VALUE SYSTEM STABILITY IN THE MASS PUBLIC
By
David Ciuk

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Political Science
2011

Abstract
VALUE SYSTEM STABILITY IN THE MASS PUBLIC
By
David Ciuk
Are individual value choices stable across different situational contexts? Traditional
psychological theory provides a positive answer to this question. But, new research in
political science concludes otherwise, causing many researchers to question whether values
have the pervasive influence on human behavior that has long been assumed to exist. In
this dissertation, I test the stability and relevance of value choices in several ways. First,
I use an experimental design and new data funded by the TESS program to test whether
value choices are subject to priming effects. Second, I use a matching design and data
from three different time points over the past 16 years to examine the distribution of value
choices at the aggregate level. Third, I return to the TESS data and test the relationships
between values and various forms of political behavior.
In short, my results suggest, at the individual level, that some values are stable and
some are not. At the aggregate level, values “move” over time, but they do so predictably.
Lastly, my results suggest that values have a rather heavy influence on evaluations of
Barack Obama, but their influence on turnout and vote choice is negligible, at best.
Taken as a whole, this dissertation speaks to a contentious body of literature that
focuses on the theoretical status of values as fundamental influences on human behavior.

To Mom and Dad, to Dawn, and to Bruce and Trent:
Thanks for all your patience, support, and love.

iii

Acknowledgments

I owe thanks to many of my graduate student colleagues at MSU. Without their
continuous support and encouragement, graduate school would have been a much longer
and much rougher road.
I also owe thanks to Steve Kautz and Chuck Ostrom for their support throughout my
five years at MSU. Their advice and suggestions helped make me a well-rounded scholar
and prepared me for life in academia.
I owe many, many thanks to Paul Abramson, Eric Chang, Sandy Marquart-Pyatt,
and Sandy Schneider for all their effort, suggestions, and advice that improved my work.
Meetings and conversations with these folks were a pleasure and always pushed my work
forward.
And last, but certainly not least, I owe a very special thanks (a lifetime worth of
thanks, in fact) to Bill Jacoby. Bill served as chair of the dissertation committee, and
without him, this dissertation would not have been possible. His thoughtful critiques,
his reading (and rereading) my work, and his ability to push, encourage, and to motivate
propelled me to successfully finish and defend this dissertation.

iv

Contents

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

viii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

1 Introduction
1.1 Past Scholarship on Attitude Structure . . . . . . . . . . . . . . . . . . .
1.2 New Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Layout of Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1
2
3
4

2 Literature Review
2.1 Previous Research on Attitude Organization
2.2 Definitions . . . . . . . . . . . . . . . . . . .
2.3 Major Players in Values Research . . . . . .
2.4 Values in Political Science . . . . . . . . . .
2.5 Assumptions, or Testable Hypotheses? . . .
2.6 A New Contribution . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

6
6
7
9
11
13
15

3 Measurement and Value System Stability
3.1 Research on Value System Dynamics . . . . . . . . . . . . . .
3.2 Value System Measurement . . . . . . . . . . . . . . . . . . .
3.3 Models of Importance Ratings & Rankings, and Assumptions .
3.3.1 A Model of Value Importance Ratings . . . . . . . . .
3.3.2 A Model of Ranked Value Preferences . . . . . . . . . .
3.3.3 Differences Between the Models . . . . . . . . . . . . .
3.4 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Research Design & Data . . . . . . . . . . . . . . . . . . . . .
3.6 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7.1 Transitivity in Value Rankings . . . . . . . . . . . . . .
3.7.2 Differences Between Experimental Groups . . . . . . .
3.7.3 Results of Ordered and Rank-Ordered Logit Models . .
3.8 Conclusion & Discussion . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

16
16
17
18
19
24
25
26
28
29
33
34
35
36
40

4 The Temporal Stability of Values
4.1 Research on the Temporal Stability of Value Systems . . . . . . . . . . .
4.2 Research Design, Data, & Methods . . . . . . . . . . . . . . . . . . . . .

54
55
57

v

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

4.3
4.4

4.5

4.2.1 A Note on Matching . .
Hypotheses . . . . . . . . . . .
Results . . . . . . . . . . . . . .
4.4.1 Matched and Unmatched
4.4.2 Rank-Ordered Logit . .
Conclusion & Discussion . . . .

. . . .
. . . .
. . . .
Data
. . . .
. . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

5 Values, Attitudes, and Behavior
5.1 Past Literature . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 How Values Affect Attitudes . . . . . . . . . . . . . . . . . . .
5.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1 A Note on Multivariate Probit with Selection . . . . .
5.4.2 Measurement . . . . . . . . . . . . . . . . . . . . . . .
5.5 Formal Hypotheses . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Results Pertaining to Policy-Based Approval for the President
5.7 Results Pertaining to Turnout and Vote Choice . . . . . . . .
5.8 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . .
6 Conclusions
6.1 Chapter 3: Transitivity, Stability, and Measurement . . . .
6.2 Chapter 4: The Effects of Contextual Salience Over Time .
6.3 Chapter 5: Values, Attitudes, the Vote, and Measurement
6.4 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Avenues for Future Research . . . . . . . . . . . . . . . . .

vi

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.

59
61
63
64
66
66

.
.
.
.
.
.
.
.
.
.

76
76
78
79
80
80
81
82
82
86
88

.
.
.
.
.

103
103
105
107
109
112

List of Tables

3.1

Percent With Clear Preferences and Consistent, by Value Pair . . . . . .

47

3.2

Percent Transitive, by Triad . . . . . . . . . . . . . . . . . . . . . . . . .

47

3.3

Difference in Mean Rating/Ranking of Equality . . . . . . . . . . . . . .

48

3.4

Difference in Mean Rating/Ranking of Economic Security . . . . . . . . .

48

3.5

Ordered Logit and OLS Models of Ratings of Equality and Econ. Security

49

3.6

Changes in Predicted Probabilities by Treatment Group for Equality Rating 50

3.7

OLS Models for Value Rankings . . . . . . . . . . . . . . . . . . . . . . .

51

3.8

Rank-Ordered Logit Model (Equality as Reference Category) . . . . . . .

52

3.9

Rank-Ordered Logit Model (Economic Security as Reference Category) .

53

4.1

Mean Rankings by Year . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

4.2

OLS Results on Matched and Unmatched Data, 1994 & 2003 . . . . . . .

72

4.3

OLS Results on Matched and Unmatched Data, 2003 & 2005 . . . . . . .

73

4.4

OLS Results on Matched and Unmatched Data, 1994 & 2005 . . . . . . .

74

4.5

Rank-Ordered Logit Results (Social Order is Reference Category) . . . .

75

5.1

Correlations Between Value Rankings and Policy Attitudes . . . . . . . .

92

5.2

Correlations Between Value Ratings and Policy Attitudes . . . . . . . . .

92

5.3

Full Ordered Logit Model, Y = Issue Scale Ratings Group . . . . . . . .

93

5.4

Full Ordered Logit Model, Y = Issue Scale Rankings Group . . . . . . .

94

5.5

Ordered Logit Model, Y = Issue Scale Full Comparisons . . . . . . . . .

95

5.6

Ordered Logit Model, Y = Issue Scale Party & Ideology Omitted . . . .

96

5.7

Predicted Probabilities from Full Model, Ratings Group . . . . . . . . . .

97

vii

5.8

Pred. Probs. with Ideol. & Party Omitted, Ratings Group . . . . . . . .

97

5.9

Pred. Probs. from Full Model, Rankings Group (Educ. = 1) . . . . . . .

98

5.10 Pred. Probs. from Full Model, Rankings Group (Educ. = 4) . . . . . . .

98

5.11 Pred. Probs. with Party and Ideol. Omitted, Rankings Group (Educ.= 1)

99

5.12 Pred. Probs. with Party and Ideol. Omitted, Rankings Group (Educ. = 4) 99
5.13 Correlations Between Value Ratings and Turnout & Vote Choice . . . . .

100

5.14 Correlations Between Value Rankings and Turnout & Vote Choice . . . .

100

5.15 Probit Selection Model of Vote Choice and Turnout . . . . . . . . . . . .

101

5.16 Logit Models of Turnout . . . . . . . . . . . . . . . . . . . . . . . . . . .

102

5.17 Logit Models of Vote Choice . . . . . . . . . . . . . . . . . . . . . . . . .

102

viii

List of Figures

3.1

Mean Ratings and Rankings . . . . . . . . . . . . . . . . . . . . . . . . .

44

3.2

Distributions of Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . .

45

3.3

Distributions of Rankings . . . . . . . . . . . . . . . . . . . . . . . . . .

46

4.1

Mean Rankings by Year . . . . . . . . . . . . . . . . . . . . . . . . . . .

70

ix

Chapter 1

Introduction

Political information is abundant, and making sense of it all is no easy task. This
information overload, along with the general lack of political knowledge of the American
public, makes it easy to think that behavior among members of the electorate is random
–that attitudes are generated with no rhyme or reason, and candidates’ sense of style is
as important as their stands on the issues. But, such is not the case. Peoples’ attitudes
are linked, belief systems are generally cohesive, and behavior is somewhat predictable.
In this dissertation, I attempt to answer why behavior is predictable by examining the
relationships between value structure and attitude formation, organization, and modification. I first ask if value systems can be a general framework by which people evaluate
objects in the world, and I test this possibility using both an experimental design and
an observational design. Second, I attempt to shed light on the degree to which value
systems influence attitudes, and I do so using a regression–based design.
In short, my answers to these questions are not black and white. First, regarding
whether or not value systems can structure attitudes, I find (1) that measurement matters
–value systems measured with rankings exhibit more stability in the face of systematic
changes in informational environment than do value systems measured with importance
ratings– and that (2) when value systems are unstable, it is not all values in the system
that fluctuate with respect to relative importance. Rather, it is a pair of values that seem
to “trade” places. Regarding the relationship between values, attitudes, and behavior, I
1

find that values have non-negligible impact on attitudes. The relationship between values
and behavior, however, is less clear cut and more complicated. On the whole, I conclude
that the question of values structuring attitudes is more complicated than a simple “yes”
or “no.” The analyses presented in this dissertation suggest that values may be one of
several devices that people use to organize attitudes, and, a story of mass public opinion
is incomplete without considering the impact of values on choice and action.

1.1

Past Scholarship on Attitude Structure
Most research on attitude structure has focused on ideology (e.g. Downs 1957). And,

while it has been shown that elites attitude structure conform to the left/right continuum
(Huitt 1954; Macrae 1952; Poole and Rosenthal 1991), the story at the mass public level
is much more complicated. Early public opinion research showed most people to exhibit
little constraint between attitudes –individual attitudes seemed to be only loosely related,
at best– and there seemed to be little use of a central organizational device (Campbell
et al. 1960; Converse 1964). More recent research on ideology suggests that the degree to
which belief systems conform to a recognizable left/right structure depends on education,
political sophistication, and political involvement (Jacoby 1988, 1991). Those that pay
less attention to politics might use ideological labels to classify attitudes (Conover and
Feldman 1981; Levitin and Miller 1979), but not as a general organizational device.
With ideology out of the questions as the universal device by which people organize
their beliefs, scholarly attention has turned toward values. According to psychological
theory, values are beliefs about desirable end-states that determine goals and decision
rules, which guide evaluations of behavior and events (Kohn 1977; Schwartz 1992). Psychological theory also states that values rarely act in isolation. For example, one can
never only take the value of equality into account. Rather, one must think of equality,
freedom, and security together, and choice is based on the relative ordering of relevant

2

values. In this way, values work in a “comparative and competitive” fashion (Rokeach
1973).
Political scientists have, for some time, incorporated values into various models of
political behavior (e.g. Feldman and Zaller 1992), information processing (e.g. Grant and
Rudolph 2003), and social group identity (e.g. Goren 2005). But, if values are a cause
of various forms of behavior –if they do create decision rules based on their relative
ordering– then these decision rules must remain the same from one situation to the next.
In other words, if value systems do provide frameworks by which people can evaluate
objects in the world, these frameworks must remain constant through various situational
contexts. The definitional assumption, “(value systems) transcend specific actions and
situations. . . ” (Schwartz 2001, 262), must hold true.

1.2

New Research
Research on the “transsituational” nature of values is inconclusive. On one hand,

scholars find consistency in value choices (e.g. Rokeach 1973). On the other, theories of
value pluralism (e.g. Tetlock, Peterson and Lerner 1996), as well as research on various
campaign effects (e.g. Grant and Rudolph 2003) and indifference (e.g. Maio et al. 1996)
find evidence suggesting value systems are not stable. In this dissertation, I take a closer
look at these contradictory results. I ask first whether or not different measurement
strategies affect conclusions on value system dynamics. I then ask whether value systems
change over time. Finally, I attempt to measure the degree to which value system ordering
affects various political attitudes. At the most general level, this dissertation is not about
whether value systems guide political behavior: it is simply about whether value systems
can guide behavior. Specific questions concern value system measurement, value system
stability, and the degree to which values affect attitudes, behavior, and choice.

3

1.3

Layout of Dissertation
The plan for the remaining chapters of the dissertation is as follows: Chapter two

reviews values research from its inception in psychology through the most recent experiments on value system dynamics. I explain why the central question addressed in this
dissertation and how this piece of research fits into the larger body of literature.
Chapter three focuses on value system dynamics and measurement. In it, I present a
model of value ratings and value rankings. I point out important differences in the models,
derive hypotheses, and test the hypotheses using a new experimental survey, fielded
in 2010, funded by Time-sharing Experiments in the Social Sciences (TESS)1 Results
suggest that value choices are susceptible to small changes in informational context, but
these effects are more apparent when relative importance is measured with importance
ratings. Rankings, though not completely stable, exhibit more stability.
Chapter four assesses the temporal stability of values at the aggregate level. Using a
matching design and data from 1994, 2003, and 2005, I find that large events (specifically,
9/11), affect value choices and value system ordering. I also find that this change is not
permanent –that it may only be due to changes in the degree to which each value is
“accessible” in individuals’ memory banks.
Chapter five examines the degree to which value system ordering affects individuals’
attitudes and behavior. I again use TESS data from 2010, and I find that the effect of
value system ordering on attitudes is significant. The effect of value system ordering on
behavior, however, is not clear. I also find that models of behavior that measure values
with rankings tend to fit data better than models in which values are measured with
ratings.
Chapter six serves as a conclusion and discussion of results. The discussion situates
1 Time-sharing Experiments for the Social Sciences, NSF Grant 0818839, Jeremy Freese

and Penny Visser, Principal Investigators.

4

where, in the body of literature, the findings presented in this paper fit. In addition
to discussing strengths and weaknesses of the presented analyses, I discuss avenues for
future research.

5

Chapter 2

Literature Review

A person forms and acts on a near infinite number of political attitudes throughout a
lifetime. While many of these attitudes are interesting in their own right, a great deal of
research across the social and behavioral sciences is directed at finding how these attitudes
are structured and organized. Given the complexity of the political world and the low
levels of political sophistication in the American public, one could believe that most of
these attitudes are unrelated and unpredictable. This, however, is not the case. Attitudes
do exhibit structure, and behavior is at least somewhat predictable; therefore, it must be
the case that attitudes are organized and recalled in a systematic manner. This chapter
lays out a comprehensive review of the literature on attitude structure. The review
focuses first on ideology and partisanship. Second, the review examines values research,
including definitions, measurement, and holes in the literature. Last, the review presents
the set of questions that are central to this dissertation.

2.1

Previous Research on Attitude Organization
Most previous research on attitude organization focused on ideology (as in Downs

1957) and the degree to which individuals’ attitudes conformed to the proposed left/right
structure. Early research out of the Michigan school painted a somewhat troubling picture of the American public suggesting not only that a small portion of the electorates’
attitudes conformed to a left/right structure (Campbell et al. 1960), but also that much
6

of the electorate seemed to generate opinions and attitudes at random, absent of any
overall structure (Converse 1964). Subsequent research on ideology in the mass public
suggests that people may not use the Downsian conception of ideology to organize attitudes; rather, they use the words “liberal” and “conservative” as tools to classify political
objects in the same manner they use party labels (Conover and Feldman 1981; Levitin
and Miller 1979). Research on the use of ideology in the mass public also suggests that
answers may not be as easy as “yes” and “no.” Rather, some people may use the left/right
continuum “better” than others, and some issues are more conformable than others. Jacoby (1988, 1991), for instance, argues that a portion of the public characterized by high
education and high political sophistication tend to apply the left/right continuum more
often, and more accurately, than others. Also, individuals more often apply the left/right
continuum to issues that are most easily defined in a liberal/conservative light (Jacoby
1990, 1994).
In short, research has shown the liberal/conservative continuum to structure attitudes
for a portion of the population, but its application is not universal. This so, there may
be a cognitive organizational tool that is more basic, and more universal, than ideology.
Research out of psychology suggests that values and value structures may be this tool.

2.2

Definitions
In defining values, and value systems, it is necessary to first clarify what values are

not. Concepts similar to, but distinct from values include attitudes, traits, norms, and
needs. Psychological theory defines attitudes as favorable or unfavorable evaluations of
an object. Values focus on ideals, not objects (Hitlin and Piliavin 2004). Relative to
attitudes, values are more central to the formation of “personhood” (e.g. personality
and individual characteristics) (Erickson 1995; Hitlin 2003), less directly implicated in
behavior (Schwartz 1996), and, theoretically, they are more stable throughout a lifetime

7

(Konty and Dunham 1997).
Traits are defined as fixed aspects of personality (Hitlin and Piliavin 2004). Traitbased behavior is often confused with value-based behavior, but a key difference between
them is that value-based behavior suggests one has cognitive control over one’s actions
(Roccas et al. 2002). For example, one may often act in an aggressive manner (a trait), but
one may not value “aggressiveness” very highly (a value) (Epstein 1989). The difference
that is most important to this project is that values, not traits, serve as standards by
which one can evaluate objects and behaviors.
Norms and values are both group-level phenomena that are based on shared agreement. Norms, however, vary by situation where values, in theory, are “transsituational.”
Norms tend to capture an “ought” sense while values capture a personal ideal (Hitlin and
Piliavin 2004). With respect to behavior, values and norms may be in conflict (Schwartz
and Bardi 2001).
Needs, like values, influence behavior, but they do so in different ways. Needs are
biologically based where values are social constructs, and often, it is argued that values
are socially acceptable ways of packaging various needs (Hitlin and Piliavin 2004). For
example, a biological need for sex can be repackaged as a need for love (Rokeach 1973).
A value may be an expression of a need, but it is not the need itself.
A value, according to Rokeach (1973), is an enduring belief that a specific mode of
conduct or end-state of existence is personally or socially preferable to an opposite mode
of conduct or end-state of existence. Also, according to Rokeach, it is rare that values
act in isolation from each other: rather, values work in a “comparative and competitive
fashion.” Thus, it is more useful to think in terms of value systems. A value system is
defined as an enduring organization of beliefs concerning preferable modes of conduct or
end-states of existence along a continuum of relative importance. For Schwartz (1992,
2001), values (1) are beliefs that (2) refer to desirable goals. They (3) transcend specific

8

situations, (4) are ordered by relative importance, and (5) serve as standards or criteria.
Finally, (6) it is the relative importance of the set of relevant values that guides action.
Psychological research often describes values as “backstops” for individual level belief
systems (Tetlock 2000). That is, an individual may prefer one candidate to another
because one candidate better embodies a value, or set of values. But, when asked to
justify the preference of one value over another, there is no response. Values, simply, are
the most fundamental building blocks of choice.

2.3

Major Players in Values Research
Modern work on human values began in psychology with the influential work of Milton

Rokeach (Feldman 2003; Jacoby 2006). The central questions of Rokeach’s research
revolved around the basic structure of value systems and how values related to various
forms of behavior. Rokeach argued that values do not exist in isolation –no single attitude
is a function of only one value (Rokeach 1973; Feldman 2003). Rather, values act in a
comparative and competitive fashion. An attitude is often a function of how values
intersect and how an individual prefers one value at the expense of another. Therefore,
attitudes are, in part, functions of the relative ordering of values in the value system.
If Rokeach’s first major contribution to values research concerns his conceptualization
of value systems, then his second major contribution is in terms of value measurement.
The Rokeach Value Survey consists of two separate 18-item scales (to differentiate between instrumental and terminal values)1 on which respondents were asked to arrange all
items with respect to personal importance, as guiding principles in life (Rokeach 1973,
27). By asking respondents to rank-order the values in each list, Rokeach effectively
avoided problems with “end-piling” (e.g. respondents rating all values highly resulting in
a measure that suffers from a lack of variance). At the same time, his scales are often
1 Rokeach (1973) describes instrumental values as those that relate to “modes of conduct”

while terminal values concern “end-states of existence” (Feldman 2003; Rokeach 1973).
9

argued to be too long and complicated to administer on a large-scale survey (Schwartz
2007). It has also been suggests that Rokeach’s lists are incomplete, and adding omissions
would significantly increase the number of items respondents would have to rank, thus
making the survey longer and more complex (Braithwaite and Law 1985; Feldman 2003).
Shalom Schwartz builds on Rokeach’s work, attempting to find a universal structure
of values that is both comprehensive and simple. Starting with the same assumption as
Rokeach, that values emerge from basic biological and social needs (Schwartz and Bilsky
1987), Schwartz argues that individuals and societies must be responsive to three sets of
basic needs: the needs of individuals as biological organisms, requisites of coordinated
social interaction, and the survival and welfare needs of groups (Schwartz 1992, 4). Using
cross-national data and a 54 item survey, Schwartz found that there are 10 “fundamental
value types.” Schwartz further theorized that values are organized along two dimensions with compatible values (e.g. security and power) occupying adjacent space and
opposing values (e.g. power and universalism) occupying space at opposite sides of the
two-dimensional plane. Using a multidimensional scaling routine, Schwartz finds his data
to conform to the expected pattern (Schwartz 1992, 1996, 2001, 2007).
Schwartz’s comprehensive theory of value structure, along with the simplicity of his
empirical findings, make his overall contribution to values research more theoretically
sound than those before him (Feldman 2003). This so, it has been noted that the individual values get lost within the “value types” (Davidov, Schmidt and Schwartz 2008;
Feldman 2003) and that the boundaries dividing values in Schwartz’s multidimensional
scaling analysis are somewhat arbitrary, based on a subjective interpretation of the multidimensional scaling output (Knoppen and Saris 2008).
Ronald Inglehart was among the first political scientists to incorporate values into political science research. Inglehart’s theory of socioeconomic conditions and value change
posits that social and economic conditions affect the relative orderings of individual’s

10

value systems. Specifically, Inglehart suggests that those in industrial societies heavily
emphasize materialist values (e.g. economic security) while those in post-industrial societies more heavily emphasize post-materialist values (e.g. self-expression) (Abramson
and Inglehart 1987, 1995; Inglehart 1971, 1977). Inglehart’s survey asked respondents
to rank-order groups of four items (two tap into materialist values and two tap into
post-materialist values). In short, Inglehart’s findings suggest younger generations rank
post-materialist values more highly than materialist values, suggesting that socioeconomic
conditions through which one is raised affects value structure (Abramson and Inglehart
1995).
While Inglehart’s analyses have taken a fair bit of criticism (see, for example, Braithwaite, Makkai and Pittelko 1996; Duch and Taylor 1993; Inglehart and Flanagan 1987),
his work remains influential in political science, and particularly in research on European
political behavior.

2.4

Values in Political Science
In political science, values have been most often used to explain various aspects of

political behavior including attitudes, decision making, and partisan and ideological orientation. Theoretical work on values in American politics and American political culture
started with McClosky and Zaller (1984), in which the authors examine opinions and
values at both the mass and elite levels. Their findings show strong support for both
democratic and capitalist values among the general public and elites, and the authors
also make the argument that many of the ideas underpinning the American welfare state
come out of the conflict between democracy and capitalism (in which two of the core
values are equality and freedom, respectively). The big picture argument, here, is that
the intersection of equality and liberty forms the crux of American political culture.
Later research in values and American politics narrowed the focus to values’ effect

11

on politically relevant decision making processes. Feldman (1988) finds that attitudes
regarding equality of opportunity and economic individualism tie tightly with ideology,
party identification, general attitudes on government (e.g. limited government and social
welfare spending), and presidential evaluations. Feldman and Zaller (1992) push a bit
farther, analyzing open-ended responses on support for social welfare policy. In short,
they find that a significant portion of the population use values central to political culture
when discussing important political issues. Further, people often recognize, and are
affected by value conflict when forming attitudes on controversial issues.2
Since Feldman and Zaller (1992), research has connected values to behavior and attitudes on specific issues, including abortion policy (Alvarez and Brehm 1995), campaign
finance (Grant and Rudolph 2003), and national security (Davis 2007), among others.
Research has also examined the conditions under which value structure is most likely
to affect behavior. Zaller (1991) suggests that values may be more like ideology than
previously thought. That is, ideology is only used as an organizational tool by a portion
of the public characterized by high sophistication and political activity (Converse 1964;
Lewis-Beck et al. 2008; Jacoby 1988). Values, according to Zaller, are applied only by a
relatively sophisticated portion of the public as well.
While a great deal of research focuses on the relationship between values and behavior,
no firm conclusions have been drawn. Research is unable to say, with a high degree of
2 Additional research on value conflict and political attitudes suggests that when people

realize the importance of two values, as they relate to a political issue, they may be
equivocal. Zaller (1992) argues that those with high levels of information, that understand opposing sides of an issue, are more likely than others to express ambivalence.
Alvarez and Brehm (1995) show that ambivalence plays an important role on equivocation on abortion policies, but not on racial policy or attitudes about the IRS (Alvarez
and Brehm 1997, 1998). Grant and Rudolph (2003) argues that the values of free speech
and equality affect attitudes on campaign finance, but this is often mitigated by group
attitudes. Jacoby (2006), on the other hand, argues that seemingly inconsistent attitudes are not usually a result of ambivalence. Rather, they are due to low political
sophistication.

12

confidence, that values are a widely used tool for forming, organizing, and modifying
political attitudes.

2.5

Assumptions, or Testable Hypotheses?
Values research has recently refocused its attention on the most basic, definitional at-

tributes of value systems in hope of determining whether or not value systems can serve
as a framework by which individuals evaluate objects in the world. Traditional psychological theory defines values as beliefs about to desirable goals that serve as standards for
evaluation, transcend specific situations, and are ordered by relative importance. Further,
traditional theory suggests that a set of values is activated when making a decision, and
it is the relative importance of the set of activated values that guides action (Schwartz
2001).
If values do provide people with a broad evaluative framework by which people determine what is good and bad in the world, then it must be the case that this framework
provides the same bases for evaluation from situation to situation. That is, value systems
cannot vary by context; they must actually transcend specific situations. And currently,
scholarship on value system dynamics is at a crossroads. Rokeach (1973), in his seminal
work, found that peoples’ value choices are largely consistent over time. Respondents’
rank-ordered value choices during the first wave of a survey correlated quite highly with
rank-ordered value choices during later waves (r ≈ .75). More recently, research shows
not only that people are consistent in their value choices, but also that a considerable
proportion of value systems, at the mass public level, exhibit a high degree of transitivity
(Jacoby 2006). Further, value choices are stable through varying informational contexts
(Jacoby 2008). On the other side of the debate, theories of value pluralism suggest that
changes in situational context affect values, as anchors of the decision-making process,
causing some values to increase in relative importance at the expense of others (Seligman

13

and Katz 1996; Tetlock 1986; Tetlock, Peterson and Lerner 1996). Similarly, research
on media effects (Ball-Rokeach, Rokeach and Grube 1984), issue framing (Grant and
Rudolph 2003; Nelson, Clawson and Oxley 1997), and campaign effects (McCann 1997)
suggest that changes in informational context change the frames of reference by which
individuals judge objects of evaluations.
Another problem for traditional values theory is indifference. Many individuals, when
asked about abstract, loaded terms like equality, freedom, or morality, are unable to adequately and consistently distinguish relative importance. And, if individuals are unable
to rank-order competing values with precision, then a stable relative ordering of a set of
values is impossible (Maio and Olson 1998). Related research suggests that people do
not often think through value choices. Instead, people make value choices quickly, then,
when asked to justify their choices, they make different value choices altogether (Bernard,
Maio and Olson 2003). If people are unable to meaningfully distinguish between values,
then relative ordering and value consistency is irrelevant to choice and behavior.
A final problem for traditional theory is the possibility that values may not, in fact,
lie at the base of belief systems. Goren (2005) finds that partisan identification is more
stable than value ratings. Pushing this result, using experimental data, Goren, Federico
and Kittilson (2009) finds that partisan cues actually drive value choices in partisans.
Again, if values are not at the base of belief systems, they cannot provide a general
framework for evaluation.
If traditional theory is correct –that value systems are invariant to situational contexts–
implications are straightforward: value systems can provide a general framework for evaluation. If, on the other hand, newer streams of research are correct, value systems do
not provide the same information from one situation to the next, and they cannot be a
universal tool used in attitude formation, organization, and modification.

14

2.6

A New Contribution
I ask two basic questions in this dissertation. First, are value systems stable across sit-

uational contexts? To answer, I use an experimental design and an observational design.
Second, is value system order an important determinant in various political attitudes and
behaviors? I use a regression-based design and new data to test this possibility.
Currently, values research finds itself at an impasse where traditional psychological
theory pushes against new empirical research. This dissertation, as a whole, speaks
directly to this debate. And, the results of this the following analyses could possibly help
values research beyond the impasse it now faces and clear a path for future research in
the area.

15

Chapter 3

Measurement and Value System Stability

If values are at the roots of attitudes and behaviors –if they are the backstops of
belief systems from which attitudes emerge– then they cannot change with the situational
context. Current research on value system dynamics is at a crossroads. On one hand,
traditional theory and research suggests people are consistent in their value choices. On
the other, new research suggests that people may have multiple value systems, that value
choices change based on the situation, and that peoples’ value choices often change after
introspection. This chapter asks why this rift between traditional theory and new research
exists, and it does so by looking at the effects of measurement on value system dynamics.

3.1

Research on Value System Dynamics
Some of the earliest research on value systems suggested people make consistent value

choices over time (Rokeach 1973). Newer research suggests not only that people make consistent value choices, but also that individuals’ value systems are often highly transitive
(Jacoby 2006). In contrast, a large body of research finds evidence suggesting otherwise. Theories of value pluralism suggest that people may have different value systems
for different situations (Seligman and Katz 1996; Tetlock, Peterson and Lerner 1996).
Also, research on media effects (Ball-Rokeach, Rokeach and Grube 1984), issue framing
(Grant and Rudolph 2003; Nelson, Clawson and Oxley 1997), campaign effects (McCann
1997), partisanship (Goren 2005; Goren, Federico and Kittilson 2009), and indifference
16

(Bernard, Maio and Olson 2003; Maio and Olson 1998) suggest that value choices are
susceptible to changes in context and informational environment.

3.2

Value System Measurement
The contradictory results can be traced back to decisions regarding value system

measurement. Studies that find evidence of value system stability analyze preferencebased, rank-ordered sets of values. Studies that find value systems to be dynamic, on
the other hand, often analyze individual values, in isolation, via importance ratings.
The difference between rank-ordered sets and importance ratings amounts to far more
than a difference in measurement, however. The different measurement strategies reflect
fundamental differences in the conceptualizations between individual values and value
systems.
The measurement of values has long been a point of debate. Rokeach’s measure,
the Rokeach Value Survey (RVS), asked respondents to rank-order two sets of 18 values,
terminal values first (which refer to desirable end-states of existence) and instrumental
values second (which refer to preferable modes of behavior), based on personal importance
(Rokeach 1968, 1973). The scales are meant to reflect the “comparative and competitive”
nature of value systems, but are often thought to be too long and too difficult to administer on large-scale surveys (Schwartz 2001, 2007). Surveys continue to use trimmed
versions of Rokeach’s ranking measures –including only those values that are immediately relevant to attitudes and behavior of interest (e.g. Abramson and Inglehart 1995;
Hofstede 2001; Kohn 1977)– but it has been suggested that these measures are flawed
because they force respondents to differentiate between potentially unimportant values
(Maio et al. 1996) and because they produce ipsative scales, which are problematic for
analysis (Alwin and Krosnick 1985; Davis and Davenport 1999).
Rating procedures, the main alternative to rankings, are also widely used in values

17

research (Schwartz 1992, 2007). Ratings alleviate concerns about ad hoc differentiations,
respondent fatigue, ipsative measures, and they are easy to administer during telephone
interviews. This so, they are unable to account for the “relative ordering” of value systems.
Empirical research on the performance of rating and ranking measures also fails to
reach a consensus. While methodological advances have made ipsative measures less of an
issue (Borg and Groenen 2005; Jackson and Alwin 1980), it has been suggested that rating
procedures produce more accurate value measurements for less sophisticated respondents
(Maio et al. 1996). On the other hand, research has shown results from rating measures
to be obscured by those who cannot differentiate between values (Krosnick and Alwin
1988) and that ranking measures possess less measurement error than other methods used
to collect data on value systems (Miethe 1985).
Each measurement strategy emphasizes certain aspect of values systems, and these
different points of emphasis come with different trade-offs. For example, the choice to use
rankings over ratings means one will be able to account for the relative ordering of each
respondent’s value system, but it also means that data collection must not be done via
telephone interview. On the other hand, the choice to use ratings makes data collection
easier, but at the expense of the “comparative and competitive” aspect of values and value
systems. The details of these trade-offs, with respect to value system dynamics, are not
entirely clear.

3.3

Models of Importance Ratings & Rankings, and Assumptions
In this section, I discuss how individuals answer survey questions that ask of the

importance of individual values in everyday life, how individuals construct make pairwise
choices between values, and how individuals create full rank-ordered value preferences
(i.e. value hierarchies). All of the models below are based on a basic random utility
model (Luce 1959; Luce and Suppes 1965) that has been used extensively in research on

18

behavior and choice (see, for example Chapman and Staelin 1982; Dow and Endersby
2004; Kumar and Kant 2007; Skrondal and Rabe-Hesketh 2003). Each model includes a
systematic component (e.g. true utility) and a random component (e.g. error). For all
models, I assume random components are distributed normal with mean zero.
3.3.1

A Model of Value Importance Ratings

The model proposed in this sections concerns the process by which an actor determines
the utilities of values and how the actor maps these utilities onto a set of survey responses.
The model takes into account both the systematic and random components in the value
ratings process. The systematic component concerns the importance that individual i
attaches to value a, denoted Iia . The random component is assumed to have a mean of
zero and a finite variance, denoted eia . The process by which one calculates the utility
of value a (denoted uia ), which is later mapped onto a survey instrument, is described
in equation 3.1. The process resembles that which is described in a traditional random
utility model (see, for example Chapman and Staelin 1982; Heiss 2002).

uia = Iia + eia

(3.1)

The model breaks Iia down into two components: (1) true importance, denoted tia and
contextual salience cia (e.g. the degree to which situational context brings a value to
“the top of one’s head”). Taking this into account, equation 3.1 becomes equation 3.2.

uia = tia + cia +eia
Iia

(3.2)

Determining the utility of each value is only half the process: the other half of the
process involves mapping these utilities onto survey ratings (R). And, this process involves determining values for the J − 1 cut points (τ ) that separate the J categories (e.g.
19

“very important,” “somewhat important,” “moderately important,”“somewhat unimportant,” “very unimportant”) and fitting the utilities into the appropriate survey category.
In short, if the respondent judges the value in question to have utility (uia ) greater than
the final “cut point” on the scale (e.g. the point that separates the “very important” and
“somewhat important” response categories), she will record her response as “very important.” If the value’s utility falls between the third and fourth cut points, the response
recorded will be “somewhat important,” and so on until the value’s utility does not reach
the first cut point, in which case the response will be recorded as “very unimportant.”
This process is described below in equation set 3.3 (where J = 5 and “very important”
values are rated 5, and “very unimportant” values are rated 1).

R=5

if

uia > τ4 + ε4

R=4

if

τ4 + ε4 ≥ uia > τ3 + ε3

(3.3)

.
.
.
R=1

if

τ1 + ε1 ≥ uia .

The ε terms represent the error an actor applies while determining the appropriate values
of the cut points, and this error affects substantive conclusions when the survey instrument is too fine or too coarse (e.g. values may appear more stable than they actually are
if the survey instrument is too coarse, or they may appear less stable than they actually
are if the survey instrument is too fine).
A change in informational context can only change one systematic element of the
model: contextual salience, cia . But, with attention to value system stability and the
nature of the survey instrument (specifically, the cut points, τj + εj , j = 1, 2, . . . , J − 1),
an inappropriate number of cut points may cause a discrepancy between observed value
system stability and actual value system stability (the difference between the manifest

20

utility of value a before and after the change in informational context, and how the values
are mapped onto the survey instrument).
A Hypothetical Example: Importance Ratings
Let ui range between 0 and 1 (0 ≤ ui ≤ 1), let uia represent the manifest utility
individual i attaches to value a before a change in the informational environment, and
let uia represent the manifest utility individual i attaches to value a after a change in
the information environment. Recall that the components that factor into the manifest
utility are tia (true importance), cia (contextual salience), and eia random error.
Allow tia to vary across individuals, but remain constant within individuals (e.g. so
true importance does not vary within individuals by context). Let contextual salience
equal cia before a change in informational environment, and cia after a change in informational environment. And let eia be distributed normal with mean zero and variance
σ2.
Turning attention away from the psychological process by which individuals determine
the utility they attach to values toward how the manifest utility maps onto the survey
instrument, let τJ−1 separate the top two categories of the survey instrument, let τ1
separate the bottom two categories of the survey instrument, and let τJ−2 through τ2
separate the remaining categories of the survey instrument that are toward the center of
the scale. And, let ε be distributed gamma with scale θ and shape k (where both θ and
k are greater than zero).
Consider a hypothetical respondent r and his evaluation of value a. Individual r’s true
importance of value a is 0.75, the contextual salience of value a is 0, and random error is
0, so ura = tra + cra + era = 0.75 + 0 + 0 = 0.75. Now, consider how respondent r maps
ura onto a survey instrument with the appropriate number of categories (for example,

21

J = 5):

R=5

if

ura > 0.8 + ε4

R=4

if

0.8 + ε4 ≥ ura > 0.6 + ε3

(3.4)

.
.
.
R=1

if

0.2 + ε1 ≥ ura .

The manifest utility of value a (ura = 0.75) maps onto the survey instrument at R = 4.
Now, consider respondent r’s evaluation of value a after a change in informational
environment. Individual r’s true importance of value a is still 0.75, but the contextual
salience (cra ) is increased to 0.1 (the error remains 0). Now, ura = 0.75 + 0.1 + 0 = 0.85.
And, using equation set 3.4, ura maps onto the survey instrument at R = 5.
Now consider how ura and ura would map onto survey instruments that were too
fine or too coarse. If, for example, the survey instrument had ten categories (J = 10),
then the way in which the respondent maps ura and ura may be different:

R = 10

if

ura > 0.9 + ε9

R=9

if

0.9 + ε9 ≥ ura > 0.8 + ε8

R=8

if

0.8 + ε8 ≥ ura > 0.7 + ε7

(3.5)

.
.
.
R=1

if

0.1 + ε1 ≥ ura .

Here, ura maps onto the survey instrument at R = 8 and ura maps onto the instrument
at R = 9.
If, on the other hand, the survey instrument is too coarse (e.g. if J = 3), ura and

22

ura would map onto the survey instrument differently.

R=3

if

ura > 0.67 + ε2

R=2

if

0.67 + ε2 ≥ ura > 0.33 + ε1

R=1

if

0.33 + ε1 ≥ ura .

(3.6)

Here, both ura and ura would map onto the survey instrument at R = 3, and the effect
of the change in informational environment would go undetected.
Finally, it is necessary to consider the effects of ε1 , . . . , εJ−1 on how ura and ura
map onto various survey instruments. If the error in determining where each of the J − 1
τ ’s falls on the underlying utility scale is constant (e.g. if the error is about the same
for each τ in a survey instrument with three response categories as it is in an instrument
with ten response categories), and if the underlying scale is the same length (e.g. if, for
example, u ranges from 0 to 1 regardless of the survey instrument), then survey induced
error could affect observed stability. If response categories overlap (e.g. if a respondent
cannot meaningfully distinguish between R = 7 and R = 8), and the manifest utility
of value a lies in that range, then it may be the case that the respondent chooses, at
random, between the seventh and eighth response categories (see equation set 3.7 for an
example). So, any “movement” between these categories may not be the result of the

23

difference between cra and cra , rather, it may be due to measurement error.

R = 10

if

ura > 0.9

(3.7)

.
.
.
R=8

if

0.78 > ura ≥ 0.68∗

R=7

if

0.71 > ura ≥ 0.60∗

.
.
.
R=1

3.3.2

if

0.1 ≥ ura .

A Model of Ranked Value Preferences

This model considers a rank-ordered set of three values that are weakly ordered (e.g.
ties are allowed) (R = [a

b

c], for example), changes in context (captured again with

c, as above), and time, denoted t. Again, the utility of each value in the set is determined
by the processes in equations 3.1 and 3.2. Changes in context, though they do not affect
v, may affect c, which plays an important role in value system ordering and (in)stability.1
Rank-ordered value preferences can be thought of as a series of pairwise choices between all pairs of individual values in the set of values under consideration. The process
by which an individual chooses one value over another depends on the perceived utility
of each value in consideration. Using values a and b for illustrative purposes, this process
is shown in equation set 3.8 (Jacoby 2008).

Choice(a)

if

Iia + eia > Iib + eib

(3.8)

and
Choice(b)

if

Iia + eia < Iib + eib .

1 This model is based that presented in Falmagne, Regenwetter and Grofman (1997).

24

If the the errors are small and randomly distributed with mean zero, and if the true
utility of each value is relatively large, and if each I term is different from the next (for
example, if Iia − Iib = 0), then errors will cancel out over successive iterations and the
I terms will drive choice.
3.3.3

Differences Between the Models

There are three main differences between the models of ratings and rankings. First,
in the model of importance ratings (equation 3.3), the quantity uia − uib is meaningful
both in terms of sign and magnitude. In the model of rankings (equation 3.8), however,
the quantity uia − uib only matters in terms of sign. Further, the quantity uia − uia
t
t−1
only matters relative to other values in the set. In short, it suffices to say that while
rankings provide less information, the extra information that is picked up in importance
ratings is possibly due to random fluctuations in .
Second, there are two steps to the ratings model and one step in the rankings model.
The additional process of determining values for τ1 , τ2 , . . . , τJ−1 adds another source of
error to the process. The rankings process only involves one source of error, and to the
extent that actors are able to clearly distinguish between values, this total error of the
rankings process is less than the total error in the ratings process.
Finally, the values of uia , uib , and uic in the model of ratings can move freely of each
other. That is, the value of uia , for example, does not depend on uib and all. The rank
of each value in the ordered set depends on uia , uib , and uic simultaneously. That is,
if Rt−1 = [a

b

c], and the actor receives a piece of information that adds utility

to value b, it could be the case that Rt = [b

a

c]. The order of the three elements

depends on the utility that is attached to each. One element of the set cannot move
without another. In this way, rankings capture the comparative and competitive element
of value systems that Rokeach, Schwartz, and others suggested was an integral part (see

25

Rokeach 1973; Schwartz 1996, for example).

3.4

Hypotheses
There are several hypotheses that can be drawn out of these models. First, to establish

whether the actual utility of each value or the associated error is driving value choices, I
will measure the degree to which value systems are transitive. Drawing on earlier research
(e.g. Rokeach 1973; Jacoby 2006), I expect to find that value choices are largely transitive
and that the I terms –not the e terms– are driving value choices. This can be shown by
rearranging the terms in equation 3.9 (as shown in Jacoby 2008, 4).

Choice(a)

if

via − vib > eib − eia

(3.9)

and
Choice(b)

if

via − vib < eib − eia .

If the I terms are well-established and the actor can easily distinguish between values,
then the I terms should drive value choices, and the actor should make the same value
choice over several iterations, thus indicating transitivity. If, on the other hand, the I
terms are not well-established or they are not easily distinguishable, then the random
fluctuations in perceived utilities (e.g. the e terms) should drive the process. And, if
the e terms are distributed normal with mean zero, they should cause inconsistent value
choices across iterations, thus indicating intransitivity.
Second, if value choices are transitive –if the I terms are well-established and easily
distinguishable– small changes in context should not affect value choices of those in the
ranking group. The small piece of information should not affect the utility of the primed
value such that Rt is different from Rt−1 .
Third, even if value choices are transitive, small changes in context may alter value

26

ratings, and this affect may be greater than that observed in value rankings. A small
change in context may act as a small shock to the actor’s value system, and it should not
affect the true importance of the relevant value so much so that it causes the rating of the
value to increase (but, it may affect contextual salience, which could affect the dependent
variable). This so, importance ratings are subject to measurement error. Rearranging
terms in equation set 3.3,

R = 1 if Iia − τ1

>

1 − eia

R = 2 if τ1 − Iia

≥

eia − 1

(3.10)

and
Iia − τ2
.
.
.

>

R = J if τj−1 − Iia

≥

2 − eia

eia − J−1

one can see that there are several additional sources of error that can effect the importance
ratings. Among the problems with importance ratings is that an actor does not have a
well-established values for the cutoff points (τ terms). Thus, the

terms are likely to be

rather large, and the actor is likely to perceive the utility for the value under consideration
as near at least two τ ’s, so the associated ’s could play a large role in how uia , for
example, maps onto the survey instrument. In addition, since importance ratings are a
“one shot” procedure, the errors cannot cancel out over iterations. So, even if an actor
has a clear Iia , the actor may not be able to clearly distinguish between τ1 , τ2 , . . . , τJ−1 ,
thus exposing the final outcome to random fluctuations in the various

terms. In sum,

even if the first and second hypotheses are true, there could be significant priming effects
in importance ratings. And, if this is the case, it is fair to conclude that measurement
choice does affect observed value system stability.

27

3.5

Research Design & Data
I test the above hypotheses with a unique new survey experiment that focuses on

the stability of rank-ordered value preferences versus importance ratings through systematically varied informational environments.2 The experiment takes into account five
values (freedom, equality, economic security, law& order, and moral traditionalism), all
important to American political culture. Freedom and equality are central components
of classical liberalism and integral to American political culture (Devine 1972; McClosky
and Zaller 1984), economic security and law and order address challenges to American
life and have been recognized as key elements in industrial and post-industrial societies,
(Davis 2007; Hochschild 1995; Inglehart and Abramson 1994; Rossiter 1962), and most
of the issues associated with the “culture war,” such as family structure, gay rights, and
abortion, center on moral traditionalism (Brewer 2003; Hunter 1991; Schwartz 2005).
Respondents in the experiment were randomly assigned to one of two measurement
groups: the “rating” group or the “ranking” group. Respondents in each group were then
randomly assigned to one of three subgroups: the control group, the “equality” group,
or the “economic security” group. Respondents in the equality subgroups and economic
security subgroups were exposed to a short prime designed to bring the respective value
to “the top of their heads.” The wording for the equality prime is as follows: “Some people
believe that we need to do a lot more to make sure everyone has an equal opportunity
to get ahead in life.” The wording for the economic security prime is as follows: “Some
people believe we should do as much as we can to provide economic security to create jobs,
improve incomes, and decrease home foreclosures.” Upon reading the prime, respondents
in these subgroups were asked about the degree to which recent political campaigns
emphasized the primed value. After reading the prime, respondents in treatment groups
2 Data collected by Time-sharing Experiments for the Social Sciences, NSF Grant

0818839, Jeremy Freese and Penny Visser, Principal Investigators.

28

are assigned a distraction task, to indicate their interest in politics using a 4-point scale.
After completing the distraction task, respondents in the rating group were asked
to rate the importance of each of the five values on a 1 to 5 scale. Value choices for
those in the ranking group were elicited using the method of triads. Over a series of ten
questions, respondents are shown all possible combinations of three values (e.g. a triad),
and asked to choose the most and least important value from each set of three. In doing
so, respondents are actually making three separate value choices in each triad (a vs. b, a
vs. c, and b vs. c). And, over all of the triads, respondents choose between the same pair
of values three times, which acts as a check against measurement error. So, rather than
the process in equation 3.8 determining value choice, a more complex process, shown in
equation set 3.11 drives value choice over m iterations (also as in Jacoby 2008, 3):
m

m
Choice(a)

if

eb

ea > mub +

mua +
m=1

m=1

m

(3.11)

m

and
Choice(b)

if

mua +

ea < mub +
m=1

eb .
m=1

Here, if the e’s are distributed normal with mean zero, they should cancel out over iterations and the I’s drive value choice, and the effects of measurement error are minimized.
The total sample size of the survey is 1250, and each subgroup has over 200 respondents, which provides sufficient statistical power to compare measurement strategies with
respect to value system stability.

3.6

Methods
I test for transitivity in value choices using the same basic methods employed by

Jacoby (2006, 2008). I define a respondent’s pairwise preferences regarding two values to
be clear if either value a succeeds value b in the preference ordering, or if value b succeeds

29

value a. Further, I define a respondent’s preferences regarding a set of three values to be
transitive if, for example, value a succeeds value b, value b succeeds value c, and value a
succeeds value c (e.g. a

b, b

c, and a

c). In other words, a respondent’s preferences

regarding a given triad are coded as transitive if there is a clear rank-ordering. I then
examine transitivity using descriptive statistics.
I analyze the data from the survey experiment in several different ways. The quantities
of interest, the priming effects, can be explained generally, as in Holland (1986), as

T = E(Yt ) − E(Yc )

(3.12)

where E(Yt ) is the expected value of the dependent variable in the treatment group,
E(Yc ) is the expected value of the dependent variable in the control group, and T is the
treatment effect (e.g. the priming effect). Specifically to this experiment, I examine four
priming effects (for the economic security prime in both measurement groups, and for
the equality prime in both measurement groups). I test whether the priming effects are
statistically different from zero using simple two-sample difference-in-means (e.g. twosample t-tests).
I also estimate the priming effects using two types of logit models and OLS regressions.
To estimate the priming effects in the rating measurement group, I use an ordered logit
model and an OLS regression. The importance ratings, I argue, are the manifestation of
the true importance of the value, y ∗ , mapped onto an ordinal scale through J − 1 cut
points (J − 1 because there are J possible responses) (as in equation 3.10 above). The
model relating various predictor variables (including the prime), x, to y ∗ , is
Yi∗ = α + x β + i ,

(3.13)

and, as the J − 1 cut points partition y ∗ , and as it is assumed that i is distributed
30

logistic, then the probability that Y ≤ yj |x (e.g. that Y falls in a category equal to or
below yj ) is
exp(τj − x β)
P r(Y ≤ yj |x) =
.
1 + exp(τj − xβ)

(3.14)

The parameters in equation 3.14 can be fit using a linear model for the log-odds ratio
using the logit link, as in equation 3.15 (Jones and Westerland 2006)

log

P r(Y ≤ yj |x)
= τj − x β, j = 1, 2, . . . , j − 1 .
P r(Y > yj |x)

(3.15)

And, one can interpret the β’s in terms of how Y “moves” as elements in x change (Jones
and Westerland 2006, 4). Another feature of the model concerns the estimates of the
J − 1 cut points. Knowing the placements of the cut points, the confidence intervals,
and the covariance between them, it is possible to determine whether all J categories are
meaningful, or whether categories overlap.
To estimate the ranking measurement group, I use a rank-ordered logit model (also
called an exploded logit model, as in Chapman and Staelin 1982) and OLS regressions.
The rank-ordered logit model, based on the conditional logit model allows one to estimate and test for differences among respondents’ preferences for the ranked items and to
test differences in preferences across subpopulations (Allison and Christakis 1994). In a
constant utility model, the probability a respondent ranks a set of items (e.g. a, b, c, . . .)
is
P r(a, b, c . . .) = P r(a|C) × P r(b|C − {a}) × P r(c|C − {a, b}) . . .

(3.16)

where P r(a|C) is the probability a respondent chooses option a from the set C =
{a, b, c, . . .}, P r(b|C − {a}) is the probability the respondent then chooses option b from
the set C, with a removed, and so on (Chapman and Staelin 1982). In the stochastic
utility model, where there is a random error component that affects the utility for a given
item in the set (as in equation 3.1 above), the probability that a respondent ranks a set
31

of items in a particular fashion is given by
Ji
P r(Ui1 ≥ Ui2 . . . , ≥ UiJ ) =

j ∗ =1

P r(Uij ∗ ≥ UiJ ,

for j = j ∗ , . . . , Ji ) ,

(3.17)

where the left side of the equation is the given rank-order (e.g. alternative 1 is preferred
to alternative 2, alternative 2 is preferred to alternative 3, all the way to alternative
J − 1 being preferred to alternative J) , and the right side defines the various events (e.g.
(Ui1 ≥ Uij : j = 1, 2, . . . , Ji ), (Ui2 ≥ Uij : j = 2, 3, . . . , Ji ), . . . , (Uij−1 ≥ UiJ )) to be
statistically independent (Chapman and Staelin 1982).
Just as the ordered logit model is similar to successive applications of the traditional
logit model, the rank-ordered logit model is similar to successive applications of the
conditional logit model (Glasgow 1997). The parameter estimates can be interpreted as
how changes in x affect the probability a respondent chooses, for example, item a over
item b (e.g. the “baseline” item). The model also allows for comparisons between groups
in the sample.
I use OLS regression to model both individual value rankings and ratings for one
main reason. While the dependent variables in this chapter are perhaps most effectively
modeled with maximum likelihood estimation techniques, the results of these various
models do not lend easily to comparison. Though OLS models of these data may produce
out-of-bounds predictions, and though OLS may miss some intricacies that non-linear
models can account for, the results from the OLS regressions are comparable. OLS
estimates are perhaps the best way to quickly and easily compare estimates of priming
effects across experimental measurement groups.

32

3.7

Results
Figure 3.1 shows the mean responses, with standard deviations, for both measurement

groups on all five values. The most obvious difference in responses across measurement
groups is the lack of differentiation among ratings. The between group variance of value
rankings is much greater, suggesting that the comparative and competitive nature of
values is brought forward by this particular question format.
Figure 3.1 also gives a good look into the aggregate level value preferences of the
American public. Freedom is, by far, the most preferred value of the five, followed
by law & order, then economic security, with equality and moral traditionalism at the
bottom of the preference hierarchy. These results are similar to prior studies on value
rankings (e.g. Jacoby 2006, 2008) in that freedom/liberty and economic security are
valued quite highly among the public, and equality is toward the “less important” end of
the spectrum. The major change across datasets concerns law & order/social order and
moral traditionalism/morality. These differences will be more thoroughly examined in
the next chapter.
The distributions of respondents’ value ratings are shown in figure 3.2. As expected,
the graphs show a high degree of “end-piling” (e.g. rating all values very highly). The
modal response for all five values is “very important,” and 30.5 percent of all those in the
rating group (191 of 626) rated all five values that category (e.g. the highest possible
category). Literature suggesting that value ratings suffer from a lack of meaningful variance (e.g. Krosnick and Alwin 1988) seems be supported by these data. Figure 3.3 shows
the distributions of respondents’ value rankings.3 Here, the distributions of responses
are spread more evenly across all the possible values. Like figure 3.1, figure 3.3 suggests
that forcing respondents to distinguish and choose between values creates a measure with
3 Note that value rankings are coded such that 5 indicates the most important value, 4

the second most important value, and so on. A value coded 1 indicates it is the least
important value in an individual’s hierarchy.
33

more variance to analyze. Whether or not these distinctions are meaningful, however,
cannot be determined by figure 3.3. To determine this, it is necessary to examine the
degree of transitivity in individual level value choices.
3.7.1

Transitivity in Value Rankings

A respondent shows preference for one value over another if she chooses one value over
another at least twice in three iterations. A respondent’s choice between two values is
consistent if the she chooses one value over the other in all three iterations. A respondent
does not show preference if she chooses, for example, value a over value b in the first
iteration, then value b over value a in the second iteration, and does not make a choice
in the third iteration. The key, however, is not whether choices are clear or consistent.
Rather, the key is the degree to which choices are transitive. That is, if value a succeeds
value b, and value b succeeds value c, value a must then succeed value c for preference
among these three values to be transitive. If preferences are largely transitive, utility, not
error, drives choice. If, on the other hand, preferences are not transitive, it is possible
to conclude that error drives choice. Table 3.1 shows the percentage of respondents who
exhibited transitivity and consistency in each of the ten pairwise value choices.
With over 90 percent of respondents exhibiting clear preferences of each of the ten
pairs of values, and with over 60 percent making consistent choices on each of the ten
pairs of values, it seems safe to conclude that much of the public has clear preferences
with respect to values. Stronger evidence, however, lies in examining the degree to which
these preferences are transitive. I define a set of preferences to be transitive if, given a
set of three values, a

b, b

c, and a

c. That is, a set of three value preferences

is transitive if there is a clear first, second, and third ranked value. Table 3.2 shows the
percentage of the sample that exhibited transitive preferences on each of the ten triads.
The results in table 3.2 paint a clear picture of of the clarity in respondents’ value

34

choices, and results support the first hypothesis stated above. With near 90 percent of
respondents expressing transitive value preferences on each of the ten triads, it does not
appear that value choices are at all random. In fact, it is evident that respondents have
a picture of each value’s relative importance. While it may be the case that there is some
measurement error in each survey question used here (judging by the rates of consistency
on pairs of values as shown in table 3.1), it does seem to be the case that this error cancels
out over several iterations enabling respondents to make clear, consistent, transitive value
choices.
3.7.2

Differences Between Experimental Groups

To reiterate, there are two main manipulations in the experiment: measurement and
changes in informational context (e.g. primes). And, there are two central questions.
First, does a small change in context affect value system ordering? And second, does
measurement affect observed value system stability? Respondents were first assigned to
one of two measurement groups, then to an experimental subgroup (i.e. control, equality
prime, economic security prime).
Tables 3.3 and 3.4 show the differences in the mean ratings/rankings of the primed
values between the treatment and control groups. The differences between the treatment
and control groups on the mean ratings and rankings for equality are displayed in table
3.3. First, and somewhat surprisingly, the treatment group rated equality lower (that is,
less important) than the control group. In the rating group, this is likely because many
of the respondents’ “default” rating for all values is “very important” (e.g. the highest
rating), and with no room to move up, the added token of information caused respondents
to rate the value lower than they normally would. Second, as expected, the difference in
means for the rating group is statistically significant while the difference in means for the
ranking group is not.

35

Table 3.4 shows the difference in means between the different groups’ ratings and
rankings of economic security. The results are similar to those shown in table 3.3. The
first main difference is in the direction of the priming effect for economic security in
the ranking group. In all other cases, the primed group rated/ranking the particular
value lower (e.g. less importantly) than the control group. Here, however, the primed
group ranks the value somewhat higher (though not significantly higher) than the control
group. The second main difference, of course, is the lack of statistical significance of these
priming effects.
At first glance, these results support the second and third hypotheses. Though the
absolute values of the priming affects are similar across measurement groups, only that
in the ratings group is statistically significant. According to these results, it does appear
to be the case that measurement choice affects observed value system stability.
3.7.3

Results of Ordered and Rank-Ordered Logit Models

The results of the models of respondents’ ratings of equality and economic security
are reported in table 3.5. The dependent variables are coded on a scale from 1 to 5
where 1 indicates the given value is “very unimportant” and 5 indicates the given value is
“very important” to the respondents’ everyday lives. The explanatory variables of interest
are the primes. For the model of the importance rating for equality, respondents that
received the relevant prime are coded 1, and all others are coded 0. For the model of the
importance rating for economic security, respondents that received the relevant prime are
coded 1, and all other are coded 0. Other independent variables in the models include
age (coded such that 1 indicates the respondent is between 18 and 24, 2 indicates the
respondent is between 25 and 34, and 7 indicates that the respondent is over 75 years
of age), gender (coded such that 0 indicates male and 1 indicates female), and party
identification (coded such that 1 indicates the respondent is a strong Republican and

36

7 indicates the respondent is a strong Democrat).4 Also in the model are a dummy
indicating whether a person is a racial minority, education (coded in four categories such
that 1 indicates the the lowest amount of formal education and 4 indicates the most),
income (coded such that 1 indicates the lowest income and 19 indicates the most), and
ideology (coded on a scale where 1 indicates the respondent is extremely liberal and 7
indicates the respondent is extremely conservative).
Looking first at the coefficients on the control variables, age, female, and party identification have significant positive impacts on the ratings of both values. The models suggest
that, on average, older respondents, female respondents, and Democrats rate both equality and economic security higher than others. The effect of the prime on equality rating
is significant and negative, suggesting that respondents in this treatment group rated
equality, on average, lower than others. The effect of the economic security prime is not
statistically different from zero. Each threshold (τ ) is statistically distinguishable from
all adjacent thresholds, but it is the case that thresholds separating categories denoting
lower importance are closer together than those separating categories denoting higher
importance. Finally, the parallel regression assumption holds in the model of ratings on
equality. The parallel regression assumption does not hold (χ2 = 21.10 on 12 degrees
of freedom, p = .05), but results from a generalized ordered logit show that the priming
effect never reaches statistical significance across any of the four binary logit models.
While some explanatory variables’ effects on value ratings do vary over the scale of the
dependent variable, the effects of the prime do not. Across all values of the dependent
variable, the impact of the prime is near zero.5
4 Party is included in the model to control for a strong correlation with value ratings.

It is not clear, however, if party ID is a causes or is an effect of value preferences
(see Goren 2005; Goren, Federico and Kittilson 2009). Because party ID is almost
completely orthogonal to the respondents’ assignments to experimental groups, the inclusion/exclusion of party does not effect the estimated priming effect.
5 The parallel regression assumption was checked with the Brant test.

37

Table 3.6 shows the changes in the predicted probabilities for ratings of equality
between the control and treatment groups. The most striking differences between experimental groups concerns the probability that a respondent deems equality “very important.” A respondent in the control group is about 11 percent more likely to say equality
is “very important.” This so, a respondent in the treatment group is about 6 percent
more likely to call equality “somewhat important” and about 4 percent more likely to
call equality “moderately important.” These results suggest not that the small change in
informational context caused the average respondent to consider equality far lower than
normal. Rather, it appears that the prime caused respondents to drop equality down
from “very important” to “somewhat important,” and, in some cases, to “moderately important.” In any case, it does appear that the equality treatment caused an aggregate
change in respondents’ ratings of equality that is separate from age, gender, party, or any
other possible confounding factor.
Turning to the estimation of the priming effects for the ranking group, the results of
the OLS regression models appear in table 3.7. Looking first at the effects of the control
variables on the ranking of equality, both age and party identification have significant
effects. On average, older respondents rank equality lower in their value hierarchy than
younger respondents. And, with respect to party, Democrats tend to rank equality higher
than independents and Republicans. Ideology also has a significant negative effect, indicating that liberals tend to rank equality higher than conservatives. The effect on the
main explanatory variables of interest, the prime, is statistically significant, which indicates that those exposed to the equality prime tend to rank equality lower (about 25
percent of a rank lower) than those not exposed to the prime.
Examining the model of respondents’ ranking of economic security, the only statistically significant effect among control variables comes on party identification and education. Here, Democrats, on average, tend to rank economic security higher in their value

38

hierarchies than independents and Republicans. And, more educated respondents tend
to rank economic security lower than less educated respondents. The effect of the main
explanatory variables of interest, the prime, are not statistically distinguishable from
zero.
Also of note in these results is model fit. The model of equality ranking fits the data
much better (R2 = 0.21) than the model of economic security rank (R2 = 0.06). The R2
statistics on the models of value rating are both very low (0.05 and 0.04, respectively).
The additional variance in the rankings measures, at least for the value of equality, is
captured by the independent variables in the model. And, perhaps the model of equality
rank fits the data better than the model of economic security rank because equality is,
in some sense, a “tipping point” for the major political parties in the United States and
identification on the left/right continuum (see, for example Feldman 1988).
The results of the rank-ordered logit models (shown in tables 3.8 and 3.9) give a look
“inside” respondents’ rank-ordered preferences. Focusing first on table 3.8, where equality
is the baseline value, and turning attention to the effects of the equality prime, it appears
that only two values (in addition to equality) are affected by the prime. The prime does
not affect the positioning of freedom and law & order, on average (the top two values in
the aggregate). The prime does, however, affect the positioning of economic security and
moral traditionalism. In each case, when exposed to the prime, respondents were more
likely to choose economic security and moral traditionalism over equality.
Turning attention to table 3.9 and the effects of the economic security prime, it appears
that the treatment affects only the probability that one chooses freedom over economic
security (the other coefficients on economic security in the table border traditional levels
of significance, but do not pass them). Interestingly, this prime positively affects economic
security: people in this treatment group tended to rank economic security more highly
than others, especially with respect to freedom.

39

3.8

Conclusion & Discussion
Examining the distributions of the value ratings and rankings, examining the dif-

ferences between the experimental groups, and examining the differences between the
ordered logit models and the rank-ordered logit models, side-by-side, reveals important
information behind the differences between value ratings and value rankings. The first
main question in this chapter concerns the distributions of value ratings versus value
rankings. And, just as previous research demonstrated, the amount of end-piling in the
ratings data is much greater than in the rankings data (Krosnick and Alwin 1988). Because the ranking format forces people to differentiate between values, the picture of
value systems as painted by the rankings data is much clearer than that produced by the
ratings data.
The second main question in this chapter concerns the degree to which respondents’
value choices, as made across the ten triads, are transitive. Transitive preferences indicate
that respondents make meaningful value choices, rather than choose at random. Further,
transitive preferences indicate that value systems (e.g. not just value choices) are meaningful psychological structures that can be seen in empirical data. As previous research
demonstrated (e.g. Jacoby 2006), respondents’ value choices are highly transitive, and it
appears that most respondents have very clear, meaningful value systems.
Establishing transitivity plays a large role in determining whether or not value systems
exist: it does not help to determine whether or not they are stable. The first tests
of stability in this chapter have to do with the experimental primes, and determining
whether the experimental groups rated/ranked equality and economic security differently
than the control groups. The results show, with relative clarity, that the equality prime in
the rating group caused a significant number of respondents to rethink they importance
they attach to equality in their everyday life. The economic security prime may have had
less of an effect because people have a clearer view of the value since the financial crisis
40

and economic downturn began to appear in the news consistently. The priming effect
for equality did not carry across measurement groups, suggesting that when a particular
value is anchored in a value system by other values (e.g. not held in place by categories
of a survey question as is the case in the rating group) response is more stable.
The results of the second tests of stability (e.g. the ordered logit models, the rankordered logit model, and the OLS models) tell a slightly different story. Most notably,
the OLS models of equality rating/ranking suggest equality may, in fact, be an unstable
value. In being prompted to think about the importance of equality in everyday life,
people attribute less importance to it. Also important is that these results hold across
measurement groups. All this suggests that changes in accessibility do affect the relative
importance people attach to equality when compared to other values.
Economic security seems to be a more stable value, relative to equality. The effects
of the economic security prime are, for the most part, statistically indistinguishable from
zero. In short, it appears that the small change in context does not affect accessibility to
the extent that it affects the relative ordering of individuals’ value systems.
The look inside respondents’ rank-ordered preferences, provided by the rank-ordered
logit model, offers valuable information concerning the trade-offs respondents make when
prompted to think about a particular value (e.g. the comparative and competitive aspect
of values and value systems). It is interesting to note that the equality prime only affected
the positioning of equality relative to economic security and moral traditionalism. Not all
values in the set are affected: only those that occupy the lower positions in the aggregate
value hierarchy. The ordered-logit model tells a different story about economic security.
Economic security, on average, goes unaffected by the prime (as shown by the OLS model
in table 3.7. Where there is a significant effect, however, is in the trade-off with freedom:
the other three values in the set are unaffected. These results, taken together, suggest
first and foremost that changes in accessibility affect value choices. Second, the results

41

are not quite as “black and white” as value system ordering is or is not stable. Rather,
it seems as though certain values exhibit more stability than others. Possibly, values at
the top of individuals’ hierarchies are more stable than those near the bottom.
These results speak to several important pieces of scholarship that address values in
political behavior. Goren, Federico and Kittilson (2009) and Goren (2005), for example,
suggest value systems to be less stable than other important political attitudes (e.g. party
affiliation). In the same vein, Maio et al. (1996) and Maio and Olson (1998) suggest that
value ratings change when people are asked to think hard about their values. And, while
these pieces of research make important contributions to our understanding of values and
behavior, the results presented in this chapter should push future research to think more
about value system measurement, trade-offs, and information that can be gleaned from
rankings and ratings. The previously mentioned research that measures values and value
system dynamics using importance ratings do not tell a complete story. First, to the
extent that values are comparative and competitive –that choice depends on the relative
order of the relevant values– relative ordering matters, and ratings cannot capture this
important aspect of values. Second, a look inside rank-ordered value preferences provides important information in determining the extent to which value systems are static
or dynamic. More specifically, when value system ordering is measured with rankings,
information regarding which values are more stable and which are less stable becomes
available.
To begin this chapter, I posed two questions: Does situational context affect value system ordering? And, does measurement technique affect empirical results? The answers,
respectively, are sort of and yes. Changes in informational context do affect value system
ordering, but it is not the case that all values in the set are affected. Rather, some values
(e.g. economic security) are more stable than others (e.g. equality). With respect to
measurement, trade-offs seem clear. It may not be the case that rankings are stable and

42

ratings are unstable, as was hypothesized. Rather, rankings provide a look inside ordered
preferences where ratings do not. And, this look inside individuals’ ordered preferences
provides important information on how people think about values, which values people
seem to “move,” and “between-value” dynamics.

43

Figure 3.1: Mean Ratings and Rankings

5.0

q
q

4.5

q
q
q

4.0

q

3.5

q
q

3.0

q

2.5

q

2.0

Freedom

Law & Order Economic Sec.

Equality

Ratings Group
Rankings Group

*Bars represent two standard errors.

44

Moral Trad.

Figure 3.2: Distributions of Ratings

80

80

60

60

60

40

40

40

20

20

20

0

0
1

2

3

4

0

5

1

Equality

2

3

4

5

1

Law & Order

80

Percent of Total

Percent of Total

80

40

20

20

0

0
1

2

3

4

5

1

Freedom

2

3

4

5

Economic Security

*Most important values coded as 5, least important coded as 1.

45

4

5

Moral Traditionalism

60

40

3

80

60

2

Figure 3.3: Distributions of Rankings

80

80

60

60

60

40

40

40

20

20

20

0

0
1

2

3

4

0

5

1

Equality

2

3

4

5

1

Law & Order

80

Percent of Total

Percent of Total

80

40

20

20

0

0
1

2

3

4

5

1

Freedom

2

3

4

5

Economic Security

*Most important values coded as 5, least important coded as 1.

46

4

5

Moral Traditionalism

60

40

3

80

60

2

Table 3.1: Percent With Clear Preferences and Consistent, by Value Pair
Freedom, Equality
Freedom, Economic Security
Freedom, Law & Order
Freedom, Moral Traditionalism
Equality, Economic Security
Equality, Law & Order
Equality, Moral Traditionalism
Economic Security, Law & Order
Economic Security, Moral Traditionalism
Law & Order, Moral Traditionalism

% Clear Choice
94.9
97.1
96.7
95.9
95.9
96.5
95.1
95.4
97.1
96.5

% Consistent
69.1
67.2
65.1
79.8
63.5
62.4
63.5
60.5
70.5
71.2

9.6 % of respondents were consistent on all 10 pairwise choices.
50 % of respondents were consistent on at least 8 pairwise choices.

Table 3.2: Percent Transitive, by Triad
Freedom, Equality, Economic Security
Economic Security, Moral Traditionalism, Law & Order
Freedom, Moral Traditionalism, Equality
Equality, Law & Order, Freedom
Economic Security, Law & Order, Freedom
Equality, Moral Traditionalism, Law & Order
Freedom, Economic Security, Moral Traditionalism
Law & Order, Equality, Economic Security
Moral Traditionalism, Economic Security, Equality
Freedom, Moral Traditionalism, Law & Order

% Transitive
90.6
91.2
88.9
89.8
89.8
89.8
91.6
90.1
89.7
92.5

74.8% of respondents exhibited transitive preferences on all ten triads.
n = 628

47

Table 3.3: Difference in Mean Rating/Ranking of Equality
Rating Group Ranking Group
Control
4.43 (n = 217) 2.59 (n = 213)
Treatment 4.26 (n = 212) 2.41 (n = 216)
Difference
0.17∗
0.18
*Statistically significant at .05 level for a two-tailed difference-in-means test.

Table 3.4: Difference in Mean Rating/Ranking of Economic Security
Rating Group Ranking Group
Control
4.57 (n = 217) 1.90 (n = 213)
Treatment 4.49 (n = 206) 2.11 (n = 199)
Difference
0.08
-0.21
*Statistically significant at .05 level for a two-tailed difference-in-means test.

48

Table 3.5: Ordered Logit and OLS Models of Ratings of Equality and Econ. Security
Equality
Ord. Logit OLS
Prime
-0.49∗∗
-0.19∗
(.17)
(.08)
Age
0.1∗
0.04
(.05)
(.02)
Nonwhite
0.21
-0.00
(.21)
(.09)
∗
Female
0.38
0.18∗
(.16)
(.07)
Educ.
0.10
0.04
(.10)
(.04)
Party
0.15∗∗
0.06∗∗
(.05)
(.02)
Ideol.
0.06
0.03
(.07)
(.03)
Income
-0.02
-0.00
(.02)
(.01)
τ1
-2.38
–
(.49)
τ2
-2.00
–
(.47)
τ3
-0.35
–
(.44)
τ4
0.96
–
(.44)
LR χ sq.
45.89∗∗∗
–
R sq.
–
0.05
n
633
633
∗ p < .05, ∗∗ p < .01

49

Econ. Security
Ord. Logit OLS
-0.02
0.00
(.18)
(.08)
0.13∗∗
0.03
(.05)
(.02)
-0.04
-0.10
(.22)
(.08)
∗
0.32
0.12
(.17)
(.07)
0.03
-0.01
(.10)
(.04)
0.10
0.05∗
(.05)
(.02)
0.09
0.05
(.08)
(.03)
0.01
0.01
(.02)
(.01)
-2.30
–
(.51)
-1.51
–
(.47)
-0.50
–
(.45)
0.82
–
(.45)
21.38∗∗
–
0.04
635
635

Table 3.6: Changes in Predicted Probabilities by Treatment Group for Equality Rating
P r(Y
P r(Y
P r(Y
P r(Y
P r(Y

=
=
=
=
=

Very Important)
Somewhat Imp.)
Moderately Imp.)
Somewhat Unimp.)
Very Unimportant)

Control
0.74
0.18
0.07
0.01
0.01

Treatment
0.63
0.24
0.11
0.01
0.02

Change (95% C.I.)
0.11 (0.04, 0.18)
-0.06 (-0.10, -0.02)
-0.04 (-0.07, -0.01)
0.00 (-0.01, 0.00)
-0.01 (-0.02, 0.00)

Cell entries in the first column indicate the probability that Y = y for respondents
in the control group with all other independent variables held at their medians (55–
64 year old female that considers herself a political independent but leans toward the
Democratic Party). Cell entries in the second column indicate the probability that Y = y
for respondents exposed to the equality prime with all other independent variables held
at their medians. Confidence intervals for the change in the predicted probabilities were
calculated using the Delta method.

50

Table 3.7: OLS Models for Value Rankings
Equality
-0.23∗
(.09)
Age
-0.07∗
(.03)
Nonwhite
0.17
(.11)
Female
-0.13
(.09)
Educ.
0.10∗
(.05)
Party
0.11∗∗
(.03)
Ideol.
0.22∗∗
(.04)
Income
0.00
(.01)
R sq.
0.21
n
626
Prime

∗ p < .05, ∗∗ p < .01

51

Economic Security
0.12
(.11)
-0.03
(.03)
0.01
(.12)
-0.01
(.10)
-0.11∗
(.06)
0.11∗∗
(.03)
0.02
(.05)
-0.02
(.01)
0.06
626

Table 3.8: Rank-Ordered Logit Model (Equality as Reference Category)
Freedom
Eq. Prime
0.15
Econ. Sec. Prime -0.10
Age
0.22∗∗
Nonwhite
-0.06
Female
-0.01
Educ.
-0.02
Ideol.
-0.03
Party
-0.06
Econ. Security
Eq. Prime
0.44∗∗
Econ. Sec. Prime
0.27
Age
0.11∗∗
Nonwhite
-0.00
Female
0.16
Educ.
-0.11
Ideol.
-0.04
Party
-0.01
LR
n

(.18)
(.18)
(.04)
(.17)
(.14)
(.08)
(.06)
(.05)

Law & Order
Eq. Prime
0.17
Econ. Sec. Prime
-0.03
Age
0.20∗∗
Nonwhite
0.00
Female
0.24
Educ.
0.02
Ideol.
-0.09
Party
-0.15∗∗

(.16)
(.17)
(.04)
(.17)
(.14)
(.07)
(.06)
(.04)

(.17)
(.17)
(.04)
(.17)
(.14)
(.07)
(.06)
(.05)

Moral Trad.
Eq. Prime
0.44∗
Econ. Sec. Prime
0.08
Age
0.16∗∗
Nonwhite
0.09
Female
0.32∗
Educ.
-0.00
Ideol.
-0.30∗∗
Party
-0.17∗∗

(.18)
(.18)
(.04)
(.18)
(.15)
(.08)
(.06)
(.05)

779.71∗∗
626

Equality is the reference category.
Cell entries are logit coefficients.
∗∗ p < .01,∗ p < .05

52

Table 3.9: Rank-Ordered Logit Model (Economic Security as Reference Category)
Freedom
Eq. Prime
Econ. Sec. Prime
Age
Nonwhite
Female
Educ.
Ideol.
Party
Equality
Eq. Prime
Econ. Sec. Prime
Age
Nonwhite
Female
Educ.
Ideol.
Party
LR
n

-0.29
-0.38∗
0.11∗∗
-0.06
-0.17
0.09
0.01
-0.06
-0.44∗∗
-0.27
-0.11∗∗
0.00
-0.16
0.11
0.04
0.01

(.18)
(.18)
(.04)
(.17)
(.15)
(.08)
(.06)
(.05)

Law & Order
Eq. Prime
Econ. Sec. Prime
Age
Nonwhite
Female
Educ.
Ideol.
Party

-0.27
-0.30
0.10∗∗
0.00
0.08
0.13
0.05
-0.15∗∗

(.17)
(.17)
(.04)
(.16)
(.14)
(.07)
(.06)
(.05)

(.17)
(.17)
(.04)
(.17)
(.14)
(.07)
(.06)
(.05)

Moral Trad.
Eq. Prime
-0.00
Econ. Sec. Prime
-0.20
Age
0.06
Nonwhite
0.10
Female
0.16
Educ.
0.10
Ideol.
-0.26∗∗
Party
-0.16∗∗

(.18)
(.19)
(.04)
(.18)
(.15)
(.08)
(.07)
(.05)

779.71∗∗
626

Economic Security is the reference category.
Cell entries are logit coefficients.
∗∗ p < .01,∗ p < .05

53

Chapter 4

The Temporal Stability of Values

The central question in this chapter, like the previous chapter, concerns whether
values can structure attitudes. However, in the previous chapter, the focus was on the
individual. Here, in this chapter, I focus on value system ordering at the aggregate level.
Where a look at value system stability at the individual level provides insight on the
degree to which people can distinguish between values and whether or not value choice is
driven by importance or error, a look at value system stability in the aggregate provides
insight on societal priorities. The same basic theory applies: values should not change
“at random” over time. If they do, values may not be the “fundamental building blocks”
of behavior and attitudes that traditional psychological theory purports them to be.
In this chapter, I use a matching design with values data from three different time
points (1994, 2003, and 2005) to examine the extent to which time, and events that
occur over time, affect value system stability in the aggregate. Briefly, I find that there
is significant value change (an increase in the importance of social order) between 1994
and 2003, likely due to the events of 9/11. I also find significant value change (a decrease
in the importance of social order) between 2003 and 2005. I conclude that this change
is not due to changes in the utility people attach to social order. Rather, it is due to
changes in the accessibility of social order.

54

4.1

Research on the Temporal Stability of Value Systems
While research on the situational stability of value systems is abundant, research on

temporal stability is quite sparse. In fact, only a small number of studies take individual
level stability into account. Among these studies is Rokeach’s seminal work on the nature
of values (1973). In short, Rokeach found correlations among rank-ordered values to be
quite high (between .7 and .8, in most cases), but wide variation between correlations of
individual items. For example, the test-retest correlations for “responsible” and “a sense
of accomplishment” were .45 and .51, respectively, while the test-retest correlations for
“equality” and “salvation” were much higher, .71 and .88 (Feldman 2003; Rokeach 1973).
By and large, this evidence suggests that value systems are stable. Individual values,
though they vary in degree, are largely stable as well.
Among the few studies to assess the temporal stability of values with a nationally
representative sample is McCann’s (1997) study that uses panel data from 1990–1992
and multiple-item scales to measure support for certain values. In short, McCann found
a relatively low correlation over panels for support for equality of opportunity (r = .41),
but a relatively high correlation for moral traditionalism (r = .81)1 In addition to the test–
retest correlations, McCann examined the extent to which support for various presidential
candidates affected value preferences. He found that those respondents that supported
Bill Clinton tended to increase their support for equality of opportunity and decrease
their support for moral traditionalism. Respondents that supported George Bush shifted
their value preferences in the opposite direction. Conclusions to be drawn from this
study are mixed, and while support for a presidential candidate may suggest values to be
endogenous, results regarding temporal stability are, at best, less than convincing.
1 It should be noted that Sheng (1995) takes measurement error into account and adds

two additional items to McCann’s “support for equality of opportunity” scale. His results show increased stability. Support for equal opportunity and moral traditionalism
correlate at .81 and 1.00, respectively.

55

Perhaps the most recent study on the temporal stability of values is Goren’s 2005 piece
in which he assesses the dynamic relationship between partisan identification and four core
values in American politics: equal opportunity, limited government, traditional family
values, and moral tolerance. Using ANES panel data from 1992, 1994, and 1996, Goren
found the effects of partisanship on values to be significant, but no effects going from
values to party. The conclusions reached in this study suggest that critical assumptions
made by traditional values theory –namely, that values are relatively stable– are untrue.
Though influential, the above studies are not without flaws. The most common critique to the Rokeach Values Survey is its length. Error is expected in ranking and
re-ranking 18 items, regardless of their significance to everyday life. This error, likely due
to respondent fatigue as much as it is due to the instability of values, influences substantive conclusions regarding value stability. Value measurement, in both McCann (1997)
and Goren (2005), is indirect. Multiple-item measures that approximate the degree to
which respondents value equality, for example, is quite different than asking respondents
to choose between two important values (e.g. freedom and equality). In addition, the
theoretical argument addressing why partisan identity and support for a particular presidential candidate, but not issue attitudes, affects value orientation is unclear. Absent
stronger theory or data that corroborate these findings, it is not clear the degree to which
these results are generalizable.
Inglehart’s work, though it concerns values and value change just as the abovementioned studies do, broadens the scope of inquiry and makes societies the focus of
the research (e.g. Abramson and Inglehart 1995; Inglehart 1971, 1977; Inglehart and
Abramson 1994). In short, Inglehart argues that values are a result of economic advancement and security, and that a look at value priorities, in the aggregate, sheds light
on what societies view as important and what is disposable. Advanced societies, those
that Inglehart describes as post-materialistic, hold such values as self-expression and en-

56

vironmentalism, for example, in high priority while industrial societies hold economic
and domestic security in high priority. By and large, value change occurs slowly as one
generation replaces the other. And, values are an important part in the description and
explanation of any society.
In this chapter, I examine value preferences and change in the aggregate, and I attempt to answer questions on the complex relationships between values, attitudes, and
psychological composition. I work to not only determine whether or not values are stable
over time, but also to determine the mechanisms that might cause value instability. I
do so in hope of bridging the gap between traditional psychological theory on values and
new research on value system dynamics.

4.2

Research Design, Data, & Methods
As stated above, true panel data on value hierarchies is scarce, thus, I focus on value

system stability in the aggregate (like Inglehart 1977, for example). In place of panel
data, I employ a matching design in which I take data from three different time points
(1994, 2003, and 2005) and measure the effects of time (and, of course, the events that
happen over time) on value choices. I also use a rank-ordered logit model to examine
value “trade-offs” over time.2
The 1994 data is from the Multi-Investigator Study (MIS) (Sniderman, Brady and
Tetlock 1994). The MIS was a national random-digit telephone survey in which the survey
population was defined as English-speaking adults 18 years of age or older, residing in
households with telephones, within the 48 contiguous states (n = 1464). The survey
included both the standard battery of questions often found on US public opinion surveys
(e.g. party identification, ideological stances, issues and policies, demographics, etc.) and
experimental manipulations unique to this data collection effort. And, among the unique
2 Please see previous chapter for an explanation of rank-ordered logit.

57

items is a battery of questions that asked respondents to make paired comparison choices
among four values: liberty, equality, economic security, and social order (Jacoby 2006).3
In creating a rank-ordering suitable for analysis, I combined responses from the six choice
questions to form a four-item set of ordered preferences. Items in the set are coded such
that higher numbers indicate the relevant value is of high importance while lower numbers
indicate low importance.
The 2003 data is from a TESS4 study that builds an experimental design into a broad
survey regarding American political attitudes (n = 9313). Data collection was internetbased, and the data collection agency (Knowledge Networks) recruited respondents via
telephone (both numbers that were linked to mailing addresses and those that were not).
In collecting the data, oversampling took place in the Los Angeles and Chicago areas, and
there is an oversample of minority households as well. The “values” item simply listed six
values (liberty, equality, economic security, social order, morality, and patriotism) and
asked respondents to choose the most important. The chosen item was then deleted from
the list, and the respondent was asked to choose the most important item remaining. This
process was repeated five times until each respondent specified a six-item rank-ordered
preference list. I make use of four values (liberty, equality, economic security, and social
order –the same four values included in the MIS) and construct a four-item preference
ordering from the original six-item list.
The 2005 data is also from a TESS5 study that combines and experimental design with
traditional survey questions often asked in public opinion surveys concerning political
attitudes (n = 649). Again, data-collection was internet-based, and recruitment was
done using telephone numbers (some that were attached to mailing addresses, and some
3 The 1994 MIS data and codebook can be found online at http://sda.berkeley.edu/archive.htm.
4 Data collected by Time-sharing Experiments for the Social Sciences, Arthur Lupia and

Diana Mutz, Principal Investigators.
5 Data collected by Time-sharing Experiments for the Social Sciences, Arthur Lupia and

Diana Mutz, Principal Investigators.

58

that were not). Here, value preferences were constructed using the method of triads
(five values were used: liberty, equality, economic security, social order, and morality).
Using the original five-item preference list, I constructed a four-item preference list (using
liberty, equality, economic security, and social order).
All three datasets, together, provide a unique opportunity to study the dynamics of
aggregate value preferences over time. And, though each survey used different measures
to capture individual level value system ordering, the relevant values were defined in such
a way that results are comparable across time points. As previously stated, the most
important value in an individual’s belief system is coded as 4, the second most important
is coded as 3, and so on. If values are tied for the first (and second) positions in the
hierarchy, each value is coded as 3.5.
The “treatment” variable in this analysis is time, and the quantity of interest is the
effect of time (and events that occur over time) on the relative order of each of the
four values in the set. Control variables include party, ideology, age, sex, race, income,
education, and age.
4.2.1

A Note on Matching

Throughout the course of any given analysis of quantitative data, a researcher will
specify a model and obtain results, then respecify the model and obtain another set of
results, and repeat the process several times. This so, it is rare that the results of all of
these various model specifications are reported in the final draft of the paper. Thus, it is
possible to say that the reported results are model dependent (Ho et al. 2007). Matching
methods aim to reduce the degree to which causal estimates are model dependent by
controlling for some or all of the confounding influence of control variables in a manner
that is free of parametric assumptions (Blackwell et al. 2009; Ho et al. 2007). In short,
the goal of any matching algorithm is to “prune observations from the data so that the

59

remaining data have better balance between the treated and control groups...” (Blackwell
et al. 2009, 526). Data that is perfectly balanced (e.g. when the empirical distributions
of the control variables are equivalent across the treatment and control groups) requires
only a difference in means test to estimate causal effects (Blackwell et al. 2009). Further
parametric adjustments on perfectly balanced data are unnecessary in estimating the
effects of the treatment on the dependent variable (Ho et al. 2007, 200).
Most matching algorithms proceed by finding, for each treated unit, at least one
control unit that has similar values on all relevant covariates. Unmatched units (the
units that cause imbalance), are pruned from the dataset and are not used in analysis.
There are two broad classes of matching methods: exact matching and inexact matching.
Exact matching pairs observation that have the same values on all relevant covariates
that differ by treatment group. And, while an exact matching solution is desirable (see
below), it is often the case that these solutions produce too few matched pairs (Blackwell
et al. 2009). Inexact matching methods define often “similarity” Mahalanobis distance (a
dissimilarity measure based on correlation structure between sets of values) (Cochran and
Rubin 1973; Rubin 1979, 1980), the propensity score (the probability that a unit is in the
treatment group, conditional on the covariates) (Rosenbaum and Rubin 1983) or some
combination of the two (Rosenbaum and Rubin 1985). The number of control units that
are to be matched is often chosen ex ante, and balance is checked ex post. If there are too
few matches, or if the balance is not at an acceptable level, the researcher must reduce
the number of control units that are to be matched and proceed with the process again
(Iacus, King and Porro 2011). The most common problem with this class of methods is
that a researcher may often have to repeat this process several times before finding both
a large enough number of matches and an acceptable level of balance (Blackwell et al.
2009).
Preprocessed datasets that achieve perfect balance (i.e. the treatment is orthogonal

60

to all other covariates) do not require the researcher to further control for confounding
covariates. Here, a simple bivariate regression of the dependent variable on the treatment
provides a causal estimate. And, though this is a very desirable property, matching on
continuous variables (or variables that have several ordered categories), often produces
a small number of matched pairs (Blackwell et al. 2009). A relatively new “flavor” of
matching, Coarsened Exact Matching, (CEM) provides a solution to this problem by
allowing the researcher to create meaningful categories for continuous or ordered variables
with several categories (e.g. break a 7-point education scale into 4 categories or break
a 19-point income scale into fewer meaningful categories), match on the coarsened data,
then analyze the uncoarsened data.
CEM proceeds in three steps: (1) coarsen the original variables in X as much as the
research is willing to, creating C(X); (2) apply exact matching to C(X), which involves
sorting the observations into strata (s ∈ S), each with unique values of C(X); (3) discard
strata with only control units, discard strata with one treatment units (or use with
extrapolated values of the control units), and estimate the quantity of interest (Iacus,
King and Porro 2011, 16).

4.3

Hypotheses
I derive hypotheses to be tested in this chapter using the model of value rankings

presented in the previous chapter. Recall, the process by which individual i calculates
the utility (or importance) of value a is determined by equation 4.1.

. uia = Iia + eia

(4.1)

The systematic component, Iia is broken down into a true importance (tia ) and contextual salience (cia ). In the previous chapter, contextual salience varied by individual. In
this chapter, because it is captured by the year in which the survey was taken, it will be
61

subscripted by a year and value (e.g. c2003a or c2005a ). So, equation 4.1 becomes 4.2.
uia = tia + cyear,a +eia

(4.2)

Iia
Also recall that the relative ordering of an individual’s value system is determined, in
essence, by a series of pairwise choices. Using values a and b for illustrative purposes, the
process by which one chooses value a over value b, or vice versa, is shown in equation set
4.3.

Choice(a)

if

Iia + eia > Iib + eib

(4.3)

and
Choice(b)

if

Iia + eia < Iib + eib .

Rearranging the terms, as seen in equation 4.4, it becomes apparent that value choices
should be consistent if there is a relatively large difference between, for example, Iia and
Iib , and if eia and eib are relatively small, then, under normal circumstances, the true
utility of the values should drive choice and choices should be stable.

Choice(a)

if

Iia − Iib > eib − eia

(4.4)

and
Choice(b)

if

Iia − Iib < eib − eia .

If, however, circumstances were abnormal –if a major event occurred that either changed
the true utilities of values or made one value particularly accessible, data may show values
to be unstable over time. Formally, breaking down Iia into tia and cia , it becomes
apparent that, if Iia and Iib are stable, and if E( ia ) = E( ib ) = 0, that contextual

62

salience can drive choice (see equation 4.5).

Choice(a)

if

E(tia − tib ) > E(cyear,b − cyear,a )

(4.5)

and
Choice(b)

if

E(tia − tib ) < E(cyear,b − cyear,a ) .

If the event caused a major change in the true importance of individual values (in the I
terms), then the change in the relative ordering of value structures should be permanent
(or, at least very long lasting). If, on the other hand, changes in accessibility caused
changes in the relative ordering of values, then the change should only be temporary.
The values data from multiple time points and a major event (9/11) occurring just
before data collection in 2003 makes for a easily definable natural experiment. 9/11, the
most deadly attack ever on American soil, reminded many people about the importance
of security and social order. Empirically, this should result in a major shift up for social
order in individuals’ value systems between 1994 and 2003. Between 2003 and 2005,
social order may stay high on value hierarchies. If so, this suggests a true change in value
systems. If, on the other hand, social order falls back to where it was in 1994, the change
may be due to increased, then decreased accessibility.

4.4

Results
A brief look at some descriptive statistics shows that the mean rankings of each of the

four values change, at least slightly, by year.6 Figure 4.1 and table 4.1 display graphically,
and in tabular form, how the mean rankings for each value change by year. First, and
perhaps most strikingly, it is quite obvious that these values, for the most part, are not
stable (at least in the aggregate). Liberty is relatively stable, increasing only slightly in
6 Note that rankings are coded such that 4 indicates the most important value in a

hierarchy, and 1 indicates the least important value in a hierarchy.
63

importance every year, but the other three values show little stability. As expected, social
order jumps in importance between 1994 and 2003, but it declines again between 2003 and
2005. And, because economic security and equality show the opposite pattern of social
order, it appears that these two values suffer in terms of importance at the expense of
social order. These descriptive statistics, just the bivariate relationships, seem to suggest
that the events of 9/11 only affect the degree to which social order was accessible in
peoples’ minds. It does not seem to be the case that 9/11 caused long term value change.
4.4.1

Matched and Unmatched Data

The linear regressions on the matched and unmatched datasets (results shown in
shown in tables 4.2 and 4.3) show much the same pattern. The coefficients on year
in table 4.2 suggest that all values move a bit over years, but some more than others.
Liberty, for example, moves up in value hierarchies, on average, between 1994 and 2003,
but only by about 20 percent of a rank. It remains the second most important value in the
aggregate (behind economic security in 1994 and social order in 2003). The coefficients
on year on the other three values seem quite a bit more impactful. In fact, it appears to
be the case that many “traded” social order, in 2003, for economic security or equality.
Economic security and equality move down approximately half a rank, on average, while
social order moves up nearly 75 percent of a rank. Controlling for various covariates, it
is clear that events between 1994 and 2003 (likely 9/11) caused the relative ordering of
values in the mass public to shift.
Value systems seem a bit more stable from 2003 to 2005 than that from 1994 to 2003
(results are reported in table 4.3). The passage of time moved liberty very little, and
according to the matched data, it did not move at all. Equality is also relatively stable
here, moving, on average, 20 percent of a rank up the hierarchy. Economic security and
social order, during this passage of time, seem to trade positions. Economic security, on

64

average, moves 60 percent of a rank up the hierarchy, and social order moves almost an
entire rank (80 percent) down the hierarchy.
Judging by the results presented in tables 4.2 and 4.3, it is clear that values are
affected by events over time. To shed light on whether the change is due to changes in
the actual utilities of values or whether accessibility is driving the change, it is necessary
to assess the extent to which values “snap back” to 1994 relative ordering after 9/11. As
previously stated, a change in the utilities of values should result in permanent change
(e.g. no decay). A change in accessibility should decay over time. Table 4.4 displays the
results from OLS regressions run in which 1994 is the “control” condition and 2005 is the
“treatment” condition. A significant coefficient on the “treatment” variable might suggest
a change in utility, and a non-significant coefficient might signify a change in accessibility
only. Looking first at the results from the unmatched data, the coefficients on the 2005
variable show only small change between intercepts in 1994 and 2005. The biggest mover,
according to these data, is liberty, which, over all three time points, is the most stable
of the four values. The second biggest mover, social order, moved a 25 percent down a
rank during the eleven-year time period. Over the first nine-year time period, however,
social order moved up 75 percent of a rank, and over the two-year period from 2003 to
2005, social order moved down about 80 percent of a rank. The two remaining values,
equality and economic security, have significant differences between their 1994 mean and
2005 mean, but, in each case, the difference is less than 20 percent of a rank.
Turning to the results from the matched data, it appears as though the change in
values’ mean rank due to time is minimal. In three of four cases, the coefficient on year is
not significant. The single coefficient on 2005 that is significant is in the model of social
order. Social order is easily the most volatile value of the four analyzed, and it seems as
though the effect of 9/11 on value system ordering, and especially on the mean ranking
of social order, lasts beyond 2005.

65

4.4.2

Rank-Ordered Logit

The results of the rank-ordered logit model, shown in table 4.5, display the log odds
of choosing a particular value over the reference value. The variables of interest, here,
indicate the year of the survey. In 1994, controlling for all other variables, respondents,
on average, chose liberty and economic security over social order, and social order over
equality. The coefficients on the 2003 variable tell a very different story. All coefficients
on the 2003 variable are negative, indicating that respondents are less likely to choose
the given value over the reference value. The coefficients on 2003, on liberty and economic security, though significant, are relatively small in magnitude. The coefficient on
2003 on equality, however, is quite large. It seems as though many respondents “traded”
social order for equality. The coefficients on the 2005 variable tell a different story again.
Here, equality stays in place, but respondents tend to choose liberty and economic security (especially economic security) over social order. In fact, it seems as though many
respondents “trade” social order for economic security.

4.5

Conclusion & Discussion
The focus of this chapter was estimating the effect of the passage of time, and events

that occur over time, on value system ordering. Psychological theory has assumed, for
quite some time, that values are relatively stable over time, but this assumption has
never been thoroughly tested. The theory of value system stability that was presented
in this chapter makes an important distinction between a change in relative ordering
based on changes in utility, and a change in relative ordering based on accessibility. In
short, historic events, like the onset of a major war, an economic depression, or an attack
like 9/11, are expected to alter the importance people attach to certain values. So, the
question is twofold: First, does the passage of time, and events that occur over time,
affect value system ordering? And second, if changes do occur, can they be attributed
66

to changes in the utility people attach to values, or is the change due to increases and
decreases in accessibility?
The “year” coefficients in tables 4.2 and 4.3 (i.e. 2003 in table 4.2 and 2005 in table
4.3) show the differences in the intercepts between 1994 and 2003, and 2003 and 2005,
respectively. The most notable results, over both tables, is on social order. Social order,
after 9/11, became far more important in value hierarchies than it was previously, and it
did so at the expense of equality and economic security. However, after only two years, the
importance of social order dropped, and economic security gained. These results suggest,
possibly, that people did not necessarily put increased importance on social order after
9/11. Rather, it became very accessible and stayed on peoples’ minds. But, while social
order decreased in importance shortly after 9/11 (e.g. between 2003 and 2005), only
economic security –not equality– regained its importance.
The analysis of matched data tells basically the same story as that of the unmatched
data, with one important exception. In table 4.4, the coefficient on the year variable
(i.e. the 2005 variable) is not significant in three of four values. The only value that
has a different intercept in 2005 than it does in 1994, according to the matched data, is
social order, which actually decreases in importance. Otherwise, all values have similar
mean rankings, suggesting that many of the changes over time are due to changes in
accessibility, not in the actual utility attached to values.
The results of the rank-ordered logit show not what is happening to each value individually. Rather, it shows “between-value” dynamics. The coefficients on the 1994 variable,
in table 4.5, suggest that, holding everything else constant, respondents were more likely
to choose economic security over social order. The other two coefficients on 1994 are not
statistically significant. The major change between 1994 and 2003, captured by the coefficients on the 2003 variable, is the significant increase in the probability people choose
social order over all other values. It seems apparent that 9/11 had a major effect on

67

the relative ordering of most peoples’ value systems. The change that occurred between
2003 and 2005 concerned three values in particular: social order, liberty, and economic
security. These results suggest that the passage of time made social order less accessible
in peoples’ minds, and that liberty and economic security nearly reclaimed their pre-9/11
positions in the relative ordering of values.
The results presented here, on one hand, corroborate with what some previous research
has found: values, over time, are not completely stable (Goren 2005; Goren, Federico
and Kittilson 2009; McCann 1997). On the other, this study pushes these previous
results a big farther in suggesting events have only a temporary effect on value system
ordering. And, this allows for two conclusions, one fairly concrete, and one less so. First,
value system ordering is affected by events –values are not completely stable over time,
especially in the wake of major events. Second, these changes in value system ordering are
seemingly due to changes in specific value accessibility rather than changes in the utility
that individuals attach to specific values. Further, it seems as though this contextual
salience effect decays over time.
With these results in mind, it is useful to turn attention back to whether or not values
can provide a general evaluative framework from which people can evaluate objects in
the world. The general instability of values, as shown in this and the previous chapter,
do not conform to traditional psychological theory. This so, it is inaccurate to say that
values are, for example, “truisms” (Maio and Olson 1998). That is, not all values are held
with the same high esteem. Further, values, when they do “move,” do so predictably.
Lastly, some values (e.g. freedom/liberty) are more stable than others (e.g. equality, law
& order/social order).
What does this mean for values as they relate to behavior and choice? Two things are
apparent. First, some values, because they are unstable, cannot organize attitudes and
guide behavior. It may not even be the case that all values relevant to a certain decision

68

guide behavior. Second, it may not be the case that all relevant values are needed to
guide behavior. It may simply be the case that individuals rely solely on a small number
of values to guide choice. The results here narrow the focus of values research. It may
not be worth examining the stability and impact of all values in behavior. To focus on
a smaller number of values that individuals hold in high regard may be enough. Further
research that asks which values are relevant to certain individuals, or certain groups, is
warranted.

69

Figure 4.1: Mean Rankings by Year

3.0

q

q
q

2.8

q

q
2.6

2.4

q

q
q
q

q

2.2

q
q
2.0

1994

2003
Liberty
Equality
Econ. Security
Social Order

*Bars represent two standard errors.

70

2005

Table 4.1: Mean Rankings by Year

Liberty
Equality
Economic Sec.
Social Order
n

1994
2.57
(1.12)
2.40
(0.96)
2.69
(1.02)
2.35
(1.06)
1464

2003
2.70
(1.00)
2.11
(0.99)
2.29
(0.87)
2.89
(0.94)
9313

Note: All differences statistically significant at .05 level.
Standard deviations in parentheses.

71

2005
2.87
(1.06)
2.26
(1.00)
2.83
(1.02)
2.04
(1.07)
649

Table 4.2: OLS Results on Matched and Unmatched Data, 1994 & 2003
Liberty
Unmatched Matched
Age
Nonwhite
Female
Educ.
Party
Ideol.
Income
2003
Constant
R sq.
n
Matched Units

0.01
(.01)
-0.44∗∗∗
(.03)
-0.15∗∗∗
(.03)
0.09∗∗∗
(.01)
-0.06∗∗∗
(.01)
-0.00
(.01)
0.04∗∗∗
(.01)
0.18∗∗∗
(.04)
2.63∗∗∗
(.08)
0.07
7100
–

0.06∗∗
(.03)
-0.79∗∗∗
(.25)
-0.19∗∗∗
(.07)
0.13∗∗
(.06)
-0.08∗∗
(.03)
0.06
(.04)
-0.03
(.03)
0.23∗∗∗
(.07)
2.41∗∗∗
(.21)
–
–
1480

Equality
Unmatched Matched
-0.05∗∗∗
(.01)
0.30∗∗∗
(.03)
0.03
(.02)
0.00
(.01)
0.08∗∗∗
(.01)
0.11∗∗∗
(.01)
-0.06∗∗∗
(.01)
-0.38∗∗∗
(.04)
1.90∗∗∗
(.07)
0.14
7100
–

-0.02
(.02)
-0.04
(.21)
0.04
(.06)
-0.05
(.05)
0.06∗∗
(.03)
0.11∗∗∗
(.04)
-0.07∗∗
(.03)
-0.52∗∗∗
(.06)
2.24∗∗∗
(.18)
–
–
1480

Economic Security
Unmatched Matched
-0.03∗∗∗
(.01)
0.18∗∗∗
(.03)
0.09∗∗∗
(.02)
-0.13∗∗∗
(.01)
0.02∗∗∗
(.01)
-0.03∗∗∗
(.01)
-0.01
(.01)
-0.40∗∗∗
(.03)
3.08∗∗∗
(.07)
0.05
7100
–

Multivariate Imbalance Measure (L1 ) on unmatched/matched data: 0.904/0.00.
Percentage of local common support on unmatched/matched data: 7.6%/100%.
CEM matched 558 of 1464 “1994” units and 922 of 9313 “2003” units.
Standard errors in parentheses.
72

-0.06∗∗∗
(.02)
0.63∗∗∗
(.22)
0.11∗
(.06)
-0.08∗
(.05)
0.01
(.03)
-0.04
(.04)
0.01
(.03)
-0.43∗∗∗
(.06)
3.02∗∗∗
(.19)
–
–
1480

Social Order
Unmatched Matched
0.07∗∗∗
(.01)
-0.04
(.03)
0.03
(.02)
0.03∗∗
(.01)
-0.05∗∗∗
(.01)
-0.08∗∗∗
(.01)
0.03∗∗∗
(.01)
0.60∗∗∗
(.04)
2.39∗∗∗
(.07)
0.09
7100
–

0.02
(.02)
0.20
(.22)
0.05
(.07)
0.00
(.05)
0.01
(.03)
-0.13∗∗∗
(.04)
0.09∗∗∗
(.03)
0.72∗∗∗
(.06)
2.33∗∗∗
(.19)
–
–
1480

Table 4.3: OLS Results on Matched and Unmatched Data, 2003 & 2005
Liberty
Unmatched Matched
Age
Nonwhite
Female
Educ.
Party
Ideol.
Income
2005
Constant
R sq.
n
Matched Units

Equality
Unmatched Matched

Economic Security
Unmatched Matched

Social Order
Unmatched Matched

-0.01
(.01)
-0.49∗∗∗
(.03)
-0.15∗∗∗
(.03)
0.10∗∗∗
(.01)
–

-0.02
(.02)
-0.44∗∗∗
(.07)
-0.14∗∗∗
(.05)
0.10∗∗∗
(.03)
–

-0.06∗∗∗
(.01)
0.40∗∗∗
(.03)
0.05∗∗
(.02)
0.00
(.01)
–

-0.02
(.02)
0.34∗∗∗
(.07)
0.07
(.05)
0.04∗
(.03)
–

-0.02∗∗∗
(.01)
0.19∗∗∗
(.03)
0.07∗∗∗
(.02)
-0.13∗∗∗
(.01)
–

-0.03∗∗
(.01)
0.12∗
(.06)
0.10∗∗
(.04)
-0.22∗∗∗
(.02)
–

0.09∗∗∗
(.01)
-0.09∗∗∗
(.03)
0.02
(.02)
0.03∗∗
(.01)
–

0.07∗∗∗
(.02)
-0.02
(.07)
-0.03
(.05)
0.07∗∗∗
(.03)
–

-0.04∗∗∗
(.01)
0.05∗∗∗
(.01)
0.11∗∗
(.05)
2.74∗∗∗
(.07)
0.06
6733
–

-0.06∗∗∗
(.02)
0.01
(.02)
-0.04
(.07)
3.02∗∗∗
(.14)
–
–
3355

0.17∗∗∗
(.01)
-0.06∗∗∗
(.01)
0.11∗∗
(.05)
1.62∗∗
(.07)
0.11
6733
–

0.16∗∗∗
(.02)
-0.04∗∗
(.02)
0.20∗∗∗
(.06)
1.35∗∗∗
(.14)
–
–
3355

-0.01∗
(.01)
-0.01∗
(.01)
0.60∗∗∗
(.05)
2.71∗∗∗
(.06)
0.05
6733
–

-0.01
(.02)
0.01
(.02)
0.64∗∗∗
(.06)
2.87∗∗∗
(.13)
–
–
3355

-0.12∗∗∗
(.01)
0.03∗∗∗
(.01)
-0.82∗∗∗
(.05)
2.93∗∗∗
(.07)
0.10
6733
–

-0.10∗∗∗
(.02)
0.02
(.02)
-0.79∗∗∗
(.06)
2.76∗∗∗
(.14)
–
–
3355

Multivariate Imbalance Measure (L1 ) on unmatched/matched data: 0.716/0.00.
Percentage of local common support on unmatched/matched data: 15.6%/100%.
CEM matched 2824 of 9313 “2003 units and 531 of 649 “2005 units.
Standard errors in parentheses.
73

Table 4.4: OLS Results on Matched and Unmatched Data, 1994 & 2005
Liberty
Unmatched Matched
Age
Nonwhite
Female
Educ.
Party
Ideol.
Income
2005
Constant
R sq.
n
Matched Units

Equality
Unmatched Matched

Economic Security
Unmatched Matched

Social Order
Unmatched Matched

0.07∗∗∗
(.02)
-0.45∗∗∗
(.08)
-0.21∗∗∗
(.06)
0.03
(.03)
–

0.13∗∗∗
(.05)
-0.28
(.22)
-0.29∗∗
(.14)
0.18∗
(.10)
–

-0.01
(.02)
0.32∗∗∗
(.07)
0.11∗∗
(.05)
0.02
(.03)
–

0.00
(.05)
0.59∗∗∗
(.22)
0.03
(.13)
-0.03
(.09)
–

-0.01
(.02)
0.24∗∗∗
(.07)
0.18∗∗∗
(.06)
-0.13∗∗∗
(.03)
–

-0.09∗
(.05)
-0.30
(.23)
0.17
(.14)
-0.26∗∗∗
(.10)
–

-0.05∗∗∗
(.02)
-0.11
(.08)
-0.08
(.06)
0.07∗∗
(.03)
–

-0.04
(.05)
-0.01
(.24)
0.09
(.14)
0.10
(.10)
–

-0.04∗∗
(.02)
0.04∗∗
(.02)
0.25∗∗∗
(.06)
2.48∗∗∗
(.15)
0.06
1358
–

0.07∗
(.04)
0.00
(.06)
0.21
(.14)
1.58∗∗∗
(.48)
–
–
560

0.04∗∗∗
(.01)
-0.06∗∗∗
(.02)
-0.18∗∗∗
(.06)
2.22∗∗∗
(.14)
0.04
1358
–

-0.04
(.04)
-0.04
(.06)
-0.02
(.13)
2.50∗∗∗
(.46)
–
–
560

-0.00
(.01)
-0.01
(.02)
0.17∗∗∗
(.06)
2.97∗∗∗
(.14)
0.04
1358
–

-0.11∗∗∗
(.04)
0.06
(.06)
0.16
(.14)
3.91∗∗∗
(.49)
–
–
560

-0.00
(.02)
0.03
(.02)
-0.24∗∗∗
(.06)
2.33∗∗∗
(.15)
0.04
1358
–

0.07
(.05)
-0.03
(.07)
-0.36∗∗
(.15)
2.02∗∗∗
(.51)
–
–
560

Multivariate Imbalance Measure (L1 ) on unmatched/matched data: 0.809/0.00.
Percentage of local common support on unmatched/matched data: 12.4%/100%.
CEM matched 354 of 1464 “1994” units and 206 of 649 “2005” units.
Standard errors in parentheses.
74

Table 4.5: Rank-Ordered Logit Results (Social Order is Reference Category)
Liberty
Age
Nonwhite
Female
Educ.
Ideol.
Income
1994
2003
2005

-0.08∗∗
-0.40∗∗
-0.17∗∗
0.06∗
0.08∗∗
0.01
0.17
-0.24∗
0.80∗∗

(.01)
(.06)
(.04)
(.02)
(.01)
(.02)
(.12)
(.11)
(.14)

Equality
Age
Nonwhite
Female
Educ.
Ideol.
Income
1994
2003
2005

-0.13∗∗
0.51∗∗
0.10∗
-0.05∗
0.30∗∗
-0.11∗∗
-0.22
-1.36∗∗
-0.27

(.01)
(.06)
(.04)
(.02)
(.02)
(.02)
(.12)
(.12)
(.15)

Economic Security
Age
Nonwhite
Female
Educ.
Ideol.
Income
1994
2003
2005
LR χ sq.
n

-0.11∗∗ (.01)
0.33∗∗ (.05)
0.12∗∗ (.04)
-0.18∗∗ (.02)
0.12∗∗ (.01)
-0.06∗∗ (.02)
0.74∗∗ (.12)
-0.26∗ (.11)
1.21∗∗ (.14)
4097.66∗∗
7623

Social order is the reference category.
Cell entries are logit coefficients.
∗∗ p < .01,∗ p < .05

75

Chapter 5

Values, Attitudes, and Behavior

Traditional psychological theory suggests that attitudes and behavior are, in large
part, the results of individuals projecting their values onto real world situations. Research finds links between values and many important attitudes and behaviors, including
ideology (Braithwaite 1982, 1994, 1997; Rokeach 1973; Schwartz 1994), career choice
(Hofstede 2001; Rokeach 1973), management policies and style (Egri and Ralston 2004;
Kumar and Kant 2007; Lenartowicz and Johnson 2003), product preference (Kahle 1996;
Murphey et al. 2006), teaching style (Bossman 1991), and volunteer work (Omoto and
Snyder 1995), to name a few. Newer research finds this relationship to be conditional
on education and sophistication (Bardi and Schwartz 2003; Maio and Olsen 2000; Zaller
1991, 1992), but the relationship is there nonetheless. In this chapter, I take a closer look
at the relationship between values and political attitudes, and I examine the extent to
which value system measurement affects empirical results.

5.1

Past Literature
Values are thought to have a pervasive influence on attitudes and behavior, and count-

less pieces of literature find evidence in favor of this claim (see, for example Rokeach 1973).
These studies, variable with respect to methodological sophistication, find relationships
between values and attitudes (Ajzen and Fishbein 1977; Maio and Olsen 2000), prejudice (Biernat et al. 1996), volunteerism (Omoto and Snyder 1995), cooperative behavior
76

(Schwartz 1996), and various political attitudes (Borg 1995; Kinder and Sanders 1996;
Miller and Shanks 1996, for example).
Focusing on research in political science, scholarship finds links between values and
various policy attitudes (Feldman 1988; Peffley and Hurwitz 1985; Pollock, Lilie and
Vittes 1993), prejudice and tolerance (Biernat et al. 1996; Kinder and Sanders 1996),
political participation (Borg 1995; Gundelach 1995), and evaluations of candidates and
parties (Feldman 1988; Knutsen 1995; Miller and Shanks 1996). Most research, though,
focuses on only one or two values at a time. For example, research finds a strong relationship between how individuals value equality and social welfare attitudes (Feldman and
Steenbergen 2001; Kluegel and Smith 1986), racial attitudes (Kinder and Sanders 1996;
Sears, Henry and Kosterman 2000), and candidate evaluations (Miller and Shanks 1996).
While many studies find direct relationships between values, attitudes, and behaviors, a number of studies find the relationship more complicated. Ajzen and Fishbein
(1977), for example, argue that values are directly related to attitudes, and attitudes are
only one factor among many that influence behavior. Scholarship argues that values are
less directly implicated in behavior for several reasons (Schwartz 1996). Among these
reasons are that values may compete with relevant norms in making decisions regarding
appropriate behavior (Bardi and Schwartz 2003), situational forces can overwhelm values
with respect to action (Maio et al. 2001), and the “right” values may not be activated in
the decision making process (Bardi and Schwartz 2003). Additional scholarship argues
further, that only certain types of attitudes –“value expressive” attitudes, as opposed to
utilitarian attitudes– are related to values (Maio and Olsen 1994). Even under perfect
circumstances –when norms are not conflicting with values, when the right values are
activated, and when the “right” values are value expressive– values may not correlate
strongly with behavior: individuals must be politically sophisticated enough to relate
politically relevant values to attitudes Zaller (1991, 1992).

77

This chapter takes previous literature into account in measuring the influence of values on various attitudes and behaviors. Using a series of ordered logit models, I measure
the effects of value system ordering on attitudes regarding the economy, foreign policy,
national security, and race relations. Together, these attitudes provide a composite measure of issue-related support for the president. Then, moving the analysis from attitudes
to behaviors, I use a nested logit model to test the effects of value system ordering on
the choice to turnout during the 2008 presidential election and vote choice.

5.2

How Values Affect Attitudes
Though countless pieces of scholarship document the relationship between values and

attitudes, the theory remains underdeveloped (Feldman 2003). Integrating past research
into a comprehensive theory, Schwartz (2004) suggests that values affect attitudes (and
behavior) through a four-step sequential process (Hitlin and Piliavin 2004): (1) values are
activated (Verplanken and Holland 2002); (2) the activated values lead to the privileging of some actions over others (Feather 1992); (3) values then influence interpretation,
attention, and perception in situations; (4) and values then influence the planning of
action.
In other words, not every situation requires one to call all values into consideration.
Any given situation may “activate” only a select number of values, and further, not every
individual will make the connection between the situational circumstances and relevant
values. Individuals that are more capable of making the connections are more likely to
complete the four-step process of linking values to behavior. Upon values being activated
and the individual in question making the connection, the individual then makes value
choices. Values at the top of the individual’s hierarchy influence how the individual sees
and analyzes each part of the situation, and based on this perception and analysis, the
individual plans behavior. In these plans, values compete with other factors (e.g. norms,

78

cost-benefit analysis, culture, etc.) in making a final decision on the appropriate behavior
and action.

5.3

Data
I again use the 2010 TESS data1 , which combines a survey experiment with more

traditional questions often seen in surveys of American public opinion that concern political preferences (n = 1264). In addition to the experimental items (discussed in the
third chapter), the dataset includes measures of issue-based presidential approval on the
economy, race relation, foreign policy, and national security. Data was also collected on
turnout and vote choice.
Measures of issue-based presidential approval are binary and coded such that 1 indicates a positive response and 0 indicates a negative response.2 I combine these four
variables into a ratings scale that measures an overall sense of presidential approval. Each
item improves the performance of the scale, and its overall reliability is high (Cronbach’s
α = 0.85).
The dependent variables for the second analysis are also binary. The decision to turn
out to vote is coded such that 1 indicates the respondent did vote and 0 if the respondent
did not vote. Vote choice in the 2008 presidential election is coded such that 1 indicates
a vote for Barack Obama and 0 indicates a vote for John McCain.
Independent variables considered in these analyses include age, race, gender, education, income, political party identification, and ideology.
1 Data collected by Time-sharing Experiments for the Social Sciences, NSF Grant

0818839, Jeremy Freese and Penny Visser, Principal Investigators.
2 The question of wording for each of these variables is: “Do you approve or disapprove of

the way Barack Obama is handling each of these issues?” Issues are then listed for the
respondent to indicate approval or disapproval.

79

5.4

Methods
To analyze the four-item scale of issue-related support for Barack Obama, I use a set

of ordered logit models (see equations 3.14 and 3.15 in chapter 3 for details). I do not
contain my analyses to one model specification, however. Because recent empirical work
suggests the causal path between values, party, and attitudes to be quite complicated,
and because I am unable to account for possible two-way causality between partisanship
and values, I run two sets of models. In the first, I account for both partisanship and
ideology. In the second, I treat both partisanship and ideology as intervening variables
–that values cause party identification and ideology, and party identification and ideology
then affect support for the president.
The dependent variables on the behavioral models are structured such that the decision for whom to vote is nested within the decision to vote in the first place. While it
is possible to treat the choice to abstain from the election as a third choice (in a nonsequential, non-nested manner), the relevant covariates are different across each process.
This so, I model both the decision to turn out to vote (the “selection” process) and the
decision for whom to vote with a multivariate probit model that accounts for the selection
process.
5.4.1

A Note on Multivariate Probit with Selection

I argue that the choice to vote is governed by the process described in equation 5.1
∗
zi = w γi + ui

80

(5.1)

∗
where one votes if zi > 0 and does not vote if otherwise. Vote choice is governed by the
process described in equation 5.2
∗
yi = x i β + i

(5.2)

∗
∗
where one votes for a particular candidate if yi > 0 given zi > 0 and the other candidate
∗
∗
if yi ≤ 0 given zi > 0. If, in the above models, the correlation between ui and i is
greater than zero, then the assumption that E[ |X] = 0 is is violated, and estimates in
equation 5.2 become inconsistent and biased (Greene 2003). In all, ignoring equation 5.1
amounts to leaving out a “shift” in equation 5.2, which, in essence, is omitted variable
bias. Taking the “shift” into account, the outcome equation becomes
∗
yi = xi β + βλ λi + vi

(5.3)

where βλ is ρσ . The multivariate probit model, which is used in this chapter, produces
unbiased, consistent estimates, and it can be used to determine whether the two processes
are, in fact, independent.
5.4.2

Measurement

I again make use of the split sample, half the sample having value system ordering
measured via importance ratings and half with rankings. Keeping in mind the results
presented in chapter 3 –the distributions of ratings being very different than the distribution of rankings– for each different model specification, I obtain estimates with both
the ratings and rankings subgroups. I then compare the fits of the models using likelihood ratio tests for nested models and the Schwarz Criterion (also called the Bayesian
Information Criterion, or BIC). The likelihood ratio test simply compares the fits of the
models using each models log-likelihood. The BIC is different from the LR test in that

81

the BIC penalizes for the addition of parameters to the model. Generally, lower BIC
scores indicate better model fit (e.g. the model with the more negative BIC fits the data
better). Difference greater than ten points can be considered very strong evidence for
superior model fit, differences between six and ten points can be considered strong evidence, differences between two and six points can be considered moderate evidence, and
difference of less than two points indicate weak evidence (Kass and Raftery 1995; Raftery
1995, 1999).

5.5

Formal Hypotheses
Most tests of theories are relatively straightforward. First, theory suggests that values

play a larger role in attitudes than in action. If so, the data should show values to have a
more pervasive influence in issue-related support for President Obama than in the choices
to vote and for whom to vote. Second, theory suggests that only “activated” values should
play a role in attitudes and behavior. If this is true, it should not be the case that all five
values show significant relationships with the dependent variables. But, the data should
also show that the inclusion of values into the models proves a useful addition. Last,
theory suggests that people more able to link values, attitudes, and behavior are more
likely to use their values in forming attitudes and making choices (Zaller 1991, 1992). This
theory can be captured in a series of multiplicative interaction terms between education
and value ranking/rating. If correct, values should have a greater effect in more highly
educated respondents.

5.6

Results Pertaining to Policy-Based Approval for the President
A look at bivariate relationships between value rankings/ratings and policy attitudes

(see tables 5.1 and 5.2) shows that the positioning of equality and moral traditionalism
seem to have the strongest relationships with each policy attitude. Generally, higher
82

rankings/ratings of equality correspond to more approval for President Obama’s policies,
and higher rankings/ratings of moral traditionalism correspond to more disapproval.
These results lead to several interesting thoughts and questions. First, it is curious
that the a particular value’s relationship with each policy attitude is almost constant.
Questions on different policies do not seem to “activate” different values. Rather, all
policy attitudes seem to be driven by the same two values. Second, it is not freedom
and equality that drive policy attitudes, as some literature suggests (e.g. Feldman 1988).
Nor is it freedom and law & order (Davis 2007). Instead, the two values are equality and
moral traditionalism. This is somewhat unexpected. But, before drawing conclusions, it
is necessary to go beyond simple bivariate relationships.
Results of the full models, with interactions terms, are displayed in tables 5.3 and 5.4.
Of interest in these models are the coefficients on the interaction terms and the results
of the likelihood ratio tests that measure the improved fit of the full models against the
restricted models. Looking first at the ratings group (table 5.3), none of the coefficients on
the interaction terms are significant. The likelihood ratio test on nested models suggests
that the addition of the three variables does not improve model fit (χ2 = 1.67, p = .64),
and the BIC scores suggest the restricted model fits the data much better than the
unrestricted model. The results on the models of the rankings group (table 5.4 tell a
bit of a different story. The coefficient on the interaction term composed of equality
rank and education is positive and significant, suggesting that the rank of equality affects
presidential approval more in respondents with more formal education. The likelihood
ratio test on the nested models suggests some mild improvement in the fit of the full
model (χ2 = 6.95, p = .07). The BIC scores, however, indicate that the restricted model
fits the data better.
Trimmed models of policy-based presidential approval are presented in table 5.5. The
models, though slightly different across measurement groups, tell the same basic story.

83

First, values matter. Considering first the model of the rating group, all three values have
significant explanatory power. Increased importance in freedom and moral traditionalism correspond with decreased support for the president, and increased importance in
equality corresponds with increased support for the president. In the model of the ranking group, increased importance of freedom and moral traditionalism correspond with
decreased support of the president (though the coefficient on freedom is not significant).
And, though the coefficient on equality is not significant in itself, the coefficient on the
interaction term is. This (partially) supports Zaller’s findings regarding the conditional
nature of the relationship of value system ordering on attitudes.3 Turning to the fit statistics of each model, it is quite apparent that the models of the rankings group fit the data
better than the models of the ratings group. Finally, turning attention to the coefficients
on the control variables, education has a significant impact on issue-based support. In
the restricted models, the coefficients suggest that more educated respondents tend to
hold more support for the president than less educated respondents. In the full models,
the same pattern holds among members of the ratings group (though the coefficient is
not statistically significant). The coefficient on education, in the model of the rankings
group, is not statistically significant, though the coefficient on the interaction term is.
Party identification and ideology have very strong effects on policy-based approval, but
it is not altogether clear that party and ideology are not intervening variables that are
actually caused by value system ordering. The results presented in table 5.6 take this
possibility into account. These models are estimated with only values and demographic
variables on the right-hand side, and the estimates represent an “upper bound” of the
effect of values on issue-based presidential support. The results are consistent with those
presented in table 5.5, but the effects are more pronounced.
3 Models with and without values were run on for the rankings group. The BIC score on

the restricted model (e.g. values variables omitted) was -2349.07. The BIC score on the
full model was -2336.39. The LR test on nested models gave a χ2 of 12.99 (4 degrees of
freedom) and a p-value of 0.01.
84

Tables 5.7 through 5.12 show changes in predicted probabilities in levels of support for
the president based on changes in the ratings/rankings of freedom, equality, and moral
traditionalism. The story told by the changes in the ratings group (tables 5.7 and 5.8)
is obvious: values matter. Looking first at the predictions from the full model (table
5.7), someone holding a “liberal” set of values (e.g. high on equality, low on freedom
and moral traditionalism) is much more likely to have an approval score of 3 or 4 than
someone with a “conservative” value set (e.g. low on equality, high on freedom and moral
traditionalism). In fact, the model predicts that someone with liberal values is almost
three times more likely to approve of how President Obama handled at least three out
of four issues (65% to 27%). Someone with conservative values, on the other hand, is
more than twice as likely to register an approval score of 0 or 1 than someone with liberal
values. Differences are much more pronounced in predictions from the restricted model
(table 5.8). Here, one with liberal values is almost six times more likely to register an
approval score of 3 or 4 than someone with conservative values, and the difference is as
stark on the other end of the scale.
The rankings data tells the same basic story as the ratings data, but with some
additional complexity. Recall that the set of multiplicative interaction terms added explanatory power to models of issue-based support for the president when members of the
rankings group were under consideration. This so, these models take into account the
interaction between education and equality (the single statistically significant interaction
term in these models, as displayed in tables 5.4, 5.5, and 5.6. Turning attention to the
differences in predicted probabilities in table 5.9, it appears as though values make no
difference in attitudes, but this model holds education at its minimum (i.e. less than
high school). The model in table 5.10 holds education at its maximum (i.e. college graduate), and the degree to which value differences affect issue-based presidential approval
becomes clear. Respondents with liberal values, on average, are more than twice as likely

85

as respondents with conservative values to highly approve of how the president dealt with
the four relevant issues, and respondents with conservative values are about three times
as likely as those with liberal values to strongly disapprove of the way the president has
handled the relevant issues. The differences in the predicted probabilities reported in tables 5.11 and 5.12 show the same basic pattern as those in the previous two tables. The
results here, however, are a bit more extreme given they are “upper bound” estimates
(party and ideology are not included in the models).
To sum up, the models presented in this section tell a very basic story about values
and political attitudes: values matter. Generally, those that hold freedom and moral
traditionalism highly and consider equality of only limited importance express limited
issue-based approval for the president. Those that value equality highly, on the other
hand, tend to express higher issue-based support for the president. And, while data from
each measurement tells the same basic story, the data from the rankings group is a bit
more complicated. Specifically, the degree to which values affect approval is conditioned
on education (especially the effect of equality). On average, those with more formal
education are more influenced by their values than those with less education. Finally,
models in which values are measured with rankings tend to fit the data considerably
better than models in which values are measured with ratings. Rankings, it seems,
more effectively capture value systems and can more accurately describe how they affect
attitudes.

5.7

Results Pertaining to Turnout and Vote Choice
Starting with bivariate correlations of values, turnout, and vote choice (see tables 5.13

and 5.14), it is apparent that values do not affect the choice to turnout. With respect
to vote choice, values seem much more important. Higher ratings/rankings of equality
correspond to voting for Obama, and higher rankings of moral traditionalism correspond

86

to voting for McCain. Before drawing conclusions, however, it is necessary to take all
other control variables into account
Estimates from the multivariate probit models with selection are shown in table 5.15.
Looking first at the turnout equation (the top half of the table), it seems as though
only economic security matters with respect to turnout, and this only holds for the
ratings group. This result is curious, however, because the sign of the coefficient is
negative: the results suggest that people are more likely to turn out if they care less about
economic security. The control variables tell a more interesting, and more predictable
story here. According to the results, older respondents, female respondents, well-educated
respondents, and respondents that have strong partisan and ideological attachments are
more likely to turn out.
Turning to the vote choice equation (the lower half of the table), the results suggest
that values play a more important role here, but only in the ratings group. Moral traditionalism is the only value that holds any explanatory power, but it pales in comparison
to the explanatory power of party and ideology. In the rankings group, only party and
ideology have explanatory power. The model performs as well with equality and moral
traditionalism omitted. Finally, the estimate of ρ, in both models, suggests that turnout
and vote choice are statistically independent processes that can be modeled with separate
binary choice models.
These models are reported in tables 5.16 and 5.17. The models of turnout tell much
the same story: values are not an important part in the process, and the usual suspects
(e.g. age, education, ideological strength, and strength with which one is attached to a
party) are important predictors of participation. Female and economic security (in the
opposite direction than expected) are also important predictors of participation in the
ratings model. The models of vote choice, in the ratings model, suggest that equality
and moral traditionalism are important in the explanation of vote choice (a high rating

87

of equality corresponds to a higher probability of voting for Obama while a high rating of
moral traditionalism corresponds to a lower probability of voting for Obama), but only in
the ratings group. The control variables, in both models, hold quite a bit of explanatory
power. Older respondents and those with more formal education, on average, tend to
vote for Obama. And, as expected, those that identify with the Democratic Party, and
those that identify as liberal, tend to vote for Obama as well.
An examination of the fit statistics for the probit model with selection and the independent logit models suggest, in contrast with the models of issue-based presidential
approval, that the models in which values are measured with ratings perform better than
those in which values are measured with rankings. In the probit model with selection (table 5.15), values add no explanatory power in the rankings model, and the trend continues
through both independent logit models. The values variables, in the ratings models, tell a
different story. Here, values do pull some weight. Economic security, though is a negative
effect, does affect the choice to turnout, and both equality and moral traditionalism affect
vote choice.

5.8

Conclusion and Discussion
In this chapter, I tested three hypotheses on the relationships between values, atti-

tudes, and behavior: (1) values are directly related to attitudes and distally related to
behavior (Ajzen and Fishbein 1977; Hitlin and Piliavin 2004; Schwartz 1996); (2) only
“activated” values influence attitudes and action (Bardi and Schwartz 2003); and (3) the
relationship between values, attitudes, and behavior is conditional on cognitive sophistication (Zaller 1991, 1992). In addition, I tested whether or not different measurement
techniques affect model fit.
Previous research finds support suggesting values strongly influence attitudes, but
only indirectly affect behavior (Ajzen and Fishbein 1977; Schwartz 1996). The results of

88

the analyses in this chapter corroborate previous findings. Equality, freedom, and moral
traditionalism affect issue-based support for the president, and they do so quite predictably. With demographics, party identification, and ideology held constant, changes
in the three relevant values have profound effects on the probability of strong or weak
support for Obama. The analyses of turnout and vote choice also corroborate the findings
of previous research. Values had only a minimal effect on the choice to vote (and, the
results are counterintuitive), and values had a limited effect on vote choice. Controlling
for political identification variables and some demographics, moral traditionalism and
equality have an independent effect on vote choice (but only in the ratings group). Looking over both analyses (the models of issue-based support, turnout, and vote choice), the
effects of values are clearer and, perhaps, a bit stronger in the models of approval. Values,
had virtually no effect on the choice to turn out, and limited influence on vote choice.
Values seem to be more closely related to attitudes than they are to behavior, at least in
a political context.
At first glance, the theory regarding “activated” values’ effect on attitudes (Bardi and
Schwartz 2003) does not seem to be supported by these analyses. The bivariate correlations in tables 5.1 and 5.2 show that two values, equality and moral traditionalism,
correspond with all four policy-attitudes analyzed in this paper. Further, there are policies that are designed to “activate” certain values (e.g. economic policy and economic
security, race relations and equality, terrorism and law & order), but it does not appear
to be the case that any of these survey items activated values other than equality and
moral traditionalism (and freedom, given the results of the various ordered logit models).
Upon further inspection, however, it may be the case that these survey items were not
actually measuring policy-attitudes. Rather, they are measuring policy-based support
for President Obama. The fact that the four-item scale has such high reliability (Cronbach’s α = 0.85) suggests that each of these items taps into how individuals evaluate the

89

president. If it is the case that each of the four policy-attitude items actually measure
different policy attitudes, then the these analyses show no support for the theory advance
in Bardi and Schwartz (2003). On the other hand, if these items all tap into an overall
evaluation of the president (and this seems more likely), then it may not be accurate to
call this a direct test of the theory.
Zaller’s theory regarding the link between values and attitudes, conditioned on sophistication, receives mixed support in these analyses. On one hand, the models in tables
5.3 and 5.4 suggest that the interaction terms do not improve model fit enough to justify their inclusion (especially in the rating group). On the other hand, the inclusion of
the interaction between equality and education, in models of the ranking group, change
substantive conclusions regarding the effect of the relative position of equality in individuals’ value systems on policy-based approval of the president. When the interaction
term is omitted from the model, results suggest that equality has no significant effect
on the dependent variable. When included, however, results suggest that respondents
with higher levels of educational attainment use equality in forming attitudes regarding
presidential approval. It should also be noted that measurement, here, matters. Value
system ordering, as measured by rankings, does interact with educational attainment.
Values, as measured by ratings, do not interact with educational attainment.
Finally, measurement technique affects model fit. Though the models of presidential
support (tables 5.5 and 5.6) told the same substantive stories (e.g. values matter in
attitudes on policy-based approval), model fit is consistently better when value system
ordering is measured with rankings. Models of behavior, on the other hand, tell a different
story. Here, models in which values are measured with ratings perform better. One
possible explanation is sampling error, but this is not definite. This discrepancy deserves
a closer look.
The results presented in this chapter suggest, at the very least, that several values

90

matter in the formation of important attitudes and behaviors. This is not to say that all
values affect evaluations and behavior all of the time –simply that values are an important
part of explanations of attitudes and some behaviors. Also, values may not affect each
individually equally throughout a given population. It is important for researchers to
consider the possibility that not all individuals can effectively connect their own values
to objects being evaluated. Effects of values on attitudes and behavior may be stronger in
individuals that are more able to make this important connection. Finally, the choice of
measurement technique is also very important –to the extent that different measurement
techniques produce different substantive conclusions. Future research on values must
consider the trade-offs that come with the choice to use on technique over the other.

91

Table 5.1: Correlations Between Value Rankings and Policy Attitudes
Economy
Freedom
-.05
Equality
.30
Econ. Security
.09
Law & Order
-.07
Moral Trad.
-.26

Race Rel.
-.05
.25
.13
-.07
-.26

Foreign Pol.
-.03
.29
.12
-.14
-.25

Terrorism
-.04
.25
.11
-.07
-.25

4-Item Scale
-.05
.32
.13
-.10
-.30

Value rankings are coded such that higher numbers indicate greater importance. Attitudes variables are binary, coded such that 1 indicates approval and 0 indicates disapproval.

Table 5.2: Correlations Between Value Ratings and Policy Attitudes
Economy
Freedom
-.05
Equality
.12
Econ. Security
.08
Law & Order
-.06
Moral Trad.
-.17

Race Rel.
-.06
.13
.03
-.02
-.12

Foreign Pol.
-.05
.17
.10
-.06
-.16

Terrorism
-.04
.10
.07
-.06
-.14

4-Item Scale
-.06
.16
.09
-.06
-.18

Value ratings are coded such that higher numbers indicate greater importance. Attitudes
variables are binary, coded such that 1 indicates approval and 0 indicates disapproval.

92

Table 5.3: Full Ordered Logit Model, Y = Issue Scale
Ratings Group

Freedom
Equality
Moral Trad.
Free. × Educ.
Eq. × Educ.
Moral Trad. × Educ.
Log
BIC
n

Full Model
-0.48
(.34)
0.58∗
(.31)
-0.12
(.28)
0.06
(.13)
-0.11
(.10)
-0.02
(.09)
-812.58
-2241.64
616

Restricted Model
-0.3∗∗
(.14)
0.27∗∗∗
(.10)
-0.16∗
(.09)
–
–
–
-813.41
-2259.25
616

∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01
LR Test of full model against restricted model yields a χ2 of 1.67 (p = .64).
Note: Control variables (race, education, party, and ideology) and cut points omitted
from table.

93

Table 5.4: Full Ordered Logit Model, Y = Issue Scale
Rankings Group

Freedom
Equality
Moral Trad.
Free. × Educ.
Eq. × Educ.
Moral Trad. × Educ.
Log
BIC
n

Full Model
-0.05
(.23)
-0.45∗∗
(.22)
-0.30
(.22)
-0.02
(.08)
0.18∗∗
(.08)
0.06
(.07)
-756.30
-2324.61
612

Restricted Model
-0.11
(.08)
0.06
(.08)
-0.14∗∗
(.07)
–
–
–
-759.77
-2336.91
612

∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01
LR Test of full model against restricted model yields a χ2 of 6.95 (p = .07).
Note: Control variables (race, education, party, and ideology) and cut points omitted
from table.

94

Table 5.5: Ordered Logit Model, Y = Issue Scale
Full Comparisons

Freedom
Equality
Moral Trad.
Eq. × Educ.
Non-white
Educ.
Party ID
Ideology
τ1
τ2
τ3
τ4
Log
BIC
n

Ratings
-0.30∗∗
(.14)
0.27∗∗∗
(.10)
-0.16∗
(.09)
–

Rankings
-0.10
(.08)
0.06
(.08)
-0.14∗∗
(.07)
–

0.34∗
(.19)
0.18∗∗
(.08)
0.41∗∗∗
(.05)
0.28∗∗∗
(.07)
1.00
1.82
2.48
3.37
-813.41
-2259.25
616

0.24
(.19)
0.33∗∗∗
(.08)
0.43∗∗∗
(.05)
0.41∗∗∗
(.07)
2.50
3.44
4.04
4.96
-759.77
-2336.91
612

∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01

95

Ratings
-0.33∗∗
(.14)
0.56∗∗
(.26)
-0.18∗∗
(.09)
-0.10
(.08)
0.36∗
(.19)
0.61
(.37)
0.41∗∗∗
(.05)
0.28∗∗∗
(.07)
2.02
2.85
3.50
4.40
-812.71
-2254.22
616

Rankings
-0.11
(.08)
-0.40∗
(.21)
-0.14∗
(.07)
0.17∗∗
(.07)
0.29
(.19)
-0.10
(.19)
0.43∗∗∗
(.05)
0.40∗∗∗
(.07)
1.31
2.27
2.86
3.80
-756.82
-2336.39
612

Table 5.6: Ordered Logit Model, Y = Issue Scale
Party & Ideology Omitted

Freedom
Equality
Moral Trad.
Eq. × Educ.
Non-white
Educ.
τ1
τ2
τ3
τ4
Log
BIC
n

Ratings
-0.30∗∗
(.14)
0.57∗∗∗
(.09)
-0.44∗∗∗
(.08)
–

Rankings
-0.17∗∗
(.07)
0.31∗∗∗
(.07)
-0.38∗∗∗
(.07)
–

0.83∗∗∗
(.18)
0.14∗
(.07)
-1.37
-0.69
-0.16
0.57
-893.92
-2111.08
616

0.61∗∗∗
(.18)
0.28∗∗∗
(.08)
-0.78
-0.03
0.41
1.11
-862.03
-2152.65
613

∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01

96

Ratings
-0.33∗∗
(.14)
0.85∗∗∗
(.25)
-0.46∗∗∗
(.09)
-0.10
(.08)
0.85∗∗∗
(.18)
0.56
(.35)
-0.38
0.30
0.83
1.56
893.15
-2106.18
616

Rankings
-0.17∗∗
(.07)
-0.27
(.19)
-0.37∗∗∗
(.07)
0.21∗∗∗
(.07)
0.66∗∗∗
(.18)
-0.26
(.18)
-2.23
-1.48
-1.04
-0.32
-856.78
-2156.71
613

Table 5.7: Predicted Probabilities from Full Model, Ratings Group
P r(Approval
P r(Approval
P r(Approval
P r(Approval
P r(Approval

=
=
=
=
=

4)
3)
2)
1)
0)

Liberal Values
0.43
0.22
0.13
0.11
0.11

Conservative Values
0.13
0.14
0.15
0.20
0.38

Difference 95% C.I.
0.30
[0.12, 0.48]
0.08
[0.04, 0.12]
-0.02
[-0.04, 0.02]
-0.09
[-0.14, -0.04]
-0.27
[-0.43, -0.11]

Cell entries in the first column indicate the probability that Y = y for a respondents with
liberal value preferences (equality set at 5, freedom and moral traditionalism set two
standard deviations below their means). Gender and education are set at their modal
categories (male and high school diploma). Ideology and partisanship are set at their
midpoints (neutral and independent). Cell entries in the second column indicate the
probability that Y = y for respondents with conservative value preferences (equality set
two standard deviations below its mean, freedom and moral traditionalism set at 5).
Other independent variables are held constant. Confidence intervals for the change in
predicted probabilities calculated using the Delta method.

Table 5.8: Pred. Probs. with Ideol. & Party Omitted, Ratings Group
P r(Approval
P r(Approval
P r(Approval
P r(Approval
P r(Approval

=
=
=
=
=

4)
3)
2)
1)
0)

Value Set 1
0.65
0.15
0.07
0.06
0.07

Value Set 2
0.07
0.07
0.08
0.14
0.65

Difference 95% C.I.
0.58
[0.42, 0.72]
0.08
[0.04, 0.11]
-0.01
[-0.03, 0.03]
-0.08
[-0.11, -0.04]
-0.58
[-0.70, -0.44]

Cell entries in the first column indicate the probability that Y = y for a respondents with
liberal value preferences (equality set at 5, freedom and moral traditionalism set two
standard deviations below their means). Gender and education are set at their modal
categories (male and high school diploma). Ideology and partisanship are omitted from
the model. Cell entries in the second column indicate the probability that Y = y for
respondents with conservative value preferences (equality set two standard deviations
below its mean, freedom and moral traditionalism set at 5). Other independent variables
are held constant. Confidence intervals for the change in predicted probabilities calculated
using the Delta method.

97

Table 5.9: Pred. Probs. from Full Model, Rankings Group
(Educ. = 1)
P r(Approval
P r(Approval
P r(Approval
P r(Approval
P r(Approval

=
=
=
=
=

4)
3)
2)
1)
0)

Value Set 1
0.12
0.13
0.13
0.23
0.39

Value Set 2
0.12
0.14
0.13
0.23
0.38

Difference
0.00
-0.01
0.00
0.00
0.01

95% C.I.
[-0.14, 0.13]
[-0.12, 0.11]
[-0.06, 0.06]
[-0.00, 0.00]
[-0.29, 0.32]

Cell entries in the first column indicate the probability that Y = y for a respondents
with liberal value preferences (equality set at 5, freedom and moral traditionalism set
two standard deviations below their means). Gender is set at its modal value (male),
and education is set at its minimum (less than high school). Ideology and partisanship
are set at their midpoints (neutral and independent). Cell entries in the second column
indicate the probability that Y = y for respondents with conservative value preferences
(equality set two standard deviations below its mean, freedom and moral traditionalism
set at 5). Other independent variables are held constant. Confidence intervals for the
change in predicted probabilities calculated using the Delta method.

Table 5.10: Pred. Probs. from Full Model, Rankings Group
(Educ. = 4)
P r(Approval
P r(Approval
P r(Approval
P r(Approval
P r(Approval

=
=
=
=
=

4)
3)
2)
1)
0)

Value Set 1
0.55
0.21
0.09
0.09
0.06

Value Set 2
0.15
0.16
0.14
0.23
0.33

Difference 95% C.I.
0.40
[0.19, 0.62]
0.05
[0.00, 0.10]
-0.05
[-0.08, -0.01]
-0.14
[-0.21, -0.08]
-0.26
[-0.41, -0.11]

Cell entries in the first column indicate the probability that Y = y for a respondents
with liberal value preferences (equality set at 5, freedom and moral traditionalism set
two standard deviations below their means). Gender is set at its modal value (male),
and education is set at its maximum (college degree). Ideology and partisanship are set
at their midpoints (neutral and independent). Cell entries in the second column indicate
the probability that Y = y for respondents with conservative value preferences (equality
set two standard deviations below its mean, freedom and moral traditionalism set at 5).
Other independent variables are held constant. Confidence intervals for the change in
predicted probabilities calculated using the Delta method.

98

Table 5.11: Pred. Probs. with Party and Ideol. Omitted, Rankings Group
(Educ.= 1)
P r(Approval
P r(Approval
P r(Approval
P r(Approval
P r(Approval

=
=
=
=
=

4)
3)
2)
1)
0)

Value Set 1
0.29
0.16
0.11
0.17
0.27

Value Set 2
0.07
0.06
0.06
0.14
0.66

Difference 95% C.I.
0.22
[0.03, 0.40]
0.10
[0.04, 0.16]
0.05
[0.02, 0.08]
0.03
[-0.02, 0.07]
-0.39
[-0.64, -0.14]

Cell entries in the first column indicate the probability that Y = y for a respondents
with liberal value preferences (equality set at 5, freedom and moral traditionalism set
two standard deviations below their means). Gender is set at its modal value (male),
and education is set at its minimum (less than high school). Ideology and partisanship
are omitted from the model. Cell entries in the second column indicate the probability
that Y = y for respondents with conservative value preferences (equality set two standard
deviations below its mean, freedom and moral traditionalism set at 5). Other independent
variables are held constant. Confidence intervals for the change in predicted probabilities
calculated using the Delta method.

Table 5.12: Pred. Probs. with Party and Ideol. Omitted, Rankings Group
(Educ. = 4)
P r(Approval
P r(Approval
P r(Approval
P r(Approval
P r(Approval

=
=
=
=
=

4)
3)
2)
1)
0)

Value Set 1
0.82
0.09
0.03
0.03
0.03

Value Set 2
0.06
0.06
0.05
0.13
0.69

Difference 95% C.I.
0.76
[0.64, 0.87]
0.03
[-0.01, 0.06]
-0.02
[-0.04, -0.00]
-0.10
[-0.13, -0.06]
-0.66
[-0.78, -0.54]

Cell entries in the first column indicate the probability that Y = y for a respondents
with liberal value preferences (equality set at 5, freedom and moral traditionalism set
two standard deviations below their means). Gender is set at its modal value (male), and
education is set at its maximum (college degree). Ideology and partisanship are omitted
from the model. Cell entries in the second column indicate the probability that Y = y
for respondents with conservative value preferences (equality set two standard deviations
below its mean, freedom and moral traditionalism set at 5). Other independent variables
are held constant. Confidence intervals for the change in predicted probabilities calculated
using the Delta method.

99

Table 5.13: Correlations Between Value Ratings and Turnout & Vote Choice
Turnout
Freedom
0.01
Equality
-0.02
Econ. Security
-0.07
Law & Order
0.08
Moral Trad.
0.01

Obama Vote
-0.06
0.15
0.09
-0.00
-0.21

McCain Vote
0.05
-0.19
-0.15
0.06
0.22

Value rankings are coded such that higher numbers indicate greater importance.
Turnout is a dummy variable coded 1 if the respondent voted, 0 if not.
Vote choice variables are dummy variables coded 1 if the respondent voted for the candidate, and 0 otherwise.

Table 5.14: Correlations Between Value Rankings and Turnout & Vote Choice

Freedom
Equality
Econ. Security
Law & Order
Moral Trad.

Turnout
0.02
-0.02
-0.08
0.04
0.05

Vote Obama
-0.03
0.31
0.10
-0.12
-0.24

Vote McCain
0.04
-0.34
-0.16
0.16
0.29

Value rankings are coded such that higher numbers indicate greater importance.
Turnout is a dummy variable coded 1 if the respondent voted, 0 if not.
Vote choice variables are dummy variables coded 1 if the respondent voted for the candidate, and 0 otherwise.

100

Table 5.15: Probit Selection Model of Vote Choice and Turnout
Ratings
Turnout
Econ. Security -0.26∗∗
Female
0.42∗∗∗
Educ.
0.59∗∗∗
Age
0.24∗∗∗
Party (folded) 0.24∗∗
Ideol. (folded) 0.16∗
Constant
-0.86
Vote Choice
Equality
Moral Trad.
Party
Ideol.
Constant
ρ
Log
Turnout n
Vote Choice n

(.11)
(.16)
(.08)
(.05)
(.09)
(.08)
(.60)

0.09
(.10)
∗∗∗ (.09)
-0.24
0.50∗∗∗ (.05)
0.27∗∗∗ (.07)
-2.29∗∗∗ (.50)
0.12
(.39)
-316.54
543
473

Rankings
-0.06
-0.01
0.17∗∗
0.15∗∗∗
0.22∗∗∗
0.16∗∗
-0.49

(.05)
(.13)
(.07)
(.04)
(.08)
(.07)
(.35)

0.10
(.07)
0.01
(.06)
0.41∗∗∗ (.05)
0.16∗∗
(.07)
∗∗∗ (.37)
-3.38
0.44
(.30)
-392.16
560
462

∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01
Dependent variable for turnout model is a dummy variable coded 1 if the respondent
voted and 0 otherwise.
Dependent variable for vote choice model is coded 1 if respondent voted for Obama and
0 if otherwise.
LR Test of a model without values vs. full model in ratings group yields a χ2 of 12.77
(p = .01).
LR Test of a model without values vs. full model in rankings group yields a χ2 of 3.40
(p = .33).

101

Table 5.16: Logit Models of Turnout
Ratings
Rankings
∗∗ (.21)
Econ. Security -0.51
-0.12
(.09)
∗∗ (.30)
Female
0.69
-0.05
(.23)
Educ.
1.14∗∗∗ (.17) 0.33∗∗∗ (.12)
Age
0.43∗∗∗ (.10) 0.27∗∗∗ (.07)
Party (folded)
0.41∗∗ (.17) 0.43∗∗∗ (.14)
Ideol. (folded)
0.35∗
(.17)
0.27∗∗ (.13)
Constant
-1.56
(1.09)
-1.03
(.61)
Log
-159.61
-236.89
BIC
-3092.48
-3025.57
n
548
560
∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01

Table 5.17: Logit Models of Vote Choice
Ratings
Equality
0.25∗
(.14)
∗∗∗ (.12)
Moral Trad. -0.34
Party
0.70∗∗∗ (.07)
Ideol.
0.23∗∗∗ (.09)
Age
0.24∗∗∗ (.07)
Educ.
0.66∗∗∗ (.11)
Constant
-6.95∗∗∗ (.83)
Log
-284.37
BIC
-3469.25
n
633
∗ p < .1, ∗∗ p < .05, ∗∗∗ p < .01

102

Rankings
0.11
(.09)
-0.06
(.09)
0.62∗∗∗ (.07)
0.34∗∗∗ (.09)
0.19∗∗∗ (.07)
0.39∗∗∗ (.11)
-6.63∗∗∗ (.70)
-286.37
-3413.21
626

Chapter 6

Conclusions

This dissertation focuses on three central questions: (1) Are value systems stable
through various informational contexts? (2) How do values relate to political attitudes
and behaviors? And (3), does choice of measurement technique affect empirical results?
In this chapter, I summarize the results presented in the previous three chapters, I discuss
implications and where these results fit in the existing body of literature, and I discuss
avenues for future research.

6.1

Chapter 3: Transitivity, Stability, and Measurement
The third chapter tests two related hypotheses regarding transitivity in value choices

and the stability of value preferences through systematically varied contexts. Transitivity
is a test of value structure –checking whether value choices exhibit structure, or whether
they are random. Recall that I defined a respondent to have transitive value choices
with respect to a particular triad of values if, for example, value a is chosen over value
b, value b is chosen over value c, and value a is chosen over value c. The results on
transitivity, presented in table 3.2, show that approximately 90 percent of respondents
exhibit transitive preferences on each triad, and nearly 75 percent of respondents exhibited
transitive preferences on all ten triads. In short, this is strong evidence that members of
the public have clear preferences between values and that random fluctuations in error
did not drive responses; rather, systematic components seem to drive choice.
103

The positive results on transitivity make primes in the survey experiment meaningful.
Given that systematic components drive value choice, the experiment sheds light on
whether the small changes in context (e.g. the primes) can manipulate these systematic
components in a meaningful way. And, the results presented in the difference-in-means
tests, the ordered logit models, and the rank-ordered logit models suggest no concrete
answer. Rather, it seems as though equality is susceptible to changes in context and
economic security is not. The reason for the mixed results, at this point, is unknown. It
may be that a large number of Americans are ambivalent about equality (Feldman 1988;
Feldman and Zaller 1992), that people do not have a clear idea of the personal importance
they attach to equality, or that the systematic components of equality, economic security,
and moral traditionalism (e.g. the values that seemed to “move” when respondents were
treated with the equality prime) are all about equal, not easily distinguishable, and thus,
unstable. In turn, economic security may appear particularly stable because of the current
economic conditions in the United States. The economic recession has been among the
leading stories in the national news for several years. This so, people have been able to
process ample information about economic security, they know where economic security
stands as a priority, and thus, they have a clear picture of the personal importance of
economic security relative to other values.
My answer to the question “do values transcend specific situations?,” based on the
results presented in the third chapter, is “some do, some don’t.” What does this mean to
values as they relate to attitudes? Recall, traditional theory requires values to transcend
specific situations if they are to provide a general framework for evaluation (Rokeach
1973; Schwartz 1996), and new research suggests that various factors cause instability in
value systems (see Goren, Federico and Kittilson 2009; Maio and Olson 1998; Seligman
and Katz 1996; Tetlock, Peterson and Lerner 1996, for example). The results presented
in chapter three suggest that the truth about value system dynamics lies somewhere

104

between these two camps. On the one hand, it is clear that small changes in context
affect the relative ordering of certain values. On the other, certain values are unaffected
by changes in context. The degree to which these dynamics feed into the debate on how
values affect attitudes depends on the importance of “movable” values. And, the results
presented in the fifth chapter suggest that one unstable value, equality, is quite important
in attitudes, but two (apparently) stable values are important as well: freedom and moral
traditionalism.
Given that only a limited number of values had any effect on attitudes, that equality
is one of these “important” values, and that equality seems susceptible to small changes
in context, the question becomes “why is equality unstable?” Does it have to do with
education? Is measurement error the cause? Perhaps some people think about equality
in one light while others think of it in a different light? Or, perhaps equality, as people
think about it, is a combination of several values? Future research might consider whether
equality is in fact a “base” value, or whether it has a more complex composition.
Turning attention to measurement, the main question concerned whether the choice
of ratings or rankings affect observed value system stability, and the results say “no.”
Equality, regardless of whether the measurement technique was ratings or rankings, appeared unstable while economic security appeared stable. So, while ratings and rankings
produce drastically different results in terms of value systems and relative ordering (see
figures 3.1, 3.2, and 3.3), the different measurement strategies produce similar results
with respect to observed stability.

6.2

Chapter 4: The Effects of Contextual Salience Over Time
The central question in the fourth chapter focused again on the stability of values.

Here, however, focus was on aggregate level values over time rather than individuals
participating in an experiment. I compared the rank-ordered preferences of four values

105

(liberty, equality, economic security, and social order) across three different time points
(1994, 2003, and 2005). Briefly, the theory under consideration (e.g. traditional psychological theory) stated that, even at the societal level, values should be relatively stable.
Movement between values may occur after major events, but values should not move at
random.
The results (descriptive statistics presented in figure 4.1 and table 4.1, OLS regressions
presented in tables 4.2, 4.3, and 4.4, and rank-ordered logit presented in table 4.5) show
relatively strong support for the theory. Respondents in all three datasets rank liberty
highly and equality toward the bottom of the hierarchy. Social order is the biggest
“mover” between 1994 and 2003, likely due to the events of 9/11, and both equality and
economic security move down in rankings (though economic security much more than
equality). And, economic security and social order again “trade” places between 2003
and 2005 (though it is important to note that equality moves too). The takeaway point
here is that contextual salience matters (denoted c in the models and represented in the
regressions by the “year” variables). But, the effects of contextual salience diminish over
time.
These results support traditional psychological theory of values, to a certain extent.
The results of this chapter, like the results presented in the third chapter, show that
values are not entirely static. That said, it does seem to be the case that when the
relative ordering of values changes, it does so predictably. The implications are not
completely clear, but there are several possibilities. First, these results may support
value pluralism theory (Tetlock 1986; Tetlock, Peterson and Lerner 1996), but rather
than having different value structures for different social situations, people may have
different value structures when confronted with different problems (e.g. they may have
certain value preferences for “normal” circumstances, a different set of preferences when
security and social order is in question, and yet another for when economic security is in

106

question). That is, peoples’ value systems may, in some sense, be reactionary. A second
possibility is that values may only guide behavior under “normal” conditions, and different
frameworks may take over under “abnormal” conditions. In any case, it does seem that
values are relatively stable, but that contextual salience and the state of a nation (or,
perhaps more accurately, the psychological state of people in a nation) adds a layer of
complexity to how values structure attitudes.
Perhaps the most obvious weakness in this chapter is the limited number of data
points. Additional data points would so much to increase the confidence of the conclusions made here. Additional data would make each value’s trends clearer. It would also
shed light on the lasting effects of contextual salience and value system dynamics in the
aggregate.

6.3

Chapter 5: Values, Attitudes, the Vote, and Measurement
The fifth chapter focused on the relationships between values, attitudes, and behav-

iors. There were three main hypotheses in the chapter: (1) values have a direct effect on
attitudes; (2) values have a limited effect on behavior; and (3) the relationship between
values and attitudes (and behavior) is conditional on education. I also examined the
effects of measurement on model fit.
The first major finding (seen in tables 5.1 and 5.2 has to do with the lack of variation
in “activated values” on different issues. By and large, only two values, equality and
moral traditionalism, correlate with each of the four issues in the survey (the economy,
race relations, foreign policy, terrorism). Surprisingly, neither economic security nor
law & order held strong correlations with the state of the economy, foreign policy, or
terrorism. The four items scaled well (Cronbach’s α = 0.85) and formed a good measure
of issue-based support for president Obama.
Tests of the first hypothesis can best be seen in the models presented in tables 5.5

107

and 5.6. In short, the results shown represent the estimated lower and upper bounds,
respectively, of the effects of values on issue-related support for the president. In the
ratings models, freedom, equality, and moral traditionalism all have significant effects on
the dependent variable in both the full and restricted models. In the rankings models,
moral traditionalism and equality (conditionally), affect the dependent variable in the
full model, and all three values have significant effects in the restricted model. All this
is fairly strong evidence that values, even when the independent effects of party and
ideology are held constant, influence attitudes.
The clearest tests of the second hypothesis are shown in tables 5.16 and 5.17. It does
not seem as though values have any effect on the choice to turnout (the one significant
coefficient on a value is in the unexpected direction, suggesting those that are more
concerned with economic security are less likely to vote). Values and vote choice have
a far more predictable relationship, but only when measured by ratings. According
to the ratings model, the same two values that best predicted issue-based presidential
support (equality and moral traditionalism) also predict vote choice. The curious result,
here, is the rankings model: values have no effect on vote choice. The contrast between
measurement techniques is stark and deserves a closer look.
Tests of the third hypothesis can be found in the third and fourth columns of tables
5.5 and 5.6. Interestingly, there is a disparity between measurement groups again. The
multiplicative interaction term does not help the fit of the ratings model. They do,
however, improve the fit of the rankings models a bit (according to the likelihood ratio
test). More importantly, the interaction term makes the results intelligible. With no
education, high rankings of equality correspond to low approval for president Obama.
With increasing education, high rankings of equality correspond to high approval for
Obama. These results not only provide support for the third hypothesis, but they make
a compelling argument that this particular interaction term is integral to understanding

108

values and attitudes.
The examination of fit statistics across models estimated with ratings and rankings
data told an interesting but unclear story. On the one hand, models of the rankings group
fit the data much better than the ratings group with respect to issue-based approval. On
the other, models of the ratings group fit the better data when it came to behavior
(especially vote choice). Again, the reason for this discrepancy is unclear and deserves a
closer look.
A potential weakness of one of the analyses in this chapter has to do with question
wording. Questions concerning issue attitudes were, in part, referenda on Obama’s performance on important issues facing the nation. This so, all four items measured the same
basic concept (i.e. issue-based approval for the president). Future research may consider
an alternative question wording in which the president goes unmentioned. These types
of questions may “activate” different values.

6.4

Synthesis
On the whole, the analyses in this dissertation make several important contributions

to value measurement and our understanding of value system dynamics. With respect
to measurement, results show two things: first, rankings do bring out the “comparative
and competitive” aspect of values, and this is born out in the data (see figures 3.1, 3.2,
and 3.3); second, despite differences in distributions of dependent variables, observed
value system stability is not dependent upon measurement technique. The discrepancies
between scholarship that upholds traditional psychological theory on values and that
which argues traditional theory must be reformed, apparently, run deeper than choice of
measurement technique.
With respect to value system dynamics, the story is a bit more complicated. Recall, traditional theory assumes (and requires) values and value system ordering to be

109

“transsituational” (Schwartz 1996), but new research finds that finds value systems to
be dynamic (e.g. Goren 2005; Goren, Federico and Kittilson 2009; Seligman and Katz
1996; Tetlock 1986), calls the importance of values into question. And, among the major
goals of this dissertation was to bridge this gap and advance values research beyond this
apparent impasse. The results presented throughout the pages of this dissertation do not
clearly favor one camp or the other. Rather, it appears as though certain values are more
susceptible to changes in context than others. Equality, in particular, seems especially
susceptible to small changes in context.
On its face, this result is strong evidence against the assumption in traditional psychological theory on values. But, to make the claim that values cannot provide a general
framework from which people evaluate objects in the world seems premature. Given that
only the equality prime had a significant effect on value choices (remember, the economic
security prime had no effect), and given that the equality prime only affected the relative
ordering of two (perhaps three) values, it may be the case that equality is an exception
and that “stable” values form a framework for evaluation. Or, it may be the case that the
(in)stability of equality is conditional on something other than education. Respondents
with differing ideological or partisan identification may be more or less receptive to the
equality prime than others. Or, as said before, it may be the case that equality is not a
base value: perhaps equality is a combination of several base values, and perhaps these
values are stable. One thing is clear, the assumption on the “transsituational” nature of
values is not completely true. To determine the degree to which it is true or false requires
more research.
The study in the fourth chapter also spoke to the gap between traditional psychological
theory on values and new research regarding the assumption of values’ “transsituational”
nature. Like the results presented in chapter three, the results presented in chapter four
also point to a gray area between traditional theory and new research. On one hand,

110

values should remain relatively stable over time if they are to guide attitudes and choice.
On the other, people should, in some sense, react to major contextual changes around
them. And, the results presented here confirm this. Values are, in some sense, reactionary
to major changes in context. The spike in the importance of social order after 9/11
and the subsequent decline in economic security and equality –then the “switch” back to
“normal” value ordering in 2005– makes sense. Social order should move to the top of value
hierarchies after a major attack on U.S. soil. And, as the threat declines, as the contextual
salience decays, individuals should again resume their “normal” value system orderings.
These results, seem to corroborate well with many of Inglehart’s studies on value change
in Europe. The main difference is that, in Europe, value change occurs slowly over time
with economic well-being. The value change observed in these datasets presented here
occurs quickly, but, the situational context changed quickly as well. According to these
results –that values may change, but they do so predictably– values are stable enough to
provide a framework for evaluation.
Turning attention from “can values structure attitudes?” to “do values structure
attitudes?” and chapter five, one set of results bolsters current ideas on the relationship
between values, attitudes, and behavior: values have a significant effect on attitudes, but
only a limited effect on behavior. The measurement aspect of chapter five points to no
clear conclusion regarding whether rankings or ratings are preferable. The main difference
between the two methods is the observed conditional relationship between values and
attitudes (conditional on education) that was seen in the rankings group. Beyond this
result, because fit statistics of rankings models were better in models of attitudes and
worse in models of behavior, it is unclear which measurement technique is “better.”
All in all, this dissertation set out to bridge the gap between traditional psychological
values theory and new research, and perhaps it has, but not in a black and white manner.
It is likely that values can structure attitudes, but not as traditional theory suggests they

111

do. Not all values are stable, nor are they all important in attitude formation and decision
making. The question is, are values that are important in attitude formation and choice
stable? And, the research presented in this dissertation points toward equality. The
results in this dissertation suggest equality is an important value in attitude formation,
but also that it is susceptible to changes in context. Unless equality is somehow an
exception (that it is not a “base” value, or that the models presented in the various
chapters are misspecified) then the conclusion must be that values do not provide the
most general framework possible for evaluation. However, further research on the nature
of equality is needed before reaching this conclusion.

6.5

Avenues for Future Research
In large part, the conclusions reached in this dissertation will be bolstered with further

research on whether or not equality is some kind of exception. Future studies might
consider a more expansive battery of primes (perhaps have six experimental groups: a
control group and a treatment for each individual value under consideration). If several
values are susceptible to priming effects, this is further evidence that values do not provide
a general framework for evaluation. If, on the other hand, only equality is susceptible to
priming effects, it may be an exception.
Additional research on the nature of equality might also prove beneficial to our understanding of values. In particular, a study might consider testing hypotheses concerning
the composition of “equality.” Is equality itself a “base” value? Or does the term “equality”
cover various, perhaps only tangentially related concepts like racial equality, economic
equality, and equality of opportunity? If, for example, racial equality, economic equality,
and equality of opportunity are all separate values, survey instruments might be more
valuable in asking about specific types of equality rather than “equality” in a general
sense.

112

Future research might also consider testing whether additional information in valueladen primes (e.g. a partisan or an ideological cue) affects the rankings or ratings of
“stable” values. This type of study might also consider manipulating measurement technique, as presented in this dissertation. Results from a study like this would help settle
an ongoing debate regarding the “primacy” of values. In short, if the additional information “moves” values, the that information, essentially, causes value placement and thus,
is more general a framework for evaluation.
Finally, future research on preference (e.g. value preference, policy preference, candidate preference, etc.) might consider rankings rather than either binary choice or the
usual “approve/disapprove” scale. Rankings, again, force people to choose between two
potentially desirable options that cannot both occupy “most preferred” territory. This
type of measure may prove useful in shedding light on issues in racial policy and voting
behavior in additional to value preference.

113

Bibliography

114

Bibliography

Abramson, Paul R. and Ronald Inglehart. 1987. “Generational Replacement and the
Future of Post-Materialist Values.” Journal of Politics 49(1):231–241.
Abramson, Paul R. and Ronald Inglehart. 1995. Value Change in a Global Perspective.
Ann Arbor, MI: University of Michigan Press.
Ajzen, Icek and Martin Fishbein. 1977. “Attitude-Behavior Relations: A Theoretical
Analysis and Review of Empirical Research.” Psychological Bulletin 84(5):888–918.
Allison, Paul D. and Nicholas A. Christakis. 1994. “Logit Models for Sets of Ranked
Items.” Sociological Methodology 24:199–228.
Alvarez, R. Michael and John Brehm. 1995. “American Ambivalence Towards Abortion Policies: Development of a Heteroskedastic Probit Model of Competing Values.”
American Journal of Political Science 39(4):1055–1082.
Alvarez, R. Michael and John Brehm. 1997. “Are Americans Ambivalent Towards Racial
Policies?” American Journal of Political Science 41(2):345–374.
Alvarez, R. Michael and John Brehm. 1998. “Speaking in Two Voices: American Equivocation About the Internal Revenue Service.” American Journal of Political Science
42(2):418–452.
Alwin, Duane F. and Jon A. Krosnick. 1985. “The Measurement of Values in Surveys: A
Comparison of Ratings and Rankings.” The Public Opinion Quarterly 49(4):535–552.
Ball-Rokeach, Sandra J., Milton Rokeach and Joel W. Grube. 1984. The Great American
Values Test: Influencing Behavior and Belief Through Television. New York: Free
Press.
Bardi, Anat and Shalom H. Schwartz. 2003. “Values and Behavior: Strength and Structure of Relations.” Personality and Social Psychology Bulletin 29(10):1207–1220.
Bernard, Mark M., Gregory R. Maio and James M. Olson. 2003. “Effects of Introspection
About Reasons for Values: Extending Research on Values-as-Truisms.” Social Cognition 21(1):1–25.

115

Biernat, Monica, Theresa K. Vescio, Shelley A. Theno and Christian S. Crandall. 1996.
Values and Prejudice: Toward Understanding the Impact of American Values on Outgroup Attitudes. In The Psychology of Values: The Ontario Symposium, Volume 8, ed.
Clive Seligman, James M. Olson and Mark P. Zanna. Mahwah, NJ: Lawrence Erlbaum
Associates pp. 153–189.
Blackwell, Matthew, Stefano Iacus, Gary King and Giuseppe Porro. 2009. “cem: Coarsened Exact Matching in Stata.” The Stata Journal 9(4):524–546.
Borg, Ingwer and Patrick Groenen. 2005. Modern Multidimensional Scaling: Theory and
Applications, 2nd Edition. New York: Springer.
Borg, Sami. 1995. Electoral Participation. In The Impact of Values, ed. Jan W. Van Deth
and Elinor Scarbrough. New York: Oxford University Press pp. 441–460.
Bossman, David M. 1991. “Cross-Cultural Values for a Pluralistic Core Curriculum.” The
Journal of Higher Education 62(6):661–681.
Braithwaite, Valerie. 1982. “The Sturcture of Social Values: Validation of Rokeach’s
Two-value Model.” British Journal of Social Psychology 21:201–211.
Braithwaite, Valerie. 1994.
“Beyond Rokeach’s Equality-Freedom Model: TwoDimensional Values in a One-Dimensional World.” Journal of Social Issues 50(4):67–94.
Braithwaite, Valerie. 1997. “Harmony and Security Value Orientations in Political Evaluation.” Personality and Social Psychology Bulletin 23(4):401–414.
Braithwaite, Valerie A. and H. G. Law. 1985. “Structure of Human Values: Testing the
Adequacy of the Rokeach Values Survey.” Journal of Personality and Social Psychology
49(1):250–263.
Braithwaite, Valerie, Toni Makkai and Yvonne Pittelko. 1996. “Inglehart’s MaterialismPostmaterialism Concept: Clarifying the Dimensionality Debate Through Rokeach’s
Model of Social Values.” Journal of Applied Social Psychology 26(17):1536–1555.
Brewer, Paul R. 2003. “The Shifting Foundations of Public Opinion about Gay Rights.”
Journal of Politics 65(4):1208–1220.
Campbell, Angus, Philip E. Converse, Warren E. Miller and Donald E. Stokes. 1960. The
American Voter. New York: Wiley.
Chapman, Randall G. and Richard Staelin. 1982. “Exploiting Rank Ordered Choice Set
Data within the Stochastic Utility Model.” Journal of Marketing Research 19(3):288–
301.
Cochran, William G. and Donald B. Rubin. 1973. “Controlling Bias in Observational
Studies: A Review.” SankhyÄĄ: The Indian Journal of Statistics, Series A 35(4):pp.
417–446.
116

Conover, Pamela and Stanley Feldman. 1981. “The Origins and Meaning of LiberalConservative Self-Identification.” American Journal of Political Science 29(4):617–645.
Converse, Philip E. 1964. The Nature of Belief Systems in Mass Publics. In Ideology and
Discontent, ed. David E. Apter. New York: Free Press of Glencoe.
Davidov, Eldad, Peter Schmidt and Shalom H. Schwartz. 2008. “Bringing Values Back
In: The Adequacy of the European Social Survey to Measure Values in 20 Countries.”
Public Opinion Quarterly 72(3):420–445.
Davis, Darren W. 2007. Negative Liberty: Public Opinion and the Terrorist Attacks on
America. New York: Sage.
Davis, Darren W. and Christian Davenport. 1999. “Assessing the Validity of the Postmaterialism Index.” American Political Science Review 93(3):649–664.
Devine, Donald J. 1972. The Political Culture of the United States. Boston: Little &
Brown.
Dow, Jay K. and James W. Endersby. 2004. “Multinomial Probit and Multinomial Logit:
A Comparison of Choice Models for Voting Research.” Electoral Studies 23(1):107–122.
Downs, Anthony. 1957. An Economic Theory of Democracy. New York: Harper & Bros.
Duch, Raymond M. and Michaell A. Taylor. 1993. “Postmaterialism and the Economic
Condition.” American Journal of Political Science 37(3):747–779.
Egri, Carolyn P. and David A. Ralston. 2004. “Generation Cohorts and Personal Values:
A Comparison of China and the United States.” Organization Science 15(2):210–220.
Epstein, Seymour. 1989. Values from the Perspective of Cognitive-experiential Self-theory.
In Social and Moral Values: Individual and Societal Perspectives, ed. Nancy Eisenberg,
Janusz Reykowski and Ervin Staub. Hillsdale, NJ: Erlbaum.
Erickson, Rebecca J. 1995. “The Importance of Authenticity for Self and Society.” Symbolic Interaction 18(2):pp. 121–144.
Falmagne, Jean-Claude, Michel Regenwetter and Bernard Grofman. 1997. A Stochastic
Model for the Evolution of Preferences. In Choice, Decision, and Measurement: Essays
in Honor of R. Duncan Luce, ed. A. A. J. Marley. Mahwah, NJ: Erlbaum pp. 111–129.
Feather, Norman T. 1992. “Values, Valences, Expectations, and Actions.” Journal of
Social Issues 48:109–124.
Feldman, Stanley. 1988. “Structure and Consistency in Public Opinion: The Role of Core
Beliefs and Values.” American Journal of Political Science 32(2):416–440.

117

Feldman, Stanley. 2003. Values, Ideology, and the Structure of Political Attitudes. In
Oxford Handbook of Political Psychology, ed. David O. Sears, Leonie Huddy and Robert
Jervis. New York: Oxford University Press.
Feldman, Stanley and John Zaller. 1992. “The Political Culture of Ambivalence: Ideological Responses to the Welfare State.” American Journal of Political Science 36(1):268–
307.
Feldman, Stanley and Marco R. Steenbergen. 2001. “The Humanitarian Foundation of
Public Support for Social Welfare.” American Journal of Political Science 45(3):658–
677.
Glasgow, Garrett. 1997. “Heterogeniety, Salience, and Decision Rules for Candidate Preference.” Working Paper.
Goren, Paul. 2005. “Party Identification and Core Political Values.” American Journal of
Political Science 49(4):881–896.
Goren, Paul, Christopher M. Federico and Miki Caul Kittilson. 2009. “Source Cues, Partisan Identities, and Political Value Expression.” American Journal of Political Science
53(4):805–820.
Grant, Tobin J. and Thomas J. Rudolph. 2003. “Value Conflict, Group Affect, and the
Issue of Campaign Finance.” American Journal of Political Science 47(3):453–469.
Greene, William H. 2003. Econometric Analysis, Fifth Edition. Delhi, India: Pearson.
Gundelach, Peter. 1995. Grass-roots Activity. In The Impact of Values, ed. Jan W. Van
Deth and Elinor Scarbrough. New York: Oxford University Press pp. 412–439.
Heiss, Florian. 2002. “Structural Choice Analysis with Nested Logit Models.” The Stata
Journal 2(3):227–252.
Hitlin, Steven. 2003. “Values as the Core of Personal Identity: Drawing Links between
Two Theories of Self.” Social Psychology Quarterly 66(2):pp. 118–137.
Hitlin, Steven and Jane Allyn Piliavin. 2004. “Values: Reviving a Dormant Concept.”
Annual Review of Sociology 30:359–393.
Ho, Daniel E., Kosuke Imai, Gary King and Elizabeth A. Stewart. 2007. “Matching as
Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal
Inference.” Political Analysis 15(3):199–236.
Hochschild, Jennifer L. 1995. Facing Up to the American Dream: Race, Class, and the
Soul of the Nation. Princeton, NJ: Princeton University Press.
Hofstede, Geert. 2001. Culture’s Consequences: Comparing Values, Behaviors, Institutions, and Organizations Across Nations. Thousand Oaks, CA: Sage.
118

Holland, Paul W. 1986. “Statistics and Causal Inference.” Journal of the American Statistical Association 81(396):945–960.
Huitt, Ralph K. 1954. “The Congressional Committee: A Case Study.” American Political
Science Review 48(2):340–365.
Hunter, James Davison. 1991. Culture Wars: The Struggle to Define America. New York:
Basic Books.
Iacus, Stefano M., Gary King and Giuseppe Porro. 2011. “Mltivariate Matching Methods
that are Monotonic Balance Bounding.” Journal of the American Statistical Association
106(2):345–361.
Inglehart, Ronald. 1971. “The Silent Revolution in Europe: Intergenerational Change in
Post-Industrial Societies.” The American Political Science Review 65(4):991–1017.
Inglehart, Ronald. 1977. The Silent Revolution: Changing Values and Political Styles
Among Western Publics. Princeton, NJ: Princeton University Press.
Inglehart, Ronald and Paul R. Abramson. 1994. “Economic Security and Value Change.”
The American Political Science Review 88(2):pp. 336–354.
Inglehart, Ronald and Scott C. Flanagan. 1987. “Value Change in Industrial Societies.”
The American Political Science Review 81(4):1289–1319.
Jackson, David J. and Duane F. Alwin. 1980. “The Factor Analysis of Ipsative Measures.”
Sociological Methods and Research 9(2):218–238.
Jacoby, William G. 1988. “The Sources of Liberal-Conservative Thinking: Education and
Conceptualization.” Political Behavior 10(4):316–332.
Jacoby, William G. 1990. “Variability in Issue Alternatives and American Public Opinion.”
Journal of Politics 52(2):579–606.
Jacoby, William G. 1991. “Ideological Identification and Issue Attitudes.” American Journal of Political Science 35(1):178–205.
Jacoby, William G. 1994. “Public Attitudes Toward Government Spending.” American
Journal of Political Science 38(2):336–361.
Jacoby, William G. 2006. “Value Choices and American Public Opinion.” American
Journal of Political Science 50(3):706–723.
Jacoby, William G. 2008. “Testing for Hierarchical Structure and Priming Effects Among
Individual Value Choices.” Presented at the Annual Meetings of the Midwest Political
Science Assocation.

119

Jones, Bradford S. and Chad Westerland. 2006. “Order Matters (?): Alternatives to
Conventional Practices for Ordinal Categorical Response Variables.” Presented at the
Annual Meetings of the Midwest Political Science Association.
Kahle, Lynn R. 1996. Social Values and Consumer Behavior: Research from the List of
Values. In The Psychology of Values: The Ontario Symposium, Volume 8, ed. Clive
Seligman, James M. Olson and Mark P. Zanna. Mahwah, NJ: Lawrence Erlbaum
Associates pp. 131–151.
Kass, Robert E. and Adrian E. Raftery. 1995. “Bayes Factors.” Journal of the American
Statistical Association 90(430):pp. 773–795.
Kinder, Donald R. and Lynn M. Sanders. 1996. Divided by Color: Racial Politics and
Democratic Ideals. Chicago: University of Chicago Press.
Kluegel, James R. and Eliot R. Smith. 1986. Beliefs About Inequality: Americans’ Views
of What Is and What Ought To Be. Hawthorne, NY: de Gruyter.
Knoppen, Desiree and Willem Saris. 2008. “Do we have to combine Values in the Schwartz’
Human Values Scale? A Comment on the Davidov Studies.” Survey Research Methods
3(2):91–103.
Knutsen, Oddbjorn. 1995. Party Choice. In The Impact of Values, ed. Jan W. Van Deth
and Elinor Scarbrough. New York: Oxford University Press pp. 460–491.
Kohn, Melvin L. 1977. Class and Conformity: A Study in Values, 2nd Edition. Chicago:
University of Chicago Press.
Konty, Mark A. and Charlotte Chorn Dunham. 1997. “Differences in Value and Attitude
Change Over the Life Course.” Sociological Spectrum 17(2):177–197.
Krosnick, Jon A. and Duane F. Alwin. 1988. “A Test of the Form-Resistant Correlation Hypothesis: Ratings, Rankings, and the Measurement of Values.” Public Opinion
Quarterly 52(4):526–538.
Kumar, Sushil and Shashi Kant. 2007. “Exploded Logit Modeling of Stakeholders’ Preferences for Multiple Forest Values.” Forest Policy and Economics 9:516–526.
Lenartowicz, Tomasz and James Patrick Johnson. 2003. “A Cross-National Assessment of
the Values of Latin America Managers: Contrasting Hues or Shades of Gray?” Journal
of International Business Studies 34(3):266–281.
Levitin, Teresa E. and Warren E. Miller. 1979. “Ideological Interpretations of Presidential
Elections.” American Political Science Review 73(3):751–771.
Lewis-Beck, Michael S., William G. Jacoby, Helmut Norpoth and Herbert F. Weisberg.
2008. The American Voter Revisited. Ann Arbor, MI: The University of Michigan
Press.
120

Luce, R. Duncan. 1959. Individual Choice Behavior. New York: Wiley.
Luce, R. Duncan and Patrick Suppes. 1965. Preference, Utility, and Subjective Probability. In Handbook of Mathematical Psychology III, ed. R. Duncan Luce, Robert R. Bush
and Eugene Galanter. New York: Wiley pp. 249–410.
Macrae, Duncan Jr. 1952. “The Relation Between Roll Call Votes and Constituencies
in the Massachusetts House of Representatives.” American Political Science Review
46(4):1046–1055.
Maio, Gregory R. and James M. Olsen. 1994. “Value-Attitude-Behavior Relations: The
Moderating Role of Attitude Functions.” 33(3):301–312.
Maio, Gregory R. and James M. Olsen. 2000. What is a “Value-Expressive” Attitude? In
Why We Evaluate Functions of Attitudes, ed. Gregory R. Maio and James M. Olsen.
Mahwah, NJ: Lawrence Erlbaum Associates pp. 249–269.
Maio, Gregory R., James M. Olsen, Lindsay Allen and Mark M. Bernard. 2001. “Addressing Discrepancies between Values and Behavior: The Motivating Effect of Reasons.”
Journal of Experimental Social Psychology 37(2):104–117.
Maio, Gregory R. and James M. Olson. 1998. “Values as Truisms: Evidence and Implications.” Journal of Personality and Social Psychology 74(2):294–311.
Maio, Gregory R., Neal J. Roese, Clive Seligman and Albert Katz. 1996. “Rankings, Ratings, and the Measurement of Values: Evidence for the Superior Validity of Ratings.”
Basic and Applied Social Psychology 18(2):171–181.
McCann, James A. 1997. “Electoral Choices and Core Value Change: The 1992 Presidential Campaign.” American Journal of Political Science 41(2):564–583.
McClosky, Herbert and John Zaller. 1984. The American Ethos: Public Attitudes toward
Capitalism and Democracy. Cambridge, MA: Harvard University Press.
Miethe, Terance D. 1985. “The Validity and Reliability of Value Measurements.” The
Journal of Psychology 115(5):441–453.
Miller, Warren E. and J. Merrill Shanks. 1996. The New American Voter. Cambridge,
MA: Harvard University Press.
Murphey, Edward F., Jr., Mark D. Woodhull, Bert Post, Carolyn Murphy-Post, William
Teeple and Kent Anderson. 2006. “9/11 Impact on Teenage Values.” Journal of Business
Ethics 69(4):399–421.
Nelson, Thomas E., Rosalee A. Clawson and Zoe M. Oxley. 1997. “Media Framing of a
Civil Liberties Conflict and Its Effect on Tolerance.” The American Political Science
Review 91(3):567–583.

121

Omoto, Allen M. and Mark Snyder. 1995. “Sustained Helping without Obligation: Motivation, Longevity of Service, and Perceived Attitude Change Among AIDS Volunteers.”
Journal of Personality and Social Psychology 68(4):671–686.
Peffley, Mark A. and Jon Hurwitz. 1985. “A Hierarchical Model of Attitude Constraint.”
American Journal of Political Science 29(4):871–890.
Pollock, Philip H., III, Stuart A. Lilie and M. Elliot Vittes. 1993. “Hard Issues, Core Values and Vertical Constraint: The Case of Nuclear Power.” Bristish Journal of Political
Science 23(1):29–50.
Poole, Keith T. and Howard Rosenthal. 1991. “Patterns of Congressional Voting.” American Journal of Political Science 35:228–278.
Raftery, Adrian E. 1995. “Bayesian Model Selection in Social Research.” Sociological
Methodology 25:pp. 111–163.
Raftery, Adrian E. 1999. “Bayes Factors and BIC: Comment on ‘A Critique of the
Bayesian Information Criterion for Model Selection’.” Sociological Methods and Research 27(3):411–427.
Roccas, Sonia, Lilach Sangiv, Shalom H. Schwartz and Ariel Knafo. 2002. “The Big Five
Personality Factors and Personal Values.” Personality and Social Psychology Bulletin
28(6):789–801.
Rokeach, Milton. 1968. Beliefs, Attitudes, and Values: A Theory of Organization and
Change. San Francisco: Jossey–Bass.
Rokeach, Milton. 1973. The Nature of Human Values. New York: Free Press.
Rosenbaum, Paul R. and Donald B. Rubin. 1983. “The Central Role of the Propensity
Score in Observational Studies for Causal Effects.” Biometrika 70(1):pp. 41–55.
Rosenbaum, Paul R. and Donald B. Rubin. 1985. “Constructing a Control Group Using
Multivariate Matched Sampling Methods That Incorporate the Propensity Score.” The
American Statistician 39(1):pp. 33–38.
Rossiter, Clinton. 1962. Conservatism in America: The Thankless Persuasion. New York:
Vintage Books.
Rubin, Donald B. 1979. “Using Multivariate Matched Sampling and Regression Adjustment to Control Bias in Observational Studies.” Journal of the American Statistical
Association 74(366):pp. 318–328.
Rubin, Donald B. 1980. “Bias Reduction Using Mahalanobis-Metric Matching.” Biometrics 36(2):pp. 293–298.
Schwartz, Shalom H. 1992. Universals in the Content and Structure of Values. In Advances
in Experimental Social Psychology, ed. Mark P. Zanna. New York: Academic Press.
122

Schwartz, Shalom H. 1994. “Are There Universal Aspects in the Structure and Content
of Human Values?” Journal of Social Issues 50(4):19–46.
Schwartz, Shalom H. 1996. Value Priorities and Behavior: Applying a Theory of Integrated Value Systems. In The Psychology of Values: The Ontario Symposium, Volume
8, ed. Clive Seligman, James M. Olson and Mark P. Zanna. Mahwah, NJ: Lawrence
Erlbaum Associates.
Schwartz, Shalom H. 2001.
“A Proposal for Measuring Value Orientations
Across Nations.” in ESS Questionnaire Development Report, available at
http://www.europeansocialsurvey.org.
Schwartz, Shalom H. 2004. Mapping and Interpreting Cultural Differences Around the
World. In Comparing Cultures: Dimensions of Culture in a Comparative Perspective,
ed. Henk Vinken, Joseph Soeters and Peter Ester. Leiden, Netherlands: Brill.
Schwartz, Shalom H. 2005. “Basic Human Values: An Overview.” available at
http://151.97.110.134/Allegati/convegno%207-8-10-05/Schwartzpaper.pdf.
Schwartz, Shalom H. 2007. Value Orientations: Measurement, Antecedents, and Consequences Across Nations. In Measuring Attitudes Cross-Nationally: Lessons from the
European Social Survey, ed. Roger Jowell, Caroline Roberts, Rory Fitzgerald and
Gillian Eva. Los Angeles: Sage.
Schwartz, Shalom H. and Anat Bardi. 2001. “Value Hierarchies Across Cultures: Taking
a Similarities Perspective.” Journal of Cross-Cultural Psychology 32(3):268–290.
Schwartz, Shalom H. and Wolfgang Bilsky. 1987. “Toward a Universal Psychological
Structure of Human Values.” Journal of Personality and Social Psychology 53(3):550–
562.
Sears, David O., P.J. Henry and Rick Kosterman. 2000. Egalitarian Values and Contemporary Racial Politics. In Racialized Politics: The Debate About Racism in America,
ed. David O. Sears, Jim Sidanius and Lawrence Bobo. Chicago: University of Chicago
Press pp. 75–116.
Seligman, Clive and Albert N. Katz. 1996. The Dynamics of Value Systems. In The Psychology of Values: The Ontario Symposium, Volume 8, ed. Clive Seligman, James M.
Olson and Mark P. Zanna. Mahwah, NJ: Lawrence Erlbaum Associates.
Sheng, Shing-Yuan. 1995. “NES Measurement of Values and Predispositions, 1984–1992.”
National Election Studies, Technical Report No. 50.
Skrondal, Anders and Sophia Rabe-Hesketh. 2003. “Multilevel Logistic Regression for
Polytomous Data and Rankings.” Psychometrika 68(2):268–287.

123

Sniderman, Paul M., Henry E. Brady and Philip E. Tetlock. 1994. “The Multiple Investigator, Common Vehicle.” Public Opinions: Newsletter of the APSA Organized Section
on Elections, Public Opinion, and Voting Behavior 1(1):10–11.
Tetlock, Philip E. 1986. “A Value Pluralism Model of Political Reasoning.” Journal of
Personality and Social Psychology 50(4):819–827.
Tetlock, Philip E. 2000. Coping with Trade-offs: Psychological Constraints and Political
Implications. In Elements of Reason: Cognition, Choice, and the Bounds of Rationality,
ed. Arthur Lupia, Matthew D. McCubbins and Samuel L. Popkin. Berkeley, CA:
University of California Press.
Tetlock, Philip E., Randall S. Peterson and Jennifer S. Lerner. 1996. Revising the Value
Pluralism Model: Incorporating Social Content and Context Postulates. In The Psychology of Values: The Ontario Symposium, Volume 8, ed. Clive Seligman, James M.
Olson and Mark P. Zanna. Mahwah, NJ: Lawrence Erlbaum Associates.
Verplanken, Bas and Rob W. Holland. 2002. “Motivated Decision Making: Effects of Activation and Self-Centrality of Values on Choices and Behavior.” Journal of Personality
and Social Psychology 82(3):434–447.
Zaller, John. 1991. “Information, Values, and Opinion.” American Political Science Review 85(4):1215–1237.
Zaller, John. 1992. The Nature and Origins of Mass Opinion. New York: Cambridge
University Press.

124