattribute
Attitude Models.” Journal of Marketing Research 12: 265-275.

Bennett, Linda, and Stephen Earl Bennett. 1990. Living with Leviathan: Americans
Coming to Terms with Big Government. Lawrence: University Press of Kansas.

Berelson, Bernard, Paul Lazarsfeld, and William McPhee. 19544. Voting: A Study of
Opinion Formation in a Presidential Campaign. Chicago: University of Chicago
Press.

Bojamic, David C and Roger J .Calantone. 1990. “A Contribution Approach to Price
Bundling in Tourism.” Annual of Tourism Research, 17: 528-540.

Brady, Henry and Paul Sniderrnan. 1985. “Attitude Attribution: A Group Basis for
Political Reasoning.” American Political Science Review 79: 1061-1078.

Braungart, Richard and Margaret Braungart. 1989. “Political Generations.” Research in
Political Sociology 4: 281-319.

Brittan, Samuel. 1975. “The Economic Contradictions of Democracy. ” British Journal of
Political Science 5: 129-59.

Campbell, Augus, Philip E. Converse, Warren E. Miller, and Donald E. Strokes. 1960.
The American Voter. University of Chicago Press.

Campbell, Donald R. and Donald W. Fiske. 1959. “Convergent and Discriminant
Validation by the Multitrait-Multimethod Matrix.” Psychological Bulletin, 56: 81-
105.

Carrnines, Edward G. and Harold W. Stanley. 1992. “The Transformation of the New
Deal Party System: Social Groups, Political Ideology, and Changing Partisanship
Among Northern Whites, 1972-1988.” Political Behavior 14: 213-240.

Carroll, J. Douglas. 1969. “Categorical Conjoint Measurement.” paper presented at
Meeting of Mathematical Psychology, Ann Arbor, MI, August.

195

----- . 1973. “Models and Algorithms for Multidimensional Scaling, Conjoint
Measurement, and Related Techniques.” in Multiattribute Decisions in Marketing,
P.E Green, Y. Wind, eds. Hindsdale, IL: Dryden Press, 335-37; 341-48.

----- . and P. E. Green. 1995. “Psychometric Methods in Marketing Research: Part I,
Conjoint Analysis.” Journal of Marketing Research 32: 385-91.

Cattin, , Philippe, and Dick Wittink. 1976. “A Monte Carlo Study of Metric and
Nonmetric Estimation Methods for Multiattribute Models.” Research Paper No.
341 , Graduate School of Business, Standard University.

Chakraborty, Goutam, R. Ettenson, and G. Gaeth. 1994. “How Consumers Choose Health
Insurance.” Journal of Health Care Marketing 14: 21-33.

Citrin, Jack. 1979. “Do People Want Something for Nothing?” National Tax Journal 32:
113-29.
Conjoint Experiment. [Online] Available

http://www.hwc.ca/datapcb/Communc/generic/09stud51.htm, March 10, 1997.

Conner, W. S. and M. Zelen. 1959. Fractional Factorial Experimental Designs for
Factors at three levels. Applied Math Series S4, Washington, DC: National
Bureau of Standard.

Converse, Philip E. 1964. “The Nature of Belief System in Mass Publics.” In Ideology
and Discontent, ed. David Apter. New York: Free Press.

----- . 1974. “Some Priority Variables in Comparative Electoral Research.” In Electoral
Behavior, ed. Richard Rose, 727-745. New York: Free Press.

Croker, Royce. 1981. “Federal government Spending and Public Opinion.” Public
Budgeting & Finance Autumn: 25-35.

Curry, Joseph. 1997a. Conjoint Analysis: After the Basics. [Online] Available:
http:/www.swatooth.com/pages, May 24, 1997.

----- . 1997b. Understanding Conjoint Analysis: Predicting Choice. [Online] Available:
http:/www.swatooth.corn/pages, May 24, 1997.

Dalton, Russell. 1988. Citizen Politics in Western Democracies. NJ: Chatham House
Publishers, Inc.

Department of Management and Budget of Michigan. 1997. Executive Budget of
Michigan, Fiscal Year 1998.

Downs, Anthony. 1957. An Economic Theory of Democracy. New York: Harper.

196

Eismeier, Theodore J. 1982. “Public Preferences about Government Spending.” Political
Behavior 4: 133-45.

Erikson, Robert S. 1979. “The SRC Panel Data and Mass political Attitudes.” British
Journal of Political Science 9: 89-114.

----- , Gerald C. Wright, Jr, and John P. McIver. 1989. “Political Parties, Public Opinion,
and State Policy in the United States.” American Political Science Review 83:
728-50.

Free, Lloyd A., and Hadley Cantril. 1967. The Political Beliefs of Americans: A Study of
Public Opinion. New York: Simon and Schuster.

Gillian, Carol. 1982. In a Different Voice. Cambridge, MA: Harvard University Press.

Gold, Steven D. 1995. The Fiscal Crisis of the States: Lessons for the Future.
Georgetown University Press.

Goldberg, Lewis R. 1965. “Diagnostics vs. Diagnostic Signs: The Diagnosis of Psychosis
vs. Neurosis from the MMPI.” Psychological Monographs: General and Applied
79: 1-27.

Green, Paul E. 1974. “On the Design of Choice Experiments Involving Multifactor
Alternatives.” Journal of Consumer Research 1: 61 -68.

----- , and V. Srinivasan. 1973. Multiattribute Decisions in Marketing: A Measurement
Approach. The Dryden Press.

----- , and V. Srinivasan. 1978. “Conjoint Analysis in Consumer Research: Issues and
Outlook.” Journal of Consumer Research 5: 103-23.

----- . and V. Srinivasan. 1990. “Conjoint Analysis in Marketing: New Developments with
Implications for Research and Practice.” Journal of Marketing 54: 3-19.

----- . and Yoram Wind. 1975. “New Way to Measure Consumers’ Judgments.”
Harvard Business Review July-August: 107-117.

----- , Carroll, .1. Douglas, and Carmone, Frank J. 1978. “Some New Types of Fractional
Factorial Designs for Marketing Experiments.” in Research in Marketing Vol. 1,
ed. J. N. Sheth, Greenwich, CT: JAI Press.

Hagerty, Michael R. 1986. “The Cost of Simplifying Preference Models.” Marketing
Science 5: 298-321.

197

Hahn, GI. and S. S. Shapiro. 1966. A Catalog and Computer Program for the Design
and Analysis of Orthogonal symmetric and Asymmetric Fractional Factorial
Experiments. Technical Report 66-C 165. Schenectady, New York: General
Electric Research and Development Center.

Hair, J. F., Anderson, R.E., Tatham, R.L. and Black, W. C. 1995. Multivariate Data
Analysis: with Readings, 4th Ed. Englewood Cliffs, NJ: Prentice Hall.

Hansen, John M. 1998. “Individuals, Institutions, and Public Preferences over Public
Finance.” American Political Science Review 92: 513-31.

Hicks, Alexander M. and Duane H. Swank. 1992. “Politics, Institutions, and Welfare
Spending in Industrialized Democracies.” American Political Science Review, 86:
658-74.

Hill, Kim Quaile and Jan E. Leighley. 1992. “The Policy Consequences of Class Bias in
State Electorates.” American Journal of Political Science 36: 351-65.

----- , Jan E. Leighley, and Angela Hinton-Anderson. 1995a. “Lower Class Mobilization
and Policy Linkage in the US. States.” American Journal of Political Science 39:
75-86.

----- , and Angela Hinton-Anderson. 1995b. “Pathways of Representation: A Causal
Analysis of Public Opinion-Policy Linkages.” American Journal of
Political Science 39: 924-35.

Hooley, Graham. 1981. “Modeling the Student University Choice Process.” European
Research 9: 158-1570.

Hu, Clark. Conjoint Analysis Library. [Online] Available
http://www.nevada.edu/~huc/html/welcome.htrnl. November 6, 1996.

Huber, Joel. 1987. “Conjoint Analysis: How We Got Here and Where We Are.” Sawtooth
Software Conference Proceedings.

Inglehart, Ronald. 1971. “The Silent Revolution in Europe: Intergenerational Change in
Post-Industrial Societies.” American Political Science Review 65: 991-1071.

Jain, Arun, Frankin Acito, Naresh Malhotra, and Vijay Mahajan. 1979. “A Comparison of
the Internal Validity of Alternative Parameter Estimation Methods in

Decompositional Multiattribute Preference Models.” Journal of Marketing
Research 16: 313-22.

Jacoby, William G. 1994. “Public Attitudes Toward Government Spending. ” American
Journal of Political Science 38: 336-61.

198

 

Jelen, Tred G. and Clyde Wilcox. 1995. Public Attitudes Toward Church and State. M. E.
Sharpe, Inc. New York.

Johnson, Richard. 1974. “Trade-off Analysis of Consumer Values.” Journal of Marketing
Research 11: 121-127.

----- . 1997. Including Holdout Choice Tasks in Conjoint Studies. [Online] Available:
http:/www.swatooth.com/pages, May 24, 1997.

Johnson, R. Burke. 1995. “Estimation An Evaluation Utilization Model Using Conjoint
Measurement and Analysis.” Evaluation Review, 19, 313-338.

Judd, Charles and Michael Milbum. 1980. “The Structure of Attitude systems in the
General Public: Comparison of Structural Equation Model.” American

Sociological Review 45: 627-43.

Kamlet, Mark S. and David C. Mowery. 1987. “Influences on Executive and
Congressional Budgetary Priorities, 1955-1981.” American Political Science
Review, 81: 155-78.

Kamakura, Wagner and Rajendra K. Srivastava. 1984. “Predicting Choice Shares Under
Conditions of Brand Interdependence.” Journal of Marketing Research 21: 420-
34.

Kennedy, Peter. 1992. A Guide to Econometrics. Third Edition. The MIT Press:
Cambridge, Massachusetts.

Key, V0. 1952. Politics, Parties, and Pressure Groups. New York: Knopf.
----- . 1959. “Secular Realignment and Party System.” Journal of Politics 21: 198- 210.
----- . 1961. Public Opinion and American Democracy. New York: Knopf.

----- . 1966. The Responsible Electorate: Rationality in Presidential Voting, 1936-60.
Cambridge, MA: Harvard University Press.

Kiewiet, D. Roderick, and Mathew D. McCubbins. 1985. “congressional Appropriations
and the Electoral Connection.” The Journal of Politics 47: 59-82.

Kingdon, John. 1995. Agenda, Alternatives, and Public Policies. 2nd Edition. Harper
Collins College Publishers.

Krantz, David H. 1964. “Conjoint Measurement: The Luce-Tukey Axiomatization and
Some Extensions.” Journal of Mathematical Psychology 1: 248-77.

199

Kruskal, Joseph B. 1965. “Analysis of Factorial Experiments by Estimating Monotone
Transformations of the Data.” Journal of the Royal Statistical Society, Series B,
27: 251—63.

Kuttner, Robert. 1980. Revolt of the Haves: Tax Rebellions and Hard Times. New York:
Simon and Schuster.

Lacy, Dean. 1998. “A Theory of Nonseparable Preferences in Survey Responses.” Paper
presented at the Annual Meeting of the Midwest political Science Association,
Chicago, IL, April 23-25, 1998.

Ladd, Everett Carl, Mayilyn Potter, Linda Basilick, Sally Daniels, and Dana Suszkiw.
1979. “The Polls: Taxing and Spending.” Public Opinion Quarterly 43: 126-35.

Lazarsfeld, Paul, Bernard Berelson, and Hazel Gaudet. 1944. The People ’s Choice. New
York: Duell, Sloan and Pearce.

Leigh, Thomas W., D. MacKay, and J. Summers. 1984. “Reliability and Validity of
Conjoint Analysis and Self-Explicated Weights: A Comparison.” Journal of
Marketing Research 21 : 456-62.

Lennan, SR. and J .J . Louviere. 1978. “On the Use of Direct Utility Assessment to
Identify the Functional Form of Utility and Destination Choice Models. ”
Transportation Research Record 673: 78-86.

LINMAP, 1989. Conjoint LINMAP manual. Bretton-Clark.

Louviere, Jordan J. 1988. Analyzing Decision Making: Metric Conjoint Analysis. Sage
Publications.

Marchant-Shapiro, Theresa and Kelly D. Patterson. 1995. “Partisan Change in the
Mountain West.” Political Behavior 17: 359-378.

McClain, John and Rao, Vithala. 1974. “Trade-offs and Conflicts in Evaluation of Health
systems Alternatives: A Methodology for Analysis.” Health Services Research, 9:
35-52.

McLean, R. and V. Anderson. 1984. Applied Factorial and Fractional Designs. New
York: Marcel Deckker.

Miller, Warren E. 1992. “Generational Changes and Party Identification.” Political
Behavior 14: 333-352.

200

----- , and Merrill Shanks. 1996. The New American Voter. Cambridge: Harvard
University Press.

Modigliani, Andre and Franco Modigliani. 1987. “The Growth of the Federal Deficit and
the Role of Public Attitudes.” Public Opinion Quarterly 51: 459-80.

Monroe, Alan D. 1979. “Consistency Between Public Preferences and National Policy
Decisions.” American Politics Quarterly 7: 3-19.

----- . 1983. “American Party Platforms and Public Opinion.” American Journal of
Political Science 27: 27-42.

Niemi, Richard and Herbert Weisberg. 1993. Classics in Voting Behavior. Washington,
DC: CQ Press.

Neslin, Scott A. 1981. “Linking Product Features to Perceptions: Self Stated versus
Statistically Revealed Importance Weights.” Journal of Marketing Research 18:
80-86.

Olshavsky, R., and Granbois D. 1979. Consumer Decision Making - Fact or Fiction?”
Journal of Consumer Research 6: 93-100.

Page, Benjamin 1 and Robert Y. Shapiro. 1983. “Effect of Public Opinion on Policy.”
American Political Science Review 77: 175-190.

----- . and Robert Y. Shapiro. 1992. The Rational Public: F ifiy Years of Trends in
Americans ’ Policy Preferences. Chicago: University of Chicago press.

Parker, Barnett and V. Srinivasan. 1976. “A Consumer Preference Approach to the
Planning of Rural Primary Health Care Facilities.” Operations Research 24: 991-
1025.

Pekelrnan, Dov and Subrata K. Sen. 1979. “Improving Prediction in Conjoint
Measurement.” Journal of Marketing Research 16: 21 1.

Peter, Paul J. 1979. “Reliability: A Review of Psychometric Basics and Recent Marketing
Practices.” Journal of Marketing Research 16: 6-17.

Peterson, Paul E. 1985. “The New Politics of Deficits.” In The New Directions in
American Politics, ed. John E. Chubb and Peterson, Paul E. Washington, DC: The
Brookings Institution.

----- , and Mark Rom. 1989. “American Federalism, Welfare Policy, and Residential
Choices.” American Political Science Review, 83: 711-28.

201

Popkin, Samuel. 1991. The Reasoning Voter. Chicago: University of Chicago Press.

Ramsing, K and J. Wish. “What do Library Users Want? A Conjoint Measurement
Technique May Yield the Answer,” Information Processing and Measurement

18: 237-242.

Reibstein, David, J. Bateson, and W. Boulding. 1988. “Conjoint Analysis Reliability:
Empirical Findings.” Marketing Science 7: 271-86.

Rentz, Joseph 0. “Generalizability Theory: A Comprehensive Method for Assessing and
Improving the Dependability of Marketing Measures.” Journal of Marketing
Research 16: 19-28.

Rubin, Irene S. 1993. The Politics of Public Budgeting. Second Edition. Chatham House
Publishers, Inc.

Sanders, Arthur. 1988. “Rationality, Self-Interest, and Public Attitudes on public
Spending.” Social Science Quarterly 69: 311-24.

Schick, Allen. 1990. The Capacity to Budget. The Urban Institute Press.

Schuman, Howard and Stanley presser. 1996. Questions and Answers in Attitude Surveys.
Sage Publications.

Sears, David 0., and Jack Citrin. 1982. Tax Revolt: Something for Nothing in California.
Cambridge: Harvard University Press.

Shamir, Michal and Jacob Shamir. 1995. “Competing Values in Public Opinion: A
Conjoint Analysis.” Political Behavior 17:107-113.

Shapiro, Robert Y. and Harpreet Mahajan. 1986. “Gender Differences in Policy
Preferences: A Summary of Trends from the 19608 to the 19803.” Public Opinion
Quarterly 50: 42-61.

Shocker, Allan D. and V. Srinivasan. 1977. “LINMAP (Version II): A FORTRAN IV
Computer Program for Analyzing Ordinal Preference (Dominance) Judgments

Via Linear Programming Techniques for Conjoint Measurement.” Journal of
Marketing Research 14: 101-103.

Shukla, PK. 1990. Measuring Faculty Preferences for Collective Bargaining: A Conjoint
Analysis with Extensions to Selected Illustrations in Educational Administration,

unpublished dissertation, UCLA.

. ----- , and James Bruno. 1991. “Use of Conjoint Analysis and Marketing Approaches in
Education Survey.” Education 112: 451-58.

202

Smith, Tom W. 1984. “The Polls: Gender and Attitudes Toward Violence.” Public
Opinion Quarterly 48: 384-96.

Sniderman, Paul M. 1993. “The New Look in Public Opinion Research.” in Ada F inifter
ed. Political Science: The State of the Discipline II. American Political Science
Association.

Sparling, J. 1977. Conjoint Scaling as a Decision Aid in Curriculum Development. Paper
presented at 1977 AERA National Conference.

SPSS. 1997. SPSS Conjoint 8.0 Manual. SPSS, Inc.

Stoper, Emily and Roberta A. Johnson. 1977. “The weaker Sex and the Better Half: The
Idea of Women’s Moral Superiority in the American Feminist Movement.” Polity
10: 192-217.

Strauss, William and Neil Howe. 1991. Generations: The History of American ’s Future,
1584 to 2069. William Morrow and Company, Inc: New York.

Srinivasan, V. and Allan D. Shocker. 1973. “Linear Programming Techniques for
Multidimensional Analysis of Preferences.” Psychometrika 38: 337-69.

Stanley, Harold W. and Richard G. Niemi. 1991. “Partisanship and Group support, 1952-
1988.” American Politics Quarterly 19: 189-210.

Stimson, James A. 1991. Public Opinion in America: Moods, Cycles, and Swings.
Boulder: Westview Press.

Sundquist, James L. 1983. Dynamics of the party System: Alignment and Realignment of
Political Parties in the United States, Revised Version. Washington, DC:
Brookings Institution.

The National Election Studies, Center for Political Studies, University of Michigan. The
NES Guide to Public Opinion and Electoral Behavior. [Online] Available
http://www.umich.edu/~nes/nesguide/nesguide.htm. Ann Arbor, MI: University
of Michigan, Center for Political Studies [producer and distributor], 1995-1998.

Turner, CF. and E. Kruss. 1978. “Fallible Indicators of the Subjective State of the
Nation.” American Psychologist 33: 456-470.

Tversky, Amos. 1967. “A General Theory of Polynomial Conjoint Measurement.”
Journal of Mathematical Psychology 4: 1-20.

203

Wald, Kenneth D. 1997. Religion and Politics in the United States. Third Edition. CQ
Press: Washington, DC.

Welch, Susan. 1985. “The ‘More for Less’ Paradox: Public Attitudes on Taxing and
Spending.” Public Opinion Quarterly 49: 310-16.

Whitmore, GA. and Cavadias, GS. 1974. “Experimental Determination of Community
references for Water Quality-Cost Alternatives.” Decision Sciences 5: 614-31.

Wildavsky, Aaron. 1988. The New Politics of the Budgetary Process. Scott, F oresman
and Company.

Wilson, L.A. II. 1983. “Preference Revelation and Public Policy: Making Sense of
Citizen Survey Data.” Public Administration Review July/August: 335-342.

Wind, Yoram and Lawrence Spitz. 1976. “An Analytical Approach to Marketing
Decisions in health Care Organizations.” Operations Research 24: 973-90.

Wittink, Dick and Philippe Cattin. 1989. “Commercial Use of Conjoint Analysis: An
Update.” Journal of Marketing 53: 91-96.

----- , D, M. Vriens, and W. Burhenne. 1994. “Commercial Use of Conjoint Analysis
in Europe: Results and Critical Reflections.” International Journal of Research in
Marketing, 11, 41-52.

Wlezien, Christopher. 1995. “The Public as Thermostat: Dynamic of Preferences for
Spending.” American Journal of Political Science 39: 981-1000.

Young, Forrest W. 1972. “A Model for Polynomial Conjoint Analysis Algorithms.” in
Multidimensional Scaling: Theory and Applications in the Behavioral Sciences
Vol. 1, R. N. Shepard, A. K. Romney, and S. Nerlove, eds. New York: Academic
Press: 69-104.

Zaller, John. 1992a. “A Simple Theory of the Survey Response: Answering Questions
versus Revealing Preferences.” American Journal of Political Science, 36, 579-

616.

----- . 1992b. The Nature and Origins of Mass Opinion. New York: Cambridge
University Press.

Zufryden, Fred. 1983. “Course Evaluation and Design Optimization: A Conjoint Analysis
Based Application.” Interfaces 13: 87-94.

204

HICHIGAN smTE UNIV. LIBRQRIES
llHIHill!!!IIIMIHNHIIlIlHllIIII\IIIIIWIINHIHIIII
31293017568058