Political Ground Truth: How

Personal Issue Experience Counters

Partisan Biases

Aaron B. Strauss

A Dissertation
Presented to the Faculty
of Princeton University
in Candidacy for the Degree
of Doctor of Philosophy

Recommended for Acceptance

by the Department of Politics
Adviser: Martin Gilens

September 2009
c Copyright by Aaron B. Strauss, 2009.

All Rights Reserved

 Abstract

Partisan attachments create pervasive bias in the way citizens process information.
Political scientists, psychologists, and recently neuroscientists find that people will
believe nearly anything if a favored politician espouses the view. Yet, even though
partisan affiliation is one of the most, if not the most, stable political attitudes, large
segments of the public switch their vote choice from one party to the other between
elections or split their tickets within a single election. This dissertation examines
one explanation for shifting political views: personal experience with specific issues.
Campaigns focus on in attracting more support, but partisan biases hinder their
efforts when predispositions lead voters to doubt statements made by disfavored
politicians. This dissertation explores the theory that campaigns can successfully
target voters who have experience on a particular political issue. The voter will use
her independent knowledge on the topic to judge, or “ground truth,” the politician’s
views; if the voter and the politician agree, the voter will hold the candidate in higher
esteem. With the advent of massive campaign databases of information on voters,
campaigns are now able to identify these crucial voter-issue linkages.
The Personal Experience Model explores why personal experience plays such a
crucial role in political judgments. This formal model is an extension of Zaller’s
Receive-Accept-Sample model. The theory behind the Personal Experience Model is
presented, related to existing theories, and supported by empirical evidence. Obser-
vational data from the 2000 presidential campaign, two survey experiments, and two
field experiments all support the model’s hypotheses. Finally, the strategic implica-
tions for campaigns, and the normative implications for democracy, are considered.

iii
 Acknowledgements

If you’re reading this, your name will appear here.

iv
To my family—Mom, Dad, and Marc—who are responsible for my love of learning.

v
Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

1 Introduction 1

2 The Personal Experience Model 6

2.1 Theory and Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Psychological Model of Personal Experience and Issue Stability 9
2.1.2 Extending Zaller by Introducing Personal Experience . . . . . 14
2.1.3 Learning Model: Issue Positions . . . . . . . . . . . . . . . . . 18
2.1.4 Voters’ Beliefs About Candidate Positions . . . . . . . . . . . 23
2.1.5 Voters’ Candidate Evaluations . . . . . . . . . . . . . . . . . . 26
2.1.6 Issue Experience as Signals . . . . . . . . . . . . . . . . . . . . 32
2.1.7 Model Extensions and Details . . . . . . . . . . . . . . . . . . 34
2.2 Concurrence With Other Theories in the Literature . . . . . . . . . . 37
2.2.1 Definitions of Concepts in the Literature . . . . . . . . . . . . 38
2.2.2 Personal Experience Yields Nuanced and Stable Opinions . . . 39

vi
 2.2.3 Evidence for Cue-Taking and the Role of Experience . . . . . 42
2.2.4 Easy Issues and Retrospection Affect Political Evaluations . . 46
2.3 Alternative Voter-Issue Linkages in the Literature . . . . . . . . . . . 48

3 Experienced Evaluations and Self-Interested Opinions 52

3.1 The Personal Experience Model and the 2000 Presidential Campaign 52
3.1.1 The Campaign About Nothing . . . . . . . . . . . . . . . . . . 52
3.1.2 Issue-Driven Vote Choice . . . . . . . . . . . . . . . . . . . . . 57
3.1.3 Data from the 2000 Annenberg Survey . . . . . . . . . . . . . 61
3.2 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2.1 Learning Over the Course of the Campaign . . . . . . . . . . . 62
3.2.2 Cue-Taking and Self-Interest . . . . . . . . . . . . . . . . . . . 65
3.2.3 Experienced Voters Judge Politicians on Issues . . . . . . . . . 70
3.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4 Evidence from Survey Experiments 75

4.1 Theory and Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Survey Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2.1 Overall Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2.2 Specifics of Princeton Survey Design . . . . . . . . . . . . . . 82
4.2.3 Specifics of Nationwide Survey . . . . . . . . . . . . . . . . . . 84
4.3 Design Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.3.1 Characteristics of the Survey Respondent Population . . . . . 87
4.3.2 Features of the Survey . . . . . . . . . . . . . . . . . . . . . . 88

vii
 4.3.3 Definitions of Issue Experience . . . . . . . . . . . . . . . . . . 93
4.4 Support for Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4.1 Candidate Evaluation . . . . . . . . . . . . . . . . . . . . . . . 96
4.4.2 Issue Opinions and Experience . . . . . . . . . . . . . . . . . . 101
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

5 Field Experiments: New Methodology and Evidence 107

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1.1 Background of the Methodological Problem . . . . . . . . . . 110
5.2 The Formal Framework of GOTV Campaign Planning . . . . . . . . 112
5.2.1 The Planner’s Decision Problem . . . . . . . . . . . . . . . . . 113
5.2.2 Data from a Randomized Field Experiment . . . . . . . . . . 117
5.2.3 The Bayesian Planner . . . . . . . . . . . . . . . . . . . . . . 119
5.2.4 Bayesian Optimal Campaign Planning at A Glance . . . . . . 121
5.3 The Optimal Nonpartisan Campaign Strategy . . . . . . . . . . . . . 123
5.3.1 The Optimization Method . . . . . . . . . . . . . . . . . . . . 123
5.3.2 The Statistical Method . . . . . . . . . . . . . . . . . . . . . . 125
5.4 The Optimal Partisan Campaign Strategy . . . . . . . . . . . . . . . 130
5.4.1 The Decision Problem . . . . . . . . . . . . . . . . . . . . . . 131
5.4.2 Data Requirements . . . . . . . . . . . . . . . . . . . . . . . . 132
5.4.3 Derivation of the Optimal Strategy . . . . . . . . . . . . . . . 133
5.5 Empirical Evaluation of the Proposed Method . . . . . . . . . . . . . 134
5.5.1 Evaluation Method . . . . . . . . . . . . . . . . . . . . . . . . 136

viii
 5.5.2 A Nonpartisan GOTV Campaign with a Single Mobilization
Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.5.3 Partisan Example: Parents and Eduction Spending . . . . . . 140
5.5.4 Partisan Example: Parents and Education Spending . . . . . . 143
5.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

6 Campaign and Normative Implications of Microtargeting 154

6.1 Campaign Microtargeting . . . . . . . . . . . . . . . . . . . . . . . . 154
6.2 Implications for Campaign Strategy . . . . . . . . . . . . . . . . . . . 158
6.2.1 Microtargeting and Message Control . . . . . . . . . . . . . . 158
6.2.2 Translating the Personal Experience Model into a Formal Game160
6.2.3 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.2.4 Best Response . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
6.2.5 Edge Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . 169
6.2.6 No Internal Equilibria . . . . . . . . . . . . . . . . . . . . . . 173
6.2.7 Model Extensions: Multiple Issues and Platform Decisions . . 174
6.2.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.3 Normative Implications . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.3.1 Heuristics and Judging Democracy . . . . . . . . . . . . . . . 176
6.3.2 Positives for Democracy . . . . . . . . . . . . . . . . . . . . . 178
6.3.3 Negatives for Democracy . . . . . . . . . . . . . . . . . . . . . 180
6.3.4 Further Discussion: Party Structure . . . . . . . . . . . . . . . 182
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

ix
7 Conclusion 185

A Appendix for Chapter 2 188

A.1 Details for Candidate Evaluation Simulation . . . . . . . . . . . . . . 188

B Appendices for Chapter 4 191

B.1 Question Wording in Nationwide Survey . . . . . . . . . . . . . . . . 191
B.1.1 Candidate Descriptions . . . . . . . . . . . . . . . . . . . . . . 191
B.1.2 Candidates’ Issue Signals . . . . . . . . . . . . . . . . . . . . . 192
B.1.3 Experience Questions . . . . . . . . . . . . . . . . . . . . . . . 193
B.2 Auxiliary Regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
B.3 Issue Experience Measures and the Hypotheses . . . . . . . . . . . . . 195

C Appendix for Chapter 5 206

C.1 Computational Appendices . . . . . . . . . . . . . . . . . . . . . . . . 206
C.1.1 Nonpartisan Case: the Knapsack Problem . . . . . . . . . . . 206
C.1.2 Partisan Case: the Stochastic Knapsack Problem . . . . . . . 207

D Appendix for Chapter 6 209

D.1 Analytical Solution for Variance of Vote Share . . . . . . . . . . . . . 209

x
Chapter 1

Introduction

Prominent research in political science (Campbell et al., 1960; Bartels, 2002) empha-
sizes the long-lasting attachments voters have to political parties. Yet large portions
of the electorate switch parties from election to election (Key, 1968) or split their
ticket within one election (Fiorina, 1996). These defections are often attributed to
the preferences of voters relative to the candidates (Downs, 1957; Enelow and Hinich,
1984; Carmines and Stimson, 1990); the debate over relative effects of partisan identi-
fication versus issue opinion continues in recent research (Ansolabehere et al., 2008).
This dissertation synthesizes the arguments from both camps of thought, demon-
strating how party identification and issue opinions interact to form individuals’ po-
litical judgments. To provide observable predictions of the theory, the model and the
evidence presented highlights the role of personal experience in political evaluations.
Two recent anecdotes illustrate this connection. From the 2004 to the 2006
elections, the percentage of Hispanics identifying as Democrats (as measured by

1
the National Exit Poll) jumped by 15 percentage points (from 43% to 58%), while
non-Hispanics showed no increase in their inclination to be identified as Democrats.
Hispanics’ responses to another question in the exit poll shed light on the underlying
reason for the massive movement: 37% of Latinos indicated that the issue of “illegal
immigration” was “extremely important” in their voting decision (compared to only
29% of the non-Latino population). While immigration policies might not personally
affect these voters (as they were already citizens), their involvement in the Hispanic
community most likely gave them firsthand experience with the immigration issue.
And despite illegal immigration falling off as a key concern in 2008, this trend of
Hispanics favoring the Democrats continued in the 2008 election. As measured by
the National Exit Poll, Obama garnered 67% of the Hispanic vote, as compared to
53% for Kerry.
As Hispanics were flocking to the Democratic party, active duty military personnel
were leaving the Republican party in droves. A 2004 mail survey of subscribers to
Military Times found that 60% of active duty personnel identified as Republicans.
The same survey in 2006 found that this percentage had dropped to 46%. Republican
affiliation among the electorate at large (as measured by exit polls) dropped only 1.5
percentage points over these two years. Almost assuredly, the wars in Iraq and
Afghanistan had a profound effect on the political views of the military. In 2008,
Obama outperformed Kerry among the active military and veterans despite the facts
that both Kerry and Obama’s opponent were veterans.
This dissertation argues in an individual’s political allegiance is likely to shift
when the political parties take identifiable positions on issues with which the in-

2
dividual has significant personal experience. When a large group of people have
experience with a prominent issue (e.g., Hispanics and immigration, the military
and the Iraq War), substantial changes can occur in the political landscape. Gen-
erally, an individual’s partisan bias weighs heavily on her political decision-making
process as she adopts the positions of ideologically-similar elites and does not ad-
just her political worldview.1 Campaigns can use recent technological advances to
identify (i.e., microtarget) the segments of the population who have experience with
certain issues and who thus may be ripe for a political change of heart.
Chapter Two presents the Personal Experience Model, a learning model based on
Zaller’s (1992) Receive-Accept-Sample (RAS) model. In this model, personal expe-
rience leads to stable views on a political issue. This postulate enables to two main
hypotheses. First, if political elites start debating an issue, citizens with experience
judge politicians on the issue, comparing their experience to the politician’s argu-
ment. In effect, an individual “ground truths” an elite’s views by evaluating how
consistent the elite’s arguments are with the voter’s first-hand experience.
The second main hypothesis is that voters with experience shift their issue opinion
less in the direction of a politician’s appeal than do individuals without experience.
Simulations illustrate examples of the model at work in the campaign setting. I
discuss how existing research is consistent with the Personal Experience Model and
how the foundations of the model differ from research on issue publics and self-
interest.
Chapter Three highlights three observational cases from the 2000 presidential
1
Throughout this dissertation the female pronoun is used to identify an individual voter. The
male pronoun is used to identify an individual candidate.

3
campaign that illustrate the model in action: Social Security privatization, the Pa-
tient’s Bill of Rights, and defence spending. Regarding the first issue, seniors, who
have experience with Social Security, judged Bush and Gore more on privatization
than did younger voters. Younger voters were more apt to adopt the position of their
preferred candidate than were seniors. An analogous phenomenon occurs among vot-
ers ensured by HMOs and the Patient’s Bill of Rights as well as voters in military
households and defense spending.
Chapter Four looks for the effects predicted by the Personal Experience Model
in a survey experiment setting. Two Internet panel surveys, one of Princeton stu-
dents (n=273) and one of the public nationwide (n=391), asked respondents about
hypothetical candidates and real-world issues. Respondents were also queried about
their experiences relating to these issues. The results are generally supportive of
the Personal Experience Model’s two main hypotheses but are noisy because of low
n-sizes.
Chapter Five develops a methodology for analyzing field experiments that iden-
tifies voters who are responsive to candidate appeals. Off-the-shelf non-parametric
data-mining techniques are applied to both the treatment group and the control
group of the experiment. For each segment of the population, the difference between
the estimated effect for the two treatment assignments is the treatment effect. Re-
source allocation algorithms are applied to these estimates to determine the optimal
campaign strategy. The method is general enough to be applicable in both non-
partisan (i.e., get-out-the-vote) and partisan settings. Examples with experienced-
voter segments demonstrate the power of this method and validate the candidate

4
evaluation hypothesis of the Personal Experience Model.
The final chapter presents the implications of the Personal Experience Model and
campaign microtargeting in general for campaign strategy and democracy. A game
theoretical model explicates the situations in which a campaign should spend its
money microtargeting voters rather than broadcasting to all voters its candidate’s
position on an issue. In general, broadcasting is a gamble that may have large returns
for a campaign, but more often may be useless or even backfire. Thus, campaigns that
are behind in the polls find this risky strategy useful, while advantaged campaigns
often prefer to microtarget.
The normative implications for democracy are mixed, depending on one’s per-
spective. The Personal Experience Model, combined with modern campaign tools,
increases the efficiency of microtargeting. As with many tools, these advancements
may be used for positive or negative purposes. Information may be disseminated to
the public in a more relevant fashion as voters are matched with issues they care
about. This information gain is viewed as a positive for the democratic process.
However, the increased party attachment that microtargeting produces in some seg-
ments of the public leads some individuals to accept the statements of political elites
more easily even if those arguments are completely fabricated.
In brief, the Personal Experience Model explains a voter-issue linkage that limits
the partisan bias that pervades the political arena. The micro-level model provides
deeper understanding of the research on issue publics and self-interest. Campaigns
can leverage the implications of the model to increase the impact and efficiency of
their appeals.

5
Chapter 2

The Personal Experience Model

2.1 Theory and Hypotheses

As demonstrated by Hispanics and the military after the 2004 election, large segments
of the population can be persuaded to change parties from one election to the next.
However, most research on political persuasion does not account for these changes,
especially when the changes are so pronounced in specific groups. The Personal
Experience Model seeks to address this discrepancy, at least among a subset of
voter-issue interactions.
The so-called “Michigan model” (Campbell et al., 1960) highlights the role of
early-life events and relationships in the formation of long-standing partisan identi-
fication. Interactions with the political system at an early age form individuals’ par-
tisan loyalties, which are dubbed the “unmoved mover” by Campbell et al.. “[O]nly
an event of extraordinary intensity can arouse any significant part of the electorate

6
to the point that its established political loyalties are shaken” (Campbell et al., 1960,
p. 151).
The mechanism for stable partisanship has been elucidated over the years and
scholars have reached a general consensus. Consistent with the psychological concept
of cognitive dissonance (Festinger, 1957), individuals are more likely to accept and
process political information and assertions that are consistent with their current
views. Prominent research (Zaller, 1992; Bartels, 2002; Taber and Lodge, 2006)
applies this concept to politics, formalizing the phenomenon and demonstrating how
the predictions are consistent with observations of public political opinions. Dozens
of micro studies, including some from neuroscience (Knutson et al., 2006), have
verified that partisan identification skews individuals’ perceptions of the world in a
self-confirming manner, leading to hardened political opinions.
One emblematic explanation for the micro-foundations of this “partisan bias”
is offered by Lupia and McCubbins’s (1998) model of persuasion. In this model,
voters are persuaded by elites with both “perceived common interest” to the voter
and “perceived knowledge” of the subject at hand. Listening to trusted politicians
leads to further agreement between the voter and the elite (i.e., cue-taking, Gilens
and Murakawa (2002)), and thus more perceived common interest. This positive
feedback loop, or “cycle of partisan bias,” begs two questions. First, what is the
origin of “perceived common interest?” Second, if politicians cannot persuade voters
to cast a ballot for previously untrusted candidates or parties (i.e., those without
perceived common interest) then why do some voters switch parties between elections
or split their votes within an election?

7
 The Michigan model clearly answers the first of these questions with the notion
that voters adopt the party identification of their parents and the social atmosphere
in which they are raised. But the answer to the second question, change in party
identification, is less clear: What is an “event of extraordinary intensity” (Campbell
et al., 1960)? The lack of theory on the determinants of party-switching has left the
door open to other approaches.
Two large areas of work that explain a shifting electorate are retrospective voting
and issue voting. Fiorina (1981) argues that voters are affected by how parties and
candidates performed their governing duties while in power. Retrospective voting,
especially on the economy, certainly plays a role in candidate and party evaluations
(Mueller, 1970). However, these types of judgments should be relatively uniform
across the populace, and do not account for heterogeneous movements, such as with
Hispanics and the military after 2004.
A second answer to the vote-switching question is presented by scholars who
analyze voters’ issue opinions. The Hotelling-Downs (Hotelling, 1929; Downs, 1957)
model presents voters as rational actors who vote for the party (or candidate) that
most closely matches their issue positions.1 Downs’ uni-dimensional, one-valued
utility is expanded into several issues of varying importance (or “salience”) by Enelow
and Hinich (1984). Certain segments of the population altering their issue opinions
or issue saliences could account for the observed shifts in the electorate’s political
views. However, if voters cue-take from preferred politicians via the cycle of partisan
1
That voters are rational actors in formalized models such as Hotelling-Downs and Fiorina does
not rule out the possibility that voters are rational when succumbing to partisan bias. It may be
rational to minimize the displeasure of believing two seemingly contradictory ideas.

8
bias, then changes in issue opinion will only polarize the electorate. No shifts from
one party to the other, or one candidate to the other, will be observed.

2.1.1 Psychological Model of Personal Experience and Issue

Stability

A mechanism for breaking the cycle of partisan bias—which emphatically does not
rule out other causes—is personal experience. Individuals create and strengthen
political opinions through personal experience on a subset of the many issues that
political elites must deal with. This subset of issues forms a benchmark against which
to judge parties, candidates, and political elites,enabling an individual to develop a
political worldview.
First I present the psychological flow of political information, and then I develop
a learning model that formalizes these flows. The flow of information is a superset
of Zaller’s (1992) Receive-Accept-Sample (RAS) model, and I call my model the
“Personal Experience Model”. Figure 2.1 presents a box-and-arrow depiction of the
Personal Experience Model.
A review of Zaller’s model is crucial for understanding the how the Personal
Experience Model fits within the current political science literature. Under RAS,
an individual incorporates information into a political opinion by first receiving the
information. As first discussed by Downs (1957), individuals can either seek out this
information or happen upon it accidentally. More politically attentive individuals
are more likely to receive political information.
Next, individuals (either consciously or subconsciously) accept or reject a piece

9
of political information based on their worldview. In Zaller’s words “People tend
to resist arguments that are inconsistent with their political predispositions” (page
44). This accept/reject process, based on what I label an individual’s “worldview”,
is the key for producing the stable partisan affiliation and partisan biases found by
Campbell et al. (1960), Bartels (2002), and several others. This worldview acts more
like an online model of important political figures, candidates, and parties. Zaller
acknowledges that voters’ evaluations of other people (rather than issues) act more
like an online model (p. 281).2 If the argument (i.e., “consideration”) is accepted
then it is stored among all other considerations. Recently accepted (or “thought
about”) considerations are most accessible to the individual.
Figure 2.1 depicts this relative availability of political arguments with a First-In-
First-Out (FIFO) queue, a computer science concept. The first political considera-
tion heard (if never thought of again) works its way down to the end of the queue
(i.e., human memory) until it is completely forgotten (or at least, inaccessible). Con-
siderations already in memory that are primed or re-accepted move to the front of
the queue and become more accessible. While some political psychologists (Weston,
2007) would take issue with this simplistic view of the brain, it serves the purpose
for this dissertation.
When an individual is asked, either by a survey interviewer or a friend, to report
her issue opinion, Zaller asserts that the accessible considerations relevant to this
issue are sampled, and the average of these considerations is reported. Opinion
2
An alternative explanation is based on cognitive dissonance rather than source evaluation (or
credibility). In this case, voters would only accept considerations that are consistent with their
existing issue beliefs. However, numerous studies find that voters often switch their position to
adopt the stance of their favored candidate. See discussions in Sections 2.2.3 and .

10
stability can be defined as the inverse of the variance of repeated sample averages
of considerations. Individuals often appear to have unstable opinions about issues
(Converse, 1964) because at any point in time very few considerations about an issue
may be accessible and the accessible considerations may depend on seemingly random
life occurrences (e.g., radio report, water cooler conversation). The model predicts
that this instability will be especially acute when an individual rarely receives and
accepts arguments about that issue and further does not perceive a link between
her political worldview and considerations about the issue. Stability increases when
issues are averaged together, since a large set of considerations is relevant. This
prediction is consistent with recent evidence from surveys (Ansolabehere et al., 2008).
Although Zaller contrasts his theory with an on-line model (Lodge et al., 1989),
the differences between the two perspectives are not significant for the purposes of this
dissertation. While memory-based and on-line models differ in the causal mechanisms
of political opinions, the outcomes are generally not in conflict. Specifically, Lodge
et al. (1989) argue that memories cannot be sampled and evaluated in an unbiased
manner, but Zaller (1992)’s Resistance Axiom incorporates this bias. Zaller argues
that for political issue opinions (as opposed to evaluations of politicians), the on-line
model does not account for the wide variance in people’s opinions over time. However,
the on-line model never specifies how prone the current evaluation is to change.
Also, despite Zaller’s “top of the head” language (p. 49), he allows the full history of
consideration to be sampled (p. 121). This time-invariant weighting of considerations
compares favorably to the on-line model’s integration of all considerations over time
into one evaluation. Thus, while these two models disagree on the innerworkings of

11
voters’ minds, they share fairly consistent predictions.
The theory outlined here is an extension of Zaller’s model rather than the on-line
model for two reasons. First, Zaller’s RAS model, especially the “accept” step of
RAS, is consistent with the cycle of partisan bias. Second, in stark contrast to the on-
line model, Zaller’s “sample” step of RAS implies that reported political judgements
are draws from random variables.
This sampling also differentiates Zaller’s model from a Bayesian updating process
(e.g., Gerber and Green, 1998). If voters were pure Bayesians, they would report a
summary statistic of their beliefs when queried on a survey. But as early psycho-
logical experiments (Grant et al., 1951) persuasively demonstrate, individuals draw
from their belief distribution when reporting opinions. For instance, in Grant et al.
(1951), laboratory participants were asked whether a light bulb would turn on or
not. They quickly learned (in one set of trials) that the light bulb would turn on
75% of the time, but instead of maximizing their probability of being correct and
always predicting the light would turn on, 25% of the time they predicted the light
would not turn on.3 Given this empirical evidence, the model presented in this dis-
sertation assumes that individuals draw from top-of-the-head considerations and are
not perfect Bayesians.
However, Bayes Rule does provide a convenient mathematical way to combine
new data into a distribution. In the following sections Bayes Rule is referenced
3
By always predicting that the light bulb will turn on, a perfect Bayesian would be correct 75%
of the time. Zalleresque individuals, sampling from a set of considerations of previous light bulb
results (the FIFO queue), would predict the bulb to turn on 75% of the time (and be correct 62.5%
of the time). In the experiments, across the entire participant group, 75% of predictions were for
the bulb to be on, meaning that there was not a mix of Bayesians and Zalleresque individuals, but
only Zalleresque individuals in the population.

12
13
Figure 2.1: The flow of political information through the mind of an individual, based on Zaller’s (1992)
Receive-Accept-Sample (RAS) model. Rectangles represent ephemeral processes that information flows
through; ovals represent stable stores of information (or sources of information). Solid arrows indicate in-
formation flow; the one dashed arrow indicates a cause and effect (i.e., partisan identification and woldview
affect which pieces of information are accepted or rejected). The FIFO symbol under “Set of Considerations”
represents the computer science term “First In-First Out” queue and is a simplistic representation of the
human brain’s memory capability.
when calculating the mean of posterior distributions of considerations. Also, I admit
that these theoretically-determined distributions are subject to human fallacies, such
as forgetting information (i.e., considerations) at variable rates. These errors are
unmodeled and do not alter the basic hypotheses.

2.1.2 Extending Zaller by Introducing Personal Experience

To answer the questions of where partisan biases originate and why certain segments
of the population shift partisan loyalties, Zaller’s model is extended by adding per-
sonal experience as a source of political considerations. In the Personal Experience
Model, a steady flow of experiences related to a political issue leads to a relatively
constant set of accessible considerations, and hence a stable political opinion. That
claim is formally stated as a postulate below; it is one of the four predictions of the
Personal Experience Model, along with two main hypotheses and one corollary.
“Personal experience” is defined as frequent interactions with a political issue,
whether this interaction is sought after (e.g., being an environmental activist) or
more incidental (e.g., being retired and receiving Social Security checks). Frequent
conversations about political issues or major life events also count as personal ex-
periences. The key factors for determining whether an interaction is considered
“personal experience” in this context is (1) the individual plays an active role, and
(2) the interaction produces an increased flow of received political considerations on
the issue. Examples include owning a gun, immigrating, having an abortion, being in
an occupation that deals with a political issue (e.g., doctor, teacher), and discussing
politics with friends who have a large stake in a political outcome (e.g., gays and

14
equal rights). An “active role” need not be a large role: cashing a Social Security
check counts as active (though barely), but listening to political news does not.
The second criterion above, the increased flow of political considerations (repre-
sented by the thick arrow in Figure 2.1) leads to stable considerations in three ways.
First, the experience is often based on repeat interactions or one important event,
so the distribution of considerations is narrow. Second, experience leads individ-
uals to process these considerations in a consistent manner (Wood, 1982). Third,
the increased frequency of acceptance of considerations leads to a larger number of
political arguments being accessible. Thus, when the set of considerations of an in-
dividual with issue experience is sampled (e.g., by a survey interviewer), a consistent
set of considerations is accessible and the resulting opinion distribution has a small
variance.
The single-interaction nature of some political experiences causes the resulting
considerations to be narrow in scope. For instance, consider a teacher who works at
a crumbling school. Experiences generated from this interaction are likely to indicate
that public schools need more funding. The single-sided considerations generate by
this personal experience contrasts with the two-sided considerations present in news
reports about school budgets or vouchers.
Not only do some individuals receive similar experiences over time, but the similar
situations in which they have these interactions are most likely to be stable. Consider
a new nurse in the health care field. As the nurse learns the ins and outs of the
hospital at which she works, she develops a better understanding of the successes
and failures of the health care system. A nurse and patient who both witness a

15
failing of the health care bureaucracy may come to different conclusions about the
system. In fact, the patient might not be able to accurately identify the underlying
the problem of the system. Individuals with a repeating personal experiences are
better able to generate political consideration from interactions.
The result of repeated, narrow experiences being processed under similar cir-
cumstances leads to a locus of considerations, often on one side of the issue. Since
the individual is experiencing these considerations herself (rather than receiving the
information from another party), the considerations are likely to be accepted. Conse-
quently, these considerations are more likely to be sampled when the issue is explicitly
raised by a politician or a survey interviewer.4 The output of the Zaller’s model under
repeated experiences is a stable (though not necessarily extreme) political opinion.
Stability Postulate: Individuals who have experience with an otherwise political
issue in their non-political lives will have more stable opinions on the issue.
All three predicates are not individually necessary for the Stability Postulate to
hold. For instance, the relationship between experience and stability holds regardless
of whether the experiences received are ideologically one-sided or two-sided. In the
former case, where the considerations are all on one side of the ideological spectrum
(e.g., owning a gun and wanting to keep it for hunting), the process of developing
a stable opinion is straightforward. However, in the case of the nurse dealing with
the health care system, that personal experience might lead the nurse to understand
both sides of the ideological struggle. In this case, the desire to hold a consistent
worldview (the mechanism behind cognitive dissonance) might lead the nurse to
4
Similar conceptions of repeated interactions are present in advertising literature, which empha-
sizes that repeating pictures or phrases affect future actions (Sawyer, 1973).

16
develop a nuanced view of health care policy. As detailed in Section 2.2.2, studies
demonstrate that personal experience often leads to a more sophisticated approach to
the issue at hand. In an alternative, but less likely scenario, an individual receiving
conflicting considerations through experience may pick one side of the issue and
reject arguments from the other side. In either case, the crucial feature of personal
experience is the increased flow of received and accepted political considerations on
an issue.
The considerations generated from personal experience contrast with those from
other sources of information, such as political news. Political information delivered
to individuals via the news is often sporadic and two-sided. Even if the individual
accepts arguments from one set of elites, the elites’ position on the issue might change
because of external factors (e.g., which party controls the White House and foreign
policy). Thus, normal political discourse is less likely to produce stable issue opinions
than is personal experience.5 .
Returning to Zaller (1992), this stability framework is referenced when he dis-
cusses attitude change (Chapter 7). When an individual receives political commu-
nication on an issue, the individuals combines this new information with existing
considerations. If the existing considerations carry a large weight relative to the
new information, then the new information will have little impact on the reported
opinion. In the next section, I further develop this idea using a learning model, and
demonstrate why issue stability is crucial in the arena of voter persuasion.
5
An exception to this general rule dealing with retrospective voting is discussed in Section 2.2.4

17
2.1.3 Learning Model: Issue Positions

Following the lead of Achen and Bartels (2006), the Personal Experience Model
represented in Figure 2.1 can be formalized using a learning model. In its simplest
form, the model includes one voter, one issue, and one candidate.6 The model has two
periods: before and after the candidate sends a signal (i.e., political communication)
on the issue. Prior to receiving a candidate’s signal, the voter has an opinion about
the issue, centered at δ1 . This opinion has a precision (Bartels, 1993), τ12 , which, if
the Stability Postulate is correct, increases with personal experience.7 For simplicity,
I assume that this prior is normally distributed; hence, in period 1, the voter’s issue
position, x1 , is the distribution,

1
x1 ∼ N (δ1 , ). (2.1)
τ12

At the end of period 1, a candidate announces a position, γ, on the issue. Not all
signals are created equal. The persuadability (ψ 2 ) of the candidate’s message (i.e.,
signal) depends on political factors such as how forcefully the candidate argues for
the position and whether voters consider the position “cheap talk.”
Voters update their beliefs about the best policy by adding the candidate’s con-
sideration to the original distribution.8 To avoid discretization of the posterior belief,
6
For clarity, I drop the usual subscript i to indicate a single voter.
7
Instead of modeling stability in terms of variance, I use precision (the inverse of variance)
because precision more closely aligns with the concepts of the model.
8
In this instantiation of the model, voters are assumed to retain all considerations. Forgetfulness
(i.e., the FIFO queue) is revisited in succeeding sections. However, forgetting period 1 considerations
can be approximated by increasing ψ 2 , thus increasing the impact of the candidate signal.

18
I model the candidate’s announcement as a normal curve centered on γ and with
precision ψ 2 . The voter’s period 2 issue opinion is drawn from a mixture distribution
with a mean and variance of

δ1 τ12 + γψ 2
E[x2 ] = (2.2)
τ12 + ψ 2
τ 2 (δ1 − E[x2 ])2 + ψ 2 (γ − E[x2 ]2 )
Var[x2 ] = 1 + 1 (2.3)
τ12 + ψ 2

The posterior’s mean (Equation 2.2) is consistent with Bayesian updating if x1 were
a prior and the candidate’s signal were a draw from a normal distribution.
A concrete example of this model, albeit outside of the campaign framework, is
Al Gore’s movie An Inconvenient Truth. Before the movie was released, the envi-
ronment was not often a topic in America’s political discourse. Even liberals, while
generally believing in environmentalism, may have held unstable beliefs about how
many resources the United States should commit to stop global warming. This un-
certainty is depicted by the wide distribution in left panel of Figure 2.2. Gore’s
documentary sends a strong signal to the left of the voter’s mean prior distribution
(center panel). The voter accepts this signal and becomes a fervent environmentalist
(right panel).
Brody and Page (1972) would label this interaction between the liberal citizen
and Al Gore’s documentary “persuasion.” However, since the liberal knew little
about environmentalism before viewing the documentary, she could not have thought
critically about the subject. Instead, the liberal accepted the cue from Al Gore, with
no ability to judge the quality of the arguments. Following other scholars (e.g. Gilens

19
 Voter's Prior Beliefs Candidate's Signal Voter's Posterior Beliefs
Density of Beliefs

Density of Beliefs

Density of Beliefs
−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0

Position on Environment Position on Environment Position on Environment

Figure 2.2: A hypothetical liberal’s response after viewing An Inconvenient Truth.

The liberal’s initial position (left panel) is moved close to the position espoused Gore
(center panel) who is a favored politician. The right panel displays the resulting issue
opinion distribution in bold.

and Murakawa, 2002), I label the shift of opinion by the liberal as “cue-taking.” I
reserve the term “persuasion” for circumstances akin to “central route processing”
(Petty and Cacioppo, 1981), in which a voter’s issue opinion shifts based on critical
analysis of the current set of issue considerations.
Because most people go about their lives generally oblivious to many political
issues, cue-taking is prevalent in politics. Campbell et al. (1960) emphasize a similar
point when they note, “For many voters the details of the political landscape may be
quite blurred until they are brought more into focus during the campaign period” (pp.
135-136). If the campaign environment induces cue-taking, then a straightforward
application of the model predicts that campaigns should polarize public opinion.
Consider a campaign environment in which both the Democratic and the Repub-

20
 lican candidates provide opposing signals to two voters (also one Democrat and one
Republican) about an issue. The voters only accept the signal from the candidate
they trust more.9 If the voters have unstable opinions about the issue (i.e., little
personal experience), then their opinions will polarize (Figure 2.3). This conclusion,
supported by existing research (Ansolabehere et al., 1994; Parker-Stephen, 2007), is
crucial: When voters do not have experience on an issue and this issue is empha-
sized during the campaign, the electorate polarizes and individuals’ existing political
worldviews are not altered.

Voters' Prior Beliefs Candidates' Signals of Issue Stances Voters' Posterior Belief
Density of Beliefs

Density of Beliefs

Density of Beliefs

−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0
Issue Opinion Scale Issue Opinion Scale Issue Opinion Scale
(Blue=Dem, Red=GOP) (Blue=Dem, Red=GOP) (Blue=Dem, Red=GOP)

Figure 2.3: Issue polarization. When two opposing candidates send signals (center
panel) to two moderate voters of different parties (left panel), the voters’ opinions
diverge in opposing directions (right panel).

Alternatively, if an individual has experience on an issue, signals from elites may

fail to confirm the individual’s current worldview. Consider the following example of
a campaign between two candidates of the major parties who are debating the issue
9
See Section 2.1.5 for a formalization of this statement.

21
of gun control. The Republican candidate has a moderate position, in contrast to the
Democratic nominee’s more orthodox, left-wing view. Of the two voters analyzed in
this example (Figure 2.4, left panel), the Democrat is a hunter who has experience
with the issue (i.e., precise prior opinion) while the Republican does not (wide prior
opinion). In between periods 1 and 2, the Democratic candidate sends a leftist signal
and the Republican candidate sends a moderate, right-of-center signal (Figure 2.4,
center panel). While the Republican voter cue-takes and moves slightly to the right
to follow the Republican candidate, the Democratic hunter does not move left to any
significant degree (Figure 2.4, right panel). In general, this phenomenon, derived
∂|γ−E[x2 ]|
mathematically as ∂τ 2
< 0, forms the first of the two main hypotheses:
Cue-Taking Hypothesis: Ceteris paribus, when a favored political elite pro-
vides a signal about his position on an issue, the segments of the population that have
the least personal experience with the issue are most likely to cue-take (i.e., alter their
issue position in the direction of the signal).
Returning to the example Figure 2.4, the Democratic voter’s opinion is closer (on
average) to the Republican candidate’s view on this issue, causing a contradiction in
the previously-loyal Democrat’s worldview. To calculate how much more negatively
the Democratic voter would rate the Democratic nominee, one must consider both
where the voter thought the candidates stood on the issues in period 1 and how the
voter judges politicians.

22
 Voters' Prior Beliefs Candidates' Signals of Issue Stances Voters' Posterior Belief
Density of Beliefs

Density of Beliefs

Density of Beliefs
−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0
Issue Opinion Scale Issue Opinion Scale Issue Opinion Scale
(Blue=Dem, Red=GOP) (Blue=Dem, Red=GOP) (Blue=Dem, Red=GOP)

Figure 2.4: Experienced voters do not cue-take. The Democratic hunter does not cue-
take from her preferred politician on this issue because of her highly-informed prior
on the issue. The inexperienced Republican voter shifts her issue opinion toward the
stance of her preferred candidate.

2.1.4 Voters’ Beliefs About Candidate Positions

Voters have beliefs about where candidates stand on the issues; these beliefs become
more certain after receiving signals about the candidates’ platforms. Before a signal is
sent, a voter’s beliefs about the candidate’s positions often suffer from “projection”
(Brody and Page, 1972), whereby the voter assumes that her preferred candidate
holds the same position that they do.10 I do not explicitly model the determinants
of projection, instead allowing the center of the distribution of prior beliefs about
the candidate’s position, µ1 , to be an unparameterized variable. Under complete
projection, µ1 = δ1 .Voter have certainty about these beliefs, υ12 , which most likely
10
In addition to projection, a voter’s prior belief on a candidate’s position may be based on
general factors such as the candidate’s party.

23
increase with political attentiveness and sophistication. Similarly to above, the prior
belief, y1 is assumed to drawn from the normal distribution,

1
y1 ∼ N (µ1 , ). (2.4)
υ12

The candidate’s signal helps the voter identify where the candidate stands on
the issue. The signal has the same center point, γ, and precision, ψ 2 , as in Section
2.1.3 . One might imagine that a candidate’s signal could be unpersuasive (i.e., not
move an individual’s issue position) yet credible (i.e., be informative about where
the candidate stands), so that learning and cue-taking are based on different inter-
pretations of a single signal; but for simplicity, I model these signal characteristics as
the same parameter. Certainly, these qualities could be correlated in certain cases.
For example, a politician who appears to be pandering would produce an un-credible
and non-persuasive issue appeal.
As with the voter’s own issue position, the voter updates her belief about the
candidate’s positions by introducing considerations generated by the candidate sig-
nal. Beliefs at period 1 are treated as priors in period 2. The updated, period 2,
beliefs about the candidate are a mixture model with mean and variance,

µ1 υ12 + γψ 2
E[y2 ] = (2.5)
υ12 + ψ 2
υ 2 (µ1 − E[y2 ])2 + ψ 2 (γ − E[y2 ]2 )
Var[y2 ] = 1 + 1 (2.6)
υ12 + ψ 2

Returning to the case of the Democratic hunter and gun control, imagine that

24
 the Republican candidate microtargets the Democratic hunter, highlighting both
candidates’ positions. Before the signals are sent, the voter might have weak priors
on the candidates’ positions according to their parties (left panel, Figure 2.5). The
Republican sends a mailing to the gun owner’s household indicating a moderate-
right stance for the Republican and an extreme liberal position for the Democrat
(center panel). The Republican has more credibility in revealing his own position
than his opponent’s because the Republican might be held to account for his stated
issue opinion if the Republican wins (Ferejohn, 1986). These signals are combined
with the voter’s prior and result in a posterior belief about the candidates’ positions
(right panel).

Prior Beliefs of Republican Cand.'s Signals of Posterior Beliefs of

Candidates' Positions Candidates' Positions Candidates' Positions
Density of Beliefs

Density of Beliefs

Density of Beliefs

−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0
Gun Control Opinion Gun Control Opinion Gun Control Opinion
(Blue=Dem Cand., Red=GOP Cand.) (Blue=Dem Cand., Red=GOP Cand.) (Blue=Dem Cand., Red=GOP Cand.)

Figure 2.5: Projection combining with one candidates’ signals about both candidates.
The voter starts out with prior beliefs about both candidates’ based on their party.
The Republican candidate signals that he has a moderate-right position and signals,
with less credibility, that the Democrat has an extreme left position. The voter
updates her beliefs about the candidates’ positions accordingly.

25
2.1.5 Voters’ Candidate Evaluations

Traditionally, candidate evaluation is modeled as utility maximization (Downs, 1957)

over a range of issues (Enelow and Hinich, 1984). These types of models, where voters
place weights on a range of issues, are consistent with the issue public literature
(Krosnick, 1990). Members of an issue public place above-average weight on a specific
issue. These models of voter utility require a lot from citizens. First, voters must
know their own position on all the issues. Second, voters must know the politician’s
view on each issue. Third, voters must assign a weight to each issue. The weight-
based model is on-line, with voters keeping track of all this information.
Instead, I propose a more Zaller-esque model of candidate evaluation. Voters
compare a candidate’s issue signal against their considerations of the issue. For
instance, when a candidate delivers a signal about an issue, the voter updates her
belief about the candidate’s position, forming the posterior belief y2 (as in Section
2.1.4). Next, to evaluate the candidate on this issue, the voter compares her prior
belief about the issue to the candidate’s perceived position. This comparison is
formalized with the Kullback-Leibler distance,

∞
x1 (z)
Z
KL(x1 ||y2 ) = x1 (z)log dz , (2.7)
−∞ y2

where z is the issue dimension and, as above, x1 is the voter’s issue position at
period 1. For gaussian distributions, the Kullback-Leibler distance has the analytical

26
solution,

    2 
1 τ1 υ2 2
KL(x1 ||y2 ) = log + + (δ1 − µ2 )υ2 − 1 , (2.8)
2 υ2 τ12

where µ2 and υ22 are the mean and variance of y2 (Equations 2.5 and 2.6). How-
ever, y2 only approximates a unmixed Guassian distribution is candidate stances are
highly credible (i.e., ψ 2 /υ 2 >> 1). In addition, Equation 2.8 is an idealized, contin-
uous version of the discrete process of comparing a voter’s considerations about an
issue to considerations of a candidate’s platform. The use of the Kullback-Leibler
distance helps clarify the distinction between a Zaller-esque model in which entire
distributions are compared to an implicitly Bayesian model in which means (or other
summary statistic) would be compared (e.g., Enelow and Hinich, 1984).
Candidate stances on issues for which the voter has a vague prior will have little
impact on the voter’s worldview. The uncertainty of the voter’s position means the
the Kullback-Leibler distance will not be extreme in either distance or closeness.
Once the voter cue-takes on this issue (Section 2.1.3), the K-L distance will shrink,
reinforcing existing predispositions.
Candidate evaluations may not be reinforced when the voter has a precise prior
on the issue in question. In this case, the voter compares her beliefs with the can-
didate’s and the difference may be stark. An apt analogy is that the voter treats
her personal experience as the “ground truth” for judging the elite’s reliability. One
incongruent signal from a preferred candidate might be enough to induce the voter’s
defection. Understanding the full voter-candidate interaction, requires the consider-
ation of multiple issues.

27
 To calculate the voter’s overall utility of a candidate, all considerations on the
top of a voter’s head must be taken into account, and these considerations may be
on multiple issues. Let the scope of J issues be represented by j = 1...J and the
positions of the voter and the candidate at time t be xjt and yjt respectively. The
voter need not have all these issues on the top of her head at one particular time.
Whatever combination of issues is on her mind determines the voter’s candidate
evaluation, which is formally,

J
X
U (yt· |xt· ) = − δj ∗ KL(xtj ||ytj ), (2.9)
j=1

where δj is the proportion of a voter’s considerations dealing with issue j. The

voter need not take an active role in consciously setting δj for all issues; for issues
the voter is unfamiliar with, δj = 0. In a two-sided issue debate, the model assumes
voters only receive (positively informative) signals from the candidate with the higher
U (yt· |xt· ).
As noted, a candidate signal on a new, unfamiliar issue would lead to a small
K-L distance after cue-taking, thus reinforcing existing predispositions. However, if
considerations this new issue replace considerations of an older issue for which the
voter also took an elite cue, candidate utility would not change at all.
The chance of a switch in candidate support is particularly acute in three situa-
tions. In the first, the voter is nearly undecided in the first period; e.g., KL(x1 ||y1D ) ≈
KL(x1 ||y1R ), where D and R superscripts represent the two candidates. In this case,

28
a small change in the relative KL distances results in a change in vote choice.
Second, priming may occur; “priming” is a traditional political science explana-
tion of why voters alter their opinions of candidates once a signal is received (Iyengar
and Kinder, 1989). The signal about a particular issue (j) causes considerations
about that issue to become more accessible in period 2. When asked to evaluate the
candidates, the voter relies more heavily on these “primed” considerations; formally,
δj increases.
Third, if the voter has a precise prior about only the one issue at hand (i.e.,
the voter has no experience on other issues), then the voter is susceptible to micro-
targeting by the non-preferred candidate. Consider the example of the Democratic
hunter. The Republican’s signals about gun control indicate that the Republican
candidate is closer to the voter on this issue. Remaining loyal to the Democratic
candidate would cause cognitive dissonance; the hunter does not cue-take from her
initially preferred candidate (the Democrat) since she is set in her ways. The equa-
tions above demonstrate that the voter will alter her perceptions of the candidates
to favor the Republican. Transforming the voter’s worldview and cue-taking on all
the other (non-experiential) issues from the Republican candidate causes no disso-
nance. Hence, this Democratic hunter is suspectable to defection by a well-targeted
Republican appeal.
This dissertation focuses on the third mechanism: Voters’ political worldviews
depend upon candidates’ stances on the issues with which the voter has experience.
This phenomenon results in the following prediction, which is the main hypothesis
of the dissertation:

29
 Candidate Evaluation Hypothesis: Ceteris paribus, after receiving a candi-
date’s signal on an issue, the segments of the population that have the most personal
experience with the issue alter their perceptions of the candidate more than the seg-
ments of the population with no personal experience.
Political worldviews originate with and change based on issues for which voters
have experience. For these issues, voters know the “ground truth” and need not rely
on political elites for an indication of what to believe.
This hypothesis might be particularly relevant during primaries, where no parti-
san differentiation exists. Consider the hypothetical example of a primary election
between two candidates with an electorate of three voters, each of whom has ex-
perience on a single issue. The 20-day campaign moves through three phases and
focuses on two issues: health care and education. The three voters are Alice the
architect, Ted the teacher, and Doris the doctor; they must decide between two the
candidates, Mandy and Nathan. Figure 2.6 tracks the hypothetical voters’ candidate
evaluations over the course of the campaign.11 They start with opposing predispo-
sitions about the candidates, with Alice favoring Mandy, Ted favoring Nathan, and
Doris undecided. These predispositions may be attributable to valence evaluations
(e.g., competence, experience) or other traits (e.g., shares my values), which in turn
may originate from identity politics (e.g., voter-candidate shared gender or race).12
Candidate traits are modeled just as issues are; for traits, voters are very certain
about where they stand (e.g., competence is good) but less certain about politicians’
11
See Appendix for the exact parameter values that generated the figure.
12
In a general election, these predispositions are often aligned with partisan allegiances and are
very strong. A primary campaign magnifies the effect of issues.

30
stances (e.g., claims of competence are cheap talk).
The left-most portion of Figure 2.6 displays these starting preferences. No issues
considerations are on the top of voters’ heads on the initial day of the campaign
(phase 1).
Relative Evaluation of Candidates (Mandy − Nathan)

More Pro−Mandy
2

Alice the
Architect
0

Doris the
Doctor
Ted the
Start of Campaign

Teacher
−2

More Pro−Nathan
−4

Health Care Phase Education Phase

5 10 15 20

Days of Campaign

Figure 2.6: A hypothetical primary campaign with three voters, two candidates, and
two issue phases (three phases total). In the first issue phase, when health care is
discussed, the initially undecided doctor (Doris) sides with the candidate (Nathan)
who concurs with her stable opinion on health policy. The other two voters polarize.
In the second phase, when Education is the hot topic, the teacher (Ted) switches
candidate allegiances because his previously preferred candidate (Nathan) takes what
Ted considers an incorrect stance.

In the next phase of the campaign (days 2-10), the issue of health care becomes
prominent and politicians stake out positions. Considerations related to health care
are added to the top’s of voters’ heads. The only voter with experience on this issue is
Doris (who is a doctor). She sides with candidate Nathan on the issue, and increases

31
her relative evaluation of him. The other two voters polarize in their opinion of the
candidate, as the presence of actual issue information (rather than trait information)
appears to solidify their respective choices.
After the candidates have finished debating health care, education policy becomes
relevant in phase 3, days 11-20. Alice and Doris cue-take from their preferred can-
didate; Ted the teacher sides with his previously unfavored candidate, Mandy. The
education debate eventually crowds out health care considerations. By this point, on
the last day of the hypothetical campaign, Ted switches candidate allegiances and
supports Mandy over Nathan.
This simulation illustrates the futility of campaigns’ attempts to alter the minds
of voters on a host of issues with which voters have no experience. In the normal
course of a campaign, two-sided elite messaging on issues unfamiliar to voters leads to
polarization, just as Alice and Ted diverge in the health care phase of the campaign.
A real-world example from the 2008 general election is the sudden emphasis on wealth
redistribution after the third debate between Obama and McCain (in which “Joe the
Plumber” was referenced 24 times). Despite this shift in campaign dialogue, the polls
barely moved.13

2.1.6 Issue Experience as Signals

In the Personal Experience Model candidate signals and issue experiences are varia-
tions on the same theme. They are both considerations that are received by voters
13
See Charles Frankin’s moving average (http://www.pollster.com/polls/us/
08-us-pres-ge-mvo.php) for the flat McCain and Obama lines before and after the Octo-
ber 16, 2008 debate.

32
and subsequently accepted or rejected (Figure 2.1). The most important distinction
between these two types of considerations is that candidate signals’ acceptance de-
pends on the voter’s evaluation of the candidate, while the acceptance of personal
experiences hinge on the proximity of the voter-issue interaction.
When a voter interacts with a non-politicized issue, the individual’s issue opinion
proceeds analogously to the logic of Section 2.1.3. A real-world experience (e.g., a
teacher that sees the roof of school cave in) and generates an additional consideration
to her issue opinion distribution (e.g., schools need more funding because they are
crumbling). How much this consideration affects the individual’s opinion depends
on the weight (i.e., precision) of the new consideration (ψ 2 ) relative to the weight of
existing considerations (τ 2 ).
With the model formally explicated, the three properties of personal experience
discussed in Section 2.1.2 can also be formally linked to greater issue stability. First,
repeated considerations on at the same point on the issue spectrum, create a narrow
distribution of experiences and a higher precision, ψ 2 . Second, easier processing of
experiences leads to higher reception of considerations, and again a tighter distribu-
tion of experiences and a larger ψ 2 . Third, since voters forget considerations over
time, repeated considerations lead to voters having experienced-based considerations
on the top of their head more often. As long as the precision of experiential consider-
ations (ψ 2 ) is greater than the background noise considerations (τ 2 ), increasing the
proportion of experiential considerations will lead to lower issue opinion variance.14
14
One type of experience that this dissertation does not focus on is a one-time, high-impact
experience. Examples include having an abortion or being (close to) a victim of violent crime.
If these considerations are forgotten at a slower rate than other consideration (e.g., mass media-
based), then the math operates in the same way as if these experiences were repeated, though more

33
 ∂Var[x2 ]
Formally, ψ 2 > τ 2 → ∂ψ 2
< 0. Infrequent, forgettable voter-issue interactions
may increase the variance of a voter’s opinion.
The process of issue opinion formation becomes more complicated when the in-
dividual has experience with an already politicized issue. For instance, assume a
conservative parent has a narrow prior on the issue of gay marriage because the issue
is often discussed in the political arena and this individual accepts only conservative
arguments. If the parent’s child “comes out of the closet”, then the parent’s personal
experience will directly conflict with their existing issue opinion. If the parent ac-
cepts these personal experience considerations, then their issue opinion distribution
becomes bimodal, and the parent may have a difficult time deciding where to stand
on the issue (Figure 2.7).15 This bimodal distribution does not generate as clear pre-
dictions for candidate evaluations as the situations discussed in Section 2.1.5 where
personal experience occurs before issue politicization. As Figure 2.7 display, personal
experience can balance political predispositions – a result found even in the highly
politicized environment of Congress (Washington, 2008).

2.1.7 Model Extensions and Details

The model, as presented in its most limited form, can be applied in alternative
situations. The model works just as well when considering political parties, news
organizations, or politically active citizens rather than candidates and is flexible
quickly forgotten experiences.
15
The initial “if” in this sentence is a “big if” in that the parent’s prior distribution would
normally mean that they would not accept liberal consideration. Only because the source of these
considerations is so close to them might these considerations be accepted, and even then perhaps
at a slow rate of acceptance.

34
 Voter's Prior Beliefs Personal Experience Voter's Posterior Beliefs
Density of Beliefs

Density of Beliefs

Density of Beliefs
−1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0

Position on Gay Rights Position on Gay Rights Position on Gay Rights

Figure 2.7: A hypothetical conservative parent’s issue consideration distribution after

her child comes out of the closet. The conservative’s initial position on gay rights (left
panel) is balanced by personal experience (center panel). The right panel displays
the resulting issue opinion distribution in bold. Normally, liberal considerations of
the center panel would be rejected by the conservative, but these consideration are
generated by a close, personal situation.

enough to be used in partisan and non-partisan situations.

As modeled, signals from non-preferred candidates do not contain negative in-
formation. As Zaller (1992) demonstrates with opinions about the Vietnam War, if
only the conservative elite is communicating to the public, then even liberal voters
will shift their opinions toward the conservative end of the spectrum. The shift of
liberal opinion is less than the shift in conservative opinion and may reflect the lower
probability of a liberal accepting a consideration from a conservative elite.
The probability of a voter receiving a signal does not appear in the Personal
Experience Model, but the effects of paying more or less attention to politics can
be captured by the precision of a candidate’s signal. In fact, small acceptance prob-

35
abilities are captured in the hypothetical election simulations by assuming weak
candidate signals (small ψ 2 ). Alternatively, an additional parameter and process
could be explicitly added to the model.
The model can be applied beyond political issues. For example, character traits
can be a powerful force in politics because voters judge individuals in their daily
lives. Consider the trait of honesty. Voters’ ideal points, xtj are all essentially
grouped at positive infinity (or the upper limit of the trait dimension)—voters want
their politicians to be extremely honest. Also, most voters know with a high degree
of certainty that they prefer honest people to dishonest people (since they have
experience dealing with both in their daily lives), so τj2 is large. Candidate cues
on this topic often lack credibility (i.e., ψ 2 is low) because politicians often cannot
credibly state “I am honest.” Thus, a credible cue from a media source that a
candidate is dishonest can have a devastating effect on voters’ evaluations of that
candidate.
Character traits are an example of an “issue” with which most of the electorate
has extensive experience. Individuals must judge the honesty, trustworthiness, and
empathy of others on a daily basis. The model’s implications extend to these widely
familiar issues. Aggregating the conclusion from the Candidate Evaluation Hypoth-
esis across the electorate implies that mass communication about issues with which
more voters are familiar (i.e., the “easy issues” of Carmines and Stimson, 1990)
will influence candidate evaluation more than communicating about complex, niche
issues.
Broadcast Corollary: Signals on familiar issues or candidate character traits

36
influence voters’ perceptions more than information about more complex issues.
The Candidate Evaluation Hypothesis and the Broadcast Corollary have direct
implications for political campaigns. If a campaign is aware that a particular voter
has personal experience with an issue, then the campaign would benefit from com-
municating to the voter on that issue (provided the voter and the candidate have
congruent opinions). However, campaigns are often unable to procure this type of
information; instead, they rely on mass communication such as television advertis-
ing. In the case of mass communication, campaigns would do well to follow the
Broadcast Corollary and advertise about issues that intersect with the lives of the
greatest number of people.

2.2 Concurrence With Other Theories in the Lit-

erature

The literature substantiates much of the the Personal Experience Model, especially
the claims about issue opinion stability (Stability Postulate), cue-taking (Cue-Taking
Hypothesis), and easy issues (Broadcast Corollary). Circumstantial evidence sup-
ports the experience-issue interaction (Candidate Evaluation Hypothesis), yet no
study directly tests the effect of personal experience on cue-taking and candidate
evaluation. Chapters 3, 4, and 5 focus on providing novel evidence for the Candidate
Evaluation Hypothesis; the following sections review the existing evidence.

37
2.2.1 Definitions of Concepts in the Literature

Over the decades, various terms (e.g., “projection,” “persuasion”) have represented
distinct concepts. Before exploring the existing research, it is useful to review and
define some of the terms that appear, noting their relationship to the Personal Ex-
perience Model.

• Issue Opinion Stability is the precision of the distribution of repeated sam-

ples of a person’s accessible considerations of an issue. A helpful way to think
about opinion stability is to imagine asking an individual to report an opinion
on a repeated basis. The variance of her reported opinions is the inverse of her
issue stability.

• Self-Reported Issue Opinion Confidence is an individual’s belief about her

stability on the issue. Confidence should not be conflated with issue stability
as individuals are often poor judges of their own characteristics.

• Candidate-projection is the phenomenon whereby a voter assumes her pre-

ferred candidate agrees with her on an issue that she has relatively little infor-
mation on. This mechanism often occurs before candidates send strong signals
on the issue. See “projection” in Markus and Converse (1979).

• Learning is the process by which the voter ascertains the true issue position
of the candidates. It occurs after the voter receives strong signals on the issue.
See Lenz (2006) for an excellent discussion.

• Cue-taking is the process whereby the voter agrees with her preferred candi-

38
 date’s issue position without any critical thinking. See “projection” in Iyengar
and Kinder (1989), “persuasion” in Markus and Converse (1979), and “periph-
eral route processing” in Petty and Cacioppo (1981).

• Persuasion on an issue opinion occurs when the voter thinks critically

about issue considerations and rejects them in favor of other considerations.
Replacement of accessible considerations over time is not considered persuasion.
See “central route processing” in Petty and Cacioppo (1981). Persuasion does
not play a large role in the Personal Experience Model.

• Priming is the mechanism by which voters increase the salience of a particular

issue in the vote decision. This salience may be raised directly by stressing the
importance of an issue or indirectly by mentioning the issue and thus increasing
the accessibility of considerations related to the issue. See Iyengar and Kinder
(1989).

2.2.2 Personal Experience Yields Nuanced and Stable Opin-

ions

The finding that personal (or “direct”) experience with an issue induces attitude sta-
bility originates in the psychology literature. Regan and Fazio (1977) employ both
a field experiment and a laboratory experiment to test the effect of direct experi-
ence on attitude stability. Their experiments randomly assigns the direct experience
treatment. In the field experiment, college students are assigned to permanent or
temporary housing. In the laboratory experiment, some participants are assigned

39
puzzles to work on, while others received only descriptions of the puzzles. In both
experiments, the reported attitudes of the participants assigned to the “direct expe-
rience” treatment were more consistent with the participant’s actions than for the
students who were assigned to the peripheral experience.
Wood (1982) examines how existing experiences affect the amount of variation
in issue opinions. Subjects in this experiment are asked to “to list specific instances
of times when they had engaged in actions relating” to the environment and to
report their opinions about environmentalism. A month later, Wood requests that
the subjects prepare a persuasive, pro-environmentalism talk in exchange for five
dollars. On average, this offer of money induces subjects to report views that were
more anti-environmentalism than their original position (presumably because the
offer of money makes the arguments appear fraudulent). However, individuals with
prior experience with the environmental movement were less likely to change their
opinion.
Observational studies demonstrate that individuals with experience on an issue
have more nuanced and stable opinions, supporting the Stability Postulate. Sotirovic
(2001) conducts a telephone survey of 395 Wisconsin adults to examine their attitudes
toward crime policy and the death penalty. The respondent’s answers to open-ended
questions are coded for number of distinct ideas and integration between ideas; the
answers are then aggregated into a “complexity” measure. Respondents who have
experience with non-trivial crimes (either personal or through friends) demonstrate
higher thought complexity.
McFadyen (1998) conducts in-depth interviews with 67 employed Britons to probe

40
their views on unemployment. The author asks participants about stereotypes of the
unemployed, the efficacy of government actions, and their thoughts about possible
solutions. To gauge the level of experience that participants had with unemployment,
McFadyen queries the subjects about their relationships with unemployed family or
“close” friends, as well as any direct experience they had with being unemployed.
A multivariate analysis finds that amount of experience with the unemployed is the
best predictor of how few stereotypes the respondents express, with experience being
more predictive than either ideology or class and leading to fewer stereotypes.
In both the crime and unemployment studies, experience does not lead to po-
larization of issue opinion. In fact, just the opposite occurs. In the crime study,
Sotirivic finds that the respondents with the most extreme views had the least com-
plex thought processes. And personal experience with the unemployed does not push
the Britons in McFadyen’s study to the left on unemployment policy; political ide-
ology is a much better predictor of issue opinion than level of issue involvement.
Consistent with these findings, the direct effect of experience on issue position is not
specified by the personal experience model.
The impact of personal experience on issue opinion stability is directly measured
by van Knippenberg and Daamen (1996). The authors poll the attitudes of the Dutch
on energy generation twice, through surveys two months apart. In each survey, re-
spondents are given a range of six methods of generating energy and are asked to pick
two. In the first survey, brief descriptions of each option are provided. Van Knip-
penberg and Daamen measured issue involvement on a “slightly modified version”
of the scale developed by Verplanken (1991), which queries respondents about their

41
activities with respect to the issue (e.g., talking with friends about the issue). While
they do not provide exact statistics, the authors report that higher involvement leads
to increased stability across the two-month period.
Taken in sum, these studies demonstrate that direct experience, personal in-
volvement, or even the experience of close friends or family members induce a more
complex thought process about an issue. The resulting opinions of individuals with
personal experience are more stable than those of the public at large. Thus, the
evidence for the Stability Postulate is strong.

2.2.3 Evidence for Cue-Taking and the Role of Experience

The psychological literature provides the basis for cue-taking. Similar to Bartels’
(2002) findings that the same information generates polarized responses between
members of opposing parties, Lord et al. (1979) show that conflicting studies on
capital punishment have a polarizing rather than moderating effect. Undergradu-
ates who hold prior beliefs in favor of capital punishment consider the pro-capital
punishment study more convincing; consequently, their issue opinions become more
extreme. The opposite holds for the participants who enter the experiment with
anti-death penalty views. Similar to the Al Gore and global warming hypothetical
in Section 2.1.3, individuals generally accept arguments that are congruent with their
existing beliefs and adjust their views accordingly.
Political science research provides several examples of cue-taking. Zaller (1992)
demonstrates how elite consensus leads to mass consensus. In contrast to Zaller’s
one-stream example, campaigns provide an example of a two-stream environment.

42
As the election was heating up in the summer of 2004, Democrats and Republicans
diverged over even the relatively factual issue of whether Saddam Hussein was in-
volved in the September 11th attacks (Parker-Stephen, 2007). In a striking example
of issue opinion following elites from one end of the spectrum to the other, Laud-
erdale (n.d.) finds that a change in the party that controls the White House alters
whether conservatives or liberals favor an activist foreign policy.
In one of the clearest examples, Lupia (1994) provides evidence of cue-taking
in the 1988 California car insurance elections. Voters who have little knowledge
of the substance of the ballot initiatives, yet know where various interests groups
stood on the measures, mirrored the behavior of high-knowledge voters. Individuals
who lacked both substantive and endorsement awareness, on the other hand, devi-
ated from the voting patterns of the more knowledgeable groups.16 Small cues from
advocacy organizations help citizens cut through the clutter of political issues.
The Cue-Taking Hypothesis, however, does not merely assert that cue-taking
occurs; it predicts that personal experience moderates the effects of cue-taking. Ex-
periments from psychology provide evidence. In Wood’s (1982) experiment on the
environment, the participants are asked to list their personal connections with an
environmental issue, e.g., membership in the Sierra Club or Audubon Society. All
subjects are then exposed to a fabricated recording of an interview with a graduate
student in which the graduate student presents several arguments against preser-
vation. Wood finds that subjects who report more personal experience with the
16
In this case, a cue from an untrusted source (e.g., the insurance industry) appears to provide
information to individuals about what not to believe. However, since two streams of information
exist (e.g., Ralph Nader countering the insurance industry), it is unclear whether the voters are
using positive or negative cues.

43
environment are less susceptible to the graduate student’s persuasion even when
controlling for subjects’ initial views.
One of the experiments used to differentiate between “central route” and “pe-
ripheral” processing speaks directly to the Cue-Taking Hypothesis. Undergraduate
participants are presented with a set of arguments for a change in university policy.
While some students are under the impression that the policy would affect their
school, others are told that the policy is under consideration at a nearby institution.
The arguments (all against the policy) vary by logical soundness and overall appeal.
Students who are personally involved in the issue (i.e., those under the “own school”
condition) discriminate between the compelling and non-compelling arguments, while
those not involved are persuaded by the speaker regardless of the argument’s quality.
Perhaps both self-interest and issue experience play a role in inducing students to
think critically about the arguments, but cue-taking is clearly at work in the “other
school” condition.
The Cue-Taking Hypothesis dovetails well with the Lupia and McCubbins (1998)
model of persuasion, which predicts that less persuasion (i.e., cue-taking) should oc-
cur when “perceived speaker knowledge” is low. If personal experience increases a
voter’s knowledge (or self-perceived knowledge) of an issue—a reasonable assumption—
then the Cue-Taking Hypothesis is supported by the experiments that Lupia and
McCubbins (1998) conduct in the labratory.
Shifting to the Candidate Evaluation Hypothesis, Mutz (1992) finds an inter-
esting interplay between personal experience with the economy and political eval-
uations. She traces the effects of experience with unemployment and knowledge

44
gained from the mass media through the perception of the economy and incum-
bent politicians (governor and president). Consistent with the Personal Experience
Model, being unemployed has an impact on state-level judgments. At the national
level, the mass media has much more influence in shaping people’s opinion about
the national economy. Specifically, Mutz writes “These overall patterns suggest that
personal experience may indeed have an indirect influence on political evaluations
at the state level by means of their influence on personal and ultimately social con-
cerns. At the national level, however, the path that translates these concerns to
political significance is incomplete.” This latter finding might appear to refute the
Candidate Evaluation Hypothesis, but being unemployed is a stronger signal of the
state economy rather than the national economy (since the reason for the unemploy-
ment might be localized). The finding that individuals act rationally by discounting
their personal experiences when making national judgments is encouraging for any
hypotheses grounded in learning models.
While the direct evidence that personal experience moderates cue-taking is estab-
lished in the psychology literature, minimal research exists on linking personal ex-
perience to political evaluations. Gilens and Murakawa (2002) characterize the state
of research by noting, “We are not aware of any research that explicitly compares
citizens’ decision-making with regard to high-involvement versus low-involvement is-
sues.” The Personal Experience Model and the empirical examples presented in later
chapters aim to fill that void.

45
2.2.4 Easy Issues and Retrospection Affect Political Evalu-

ations

Existing research also supports the Broadcast Corollary. Kelleher and Wolak (2006)
combine presidential approval ratings and content analysis of media stories to ex-
amine the priming effects of easy and hard issues. Since the dependent variable
(presidential approval) is dichotomous, the Broadcast Corollary would predict that
changes attributed to priming would be observable only for familiar (or “easy”) is-
sues. On complex (or “hard”) issues, individuals who approve of the President would
take the President’s side and continue to approve of him; individuals who do not ap-
prove of the President would act analogously (assuming two issue streams). Thus,
even if considerations of the complex issue are more likely to be on the top of a
person’s head (Zaller, 1992), there would be no change in the correlation between
issue opinion and presidential approval before and after the increase in media stories.
On the other hand, individuals’ opinions of familiar issues are not as susceptible to
cue-taking; thus, issue priming by the media might alter their perceptions of the
president.
Kelleher and Wolak’s findings are consistent with this line of reasoning. When the
media reported on easy issues (e.g., the economy, the President’s character), Kelleher
and Wolak observe the priming effects. On the other hand, their data do not exhibit
the effects of priming when the media emphasizes hard issues (e.g., foreign policy).
Lenz (2006) finds similar results when he examines over a dozen cases of suspected
priming and determines that priming occurred in only a handful of cases. Of the
13 examples of supposed priming, Lenz argues that just four are actual examples

46
of priming: two character issues, the economy in 1992, and defense spending after
9/11. Scholars consistently label character issues and the economy as easy issues.
The priming of defense spending is trickier: perhaps the lack of a competing stream
of information prevented Democrats from polarizing away from the President.
When elite signals on a topic are one-sided a similar phenomenon to easy issues
occurs. Often, one-sided news occurs because of the resolution of an issue, such as
when there is consensus that a policy succeeded or failed (Canes-Wrone et al., 2001).
This consensus fills the tops of all voters’ heads with one-sided considerations; if these
considerations are prevalent (such as constant reporting on the war or economy), the
resolved issue mimics an easy issue. The Broadcast Corollary predicts that these
resolved issues have large effects on vote choice.
Empirical research confirms this prediction. Mueller (1970) finds that retrospec-
tive voting on these oft-reported issues (in the spirit of Fiorina, 1981) have a consis-
tent effect on presidential vote, which is a highly partisan contest. Recent research
(Achen and Bartels, 2005) demonstrates that this effect can be so strong that the out-
comes need not be logically connected to actual policy. As the Broadcast Corollary
indicates, candidates who can take advantage of issues that resolved in their favor
are probably better off focusing on those issues rather than microtargeting certain
experienced voters.
The existing psychology and political science literature provides solid support
for the Stability Postulate, the Cue-Taking Hypothesis and the Broadcast Corol-
lary. There is also circumstantial evidence for the Candidate Evaluation Hypothesis,
although the political science literature is lacking on this subject.

47
2.3 Alternative Voter-Issue Linkages in the Liter-

ature

While much of political science literature focuses on the characteristics of a voter

(e.g., attentiveness) or an issue (e.g., complexity) separately, strains of the literature
emphasize linkages between voters and issues. Specifically, research on self-interest
and issue publics theorize why voters might rely on certain issues more than others
in their political decisions. However, studies demonstrate that self-interest does not
affect vote choice in a consistent manner.
The Personal Experience Model generally captures the evidence of the self-interest
literature. With regard to issue publics, the Personal Experience Model provides a
logical foundation for the conclusions reached by the issue publics literature; this
logic also broadens the scope of the theory to include both issues and character
traits.
One perspective on issue-voter interaction is that voters base their political deci-
sions on self-interest. Chong et al. (2001) demonstrate that self-interest in a policy
can be primed or unprimed depending on the political communication delivered. In-
dividuals with a vested interest in a certain policy change (e.g., domestic partner
health benefits) display a higher propensity to favor the change after their self-
interest in the change is highlighted. Although the authors do not measure voters’
evaluations of candidates who take a position on the policy changes tested in the
experiment, one logical conclusion to the research might be that politicians should
communicate to voters on issues in which the voters have a self-interest and that

48
these communications should emphasize that connection. On the other hand, Chong
et al. also find that self-interested voters can be swayed against their direct economic
interests with sociotropic arguments. In a climate with message streams (one that
primes self-interest and one that primes sociotropic considerations), it is unclear how
much self-interest would affect political judgments.
Consistent with Chong et al.’s finding that political communication can reduce
the salience of self-interested consideration, Sears and Funk (1990) find that self-
interest is rarely a factor in vote decision; instead, symbolic politics often drives
individuals’ policy stances. The Personal Experience Model predicts the influence
of symbolic politics: Voters cue-take from their respective group’s leaders, polarize,
and do not alter their political evaluations. In some cases, however, Sears and Funk
do find that self-interest plays a role in politics—specifically when the benefits are
clear or the stakes are high (i.e., easy issues). By focusing on which issues are easy
for which voters, the Personal Experience Model extends the findings of Sears and
Funk and yields more nuanced implications.
In a second thread of political science research, Krosnick (1990) develops the
idea of issue publics first presented by Converse (1964). Although the vast majority
of voters do not closely follow the political debate on most issues, many voters
find one or two issues to be important. Krosnick emphasizes self-interest, group
identification, and personal values as the bases for how voters choose issue publics.
Voters in an issue public have more stable opinions about the issues and base their
political decisions on these issues; Gershkoff (2006) confirms these findings in an
extensive analysis of issue publics. Malhotra and Kuo (2009) demonstrate that, in

49
the aftermath of Hurricane Katrina, increased personal importance of the hurricane
devastation moderates the effects of partisan polarization and cue-taking, analogous
to the Cue-Taking Hypothesis for personal experience.
The Personal Experience Model shifts and extends the idea of issue publics in two
ways. First, the level of personal experience explains why a person enters an issue
public. For instance, in the preface of Gershkoff’s (2006) dissertation, she relates
a story about a Russian hairdresser who was in the immigration issue public. Left
unsaid is the high likelihood that the hairdresser was in the immigration issue public
because of her personal experience emigrating. Second, the Personal Experience
Model provides does not rely on large differences in issue salience to affect political
preferences.
In the issue publics framework, certain issues are crucial for political judgments
because of the conscious weight that a voter places on them. These issues are most
likely to be reported as “most important” on a political survey. In contrast, in the
Personal Experience Model, the mechanism by which some issues are more predictive
of political preference than others—the desire to avoid cognitive dissonance—is the
result of life-happenstance. Voters often have these experiential issues foisted upon
them (e.g., receiving Social Security checks, emigrating from Russia) rather than
choosing to be a member of an issue public. The Personal Experience model focuses
on issues—big or small, inherently political or not—that form the bases for voters’
“ground truth” for judging politicians.
Psychology research posits an additional alternative hypothesis. Fazio and Zanna
(1978) propose that confidence is an intervening factor between the experience-issue

50
stability link. Individuals with direct experience with a topic are more likely to
hold higher confidence in their attitudes as well as to be more likely to match their
behavior to their attitudes. More pertinently, when Fazio and Zanna exogenously
increase confidence in an attitude (independent of how the attitude was formed),
the correlation between attitude and subsequent behavior increases as well. Thus,
voters who have more confidence in all their opinions (regardless of whether this
confidence is justified) may exhibit less cue-taking than would be otherwise expected.
If this research is correct, while the Personal Experience Model would still hold as
presented (because direct experience would lead to confidence which would moderate
cue-taking), part of the picture would be unmodeled (i.e., more confident people are
less prone to cue-taking). This causal mechanism is explored in Chapter 4.
Undoubtedly, further alternative hypotheses to the personal experience model
exist. But, having dealt with the two main alternatives in the political science
literature, this dissertation now focuses on finding direct support for the hypotheses
and implications of the Personal Experience Model.

51
Chapter 3

Experienced Evaluations and

Self-Interested Opinions:
Observational Evidence from the
2000 Presidential Campaign

3.1 The Personal Experience Model and the 2000

Presidential Campaign

3.1.1 The Campaign About Nothing

From the perspective of political practitioners, the 2000 presidential campaign oc-
curred at just the wrong moment in history. While it would be another two to

52
six years before microtargeting became prevalent in national politics, the 2000 elec-
torate was highly microtargetable because there was no single issue that dominated
the campaign discourse. Gore and Bush each focused their attention on several is-
sues, sometimes talking past each other. While the campaigns delivered competing
talking points about Social Security privatization, Bush also emphasized military
readiness (with Gore largely silent), and vice versa on the Patients’ Bill of Rights.
These three campaign issues do have one common feature: they each exhibit evidence
of issue-driven vote preference among the segment of the population with experience
on the issue. This main finding is based on the Annenberg 2000 dataset, which
includes panel data recorded before and after the parties’ nominating conventions.
The key finding of this Chapter is that specific segments of the electorate—
individuals with personal experience on a political issue—were susceptible to target-
ing by the campaigns. When Bush or Gore emphasized an issue in their convention
speech, experienced voters learned about that the candidates’ stances on that issue
faster than the rest of the public and judged the candidates that issue. This evi-
dence provides an important caveat to research that discounts the issue-driven effects
(Lenz, 2006) in a manner consistent with the Candidate Evaluation Hypothesis of
Chapter 2.
Voters in the 2000 electorate could not agree on a single most-important issue. In
one poll, the modal response (representing 20% of responses) to the question, “What
is the most important problem facing the country” was education—an area in which
the federal government has little say.1 This result presents a contrast to the 2004
1
ABC News. November 1, 2000. http://abcnews.go.com/images/pdf/836a16Tracking16.
pdf

53
campaign, when the war on terrorism and the Iraq war combined to become the
single most important issue for a third of the electorate, with the economy not far
behind. In the 2008 campaign, after the mid-September financial crisis the economy
mattered most to over a third of the public.
In this varied issue landscape, it is not surprising that Bush’s August 3, 2000
convention speech was largely devoid of issues while Gore focused on several small
issues in his August 17 convention speech. Bush spent the bulk of his speech defining
who he is. For instance, he explained the term “compassionate conservative” with
multiple vignettes. This tactic appeared to work, as Bush opened up a double-
digit point lead after the convention. However, the Republican nominee’s speech did
touch upon four concrete policy proposals: education, Social Security privatization,
military readiness, and taxes. On military readiness, Bush said in part,

Our military is low on parts, pay and morale. If called on by the Com-
mander in Chief today, two entire divisions of the Army would have to
report, “Not ready for duty, sir.” ... The world needs America’s strength
and leadership, and America’s armed forces need better equipment, bet-
ter training, and better pay.

Bush made an especially forceful push for Social Security privatization.

Social Security has been called the “third rail of American politics,” the
one you’re not supposed to touch because it shocks you. But if you don’t
touch it, you can’t fix it. And I intend to fix it. To seniors in this country:
You earned your benefits, you made your plans, and President George W.

54
 Bush will keep the promise of Social Security. No changes, no reductions,
no way. Our opponents will say otherwise. This is their last, parting
ploy, and don’t believe a word of it. Now is the time for Republicans and
Democrats to end the politics of fear and save Social Security, together.
For younger workers, we will give you the option—your choice—to put a
part of your payroll taxes into sound, responsible investments.

Gore, in direct contract to Bush, touched on dozens of issues in his nomination

acceptance speech, some rather minor (e.g, tobacco advertising to children, Internet
privacy). The convention as a whole was a success, giving Gore a lead in the polls
that he would hold until to the first debate in October. Perhaps the lasting legacy
of the convention was Gore’s coining of the oft-parodied term “lockbox”:

At a time when most Americans will live to know even their great-
grandchildren, we will save and strengthen Social Security and Medicare—
not only for this generation, but for generations to come....

And to me, family values means honoring our fathers and mothers, teach-
ing our children well, caring for the sick, respecting one another—giving
people the power to achieve what they want for their families.

Putting both Social Security and Medicare in an iron-clad lockbox where

the politicians can’t touch them—to me, that kind of common sense is a
family value.

Even though the media focused on the “lockbox” and hence Social Security, Gore
actually devoted more of his speech to vilifying health maintenance organizations

55
(HMOs) and to rallying support for a Patients’ Bill of Rights:

And I will never forget a little boy named Ian Malone—who suffered from
a medical mistake during childbirth, and needs full-time nursing care for
several years. I met him and his parents in Seattle, near their home
in Everett, Washington. Their HMO had told the Malones it would no
longer pay for the nurse they needed, and then told them they should
consider giving Ian up for adoption.

That’s when his mom and dad got really mad. They told their story in
public, and the HMO was embarrassed. Because they fought for their
baby, today Ian has the care he needs to stay alive. But no family in
America should have to go on national television to save their child’s life.

Dylan and Christine Malone are here with us tonight. Ian’s here, too.
And I say to them, and to all the families of America: I will fight for a
real, enforceable Patients’ Bill of Rights.

It’s just wrong to have life and death medical decisions made by bean-
counters at HMOs who don’t have a license to practice medicine, and
don’t have a right to play God. It’s time to take the medical decisions
away from the HMOs and insurance companies and give them back to
the doctors, the nurses, and the health care professionals...

Big tobacco, big oil, the big polluters, the pharmaceutical companies, the
HMOs. Sometimes you have to be willing to stand up and say no - so
families can have a better life.

56
In the week after the convention, Gore produced two television commercials about
the Patients’ Bill of Rights, one of which focused on the case of Ian Malone.

3.1.2 Issue-Driven Vote Choice

Political science has its own version of the chicken and egg question: Which comes
first, the vote preference or the issue opinion? Researchers often attempt to find
evidence of “priming”: issue-driven effects of a campaign (or other political elites)
emphasizing an issue and thus altering voters political judgments through the issue’s
increased salience (e.g., Iyengar and Kinder, 1989). As discussed in depth by Lenz
(2006), reverse causation is a troubling problem when attempting to discern the ef-
fects of issue opinion on vote preference. As with cue-taking (discussed in Chapter 2),
voters often adopt the issue opinions of their favored candidate; therefore, analyzing
the correlations of simultaneous issue opinion and vote choice is problematic.
A more useful approach involves correlating previous issue opinion with change
in vote using panel data. If the correlation increases over time, then issue opinion
must be driving vote choice and not vice versa. A change in vote choice in July, for
example, cannot alter a voter’s issue opinion in June. One caveat is that if the issue in
question already dominates the June vote choice and continues to dominate in July,
the analysis will not detect an effect even though the issue opinion is affecting vote
choice. For analytic purposes, the best issues to examine are those which increase
in salience in between waves of a panel data set. This empirical need dovetails with
the Personal Experience Model, as the model produces cleaner predictions when the
issue starts with few elite-driven considerations (see discussion in Section 2.1.6).

57
 The issues of military readiness, Social Security, and the Patients’ Bill of Rights
were selected for analysis for two reasons. First, they meet the requirements for
being lesser-known issues that were emphasized during the 2000 campaign. Other
issues, such as the licensing of handguns, also fit that description. However, data on
personal experience for these other issues is lacking. For instance, the Annenberg
survey did not ask respondents about gun ownership. The three issues highlighted
here also meet the requirement that data on personal experience be available.
Each of the three issues fits the assumptions of the Personal Experience Model and
can be used to test the model’s predictions. The model assumes that there is little-
to-no political discussion of the issue in period 1 and that subsequently a candidate
delivers a signal about the issue. For each issue, I verify that the public entered the
convention phase of the campaign with little knowledge of where candidates stood on
the issue and that the electorate learned about these stances over the post-convention
weeks. The text of the convention speeches provides evidence that Bush sent signals
about Social Security and military readiness and that Gore emphasized his positions
on Social Security and the Patients’ Bill of Rights.
The Personal Experience Model also requires an observable measure of voters’
personal experience with the issue. Voters who live with active duty members of
the military or veterans are classified as having experience with the military readi-
ness issue. The Annenberg survey does not provide a distinction between various
subgroups of this class (e.g., active duty, veteran, military household).
For Social Security, it is assumed that voters over the age of 65 have experience
with Social Security since they receive a check from the Social Security Administra-

58
tion each month. They have experience with how reliable the government agency is
and how much the money helps them pay for expenses. This measure is of debatable
quality since these voters do not have experience with Social Security privatization.
For the Patients’ Bill of Rights, it is assumed that individuals with health care
through an HMO have experience with this issue. These voters know how helpful
additional rights (such as the ability to sue) would be to them in their dealings with
the HMO. While the Patients’ Bill of Rights would apply to all citizens, (not just
the HMO-insured), Gore closely tied the Bill of Rights to HMOs, believing it was
most relevant for the HMOs’ customers.
With these assumptions validated, the Personal Experience Model makes two
testable predictions. First, experienced voters do not cue-take from politicians as
much as unexperienced voters. This prediction is difficult to verify with the data at
hand because of the self-interest connection between issue experience and issue posi-
tion. Military households would benefit from more military funding (either through
salary or Veterans Affairs). Retirees might believe that their Social Security checks
are at risk under a privatization plan. HMO customers would benefit from further
rights.
Thus, when a candidate emphasizes an issue, an experienced voter with a self-
interest in one side of the issue may be pulled in two different directions. On one hand,
the voter’s issue considerations from experience may help outweigh the candidate’s
argument. On the other hand, the candidate’s point may be of particular importance
to the voter (Chong et al., 2001).
The best example of this second phenomenon is Social Security privatization.

59
Seniors have no more experience with how privatization would be implemented than
other voters have. Thus when Gore claims that privatization would costs seniors their
benefits, these older voters have no experiential considerations to counter this argu-
ment. Naturally, some of them shift their issue position to be against privatization.
This pattern is evidenced in the data.
The second prediction of the model is that experienced voters judge politicians
on the issue for which they have experience (when a politician sends a signal on
that issue). The empirical test of this hypothesis is whether the correlation between
first period (in this case, pre-convention) issue opinion and change in vote preference
is greater for those with personal experience than for those without. This effect is
labeled “differential evaluation”—candidate evaluations among a certain segment of
the population. Differential evaluation is a result of experience voters learning the
candidates’ positions on an issue and, possibly, the issue becoming more salient, or
primed. This latter effect is not required, and the analysis finds only weak support
for a priming separate from learning.
Crucially, the self-interest effect counters the effects of observable differential
evaluation. A senior who shifts her opinion from pro- to anti-privatization and from
Bush to Gore reduces the correlation between prior issue opinion (pro-privatization,
the Republican stance) and change in vote (toward Gore, the Democrat). The key,
verifiable assumption is that experienced voters who shift their issue opinions in the
direction of the Democratic position do not, at the same time, switch their vote
choice toward Bush (or vice versa). If that consistency holds, evidence of differen-
tial evaluation among experienced, potentially self-interested voters is likely to be

60
understated.

3.1.3 Data from the 2000 Annenberg Survey

To isolate these effects, I analyze the Republican and Democratic National Conven-
tion panel from the Annenberg 2000 presidential survey. The Annenberg Center
initially interviewed 1,197 respondents during July as a part of their normal cross-
sectional study. The Republican convention was held from July 31 to August 3 and
the respondents were re-interviewed from August 4 through 13. For the Democratic
National Convention, 1,230 respondents were interviewed first from August 4 to 13
and then after the Democratic National Convention, which lasted from August 14
to 17.
Ideally, those labeled as having experience with Social Security would be respon-
dents who receive a Social Security check, but that question is not directly asked on
the survey. Instead, respondents 65 years old or older are assumed to receive Social
Security checks. This group comprises 16.6% of the cross-sectional national study
and about the same proportion in the panel studies. In the pre-convention period
(May through July), seniors are more likely to support privatization than younger
voters (39% to 61%) are.
The experience measure for the Patients’ Bill of Rights issue is whether the re-
spondent self-reports having their “main insurance through an HMO.” This group
encompassed over 43% of survey respondents in the national crosssection. Prior to
the nominating conventions, HMO customers were slightly more likely to favor “giv-
ing patients the right to sue their health maintenance organization or HMO” than

61
non-HMO customers by 78% to 75% (p < 0.01).
For the issue of military spending, experience is measured by the respondent’s
answer to the question, “Has anyone in your household ever served in the U.S.
military? This includes the National Guard and the Reserves.” About 37% of
the electorate belongs to a military household. These voters are much more likely to
believe the federal government should spend more on the military (54%) than voters
who do not belong to a military household (36%) are.

3.2 Results and Analysis

3.2.1 Learning Over the Course of the Campaign

Data on Social Security and the Patients’ Bill of Rights indicate that voters learned
about these issues over the course of the campaign. Learning was especially prevalent
among experienced voters. For military spending, the necessary questions were not
included in the Annenberg survey, and thus a more circumstantial case is made that
voters learned about this issue as well.
Voters started largely ignorant of Gore’s opposition to Social Security privatiza-
tion. In the first month that the survey asked about Gore’s position (May 19 to June
15, n=605), only 41% of respondents answered “no” to the question “Al Gore–do you
think he favors or opposes allowing workers to invest some of their Social Security
contributions in the stock market?” By contrast, a majority (51%) of respondents
correctly reported that Bush favored personal investment for the retirement program.
Seniors, during this early stage of the campaign, were not statistically significantly

62
more knowledgeable than the rest of the population, though slightly more (56%)
knew that Bush opposed privatization (n=95).
Voters learned about the candidates’ positions during the campaign, through
nominating conventions, debates, and advertisements. In the last three weeks before
the Election, 61% of respondents knew that Gore opposed privatization and 67%
report that Bush favored the policy change (n=2,753). Seniors learned about the
candidates’ platforms to a larger extent, with 72% and 77% correctly reporting Gore’s
and Bush’s positions (n=460). The knowledge differences on this issue between
seniors and younger voters are statistically significant.
Figure 3.1 illustrates learning over the course of the campaign. There was slow
and steady learning through the summer, with an acceleration during the nominating
conventions. Seniors learned faster during the convention period than the rest of the
population. Then, after the conventions, voter knowledge stagnated, only to rise
again—very quickly—starting with the first debate in the beginning of October.
As with to Social Security, the campaign began with voters generally ignorant of
the candidates’ stances on the Patients’ Bill of Rights. The Annenberg survey asked
voters whether they though Bush (and, in a separete question, Gore) supported a
patient’s “right to sue” her HMO. Bush’s position on this topic is genuinely ambigu-
ous as he vetoed a Patients’ Bill of Rights in Texas in 1995, but let a subsequent bill
become law without his signature when a veto-proof majority of the Texas legislature
passed it. However, in the third debate, Bush claimed credit for this law and touted
its provision that gives patients the right to “sue an HMO for denying you proper

63
 Bush's Stance

Dem. Conv.
Rep. Conv.
80%
Gore's Stance Seniors
Only
Percentage of Correct Respondents
Both Stances
70%
60%
50%
40%

All
Respondents
30%

June 1 August 1 October 1

Date

Figure 3.1: Loess smoothed curves of voters’ knowledge about Bush’s and Gore’s
positions on Social Security. While both young and old voters learn about the during
the course of the campaign, seniors gain knowledge faster.

coverage.”2
Gore’s position on the Patients’ Bill of Rights, however, was unambiguous: he
supported the right of patients to sue their HMOs. Despite this clear position, in the
first month of the Annenberg survey, only 42% of respondents knew Gore’s position
(with 38% claiming not to know). By the final two weeks of the campaign, this figure
had increased to 48% (with 31% “don’t know”). Learning among voters who used an
HMO for health insurance was slightly faster, increasing from 42% to 51% over the
2
As President, Bush’s Department of Justice successfully argued in Aetna Health, Inc. v. Davila,
542 U.S. 200 (2004) that Federal law superceded the Texas Patients’ Bill of Rights, thus disallowing
patient lawsuits.

64
same time period. The difference in knowledge at the end of the campaign between
voters experienced with HMOs and those who are not is statistically significant.
Questions about whether Bush or Gore favored increasing military spending were
absent from the Annenberg Survey. Thus, a rigorous evaluation of learning is im-
possible. However, circumstantial evidence shows that voters became more aware
of the issue of defense spending during the Republican National Convention. As
subsequently demonstrated in more detail in Section 3.2.2, the correlation between a
voter’s party and her opinion of the appropriate level of defense spending increased
over the course of Bush’s convention (p < 0.1, one-tailed). Thus, the convention
speeches and coverage must have dealt with this issue. However, the pre-convention
correlation (at 0.3) is the highest of all four issue-convention pairs, so perhaps this
issue was relatively poltically polarized before the campaign started.

3.2.2 Cue-Taking and Self-Interest

Cue-taking is clearly evident in the 2000 campaign, even among experienced voters.
Other evidence lends credence to the assumption made in Chapter 2 that source
evaluations rather than cognitive dissonance is involved in filtering signals. The
convention data finds little support for the Cue-Taking Hypothesis. Rather, Chong
et al.’s (2001) experiments on self-interest are validated. This finding implies that
the results of Section 3.2.3 that support the Candidate Evaluation Hypothesis are
biased downward.
If, as assumed in the Personal Experience Model, that voters accept signals only
from favored politicians, then as issues become politicized, voters would adopt the

65
positions of these politicians. If, on the other hand, voters only accepted congruent
signals on issues, then no voter would alter their issue position. The data from the
2000 campaign supports the former hypothesis.
For each of the four issue-convention pairs, the correlation between pre-convention
vote preference (trichotomous) and issue position increases over the course of the
convention. The panel nature of the data ensures that reverse causation does not
affect these results (i.e., post-convention vote preference is ignored). Only the Social
Security correlation boost during the Republican convention meets standard levels
of statistical significance (p = 0.10, two-tailed). However, as the left panel Figure
3.2 shows, all four shifts in correlation are in the expected direction, and jointly
the analysis is significant (p = 0.05). The right panel of Figure 3.2 displays the
correlations between the more granular 7-point party affiliation and issue opinion;
revealing a similar pattern.
Evidence for the Cue-Taking Hypothesis is weak; in most cases experienced voters
are swayed by self-interest. Of the four issue-convention pairs, three of them have
the expected sign of the Cue-Taking Hypothesis, though only one (Patients’ Bill of
Rights) is statistically significant (p = 0.08). Futher, for the Republican convention,
seniors are less likely than other voters to have stable opinions. This phenomenon
is a result of seniors shifting, in a self-interested manner, from a pro- to an anti-
privatization stance.
Support for the Cue-Taking Hypothesis is measured with a logistic regression
of post-convention issue opinion on pre-convention issue opinion with an experience
interaction term. If the Cue-Taking Hypothesis were true, then experienced voters’

66
 0.4

0.4
Pre Conv. Party ID−Issue Opinion Correlation
Pre Conv. Vote−Issue Opinion Correlation

● ●

●
0.3

0.3
●
●
● ●
●
● ●
●
0.2

0.2
●

●
0.1

0.1
●
●
●
0.0

0.0
Social Security Military Spending Social Security Patients' B.O.R. Social Security Military Spending Social Security Patients' B.O.R.
GOP Conv. GOP Conv. Dem Conv. Dem Conv. GOP Conv. GOP Conv. Dem Conv. Dem Conv.

Issue−Convention Panel Pairing Issue−Convention Panel Pairing

Figure 3.2: Change in predisposition-issue correlations before and after the con-
ventions. Predispositions are measured by trichotomous vote choice (left panel) or
7-point party affiliation (right panel). Issue opinion measures have three, four, and
two possible values for Social Security, Miliary Spending, and Patients’ Bill of Rights,
respectively. About 1200 cases in each correlation.

initial issue position would be more informative for their post-convention position
and the interaction term would be positive. Figure 3.3 displays the interaction term
coefficients and standard errors. (Issue opinions are dichotomized for ease of com-
parison.)
The positive result for the Patients’ Bill of Rights holds even when controlling
for alternative explanations. Potentially, voters who watched the convention might
have been more likely to shift their opinions, but such an interaction term was in the
opposite direction and near zero. Also, more informed voters may have had more
stable opinions; the evidence supports this idea but does not diminish the important

67
or significance of the experience (coefficient increases slightly, p = 0.07). The positive
coefficient is not solely an artifact of there being fewer non-HMO individuals who are
against a Patients’ Bill of Rights as fewer experienced individuals defected from that
position than did unexperienced individuals.The result also disappears when other
issues (such as Social Security or military spending) are entered into the regression
in place of support for a Patients’ Bill of Rights, thus eliminating the possibility that
HMO-experienced individuals happen to have more stable opinions across the entire
political spectrum.
Interaction Coefficient for Experience Moderating Cue−Taking

1.5
1.0

●
0.5

●
0.0
−0.5

●
−1.0
−1.5

Social Security Military Spending Social Security Patients' B.O.R.

GOP Conv. GOP Conv. Dem Conv. Dem Conv.

Issue−Convention Panel Pairing

Figure 3.3: Logistic regressions’ coefficients for effect of the interaction of experience
and pre-convention issue opinion on post-convention issue opinion. Issue opinion is
dichotomized. About 1,200 cases in each correlation. Point estimates (circles), one
standard deviation (bold lines), 95%-confidence interval (bold lines) displayed.

Social Security recipients, on the other hand, exhibit less cue-taking moderation

68
or none at all. This finding is especially evidenced by the Republican convention
panel data, in which a significant proportion of seniors were alienated by Bush’s anti-
privatization message. Most likely, these seniors believed that their Social Security
checks were endangered by Bush’s plan, even though Bush explicitly protected seniors
in his policy proposal.
For the Democratic convention data, even with a positive interaction term, shades
of self-interest are apparent. Over a quarter of seniors (27%) who begin the conven-
tion period in favor of privatization, switch their opinion by the end of the conven-
tion. Less than a tenth of younger voters do the same. The interaction coefficient
displayed in Figure 3.3 is positive for this issue-convention pair because among those
initially anti-privatization, more seniors stick with that position. Military families
also experience similar, though less stark, movements in favor of increased defense
spending.
The anti-privatization movement among seniors and the pro-defense spending
shift among the military are consistent with the Chong et al.’s (2001) experiments
on priming self-interest. The Annenberg survey asks respondents whether they agree
with the statement, “Allowing younger people to invest some of their Social Secu-
rity contributions in the stock market will reduce Social Security benefits for older
people.” Even at the end of the campaign (which is the only period the question
is put to respondents), after so much information about Social Security had been
disseminated, 40% of voters agreed that seniors’ benefits would be reduced under
privatization.3
3
The Gore campaign ran an advertisement claiming, perhaps disingenuously, that Bush’s plan
“could cut benefits for seniors.”

69
 As discussed in Section 3.1.2, the finding that experienced voters are also influ-
enced by self-interest biases the results in section 3.2.3 downward. While self-interest
is the likely cause for the issue opinion defection pattern of experienced voters, this
bias would exist no matter what the mechanism. The sufficient condition is that is-
sue opinion shifts are accompanied by like-minded voter preference shifts. The data
clearly show that they are. For example, Gore’s margin increases by 29 percentage
points among seniors who shift their issue opinions against privatization during the
Democratic Convention.

3.2.3 Experienced Voters Judge Politicians on Issues

Consistent with the Candidate Evaluation Hypothesis, the Annenberg data demon-
strates that voters judge candidates on issues with which they have experience. Indi-
vidual issue-convention pair findings near statistical significance; jointly, these data
strongly support the Personal Experience Hypothesis.
Support for the Candidate Evaluation Hypothesis is provided by variation on
one ordinal regression. Change in vote from the pre-convention period to the post-
convention period is regressed on issue experience, pre-convention issue opinion and
the interaction between these two terms. A positive interaction term indicates that
experienced voters relied on their pre-convention issue position to evaluation politi-
cians more than did voters without experience.
Figure 3.4 displays the point estimates (circles) and standard errors for this in-
teraction term for each issue-convention pair. Three of the four have a t-statistic
greater than unity and together these results are significant (p = 0.05, two-tailed).

70
For example the Republican convention data indicate that the average undecided
voter who enters the convention in favor of privatization has a 2% chance of shifting
her vote to Bush. A similarly situated seniors has a 6% chance of altering her vote
preference for Bush. For the Democratic convention, an undecided voter who favors
a patient’s right to sue has a 4% chance of altering her vote in favor of Gore; a sim-
ilarly situated voter who belongs to an HMO health care plan has an 11% chance.
3
Interaction Term of Exp.and Pre−Conv. Issue Position

2
1

● ●

●
0
−1

● All voters
No Issue Shifters
No Issues Shifters or Learners
Social Security Military Spending Social Security Patients' B.O.R.
GOP Conv. GOP Conv. Dem Conv. Dem Conv.

Issue−Convention Panel Pairing

Figure 3.4: Interaction coefficients from ordinal regression of the effect of experi-
ence crossed with pre-convention issue opinion on change in vote choice. Dependent
variable is the subtraction of two trichotomous variables and thus has five levels.
Explanatory variables are dichotomous. Point estimates (circles, squares, triangles),
one standard deviation (bold lines), 95%-confidence interval (bold lines) displayed.

As discussed above, the shift of experienced voters’ issue position due to their
self-interest biases these results downward. Restricting the regressions to voters

71
who stick with their issue position throughout the convention (Figure 3.4, squares),
demonstrates that voters with experience and stable opinions are much more likely
to evaluate politicians on these issues than their stable, inexperienced counterparts.
All coefficients are larger than their standard errors.
These analyses demonstrate that voters’ evaluations hinge on issues for which
they have experience. They do not, however, demonstrate an increase in the salience
of the issue, or priming. Lenz (2006) raises the point that learning, not priming may
be driving vote choice. The Personal Experience Hypothesis takes a similar view—
learning drives both priming and changes in vote choice. To check for the possibility
of priming in the absence of learning, the previous regression is restricted to only
those voters who answered the same number (or fewer) of “candidate position” ques-
tions correctly.
The results from the Democratic convention indicate that no priming occurred
apart from learning (Figure 3.4, triangles). In the case of the Republican convention,
the coefficient for Social Security increases and is statistically significant. Bush’s
convention speech was the first major general election event that mentioned Social
Security privatization, the easiest time to observe priming effects as the issue had yet
to be politicized. Priming effects cannot be tested for the issue of defense spending
because candidate position questions are not asked.
The regression results are robust to alternative hypothesis; specifically, the no
issue shifters condition is tested. The coefficients remain at approximately their
initial values when political knowledge and convention watching are controlled for.
Experience doe not cross issues: for instance the interaction of HMO customers and

72
social security privatization during the Democratic convention has a coefficient of
nearly exactly zero.

3.2.4 Discussion

Observational evidence from the 2000 presidential campaign indicates that experi-
enced voters act differently than voters who lack experience with a given political
issue. Experienced voters learn about an issue faster and as they learn, they judge
politicians on the issue. This learning may lead to priming of the issue, though
priming apart from learning is only observed at the beginning of the campaign.
The 2000 campaign also provides small support for the possibility that issue
priming can occur apart from learning. Early in the election season, political aware
individuals may know where the candidates (or parties) stand on a host of issues,
but may not know which issues will become central to the campaign. The voters
might susceptible to priming by the candidates.
Self-interest also plays a role in the issue-vote choice interaction. Candidates can
emphasize a voter’s self-motivated preference for an issue position, making that voter
more likely to adopt that position. Since self-interested voters are often experienced
voters this effect is in tension with the Cue-Taking Hypothesis. The next Chapter
deals with this concern by identifying voters who have experience with an issue but
no self-interest in the outcome.
As an alternative to experienced voters “ground truthing”, self-interested voters
may want to judge candidates on what the candidates have done for them lately.
The observational data is not helpful in teasing out these distinctions in the mecha-

73
nism behind differential candidate evaluation. The evidence does show, however, the
cleavages in issue-driven vote choice that warrant further investigation by political
science.

74
Chapter 4

Survey Experiments of Princeton

Students and Adults Nationwide

A central aspect of many political campaigns is persuasion: convincing voters who

would otherwise cast a ballot for the opponent to support the campaign’s candidate.
Several micro-theories of American politics (Campbell et al., 1960; Zaller, 1992; Lu-
pia and McCubbins, 1998) detail the difficulty of changing an individual’s partisan
predispositions and imply that campaigns’ persuasive efforts face an uphill battle.
The Personal Experience Model predicts that campaigns’ emphasis on issues with
which voters have personal experience can alter the voters’ evaluations of the can-
didates counter to the voters’ predispositions. Two randomized survey experiments
with hypothetical candidates and real-world issues support the model’s predictions.
This analysis moves away from the partisan biases literature and focuses on why
citizens might change their vote choice during of a campaign. The literature on par-

75
tisan biases has expanded from the fields of political science (Markus and Converse,
1979) political psychology (Taber and Lodge, 2006) and neuroscience (Knutson et al.,
2006). Dozens of studies find that partisans react more positively to arguments from
elites of their own party (see Bartels, 2002, for a review). If Democratic-leaning
voters listen only to Democratic elites and Republican-leaning voters listen only to
Republican elites, then the electorate polarizes. Indeed, this phenomenon occurred
in the 2004 campaign, over the issue of whether Saddam Hussein was involved in the
September 11 attacks (Parker-Stephen, 2007).
Partisan biases abound, but they are not unhelpful for campaigns seeking to
persuade voters to cast a ballot for a specific candidate. The goal of shifting the
electorate (rather than polarizing it) is more relevant for a campaign that is under-
performing in public opinion polls, but even leading candidates may want to increase
their win percentage to scare off challengers (Levitt and Wolfram, 1997). For a cam-
paign to increase its vote share, it must convince voters who would not have voted
for its candidate to cast a ballot for that candidate. Convincing these voters requires
that the campaign increase the standing of its candidate in the voters’ eyes relative
to the other candidate(s).1 Survey experiments demonstrate that certain subgroups
of voters (e.g., veterans) respond strongly to candidate signals on issues with which
they have personal experience (e.g., the Iraq war), and that these interactions in-
fluence voters’ candidate evaluations even if they run counter to existing partisan
biases.
1
These claims assume an electorate of fixed size, but in models that correlate uncertainty of
choice with abstention (Sanders, 2000), increasing a candidate’s standing among the electorate will
increase turnout as well as vote share.

76
4.1 Theory and Hypotheses

Since at least as far back as Downs (1957), political scientists have modeled political
preferences as a function of ideology. This ideology can be thought of as a com-
bination of issues, with voters preferring those candidates whose issue opinions are
closest to their own opinions. Thus, when an individual receives information (i.e., a
“signal”) from a candidate that indicates that he holds a similar position to the indi-
vidual, the individual should improve her evaluation of the candidate. Analogously,
her evaluation should deteriorate when the respondent receives an incongruent cue.
Lemma 1 : A congruent signal from a candidate will improve the respondent’s
evaluation of the candidate. An incongruent signal will lessen the respondent’s eval-
uation. (A congruent signal is defined as information that informs an individual that
the candidate supports the same side of the issue that the individual supports.)
This Lemma assumes that voters receive and incorporate signals about politicians’
ideologies into their issue beliefs. However, Zaller (1992) persuasively demonstrates
that voters filter out a great proportion of political information based on their par-
tisan predispositions. Research on deference to elites (Lupia and McCubbins, 1998)
and cue-taking (Gilens and Murakawa, 2002) shows how voters fundamentally alter
their issue positions based on the stances of their preferred elites. For example, a
large proportion of Democrats were pro-Social Security privatization (and vice versa
for Republicans) before George W. Bush and Al Gore took public stances during the
2000 campaign (Richard Johnston, 2004); after the candidates made public state-
ments, the partisans sorted into their respective positions.
As elaborated in Chapter 2, personal experience on an issue provides a basis on

77
which a voter can judge politicians independent of partisan biases. This political
“ground truthing” of candidates’ statements leads voters to critically evaluate politi-
cians on issues with which they have experience. Examples include retirees who
receive Social Security checks, hunters and guns, Hispanics about immigration, and
veterans about war policy. Recasting the Candidate Evaluation Hypothesis from
Chapter 2 into a prediction specific to candidate statements yields the first of two
main hypotheses of this chapter:
Hypothesis 1 (Candidate Evaluation): Individuals with personal experience with
an issue are more responsive to a candidate signal on that issue than are those with
no personal experience. This responsiveness will be present even when signals run
counter to existing partisan biases (or candidate preferences).
Personal experience also moderates cue-taking from politicians on issues because
the voter’s experiences crowd out elite-generated considerations. For example, as
shown in Chapter 3, during the 2000 election, Republican retirees did not become
ardent supports of Social Security privatization because in their experience the sys-
tem worked well for them. This idea is formalized in the learning model presented in
Chapter 2. Modifying the Cue-Taking Hypothesis to fit the experiments presented
in this chapter yields:
Hypothesis 2 (Cue-Taking): Individuals who have personal experience with a
political issue, when presented with information about their preferred candidate’s
stance on the issue, alter their issue opinion to match their preferred candidate’s
opinion (i.e., cue-take) less than those individuals who lack personal experience.
These hypotheses are extremely important for campaigns that seek to alter the

78
preferences the electorate. If a losing campaign shifts the political discussion to an
issue that people know very little about (even if initial public opinion polls show
support for one position), voters will sort by their existing predispositions. Given
these hypotheses, campaigns would be more successful targeting voters with messages
on issue with which they have personal experience.

4.2 Survey Design

4.2.1 Overall Goals

To test these hypotheses, two two-wave survey experiments are conducted: one pilot
experiment of Princeton undergraduates (n=273) and one survey of adults nation-
wide (n=391). Both experiments are very similar in nature, with respondents of the
first wave being re-interviewed about a month later. All interviews are conducted
online.
The two hypotheses tested have different dependent variables of interest. The
Candidate Evaluation Hypothesis requires that respondents judge politicians and the
Cue-Taking Hypothesis requires that respondents report issue positions. The survey
asks respondents to evaluate fictional candidates and real issues. Fictional candidates
are used for two primary reasons: (1) to limit the amount of prior information,
biases, or predispositions the respondent has about the candidate, and (2) so that
the candidate can plausibly take either side of issues. Since the survey tests the
interaction of personal experience with political judgments, actual issues are used.
Both surveys test two hypothetical candidates and three real issues.

79
 The basic structure of the surveys is: (1) respondents rate hypothetical candi-
dates based on paragraph descriptions, (2) respondents report issue opinions, (3)
respondents are queried on their issue experience, (4) a month later the same candi-
date paragraphs are presented with the candidates taking stances on different issues,
and (5) respondents rate the candidates and report issue positions again. This setup
allows the effect of candidate position-taking on both candidate evaluation and issue
opinion, and the interaction of personal experience with these effects, to be tested.
The month lag between waves limits the respondent’s ability to remember her wave
one responses and lowers the possibility of the respondent attempting to game the
survey.2
Because the candidates are hypothetical and the issues are real, change in can-
didate evaluation is more likely than change in issue opinion. Thus, the surveys are
created with the goal of fostering cue-taking. (See Appendix B.1.1 for candidate de-
scriptions.) First, the candidate paragraphs attempt to mimic the partisan division
of the real political arena. If Democrats feel an affinity to the hypothetical candi-
date, for example, they are more likely to cue-take. Thus, every respondent sees the
same hypothetical match-up between one Democrat and one Republican. Second,
although the candidate signals include no new information on the issue, they are
attempts to reflect the actual debate and are designed to be forceful enough to move
respondents. Third, the evaluations and issue opinions are reported on 101- and
21-point scales respectively to detect small movements. The web format facilitates
2
For instance, if no lag between the waves were included, respondents could purposely report
the same issue positive in both waves to prove that they hold a consistent opinion regardless of the
cues they receive.

80
this granularity by allowing respondents to choose their response via a slider (Figure
4.1.

Figure 4.1: Screenshot of survey question. Respondents report their issue opinion
by sliding a shuttle left or right across the screen. The web browser informs the
respondent of the shuttle current position (which, in the example above, is B:2, or
2 out of a maximum 10 on the B-side of the issue). Candidates are evaluated on a
similar scale, though with more granularity. For both issue opinion and candidate
evaluation, each respondent uses the vertical slider to indicate her certainty of her
reported belief.

In the second wave, the respondent views the entire candidate paragraph de-
scription again, but at the end of the paragraph sees the statement, “[Candidate
name] recently made the following comments about the debate over [Issue]: [Issue
statement].” Each issue has two issue statements related to it, one liberal and one
conservative. Candidates are assigned issues and issue statements randomly with one
constraint: candidates cannot take two stances on the same issue. Thus, for every
respondent, the Democratic candidate takes a stance on one issue, the Republican
candidate sends a signal on a different issue, and the third issue is left unmentioned.

81
Half the time the two candidates send two liberal or two conservative signals (though
again on different issues).

4.2.2 Specifics of Princeton Survey Design

The first wave of the Princeton survey was distributed to 1,600 students on December
6, 2007. A $100 gift certificate to either amazon.com or iTunes was offered as an
incentive to participate, with the winner to be chosen at random from students who
completed the survey. A reminder about the survey was sent out a week after the
initial email, with a second incentive (a $50 gift certificate) for the 400th respondent.
The survey was open for four weeks, during which 418 undergraduates completed the
survey (26% response rate). Students who completed the survey are referred to are
“respondents.”
The candidate evaluation pitted a former governor (Richard Miller, Democrat)
against a Retired Admiral (Mark Jones, Republican) in a race for a U.S. Senate
seat of an unspecified state. Basic biographical information and endorsements were
mentioned in the candidate paragraphs.
Respondents were asked for their opinions on three issues: gay marriage, environ-
mental regulation, and foreign aid. The gay marriage question informed respondents
that their state supreme court had ruled that the state must offer same-sex couples
either civil unions or marriage. The respondent was asked to report which one of
these options she preferred and how strongly. The question about environmental
regulation asked respondents to choose between two solutions to global warming: a
cap-and-trade system or a less restricted, incentive-based solution. With respect to

82
foreign aid, respondents were asked to indicate whether they favored doubling or
halving the foreign-aid budget.
The order of candidates and issues was not random, but differed by class year
and remained constant across both waves. For issues, the relative position of the
traditionally liberal side and conservative side was constant across the three issues
and waves. No statistically significant order effects were found.
To measure issue experience, respondents were asked about their personal habits.
To evaluate experience with gay marriage, respondents reported their frequency of
church attendance and how often they communicate with any gay and lesbian friends
and family. With regard to the environment, students were queried about the num-
ber of outdoor excursions they went on in the past year. To gauge experience in
foreign policy (specifically foreign aid), respondents indicated how much time they
had spent abroad (excluding Europe) in the past five years. Finally, respondents were
asked standard demographic questions about their gender, age, party affiliation, and
ideology, as well as how closely they follow politics.
On February 11, 2008, two months after the initial wave, an email alerted first
wave respondents that the second-wave survey would be available within a day. This
email also provided (with permission) the names of the gift certificate winners of the
first survey. The second-wave survey was delivered to students’ inboxes the following
evening and again was open for a month. Gift certificate incentives were provided,
as with the first survey. Of the 418 first-wave respondents, 283 completed at least
some questions in the second wave (273 finished the entire survey).

83
4.2.3 Specifics of Nationwide Survey

The first wave of the nationwide survey was distributed to 18,812 people on Survey
Sampling, Inc.’s email list on July 28, 2008. A standard incentive (points in Survey
Sampling Inc.’s program) helped motivate individuals to take the survey. A reminder
was send out two days later. Of the recipients, 540 completed the survey within a
3
week of the survey’s launch, for a 3% response rate. At the end of the survey, the
respondents were told that a second survey would be coming “next month.”
The candidate evaluation pitted a former governor (Richard Miller, Democrat)
against the current attorney general (Mark Jones, Republican) in a race for a U.S.
Senate seat of an unspecified state. In addition to biographical information and
endorsements, each description ended with a reason why opposing party leaders
dislike the candidate in question.
After completing the candidate evaluations, respondents were asked for their
opinions on three issues: immigration, the Iraq war, and gay marriage. For im-
migration, the choices were between deporting illegal immigrants and providing a
path to citizenship. On the Iraq war, respondents chose between keeping troops in
Iraq until that country is a “stable, functioning democracy” and withdrawing troops
“as quickly as the Defense Department believes is safe.” For the issue of gay mar-
riage, the respondents were presented with the situation in which a state supreme
court recently mandated legalizing same-sex marriage. The respondents then indi-
cated whether they supported passing a constitutional amendment to overturn the
3
About 25 respondents who completed the survey were eliminated from the analysis because
their responses and the amount of time it took them to record their responses indicated that they
had not read the questions.

84
decision or to recognize gay marriage.
The nationwide survey included a few additional features. First, for each evalu-
ation and opinion, respondents were asked to report the certainty of their response
on a 0 to 10 scale (Figure 4.1). Also, after the candidate descriptions, respondents
were asked to guess what each candidate’s opinion would be on one issue (the same
issue the candidate would take a stance on in wave two).
Several measures of issue experience were recorded. First, respondents were asked
to think about their social network, specifically, those people they talk to when “im-
portant issues or problems come up in life.” After reporting how many people were
in this group of close friends and family, respondents gauged what percentage of their
important conversations were with: individuals currently serving in the military, for-
mer military personnel, immigrants, Hispanics, and gays or lesbians. Conversations
with those currently serving in the military are probably more likely to involve Iraq
than conversations with those who do not serve.
Other measures for issue experience were tested as well. With respect to expe-
rience with immigration, Survey Sampling, Inc. provided the respondents race and
ethnicity (i.e., whether the respondent was of Hispanic descent) and the respondent’s
zip code, which was matched with the percentage of Hispanics in the respondent’s zip
code according to the U.S. census. On wave two, respondents were asked what pro-
portion of their co-workers were Hispanic. For gay marriage, respondents were asked
about their church attendance, whether they considered themselves born-again, and
(in wave two) whether they considered themselves gay or lesbian. For additional
measures on experience with the Iraq war issue, respondents were asked if they are

85
currently serving (or had served) in the military or whether this applied to another
person in their household.
Finally, respondents were asked standard demographic questions about their
party, ideology, and political attentiveness. To determine whether the respondents
represent potential voters, they were asked whether they voted in 2006 and their
intentions to vote in 2008. At the end of the survey the respondent was thanked for
her time and told that a follow-up survey would be delivered to her inbox in about
three weeks.
On August 18, 2008, exactly three weeks after the initial wave, an email alerted
first-wave respondents to the availability of the new survey. Of the 540 first-wave
respondents, 397 completed at least some questions in the second wave (391 finished
the entire survey). These numbers translate into a response rate of 74% on the
second survey.
Finally, respondents were asked to evaluate and comment on the survey. Specifi-
cally, respondents were asked if they felt “forced” to choose an option they “didn’t
particularly like.” Of the 391 respondents who reached that question, four objected
to the issue questions in general, two to immigration specifically, one to the Iraq War,
and two to gay marriage.option for the gay marriage question. Also, 14 respondents
specifically praised the range of options provided for the issues.
The real-world political scene during the period between the surveys was relatively
quiet. The biggest news during the first wave of the survey was the residual effect
of Barack Obama’s Europe trip and the launching of John McCain’s “celebrity”
advertisement (July 30). The second wave closed just as Obama announced the

86
selection of his running mate, Joe Biden. No major events occurred that were related
to any of the three issues tested in the survey.

4.3 Design Checks

4.3.1 Characteristics of the Survey Respondent Population

The characteristics of the Princeton survey’s population are constrained by the fact
that the respondents are all undergraduates at Princeton University. The population
is 54% female and a plurality are first-year students. Only 7% indicate that they
follow politics “a great deal,” with a plurality placing themselves in the middle of
the scale (“some”).
For the national survey, the population is not restricted to registered voters or
likely voters because individuals’ decisions to register to vote or to cast a ballot are
ongoing decisions. For instance, someone might want to register or decide to vote only
after receiving positive information about a candidate (Sanders, 2000). That said,
one goal of conducting a nationwide survey is to produce a respondent population
that is similar to the national electorate. Certainly very good social science can be
performed in a laboratory with a participant group that looks very different from
the population at large. However, treatment effects observed in one segment of the
population might be counterbalanced by opposite effects in other segments of the
electorate who are not included in the experiment.
As a reference point for this, I compare the breakdown of survey respondents
across demographics to similar questions asked in the National Election Pool’s 2004

87
general election exit poll.4 Overall, the demographics of the two groups are strikingly
similar (Table 4.3.1), especially considering the nature of an Internet survey. One
might expect older, poorer, and minority individuals to be underrepresented, as those
groups traditionally use the Internet less. However, these demographics are well-
represented in the survey, meeting or exceeding their percentages in the electorate.
The only troubling statistic is the overrepresentation of non-churchgoers in the survey
experiment: 32% of respondents report that they never attend church, while 15% of
the electorate reports the same. The overall similarity between the two populations
is a benefit of the fact that Survey Sampling, Inc. is a professional organization that
produces nationally representative surveys.
Because Democrats have gained significantly in party identification since Novem-
ber 2004, Pew Research Center’s August 2008 poll is used to compare partisanship.
The two surveys report approximately the same proportion of Democrats and Re-
publicans, although the survey experiment includes more self-reported Independents
than Pew reports. This result may be due to differences in question wording (all
options, including “lean Democrat/Republican,” are immediately presented to the
respondent in the survey experiment, unlike in Pew).

4.3.2 Features of the Survey

Candidate Descriptions. To foster cue-taking in the second wave of the survey,

the paragraph descriptions of the candidates are designed to elicit both positive
evaluations and partisan differences. The nationwide survey fixes some deficiencies
4
As of June 2009, individual data for the 2008 exit poll is not available.

88
 2008 2004 2008 2004
Internet Electorate: Internet Electorate:
Attribute Survey Exit Poll Diff. Attribute Survey Exit Poll Diff.
Gender Church Attendance
Male 43% 46% -3% > Weekly 13% 16% -3%
Female 57% 54% 3% Weekly 19% 26% -7%
≤ Monthly 9% 14% -5%
Race* A few times
White 79% 77% 2% a year 26% 28% -2%
Afr-Amer 14% 11% 3% Never 32% 15% 17%
Latino 13% 8% 5%
Asian 4% 2% 2% Whites: Are you born again/evangelical?
Other 10% 2% 8% Yes 29% 23% 6%
Refused 4% 0% 4%
Have you ever served in the military?
Age Yes 16% 18% -2%
18-29 16% 17% -1%
30-44 25% 29% -4% Are you currently married?
45-59 31% 30% 1% Yes 59% 63% -4%
60+ 28% 24% 4%
Are you gay, lesbian, or bisexual?
Party ID† Yes 6% 2% 4%
Democrat 31% 37% -6% No 92% 45% 47%
Republican 25% 28% -3% Refused 2% 54% -52%
Ind/Oth 44% 35% 9%
Income
Education under $30K 24% 20% 4%
< H.S. Grad 2% 4% -2% $30K-50K 22% 18% 4%
H.S. Grad 18% 22% -4% $50K-75K 20% 21% -1%
Some College 38% 32% 6% $75K-100K 10% 12% -2%
Coll Grad 27% 26% 1% $100K-150K 7% 10% -3%
Post Coll 15% 16% -1% $150K+ 3% 7% -4%
Refused 15% 11% 4%
Religion
Protestant‡ 33% 37% -4% Region
Catholic 22% 19% 3% East 22% 22% 0%
Mormon/LDS 1% 1% 0% Midwest 20% 26% -6%
Jewish 4% 2% 2% South 39% 32% 7%
Muslim 0% 0% 0% West 20% 20% 0%
Other/None 40% 11% 29%
Refused 0% 29% -29%
*Internet survey allowed more than one race to be specified
†From Pew (August 2008) instead of exit poll
‡Includes other Christians

Table 4.1: Demographics of Survey Experiment are Similar to those of the November
2004 electorate.
89
of the Princeton survey to achieve these goals.
For the Princeton survey, respondents of all political affiliations are likely to
rate both candidates positively, although Democrats rate the Democratic candidate
more positively than the Republican candidate (and vice versa for the Republicans).
Independents fall in between the partisans, rating the Democratic nominee more
highly—although this difference does not reach statistical significance (p = 0.15).
On a scale from -10 (worst) to 10 (best), the average rating for the Democratic
candidate is 5.2 and for the Republican candidate is 3.3 (p < 0.01). The correlation
between the two ratings is positive (ρ = 0.14), meaning that the descriptions did
not elicit strong partisan views. This correlation indicates that partisan feelings
are often trumped by the amount of “trust” (or some other valence characteristic)
that respondents afford to hypothetical candidates for which they have only a brief
description.
Across all respondents, the hypothetical candidates were rated slightly positively,
with means of 58 and 56 for the Democrat and Republican respectively (Figure 4.2).
When regressed on 7-level partisanship, the relative evaluation of the Democrat at the
intercept is only 0.2 points above the Republican (a statistically insignificant differ-
ence). Thus, controlling for the partisanship of the respondent, the candidate profiles
are equally positive. Within respondent, the correlation between the Democratic and
Republican evaluations is negative and statistically significant (ρ = −0.16), mean-
ing that in rating the candidates, partisan feelings trump any propensity to rate
hypothetical candidates positively or negatively in general.
Issue Opinion. In both surveys, opinions are relatively split across all three

90
 Mean Evaluation of Hypothetical Candidates
By Respondents' Party Affiliations

100
All Respondents

Mean Candidate Evaluation (0 to 100 scale)

Democratic Respondents
Independent Respondents

80
Republican Respondents
20 40 600

Democratic Candidate Republican Candidate

Figure 4.2: Mean evaluations of hypothetical candidates by party for nationwide

survey. N-size is 397 for all respondents, 258 for Democrats, 66 for independents and
90 for Republicans. Thick lines show plus or minus one standard error; thin lines are
plus or minus two standard errors.

issues. In the Princeton survey, no side of any issue garners more than 60% support.
Students support gay marriage over civil unions 55% to 36% (with 10% neutral).
A smaller plurality supports a cap-and-trade system to regulate greenhouses gases
(49%) over encouraging businesses to be green while letting the market rule (42%;
8% neutral; 1% refused). On the issue of foreign aid, 59% of respondents want to
increase the budget for foreign aid, 21% want to decrease it, and 20% are neutral.
Similar splits are seen in the nationwide survey. On a +10 (most liberal) to -10
(most conservative) scale, the mean responses are +1.4, +1.7, and -1.0 for immi-
gration, the Iraq war, and gay marriage, respectively. Across all issues, the highest
proportion of respondents on one side of the issue is 62% favoring a path to citizen-

91
ship for immigrants. All three issues correlate as expected with party; the Iraq war
polarizes partisans the most (Figure 4.3).

Mean Issue Opinion

By Respondents' Party Affiliations
10
Mean Issue Opinion Liberalness (−10 to 10 scale)

All Respondents
Democratic Respondents
Independent Respondents
Republican Respondents
−10 −5 0 5

Immigration Iraq War Gay Marriage

Figure 4.3: Mean issue opinions by party for the nationwide survey. N-size is 397
for all respondents, 258 for Democrats, 66 for independents and 90 for Republicans.
Thick lines show plus or minus one standard error; thin lines are plus or minus two
standard errors.

Candidate Signals. Hypothetical candidates’ attempts to persuade voters

through a static text cannot be expected to move voters’ issue opinions very far.
Two other factors lower expectations even further. First, the persuasion was not
directly aimed at the respondent; the candidate only states an issue opinion for him-
self. Second, half of the signals emanate from disliked candidates. While the signals
generally work as expected, these low expectations are met in both surveys.
In the Princeton survey, candidate signals prove somewhat effective in two of the
three issues: gay marriage and the environment. Despite the entire survey population

92
moving more conservative on gay marriage and more liberal on the environment,
candidate signals move respondents about 0.5 points (on average) on a 21-point scale
in the direction of the signal (beyond any overall survey population movement). A
similar magnitude is seen for the effect of the conservative foreign aid signal, but the
liberal signal also pushes respondents in the conservative direction.5
For the nationwide survey, the most any signal moves voters is 1 point (on aver-
age) on a 21-point scale (Figure 4.4). Immigration is an ideal case, with the electorate
not shifting opinion between waves and the signals moving opinion by about a point.
The survey population shifts to the left on the Iraq war between waves, but the con-
servative signal dampens this migration. Across the three issues, the liberal signal on
gay marriage is the only signal that does not shift opinion in the expected direction
when compared to the group who does not receive a signal. However, the liberal
gay marriage signal proves effective when the the signal is delivered by a preferred
candidate.

4.3.3 Definitions of Issue Experience

Although the definition of issue experience is somewhat arbitrary, the surveys at-
tempt to elicit the amount of interaction between the respondent and a political
issue. These survey measures can certainly be called into question, especially for
the Princeton survey. The nationwide survey attempts to gauge issue experience
more accurately via either life experience (e.g., veteran and Iraq war) or contact
with people directly affected by the issue (e.g., having close gay friends or family and
5
Foreign aid is consistently an ineffective issue in the Princeton survey, which lead to its omission
in the nationwide survey.

93
 Issue Opinion Shift
By Signal Direction

3
All Respondents
Liberal Signal

Shift in Issue Position (positive=liberal)

Conservative Signal

2
No signal
−2 −1 0 −3 1

Immigration Iraq War Gay Marriage

Figure 4.4: Mean shift in the liberal direction of issue opinions from wave one to
wave two in the nationwide survey. Although the relative effect of the liberal and
conservative signals are in the expected direction in nearly all cases, none of the
signals are particularly powerful.

same-sex marriage).
The Princeton survey has admittedly poor measures of issue experience. For
example, number of outdoor excursions in the past year is the chosen proxy for
experience with the environment. Potentially worse, amount of time spent abroad
(outside North America and Europe) in the past five years is the measure for ex-
perience with foreign aid. The gay marriage measure is probably a more accurate
depiction of experience; respondents are asked how often they communicate with gay
and lesbian friends and family.
The nationwide survey contains several measures of experience, with an emphasis
on conversations with individuals affected by the issue in question. (See Appendix

94
B.1.3 for question wording.) These respondent characteristics are combined into one
measure of experience for each issue. For the issues of immigration and the Iraq
war, the disparate potential sources of issue experience are amalgamated into one
measure using factor analysis. (Factor loadings are described in Table 4.3.3.) For
gay marriage, two factors are present: religiousity (church attendance and born-
again) and interaction with gays and lesbians. Since the latter factor is closer to
the Chapter 2’s definition of personal experience, that measure is analyzed here.
Specifically, listening to sermons on gay marriage in church does not fit the criterion
of active participation, while engaging in conversations with gays and lesbians does.
6
For measures in both surveys, when issue experience is dichotmized for easier
display in figures or hypothesis testing, the top 30% of respondents (approximately)
in each of the three experience measures are labeled “experienced.”
In general, the confounding property of self-interest is avoided. For immigration,
no respondent is a Hispanic non-citizen, thus immediate self-interest is avoided. With
respect to gay marriage, 24 respondents self-identify as gay or lesbian; there are not
removed them from the analysis, but the self-report is not included as additional
experience. Similarly, for the Iraq war, 20 respondents are currently serving in the
military; they are not excluded, but this characteristic is not included in the factor
analysis.
6
The gay marriage experience variable (proportion of important conversations with gays and
lesbians) is standardized just as the factor scores are for the other two issue experience measures.

95
 Immigration Iraq War
Variable Loading Variable Loading
Perc. of convs. with Hispanics 0.88 Perc. of convs. with military .50
Respondent is Hispanic .55 Perc. of convs. with veterans 0.69
Perc. of convs. with immigrants .52 Someone in household serves 0.5
Perc. Hispanic in zip code 0.5
Perc. Hispanic in workplace .38

4.4 Support for Hypotheses

4.4.1 Candidate Evaluation

The changes in candidate evaluations are the crucial piece of this analysis; these
shift indicates how the voters might take a second look at a previously non-preferred
candidate in an election. The signals work well, producing large shifts predicted by
Lemma 1: When the candidate announces an issue stance close to a respondent’s po-
sition, that respondent evaluates the candidate more highly.7 The opposite happens
when the candidate announces an issue stance far from a respondent’s own view.
These two situations are described as the candidate delivering a “congruent” or an
“incongruent” signal. The large (and statistically significant) difference between the
effects of these two types of signals for the nationwide survey is displayed in Figure
4.5. Similar effects are present in the Princeton survey.
Voters with personal experience on related issues exhibit larger swings in their
candidate re-evaluations as the Candidate Evaluation Hypothesis predcits. The data
7
When signal congruity appears in a dichotomous context, congruous signals are only those on
the same side of the issue as the respondent’s wave-one opinion. When the concept is measured
as a continuous variable, it is operationalized as the distance between the respondent’s wave one
opinion and either +10 (for liberal signals) or -10 (for conservative signals).

96
 Candidate Evaluation Shift
By Signal Congruency and Candidate Party

Change in Candidate Eval. (poss. rage −100 to +100)

20
All Respondents
Received Congruent Signal
Received Inconguent Signal

−20 −10 0 10

Democratic Candidate Republican Candidate

Figure 4.5: Mean difference between candidate evaluations in wave one and wave
two. Respondents who are neutral on an issue are mixed in with respondents who
received a congruent cue. N-sizes are 397 overall, with 200 receiving congruent cues
from the Democrat (197 incongruent) and 214 receiving congruent cues from the
Republican (183 incongruent). Thick lines show plus or minus one standard error;
thin lines are plus or minus two standard errors.

can be analyzed two ways: by candidate (and combining issue signals) and by issue
(and combining candidate evaluations). As and example of the former: in the Prince-
ton survey, the average difference in the change in Democratic candidate evaluation
is 2.6 points (21-point candidate scale). For individuals experienced on the issue
on which the Democratic candidate delivered his signal, this change in evaluation is
higher: 3.75 points (p = 0.18, all reported p-vales are two-tailed), as the Candidate
Evaluation Hypothesis predicts. For the Republican candidate’s evaluation, a similar
effect is apparent: a change of 4.1 points for experienced individuals and 2.8 points
for non-experienced individuals (p = 0.32). These results are noisy, but supportive

97
of the hypothesis.
When the evaluations of the Democratic and Republican candidates into one
evaluation margin (e.g.,Democratic evaluation minus Republican evaluation) for each
issue, issue-specific effects are quantified. For instance, the average difference in
Democratic margin from wave 1 to wave 2 when the Democrat delivers a congruent
signal on gay marriage or the Republican delivers an incongruous signal is 5.1 points
(21-point candidate scale).8 For those with experience on the issue, however, this
margin jumps to 7.9 points (p = 0.06). Analogously, for the environmental issue,
experienced respondents alter their candidate margin by 2.9 points in the direction
of the congruity of the signal, while non-experienced respondents alter their margin
by 1.8 points (p = 0.62). Foreign aid exhibits the characteristic of a weak experience
variable, with the margin difference increasing from 3.1 for respondents without
experience to 3.9 for respondents with experience (p = 0.82). The noise in these
supportive results is from the issues of environment and foreign aid.
The nationwide experiment produces similarly supportive but noisy results. As
Figure 4.6 shows, for each issue, experienced voters adjust their evaluations of candi-
dates in the direction of the signal (i.e., upward for congruent signals, and downward
for incongruent signals) further than inexperienced respondents. Assuming the tests
are independent, (implied by the randomized treatment assignment) the relationship
is significant at the 10% level (two-tailed) across all (but not within each) issues.
Regressions (one for each issue) with one interaction term (congruency crossed with
experience), the requisite main effects (in their continuous form), and no controls
8
This figure also includes the change in Republican margin when the Republican delivers a
congruent signal on gay marriage or the Democrat delivers an incongruous signal.

98
yield similar, though slightly less noisy, results.

Candidate Evaluation Shift

By Issue and Respondent Experience

30
Change in Candidate Eval. in the Direction of Signal All Respondents
25 No Issue Experience
Issue Experience
0 5 10 15 20

Immigration Iraq War Gay Marriage

Figure 4.6: Mean difference between candidate evaluations in wave one and wave
two, coded so that the direction of the shift matches the direction of the signal.
Experienced respondents are more responsive to the candidate signals than inexpe-
rienced respondents. Thick lines show plus or minus one standard error; thin lines
are plus or minus two standard errors.

These results validate the main hypothesis, but two additional analyses are crucial
to bolstering the claim. First, respondents who are reacting to traditional partisan
biases must be weeded out so that potential vote switchers are isolated. Second,
alternative hypotheses, such as political attentiveness, should be considered.
A signal is said to go against partisan biases if: (1) it is a congruent signal
delivered by a candidate that the respondent did not favor in the first wave, or (2)
it is an incongruent signal delivered by a candidate that the respondent did favor
in the first wave. An example would be a Democratic respondent who follows her

99
party affiliation in wave one and rates the Democratic candidate more highly but
receives a signal in wave two that the candidate opposes ending the Iraq war. As
the Iraq war is the most polarizing issue, the Democratic respondent mostly likely
assumed that the Democratic candidate supported ending the war. She now lowers
her opinion of the Democratic candidate, and, depending on the signal sent by the
Republican candidate, might even rate the Republican higher.
When restricting the analyses to these type of signals, the result is similar to
Figure 4.6 but noisier. Regression analyses of the two candidate evaluations, limited
to anti-partisan bias signals, finds positive and significant effects in the Princeton
survey (joint p = 0.04).9 The national survey analyses also show both candidate
evaluation regressions with the expected sign on the interaction term, but the data
are more noisy (joint p = 0.46).
The data indicate that the alternative mechanisms for increased candidate evalua-
tion responsiveness to very congruous or incongruous signals—political attentiveness
and opinion certainty—operates as expected. The more attentive a respondent is (or
the more certain she is about her wave one issue position), the more likely they are to
respond to a signal. A regression analysis (Table B.1, Appendix B.3) demonstrates
that these two mechanisms are typically stronger than the increased responsiveness
due to issue experience. Nevertheless, in two of the three issues (the Iraq war being
the exception) the coefficient for the interaction between issue experience and signal
congruity holds the correct sign (with a t-statistic above unity). The Candidate
Evaluation Hypothesis—that issue experience leads to higher responsiveness—holds.
9
Only the two main effects, congruent signal and experience, and the interaction term are in-
cluded in these regressions.

100
4.4.2 Issue Opinions and Experience

The Cue-Taking Hypothesis—issue experience moderates cue-taking—is supported

by the evidence. The data is broken down by issue and whether the signal is received
from the more favored candidate (as defined by the evaluations in wave one), yielding
six tests. The Cue-Taking Hypothesis predicts that in each case, but especially in
signals from favored candidates, individuals without personal experience on the issue
will shift their opinion in the direction of the candidate signal more than those with
experience. In both surveys, this relationship is observed in five of the six cases.
The signals in the Princeton survey all produced (on average) very small move-
ments in issue opinion. The largest amount of cue-taking produced is among inexpe-
rienced respondents on the gay marriage question; these respondents shifted less than
half a point (on a 21-point issue scale) in the direction of a favored candidate’s signal.
When signals are delivered by favored candidates, the average shift is always toward
the candidate’s position except for experience respondents on the environment issue,
who counterargued against the signal. Only signals from unfavored candidates on
the issue of foreign aid generated an unexpected relationship between experience and
cue-taking; that reverse relationship is far from statistically significant.
In the nationwide survey, the immigration signal from a favored candidate yields
a stark contrast between individuals with experience (who counterargue the signal)
and those without experience (who shift toward the candidate’s position by over
one point on the 21-point issue scale). All six relationships are displayed in Figure
4.8; only a signal from a favored candidate on the Iraq war produces an unexpected

101
 Cue Taking By Issue−Experience Pair
1.0

All
Change in Issue Opinion in Direction of Cue

Cue from favored; no exp.

Cue from favored; exp.
Cue from unfavored; no exp.
Cue from unfavord; exp.
0.5
0.0
−0.5

Gay Marriage Environment Foreign Aid

(Friends) (Excursions) (Trips Abroad)

Figure 4.7: Mean shift, in the direction of the received signal, of issue opinions from
wave one to wave two in the nationwide survey. In five of six cases, inexperienced
respondents shift their opinions in the direction of the signal further than experienced
respondents.

102
relationship.10 More counterarguing of signals from unfavored candidates is appar-
ent in the nationwide survey than in the Princeton survey, perhaps because more
partisan feelings were aroused up by the candidate descriptions (which include more
partisan language). Table B.3 in Appendix B.3 breaks down these effects by specific
experience measures.

Issue Opinion Shift in the Direction Of Signal

By Signal Source and Respondent Experience
10

All who Rec'd Signal

Shift toward signal (poss. rage −20 to +20)

From favored/No exp.

From favored/exp.
From unfavored/No exp.
5

From unfavored/Exp.
−5 0 −10

Immigration Iraq War Gay Marriage

Figure 4.8: Mean shift, in the direction of the received signal, of issue opinions from
wave one to wave two in the nationwide survey. In five of six cases, the inexperienced
respondents demonstrated cue-taking more than the experienced respondents.

In contrast to the clear link between cue-taking and experience, there is much less
of a relationship between cue-taking and attentiveness. The phenomenon of atten-
tiveness mediating cue-taking is not statistically significant, and for the issue of the
Iraq war the relationship runs counter to the hypothesized relationship (dichotomized
10
When experience is measured dichotomously and favored/unfavored signals are combined, t-
tests are significant at the 5% level for immigration and gay marriage.

103
results: Figure 4.9). The findings in support of the Cue-Taking Hypothesis are ro-
bust to alternative explanations, as demonstrated by the regression analysis in Table
B.2 in Appendix B.3.

Issue Opinion Shift in the Direction Of Signal

By Signal Source and Respondent Political Attentiveness
8

All who Rec'd Signal

Shift toward signal (poss. rage −20 to +20)

From favored/Inattentive
From favored/Attentive
6

From unfavored/Inattentive
From unfavored/Attentive
−2 0 2 −4 4

Immigration Iraq War Gay Marriage

Figure 4.9: Mean shift, in the direction of the received signal, of issue opinions from
wave one to wave two. Attentive respondents are those who read the a presidential
campaign story at least daily (n=193).

4.5 Discussion

The results of these two survey experiments generally confirm the Candidate Eval-
uation and Cue-Taking Hypotheses: personal experience with issues plays a large
role in the formation of political judgments. To an extent, campaign practitioners
already use this result when they microtarget voters. If a large swath of the elec-
torate cue-takes and hardens (or does not change) their candidate evaluations, then

104
campaigns need not waste money by targeting these individuals with mailings or
phone calls.
Campaigns should especially take note of the support for the Candidate Evalua-
tion Hypothesis. In both the Princeton and nationwide surveys, the average increase
in the magnitude of candidate evaluation shift due to experience is large enough to
switch the preferences of an additional three percent of respondents. In many close
races, three percent is the difference between winning and losing.
Many of the individual results are not statistically significant at conventional lev-
els. Much of this noise is probably due to the use of hypothetical candidates being
used as proxies for real-world politicians. Interestingly, however, the cue-taking re-
sults are less noisy and less susceptible (at least in the nationwide experiment) to
alternative hypotheses than the candidate evaluation results. This difference runs
counter to an intuitive prior belief that because it should be easier to move individu-
als’ opinion of hypothetical candidates, the differences in candidate evaluation among
subgroups should be easier to identify. While the premise of that belief is correct
(both surveys showed that as a percentage of the scale, candidate evaluations shifted
more on average than issue opinions), the implication does not follow. Furthermore,
while many of the results do not independently reach statistical significance, nearly
all point in the direction predicted by the hypotheses.
A secondary finding of note is that inexperienced respondents did not counter-
argue signals from unfavored politicians. The theory on this topic is mixed, with
Zaller (1992) claiming that in a one-issue environment all voters take cues (even
from out-party politicians) and Lupia (1994) finding that voters can glean negative

105
information from endorsements by ideologues on the other end of the spectrum. The
evidence here lends some support to Zaller, but perhaps the one-issue environment
11
tilts the playing field unfairly in that direction.
Another more tentative finding (Table B.1) is that self-reported certainty about
one’s issue position may be the proximate cause of the Personal Experience Model’s
predictions. Here, “self-reported certainty” is an individual’s belief about hew own
issue opinion stability. Psychological studies indicate that personal experience in-
creases self-reported certainty (Fazio and Zanna, 1978). This certainty may provide
the voter the necessary confidence to “ground truth” the politician on particular
issues and not cue-take from elites.

4.6 Conclusion

The hypotheses supported by these survey experiments indicate that campaigns

should identify the issues with which specific voters have personal experience. These
are the issues that significantly affect voters’ evaluations of candidates. Perhaps a
more striking finding relates to cue-taking by voters who do not have experience with
the issues. With cue-taking, even hypothetical candidates are able to cause shifts in
an individual’s issue opinions. Campaigns that emphasize issues with which voters
have little experience risk polarizing the electorate and causing little or no change in
elections results.

11
Lupia’s finding comes from a ballot initiative where even if only one side of the argument is
heard, clearly there are two active sides to the referendum (otherwise one side would not be spending
money to put out its message). In the case of the Princeton and nationwide survey experiments,
each issue receives at most one signal, thereby eliminating the possibility of two-sided debate.

106
Chapter 5

Randomized Field Experiments:

Optimizing Campaign Strategy
and Evidence for the Personal
Experience Model1

5.1 Introduction

For campaigns or political organizations to apply the Personal Experience Model,

they must be able to identify and target individuals with personal experience. The
widening scope of campaign databases, which now include not just basic information
1
Large sections of this chapter are adapted from: Imai, Kosuke, and Aaron Strauss. 2009.
“Planning the Optimal Get-out-the-vote Campaign Using Randomized Field Experiments.” http:
//imai.princeton.edu/research/choice.html

107
such as age and gender, but also commercial (e.g., magazine subscriptions), gov-
ernmental (e.g., fishing licenses), and census (e.g., percent poverty in census block)
data, is both a help and a hindrance to campaigns. While campaigns have more
granular information at their fingertips they may not be equipped to sift through the
mounds of data. This chapter develops a methodological process for identifying vot-
ers who are responsive to campaign appeals by utilizing and discriminating among
information gathered from randomized experiments. Second, this chapter examines
two cases in which the results of field experiments are consistent with the Personal
Experience Model.
Over the last decade, political scientists have shown renewed interest in the use
of randomized field experiments to study voter turnout and persuasion (e.g., Gerber
and Green, 2000; Nickerson, 2007, among many others). Building on work from more
than a half century ago (e.g., Gosnell, 1927; Hartmann, 1936; Eldersveld, 1956), these
researchers have developed creative methods for conducting field experiments with a
large number of voters in real electoral environments as a way to test various theories
of voter turnout (e.g., Gerber et al., 2003; Nickerson, 2008). The empirical findings
of these studies have the potential to significantly affect the practice of get-out-the-
vote (GOTV) efforts and persuasion campaigns in the real world (Green and Gerber,
2008).
Political science research that reports the results of field experiments tends to
focus on the statistical significance of the estimated overall average treatment effect
(ATE) of each campaign appeal. In addition, many researchers implicitly assume
the constant additive treatment effect across individual voters when using regression

108
models for statistical analysis. This leads to the standard practice of reporting a
single estimate (or at most a small number of estimates) summarizing the efficacy of
each mobilization method.
The findings based on such an approach may not be of much use for campaign
planners for two reasons. First, a planner must consider the problem of treatment
effect heterogeneity where each of the available campaign tactics may mobilize dif-
ferent voters to a varying degree. The Personal Experience Model predicts one type
of heterogeneity: Voters with experience on an issue will be more responsive to cam-
paign appeals on that issue. The second and related problem is that a planner faces
budget constraints and must evaluate the cost-effectiveness of available strategies,
which may vary across different voters. Thus, uniformly applying a single appeal to
the entire target population is at best suboptimal and sometimes is not even feasi-
ble. This means that the standard practice of academic researchers reporting only
the estimated overall ATE may not provide useful information from a practitioner’s
perspective. Similarly, the common assumption of constant additive treatment effect
is too restrictive and does not serve the purpose of campaign planners.
This gap between academic research and policy making can be addressed through
an essential and yet missing methodological tool. A formal decision-theoretic frame-
work (Imai & Strauss, 2009) allows campaign planners, both partisan and nonparti-
san, to use the results of field experiments effectively when planning their mobiliza-
tion campaigns. In the proposed Bayesian-decision theoretic framework, a partisan
planner maximizes the posterior probability that the party’s own candidate wins the
election subject to a budget constraint, whereas the objective function of a nonpar-

109
tisan planner is the posterior expected turnout among a target population of voters.
A planner starts with a prior belief about the effectiveness of each mobilization tech-
nique under consideration for voters with different characteristics and updates this
belief based on the available experimental data. With this framework, campaigns
can use the data from randomized field experiments to develop the optimal campaign
strategy for each target population.
This procedure is applied in three cases. First, in a nonpartisan setting, the non-
parametric method finds that recipients of mobilizing text messages in 2006 (Dale
and Strauss, 2008b) who were between the ages of 20 and 24 were most responsive
to the GOTV appeal. Second, in the first of two examples that relate to the Per-
sonal Experience Hypothesis, a mailing campaign about education attempting to
persuade individuals to vote for a Kentucky gubernatorial candidate had the most
effect on parents. And finally, an anti-McCain 2008 mail and robo-call program that
emphasized the economy had the most effect on individuals living in economically
distressed areas.

5.1.1 Background of the Methodological Problem

The problem of optimizing campaign tactics across voter characteristics was first
introduced by Kramer (1966) more than forty years ago. In that research, a cam-
paign manager is assumed to face the decision of whether to conduct blind or se-
lective canvassing across precincts subject to a budget constraint. After illustrating
his decision-theoretic approach with artificial data about the partisan make-up of
precincts, Kramer (1966) concludes that “the use of quantitative methods for policy

110
analysis has proved to be fruitful in many different fields, and these methods deserve
to be more widely known, and used, in political science (p.160).”
Unfortunately, few scholars who conduct field experiments have followed up on
Kramer’s proposal to inform policy makers with a formal decision-theoretic analysis.
One exception is a small number of researchers who have examined the related ques-
tion of which subgroups exhibit larger treatment effects (e.g., Gerber, 2004; Nickerson
and Arceneaux, 2006; Arceneaux and Kolodny, 2007). Campaign managers can use
these findings as a basis for planning their GOTV or persuasion campaign.
However, an important and well-known methodological problem is that if sub-
groups are formed after the experiment is conducted in an ad hoc manner, the anal-
ysis runs a risk of finding statistically significant results when no true relationship
exists (Pocock et al., 2002). Moreover, these studies do not examine the issue of
cost-effectiveness in the presence of a budget constraint, which is an essential consid-
eration for campaign planners. Thus, the planner needs a principled and systematic
approach for deriving the optimal campaign strategy from experimental data while
avoiding the post-hoc subgroup analysis and related statistical problems.
The proposed methodology extends Kramer’s pioneering work by placing the cam-
paign planner’s problem in the formal framework of statistical decision theory and
applying modern statistical methods. The decision-theoretic framework is based on
literature about treatment choice (e.g., Manski, 2005). This framework differs from
the literature by adopting a Bayesian approach rather than a frequentist approach
based on the maximin or minimax-regret criteria (see Dehejia, 2005, for a notable
exception). The standard linear programming algorithm can be used to derive the

111
optimal campaign strategy within this framework.
To address the subgroup analysis problem mentioned above, which has been some-
what neglected in the treatment choice literature a new variable selection algorithm
is presented and then used in combination with nonparametric methods and cross-
validation procedures in order to avoid the over-fitting of statistical models. These
methods are then extended to the cases of optimal partisan campaign planning. This
set of methods can be used to derive the optimal campaign strategy from randomized
field experiments.

5.2 The Formal Framework of GOTV Campaign

Planning

The problem of deriving the optimal nonpartisan GOTV campaign strategy from
experimental data can be formalized as a statistical decision problem where the
choices are treatment applications (e.g., Manski, 2005). This framework can then
be extended to the optimal partisan persuasion (and possible partisan GOTV) cam-
paign. Rather than taking a frequentist approach, which is dominant in the treatment
choice literature, this method is based on Bayesian statistical decision theory (Berger,
1985) and assumes that a GOTV campaign planner learns about the effectiveness
of various mobilization methods for different voters by analyzing randomized field
experiments. The planner then maximizes the posterior expected turnout among a
target population of voters subject to the budget constraint.2
2
The following sections describe the proposed methodology as if a campaign planner both ana-
lyzes the data and makes the decision. The methods apply equally to the situation where a planner

112
5.2.1 The Planner’s Decision Problem

The first step of the approach is to formally state the decision problem faced by a
nonpartisan GOTV campaign planner. Let P denote this planner’s target population
of voters where this population is assumed to be finite and of size N . Typically, the
target population is the registered voters in the electoral district, whose complete
list is available to the planner (at least in the U.S.). Then, the planner’s decision
problem is to assign one of K available mobilization techniques (i.e., treatments)
to each member i of this population. If a GOTV campaign can be planned at the
level of households or precincts, then i indexes appropriate units rather than indi-
vidual voters. Although the units of analysis are represented as individual voters for
notational simplicity, the proposed methodology is directly applicable to aggregate
units.
An unordered set T ≡ {0, 1, . . . , K − 1} where K ≥ 2 denotes the finite set of
mobilization techniques from which the planner makes a selection for each voter in P.
Note that T = 0 represents the strategy of not mobilizing (i.e., doing nothing). For
example, the planner may consider three strategies (i.e., K = 3) where T = 2 and
T = 1 represent a GOTV method based on a phone call and a postcard, respectively,
and T = 0 denotes a status quo strategy that involves no such phone call or mailing.
Alternatively, mobilization techniques may differ in their frequency, timing, content
of messages, and other aspects.
Next, suppose that the planner observes the J-dimensional (pre-treatment) co-
variates X for each member of this population P. Since P is a finite population, this
makes a decision based on the recommendations of a data analyst.

113
means that the planner knows the population distribution of X, i.e., P (X), whose
support is denoted by X . For example, if the target list of voters is obtained from the
voter roll, such covariates may include age, gender, voting history, party registration,
and the zip code. Since the values of these covariates are observable for every voter
on the list, the distribution of X is known to the planner.
Finally, following the statistical literature on causal inference, let Yi (t) represent
the potential turnout of voter i that will be realized if the planner applies the mo-
bilization technique t to this voter where i ∈ P and t ∈ T . Then, the observed
turnout is given by Yi = Yi (Ti ). The outcome variable is binary and is equal to
1 if voter i casts a ballot and is equal to 0 if the voter abstains. For the sake of
simplicity, any interference between voters is ignored (Cox, 1958; Rubin, 1990). But,
this assumption can easily be relaxed. If a voter’s turnout decision depends on the
treatment status of other voters within the same household (Nickerson, 2008), for
example, then an analysis needs to be conducted at the household level.
Given this setup, the planner’s mobilization strategy is characterized by the func-
tion,

δ(·, ·) : (T × X ) 7→ [0, 1], (5.1)

where the mobilization strategy δ(t, x) denotes the probability of receiving treatment
t ∈ T for a voter with X = x. Alternatively, δ(t, x) may represent the fraction of
voters with the observed covariate X = x who are contacted by the planner using the
mobilization method t. These two definitions become essentially identical when the
number of voters is large, but have different implications for the way the optimization

114
is conducted. In either case, voters with the same value of X are assumed to be
exchangeable since the planner does not have additional information to distinguish
between them. Thus, the planner is interested only in determining the value of the
function δ(t, x) for each value of t ∈ T and x ∈ X . Given this definition, the set
of feasible mobilization strategies, ∆, is the collection of functions δ(·, ·) that satisfy
the complement property constraint,

( K−1
)
X
∆ ≡ δ(·, ·) : δ(t, x) = 1 for every x ∈ X . (5.2)
t=0

A nonpartisan GOTV campaign planner’s goal is to derive the optimal mobiliza-

tion strategy to maximize turnout of the target electorate. The planner can achieve
this by deriving the strategy that maximizes the expected turnout given the observed
covariate information about X. Then, the planner’s objective function can be writ-
ten as a function of the mobilization strategy as well as the probability of a voter’s
turnout given the values of covariates and the actual mobilization strategy applied
to the voter,

N
! K−1
X X X
g(δ, ρ) ≡ E δ(t, Xi )Yi (t) X = N P (X = x) δ(t, x) ρ(t, x),(5.3)
i=1 x∈X t=0

where the turnout profile is denoted by ρ(·, ·) : (T × X ) 7→ [0, 1] with ρ(t, x) ≡

Pr(Y (t) = 1 | X = x). The turnout profile represents the turnout probability given
the characteristics of a voter and the mobilization strategy applied to this voter.
Note that the function ρ(·, ·) is unknown to the planner. If the turnout profile is
known (and there is no budget constraint), then the planner applies the mobilization

115
technique t that yields the highest value of the function ρ(t, x) given each voter’s
covariate characteristics x. However, since ρ(·, ·) is unknown in practice, the planner
must make the decision under uncertainty by learning about ρ(·, ·) from experimental
data.
A typical voter mobilization method usually has a small effect on an individual’s
turnout probability relative to the individual’s baseline predisposition to vote. For
this reason, planners may cast the objective function of equation 5.3 in terms of
treatment effect rather than turnout profile. In this framework, the planner maxi-
mizes,

N
! K−1
X X X
g(δ, ρ) = E δ(t, Xi )(τi (t) + Yi (0)) X ∝ N P (X = x) δ(t, x) τ (t,
(5.4)
x),
i=1 x∈X t=1

where τi (t) ≡ Yi (t) − Yi (0) is the treatment effect of mobilization strategy t on voter
i and τ (·, ·) : (T , X ) 7→ [−1, 1] with τ (t, x) ≡ Pr(Y (1) = 1 | X = x) − Pr(Y (0) =
1 | X = x). This setup is mathematically equivalent to equation 5.3. The decision
whether to maximize turnout profile (equation 5.3) or treatment effect (equation 5.4)
hinges on whether the planner can more easily elicit a prior for the turnout profile
or treatment effects (see Section 5.3.2).
In practice, the planner cannot maximize the expected turnout without consid-
ering the differing costs of various GOTV techniques. This concept is formalized by
assuming that the planner faces the following budget constraint,

K−1
( ) K−1
X X X X
1 δ(t, x) 6= 0 κ(t) + N P (X = x) δ(t, x) ξ(t, x) ≤ C, (5.5)
t=1 x∈X x∈X t=1

116
where 1{·} is the indicator function and C is the fixed positive constant representing
the maximum cost allowed for the GOTV campaign. In this formulation, the planner
needs to consider two kinds of costs. The first is the fixed overhead cost denoted
by the function, κ(·) : T 7→ [0, ∞) for each mobilization technique t. This cost
is incurred so long as at least one voter is assigned to the mobilization method.
The second component is the cost per voter, i.e., the cost of contacting each voter
with covariate value x ∈ X , which is represented by a known cost function ξ(·, ·) :
(T × X ) 7→ [0, ∞). Thus, the planner must determine which mobilization technique
is most cost-effective for different voters, and whether the differences are large enough
to warrant using multiple mobilization techniques. Finally, since t = 0 corresponds
to the status quo strategy of not mobilizing, both the overhead and per voter costs
are zero for this strategy, i.e., ξ(0, x) = κ(0) = 0 for all x ∈ X . This is why the
summation in equation 5.5 is taken with all possible values of t except t = 0. For
the other mobilization strategies, the cost per voter is assumed to be positive, i.e.,
ξ(t, x) > 0 for all x ∈ X and t > 0, whereas the overhead cost is non-negative, i.e.,
κ(t) ≥ 0 for all t > 0.

5.2.2 Data from a Randomized Field Experiment

Using the decision-theoretic framework, the planner can analyze the data from a
randomized field experiment to derive the optimal GOTV campaign. Certain as-
sumptions are required to link a field experiment with a planner’s decision problem.
First, the planner must assume that the experiment is conducted on a representative
sample of size n taken from the target population P. Without such an assumption,

117
the planner would be forced to model the non-random sample selection mechanism to
infer characteristics about P from the experiment. The observed data is denoted by
ei }ni=1 where Yei is the binary turnout variable, Tei ∈ T is the treatment
D = {Yei , Tei , X
ei ∈ X is the same set of covari-
variable representing mobilization techniques, and X
ates used earlier. The potential outcomes are denoted by Yei (t) where Yei = Yei (Tei ) for
t∈T.
Another key assumption required for the planner to apply the results of a ran-
domized field experiment to the decision problem is that the joint distribution of
potential outcomes and covariates does not change, i.e., P (Ye (t), X)
e = P (Y (t), X).

This assumption may be invalid if, for example, the election in which the experiment
was conducted differs significantly from the election for which the planner is design-
ing the GOTV campaign. Although in real world settings this assumption may hold
only approximately, it is essential for learning about the planner’s decision problem
from a field experiment.
If these assumptions hold, the randomization of treatments in field experiments
imply that the turnout profile, ρ(·, ·), is identified, i.e., ρ(t, x) = P (Ye (t) = 1 | X
e = x),

for all t ∈ T and x ∈ X . Although such large sample identification results are
important, in practice the planner must estimate ρ(·, ·) from a finite sample and
make the decision under uncertainty. (Similarly, the planner can identify τ (t, x) but
must estimate it from a finite sample.) This problem can be addressed by deriving
the optimal nonpartisan GOTV campaign strategy in this setting, where the planner
is assumed to be Bayesian.

118
5.2.3 The Bayesian Planner

The framework of this method assumes that the planner is Bayesian and has a prior
belief on the space of functions of ρ(·, ·). This prior distribution is denoted by π(ρ).
The Bayesian planner will update her belief via Bayes rule after observing the data
from the randomized field experiment. This posterior belief about the turnout profile
is represented by π(ρ | D). In the Bayesian statistical decision framework (Berger,
1985), the optimal nonpartisan GOTV strategy δ ∗ maximizes the posterior mean of
the expected turnout,

Z
∗
δ = argmax g(δ, ρ) dπ(ρ | D), (5.6)
δ∈∆

where the optimization is subject to the budget constraint given in equation 5.5.
There are several reasons why the Bayesian optimality criteria are used rather
than a frequentist approach based on maximin or minimax-regret criteria that is
popular in the treatment choice literature. First, the Bayesian decision has a fre-
quentist justification. To see this, consider an alternative class of GOTV mobiliza-
tion strategies that depend directly on experimental data as well as the values of
observed covariates. Such strategies are called “statistical treatment rules” and are
characterized by the function δ(·, ·, ·) : (T × X × D) 7→ [0, 1] (Manski, 2005). Thus,
the set of feasible strategies ∆ equals the set of all functions δ(·, ·, ·) that satisfy
PK−1
t=0 δ(t, x, D) = 1 for all (x, D) ∈ (X × D). Under this setting, the frequentist

119
objective function (i.e., risk) is given by,

Z X K−1
X
g(δ, FD , ρ) = N P (X = x) δ(t, x, D) ρ(t, x) dFD (5.7)
D x∈X t=0

with the following budget constraint,

K−1
( ) K−1
X X X X
1 δ(t, x, D) 6= 0 κ(t) + N P (X = x) δ(t, x, D) ξ(t, x) ≤ C,
(5.8)
t=1 x∈X x∈X t=1

for all D ∈ D. Again, the costs for the status quo strategy t = 0 are zero and are
excluded from the budget constraint. It has been shown that in most practically
relevant situations, the Bayesian decision δ ∗ defined in equation 5.6 agrees with
the decision that maximizes the expected value of g(δ, FD , ρ) averaging over the
prior distribution of D on D (see Berger, 1985, p. 159; Manski, 2005, p. 59).
Also, if the size of the experimental data is large (as in many GOTV randomized
field experiments) and little prior information is available, the Bayesian decision is
essentially equivalent to the strategy that maximizes the expected turnout. Thus,
the Bayesian decision can be justified from a frequentist perspective.
Second, an alternative optimality criterion is Wald’s (1950) minimax regret prin-
ciple (see Savage, 1951; Manski, 2005). One important advantage of the minimax
regret criterion is that it avoids the subjectivity of Bayesian optimality because it
does not require the use of prior information. On the other hand, unlike Bayesian
decision theory, frequentist theory based on the minimax regret criterion typically
does not lead to the unique optimal decision, which practitioners may find problem-
atic. In addition, the strategies that meet the minimax regret criterion can include

120
no-data rules, which do not depend on the data at all (Stoye, 2009). Such strategies
do not allow the planner to learn anything from the available experimental data.
Furthermore, a minimax regret rule can be viewed as a Bayes rule with a prior (i.e.,
a least favorable prior) distribution (Berger, 1985, Chapter 5). Thus, depending on
the planner’s subjective belief, the Bayesian GOTV strategy can meet the minimax
regret criterion.

5.2.4 Bayesian Optimal Campaign Planning at A Glance

Figure 5.1 depicts the process by which a planner arrives at the optimal strategy via
the proposed Bayesian decision theoretic framework. The planner must determine
the costs of each mobilization strategy (both overhead and per voter) as well as the
prior belief about their effects on voters with different characteristics. Since the cost
function inputs are often exogenously determined (e.g., the cost of postage and phone
calls), the planner’s only meaningful decision might be determining a prior belief. In
many cases, the planner might use a diffuse prior centered around a belief that there
is no a priori difference in effects of a mobilization method across different voters.
This is especially appropriate if the mobilization technique has not been empirically
tested (e.g., airplanes with reminder-to-vote advertisements). If the treatment is an
oft-used mobilization technique that has been extensively studied in the past (e.g.,
canvassing), then the planner might center a prior around the estimated effects in
previous experiments. Note that the influence of prior belief diminishes as the size
of experimental data increases.
Once a prior belief is elicited and a randomized field experiment is conducted,

121
 Cost Function Input Values

Prior Belief Per Voter

Planner’s Inputs: Overhead Costs
π(ρ) κ ξ
Budget
Constraint

Posterior Optimize Optimal

Experimental Belief Campaign
Data Bayes Rule Expected
D π(ρ | D) Turnout δ∗
Figure 5.1: An Overview of the Bayesian Optimal Campaign Planning Process. In-
puts over which the planner has direct control are represented by hexagons and are:
(1) the prior belief about the effects of various mobilization strategies on different
voters, π(ρ), (2) the overhead costs of each mobilization method, κ, and (3) the cost
per voter for each strategy, ξ. Data from randomized field experiments, D, are repre-
sented by the oval. These data and the planner’s prior distribution are combined via
Bayes Rule to produce a posterior belief about the effects of mobilization strategies
π(ρ | D). Finally, the proposed optimization method uses this posterior belief and
the exogenous costs, ξ, to find the optimal campaign strategy, δ ∗ , for the planner.

these two sources of information are combined via Bayes Rule to obtain the posterior
belief about the effects of mobilization strategies on different voters. The planner
now has an updated belief about the most cost-effective way to mobilize each voter.
Taking into account overhead costs, the optimal strategy may be to implement a
subset of the available mobilization techniques, even if every technique is marginally
optimal for at least one voter.

122
5.3 The Optimal Nonpartisan Campaign Strategy

Using the decision-theoretic framework, the optimal GOTV campaign strategy, as

defined in equation 5.6, can be derived. One relative advantage of the proposed
Bayesian framework over a frequentist’s approach is that the planner can completely
separate the derivation of the optimal campaign strategy from the analysis of experi-
mental data. That is, the planner first uses statistical models to obtain the posterior
belief of the turnout profile, ρ(·, ·). Conditional on this posterior turnout profile,
the planner determines the optimal campaign strategy by solving an optimization
problem subject to a budget constraint.

5.3.1 The Optimization Method

Before describing the method to obtain the posterior turnout profile, this section
describes how to obtain the optimal nonpartisan GOTV campaign strategy given
the posterior turnout profile. Let ρ̃(t, x) be the posterior turnout profile for each
t ∈ T and x ∈ X . Then, the optimal campaign strategy can be obtained by solving
the constrained optimization problem,

X K−1
X
∗
δ = argmax N P (X = x) δ(t, x) ρ̃(t, x), (5.9)
δ∈∆ t=0
x∈X

subject to the budget constraint given in equation 5.5.

To solve this optimization problem, first consider the case of no budget constraint.
In this case, the solution is trivial because the most effective strategy for each stratum

123
defined by X is used. That is, the optimal campaign strategy is given by,

 1 if t = argmax
s∈T ρ̃(s, x),

∗
δ (t, x) = (5.10)
 0 otherwise,


for any given x ∈ X . However, in many cases, the budget constraint may prevent
the planner from employing campaign strategy δ ∗ .
Next, consider the case where the budget constraint is binding so that the strategy
defined by equation 5.10 is not feasible and yet there is no overhead cost. Unfortu-
nately, in this case, the derivation of the optimal strategy is no longer trivial. Thus,
as a general strategy, the planner solves the constrained linear optimization problem,

X K−1
X
∗
δ = argmax N P (X = x) δ(t, x) ρ̃(t, x), (5.11)
δ∈∆ t=0
x∈X

δ(t, x) ≥ 0 for all x ∈ X and all t ∈ T ,





PK−1
subject to t=0 δ(t, x) = 1 for all x ∈ X ,
(5.12)



 NP PK−1
x∈X P (X = x) t=1 δ(t, x) ξ(t, x) ≤ C.


The standard linear programming algorithm can then be used to obtain the optimal
strategy, δ ∗ . If δ(t, x) represents the proportion of voters rather than the probability
of treatment assignment, then this can be formulated as a mixed integer program-
ming problem, which is more difficult but can be solved by applying an appropriate
algorithm. Moreover, as the sample size increases, this difference vanishes.
Finally, when the mobilization methods involved overhead costs, consider a case
where only a subset of mobilization techniques is applied to at least one voter. In this

124
case, the planner can solve the optimization problem in the same manner as in the
case without overhead costs except that the total overhead costs are subtracted from
the maximum budget allowed for the campaign, i.e., C− K−1
P P
t=1 1 x∈X δ(t, x) 6= 0 κ(t).

The optimal strategy is obtained by considering all possible subsets and their corre-
sponding optimal strategy, and then choosing the strategy that yields the greatest
posterior expected turnout.
Although the proposed general strategy based on linear programming is easy to
implement, in some cases an approximate method, which is more computationally
efficient, may be preferred. Appendix C.1.1 details a fast approximate solution to
the nonpartisan planner’s optimization problem, which is used in the nonpartisan
3
application of Section 5.5.

5.3.2 The Statistical Method

Numerous statistical models can be used to obtain the posterior distributions of the
turnout profile. When choosing a model, certain statistical issues need to be ad-
dressed. First, a statistical model needs to be flexible to avoid strong functional
form assumptions. In particular, a nonparametric model is useful because a planner
does not know the exact functional form of the turnout profile a priori. Second,
since many of the covariates available in GOTV field experiments are discrete (e.g.,
party registration, previous turnout history), a statistical model should be able to ac-
commodate discrete covariates. This excludes the use of some binary nonparametric
regression models that require covariates to be continuous. Third, the model fitting
3
The strategy obtained with this approximation is still labeled “optimal” for ease of language.

125
procedure should require a minimal number of arbitrary decisions from a planner (or
the data analyst who is working for the planner).
Finally, a model should not be overfitted to the data at hand and thus an appro-
priate set of pre-treatment covariates must be carefully selected. Overfitting tends
to yield a model with poor performance in the actual election to which the derived
mobilization strategy will be applied. This is the main problem of the post-hoc sub-
group analysis described in Section 5.1. In practice, this consideration is important
because the sample size may not be large enough to accommodate a high number of
pre-treatment covariates, which are potentially useful for deriving the optimal cam-
paign strategy. Thus, an appropriate variable selection algorithm is needed as a part
of the model selection procedure.
Moreover, as Gunter, Zhu, and Murphy (2007) point out, most variable selec-
tion algorithms are developed for improving prediction rather than decision making.
These concepts are closely related, but are not the same. For example, a power-
ful predictor of the outcome, i.e., a predictive variable, is not necessarily useful for
decision making if its effect on the outcome is constant between the treatment and
control groups. Similarly, a variable that explains a significant portion of treatment
effect heterogeneity, i.e., a prescriptive variable, may not be selected by standard
variable selection procedures if it does not predict the observable outcome (rather
than the potential outcome) as well as other variables.
The proposed algorithm meets these criteria relatively well. The method is similar
to that proposed by Gunter et al. (2007) but differs from their algorithm in that it
uses a tree-based method (Breiman et al., 1984) rather than Lasso (in part because

126
the outcome variable in political applications is categorical) and the measure of
importance for prescriptive variables is somewhat different. It is also possible to use
Bayesian regression tree models (Chipman et al., 2008; Hill and McCulloch, 2008).
A Bayesian approach is used to model the turnout among voters with the same
characteristics of covariates via the binomial distribution. For the moment, assume
that the sample size is sufficiently large and thus there is no need for variable selec-
tion. Using a conjugate prior, this turnout model is,

Wtx | T = t, X = x ∼ Binom(ntx , ρ(t, x)), (5.13)

ρ(t, x) | X = x ∼ Beta(atx , btx ), (5.14)

for each t ∈ T and x ∈ X where Wtx is the number of voters with Ti = t and Xi = x
who turned out, ntx is the total number of such voters, and (atx , btx ) are the prior
parameters. This model yields the familiar posterior distribution,

ρ(t, x) | Yi , Xi = x ∼ Beta(Wtx + atx , ntx − Wtx + btx ), (5.15)

where the posterior mean of ρ(t, x) is given by (Wtx + atx )/(ntx + atx + btx ).4
Although its simplicity is attractive, this model is unlikely to work well in practice
if the sample size is small relative to the number of unique values the observable
covariates X take. In particular, if one conditions upon irrelevant covariates, then
the sample size within each subgroup will be too small to yield informative inferences
4
Alternatively, the planner may formalize a prior belief in terms of treatment effects. If a
normal prior distribution is used, then the posterior of τ (t, x) can also be approximated by a
normal distribution.

127
 about ρ(t, x). Such overfitting will then necessarily lead to a mobilization strategy
that will perform poorly in the actual election. On the other hand, if important
covariates are not used to define subgroups, the planner will fail to differentiate
across voters and will choose a suboptimal campaign strategy. Thus in most practical
cases, we seek a principled way to select relevant variables and form subgroups before
applying the above standard Bayesian model.
The proposed solution to this problem involves three steps. First, a variable se-
lection algorithm is applied to decide which variable needs to be conditioned upon
when deriving the optimal campaign strategy. Next, given the selected variables,
a tree-based classification method (Breiman et al., 1984) is fitted to each treat-
ment/control group to identify relevant subgroups within the group.5 The data is
cross-validated to avoid overfitting. Finally, once all subgroups are identified in this
way, the Binomial-Beta model is applied within each subgroup to obtain the poste-
rior distribution of ρ(t, x) for all t and x. This approach is labeled semi-Bayesian
because the data are used twice — once to form subgroups and again to calculate
the posterior distribution. The details of the proposed method are:

Step 1: (Selection of Predictive Variables) Fit a classification tree to the entire

sample using all pre-treatment covariates and the treatment variable. Use K-
fold cross validation on the misclassification rate to determine the value of
the parameter that controls the complexity of the tree, e.g., the complexity
parameter in rpart() implementation in R (Ripley, 1996, Chapter 7). Denote
the predictive variables that are used in the final model by V , i.e., V ⊂ X.

Step 2: (Importance of Prescriptive Variables) Order each pre-treatment covari-

∗ ∗ ∗
ate, Xj for j = 1, 2, . . . , J, based on the statistic, Sj ≡ gj1 − gj0 , where gjk
5
This tree-based classification method is one of many nonparametric models. One disadvantage
of this approach is that gradual changes in treatment effects across covariate groups are modeled
as sharp discontinuities rather than smooth functionals.

128
 is the optimal overall turnout using the turnout profile ρ̂jk (t, x) for k = 0, 1.
Obtain ρ̂j1 (t, x) by fitting a classification tree for the treated subset of the data
(i.e., Tei ≥ 1) and using V and Xj as covariates. Similarly, ρ̂j0 (t, x) is obtained
by fitting a classification tree on the untreated subset of the data (i.e., Tei = 0).
The value of the complexity parameter from Step 1 is also used to fit the trees
in Step 2.

Step 3: (Model Fitting) For each j = 1, . . . , J with Sj > 0:

(a) Randomly divide the sample into K subsets for K-fold cross validation.
(b) Using K − 1 training sets, fit classification trees (without pruning) sepa-
rately to the treatment and control groups using V as well as the j most
important prescriptive variables in both models. Select the values of the
complexity parameters for the two models based on the mean of the opti-
mal overall turnout across K validation sets. Denote the optimal overall
turnout and the optimal campaign strategy corresponding to the selected
values of complexity parameters by gj∗ and δj∗ , respectively.

Step 4: (Derivation of Optimal Strategy) The optimal overall turnout is given by

g ∗ = max1≤j≤J gj∗ , whereas the optimal campaign strategy is given by δ ∗ =
∗
δargmax ∗.
1≤j≤J gj

The first step of the algorithm selects predictive variables using a standard fit-
ting procedure of tree-based methods. The second step orders each pre-treatment
covariate (including those identified as predictive variables in Step 1) according to its
importance as a prescriptive variable. The statistic, Sj , is designed to measure how
much a planner can increase the optimal overall turnout by interacting the value of Xj
with the treatment. This statistic provides a measure of the ability of Xj to explain
heterogeneous treatment effects. The third step uses the K-fold cross-validation pro-
cedure, given all predictive variables and different subsets of prescriptive variables,
in order to select the values of complexity parameters for classification trees fitted
separately to the treatment and control groups. This is done by comparing the opti-

129
mal overall turnout corresponding to different values of the complexity parameters.
Finally, step 4 selects the final model among the ones chosen in Step 3 by again
comparing the resulting optimal overall turnout and thus determines the optimal
campaign strategy.
A main advantage of this semi-Bayesian approach is that it inherits the sim-
plicity of tree-based methods. In particular, practitioners can interpret each of the
subgroups that result from the final model. They can thus use available prior infor-
mation within subgroups by specifying the parameters of the beta prior distribution.
The proposed approach also addresses three key issues highlighted earlier. First, the
tree-based classification models are nonparametric and can handle discrete covari-
ates effectively. Two, the use of cross validation procedure avoids overfitting. Three,
transparent algorithms such as the one proposed here prevent planners from making
arbitrary decisions when deriving the optimal campaign strategy.

5.4 The Optimal Partisan Campaign Strategy

The proposed decision theory framework and the statistical and optimization meth-
ods can be applied to the case of a partisan persuasion and GOTV campaign plan-
ning. A Bayesian planner can derive the optimal campaign strategy using random-
ized field experiments to maximize the (posterior) expected chance of winning the
election. This case assumes that two major candidates are competing for the of-
fice. (Minor party candidates may exist their probability of winning the election is
assumed to be negligible.)

130
5.4.1 The Decision Problem

Using the notation introduced in Section 5.2, the decision problem of the partisan
campaign planner is to assign one of K different mobilization methods (including the
status quo strategy of doing nothing, which is denoted by Ti = 0) to each member
of the target population P of finite size N . Again, the planner is assumed to know
the distribution of a certain set of covariates P (X). Thus, the planner’s mobilization
strategy can be characterized by δ(·, ·) (see equation 5.1) and the set of feasible such
strategies is equal to ∆ (defined in equation 5.2).
Unlike a nonpartisan campaign planner, a partisan campaign planner seeks the
mobilization strategy that will lead to electoral victory. For this decision problem,
the outcome variable Yi needs to be redefined. Let Yi (t) represent the potential voting
behavior of voter i that would be realized if the planner assigns mobilization method
t to this voter where i ∈ P and t ∈ T . The variable Yi (t) can take three different
values; it equals 1 if voter i casts a ballot for the candidate of the planner’s party, −1
if she votes for the opponent, and 0 otherwise (she votes for a third party candidate
or abstaining etc.). Then, the planner’s ultimate goal is to win the election, which
can be represented as the indicator function,

( n ) ( K−1
)
X X X
h(δ, V ) ≡ 1 δ(t, Xi )Yi (t) > 0 = 1 P (X = x) δ(t, x)ν(t, x) >(5.16)
0 ,
i=1 x∈X t=0

P PN
where ν(t, x) ≡ i∈{i0 :Xi0 =x} Yi (t)/ i=1 1{Xi = x} is a random variable representing
the vote share differential for the candidate that will result among voters with covari-
ates X = x if the planner assigns mobilization method t to them. Clearly, h(δ, V )

131
is equal to 1 if the candidate of the planner’s party wins the election, is equal to 0
if he loses. In the statistical decision theory literature, such an objective function is
called “0 − 1 loss function.”
Finally, the partisan planner typically faces a budget constraint similar to that
confronted by the nonpartisan planner, and therefore equation 5.5 also applies to the
partisan planner’s situation.

5.4.2 Data Requirements

As is the case for a nonpartisan GOTV campaign planner, certain assumptions are
required for a partisan campaign planner to be able to use randomized field experi-
ments for the optimal decision. These assumptions are essentially identical to those
described in Section 5.2.2: (1) a field experiment is conducted on a representative
sample from the same target population of voters P, and (2) the joint distribu-
tion of potential outcomes and covariates, P (Y (t), X), remains identical between
the experiment and the actual election. However, one important difference is that
the derivation of the optimal partisan campaign requires vote choice data as well as
turnout data for the voters who are subjects of field experiments; recall that in the
case of a partisan campaign, Y (t) represents a trichotomous variable rather than a
binary variable. For example, in the United States, vote choice data, unlike turnout
data, are not publicly available and cannot be verified for each voter. This means
that a sample survey needs to be conducted to derive the optimal partisan mobiliza-
tion strategy (unless the entire analysis and strategy planning are conducted at an
aggregate level where validated election results are available).

132
5.4.3 Derivation of the Optimal Strategy

The optimal partisan campaign is derived via Bayes theorem strategy, as in the case
of a nonpartisan GOTV campaign. In particular, the optimal strategy maximizes
the posterior probability of winning the election,

Z
∗
δ = argmax h(δ, ν) dπ(ν | D), (5.17)
δ∈∆

subject to the budget constraint given in equation 5.5 where π(ν | D) is the posterior
distribution of the vote share differential ν(·, ·). Using the classification method and
variable selection algorithm similar to those described in Section 5.3.2, the planner
can estimate ν(·, ·), except that the outcome variable is now trichotomous rather than
binary. Most classifiers including tree-based methods can handle such categorical
variables even when the number of categories is greater than two.
The problem, however, is that the optimization in equation 5.17 is not trivial
for two reasons. First, the integration cannot be explicitly solved. Second, the ob-
jective function is an indicator function that is not continuous. These difficulties
are often amplified by the reality that the optimization must be conducted over a
high-dimensional space if the number of treatments and/or the number of subclasses
is large. These computational considerations prevented Kramer (1966) from using
the probability of winning as the objective function of a partisan campaign plan-
ner. Instead, he used the expected plurality of votes as the objective function while
acknowledging that it may not be appropriate. Kramer (1966) noted that “the prob-
abilistic objective is the more realistic. However, this formulation is computationally

133
quite difficult to work with” (p. 141). Indeed, to maximize the expected plurality,
the partisan planner can solve the following constrained optimization problem by
applying the standard linear programming algorithm, as explained in Section 5.3.1,

X K−1
X
δ ∗ = argmax N P (X = x) δ(t, x)E(ν(t, x) | D), (5.18)
δ∈∆ t=0
x∈X

δ(t, x) ≥ 0 for all x ∈ X and all t ∈ T ,





PK−1
subject to t=0 δ(t, x) = 1 for all x ∈ X ,
(5.19)



 NP PK−1
x∈X P (X = x) t=1 δ(t, x) ξ(t, x) ≤ C.


To overcome this computational difficulty, a fast and approximate solution to

the partisan planner’s optimization problem can be applied. This method, detailed
in Appendix C.1.2, is not guaranteed to yield an optimal campaign strategy, but is
relatively fast and closely approximates an optimal strategy.

5.5 Empirical Evaluation of the Proposed Method

To asses how effective is the proposed method of deriving the optimal GOTV cam-
paign work in real world applications the method is applied to three data sets of
randomized field experiments. The intent is to use a randomly selected subset of
the data as a test data set and to obtain an unbiased estimate of the actual turnout
(or the probability of winning) by applying the resulting optimal campaign strategy
derived from the rest of the data to this test data set. This procedure mimics the
real world situation by using the test data set as the actual election to which the
optimal campaign strategy is applied. Since the treatment is randomized and the

134
test data set is not used to derive the optimal strategy, the procedure results in an
unbiased evaluation of the empirical performance of the proposed methodology.
The three applications demonstrate the power and effectiveness of this method.
All examples consist of a single treatment: for more vaired applications see Imai &
Strauss (2009). In each case, the decision for the planner is which voters to treat with
the appeal. The first application is a nonpartisan text messaging GOTV experiment
from the 2006 Congressional election. The second example is a randomized mail
experiment from the 2007 Kentucky gubernatorial campaign. One of three persuasive
mailings was sent to potential voters; two of the mailings are used a “control group” to
analyze the heterogeneous effects of the third mailing, which was on education. The
third application is a program of anti-McCain robo-calls and mailings that focused
on the economy, which were sent to potential voters in 2008. A randomized control
group did not receive any of the treatments. In both partisan examples (the 2007
and 2008 experiments), the outcome measure (candidate support) was gathered via
a survey after treatment delivery.
The partisan applications demonstrate the powerful implications of the Personal
Experience model in the real world. As detailed in Sections 5.5.3 and 5.5.4, parents of
school-aged children were more likely to re-evaluate the candidates once they learned
their stance on education and individuals most likely affected by the economy were
more likely to re-evaluate McCain. In addition to using the nonparametric method,
these results are verified using more standard generalized linear models. The findings
show strong support for the Candidate Evaluation Hypothesis presented in Chapter
2. Matching voters to issues for which they have experience increased the candidate’s

135
margin by 10-15 percentage points.

5.5.1 Evaluation Method

To assess the effectiveness of the proposed method, an additional level of cross-

validation is added to the procedure described i Section 5.3. The aim is to cross-
validate the whole procedure (consisting of the three steps described in Section 5.3.2)
and obtain an unbiased estimate of the resulting turnout under the optimal strategy
from test data that are not used for the derivation of the strategy. After randomly
dividing the sample into L subsamples, one subsample is set aside as a test set
and apply the proposed methodology is applied to the rest of the data. The derived
optimal strategy is then applied to the test set in order to obtain an unbiased estimate
of the resulting overall turnout under this strategy. The random assignment of
treatments and the random subsampling of the test set make the unbiased estimation
possible. The entire procedure is repeated L times using each subset as a test set.
Finally, the average value of the L estimated optimal turnout rates is taken as an
estimate of the turnout that would result under the proposed methodology.
Each application uses a normal-normal conjugate prior for the treatment effect τ
based on the setup defined in equation 5.4. The prior for each subgroup treatment
effect distribution is Gaussian and centered on the estimated population average
treatment effect. The value of the prior variance is chosen so that it increases in
proportion to the per capita budget constraint. In each case, a grid search is imple-
mented to approximate the optimal complexity parameter. A 10-fold cross-validation
procedure is used for Step 3 of Section 5.3.2 to determine the optimality of each com-

136
plexity parameter. In addition, the optimization problem is solved using algorithms
that yield approximate (but fast) solutions (Appendices C.1.1 and C.1.2).

5.5.2 A Nonpartisan GOTV Campaign with a Single Mobi-

lization Method

During the 2006 election, two nonpartisan organizations contributed the cell phone
numbers of newly registered individuals to an experiment that tested the efficacy
of text messages to mobilize voters (see Dale and Strauss, 2008a, for details). The
election was of moderate interest, with at least one gubernatorial or senatorial cam-
paign on the ballot in most states. Subjects were included in the experiment when
they registered to vote with a campus representative of the Student PIRGs or when
they registered online with Working Assets. About 8,000 subjects nationwide were
randomly assigned, with 50 percent probability, to either the treatment group or
control group.
The treatment group received a short text message the day before Election Day.
An example text message read “A friendly reminder that TOMORROW is Election
Day. Democracy depends on citizens like you-so please vote! -PIRG.” The text mes-
sage appeals were varied slightly, but these differences are ignored for this analysis.
Subjects were matched to the voter file using information on their registration forms.
The outcome variable is dichotomous: one for having voted in 2006, and zero for
not voting. The estimated average treatment effect, or more precisely the overall
intent-to-treat effect, is 3.0 percentage points with a standard error of 1.1. Available
covariates for the analysis include: gender, age, race, past voting history, log of

137
county population density, and registering organization.
Following the procedure outlined in Section 5.3, the optimal campaign strategy
is derived using this experimental data set. The classification tree produced in Step
1 chooses the following variables as having predictive power (i.e., ν): age, log of
county density, registering organization, whether the subject had voted in a previous
election, and gender. The prescriptive variables chosen in Step 2 are age, density,
and Hispanic (in decreasing order of Sj ). Not all these variables are included in the
final classification tree produced, however, as searching the complexity parameter
space often finds that models with fewer variables result in higher overall turnout on
validation sets.
The resulting final trees are presented in Figure 5.2. The tree for the control
group is on the left; the tree for the treatment group is on the right. Left branches
represent voters who meet the criteria of the nodes; right branches represent voters
who would falsify the nodes’ inequalities. The leaves show the predicted probability
of voting conditional on their covariates; higher probabilities are to the right at each
node. The control group tree demonstrates that, in this group, voters above the age
of 24 are predicted to vote at a rate of 62%; whether the participant’s age is known
and county population density are important determinants of the voting rates for
voters aged 24 and under. In the treatment group the age cut-point is in between 19
and 20, with population density and voting history also providing information about
turnout rates.
Much of the heterogeneous treatment effect can be explained by the voters’ ages.
Potential voters between the ages of 20 and 24 are very responsive to the treatment.

138
In the control group (left panel of Figure 5.2), these individuals vote at a rate of 62%.
In contrast, the classification tree for the treatment group assigns them a probability
of voting of 66% (the right branch of the right panel of Figure 5.2) – a four percentage
point increase above the average treatment effect. Also, the treatment model predicts
a negligible or negative treatment effect for 18- and 19-year-olds, as their probability
of voting is assigned at most 42% under the treatment unless they live in a county
with a density within a narrow range. (Such age-ranges could not be identified by a
classic model such as logit with linear explanatory variables.)
Figure 5.3 displays the performance of the proposed methodology given the max-
imum proportion of voters that could be contacted. The turnout that would result
under the optimal strategy is estimated using the difference in means estimator be-
tween the treated and untreated voters (solid lines with solid circles). The average
turnout is computed for the treated voters who are assigned to the treatment group
as well as for the untreated voters who are assigned to the control group under
the optimal strategy. The latter is then subtracted from the former to yield the
estimated overall turnout under the optimal strategy. The turnout rate achieved
with the proposed method compares favorably with the turnout rate achieved using
a naive strategy where randomly selected voters are contacted. This less-informed
strategy, which is based solely on the estimated overall average treatment effect
(“ATE strategy”), completely ignores covariate information and thus assumes zero
treatment effect heterogeneity.
Figure 5.3 shows that the proposed method results in a higher overall turnout
than an ATE strategy (at least on average), if an organization can afford to treat

139
only a subset of the population. For a campaign that can afford to treat 10% of the
population, for example, the optimal campaign strategy achieves an overall turnout
increase on average over three times greater than the turnout increase under the
ATE strategy. Because the nonparametric procedure did not isolate individuals who
reacted negatively toward the treatment, no gain over the ATE strategy is seen
when treating the entire population. This lack of negatively-responsive individuals
is not surprising given that little backlash was found to the text messages in a
post-treatment survey (Dale and Strauss, 2008b). Overall, for the text messaging
application, organizations operating under tight budgetary constraints would be wise
to use the proposed nonparametric procedure.

5.5.3 Partisan Example: Parents and Eduction Spending

In September 2007, a union delivered three mailings to its members in Kentucky in

a randomized field experiment. Each participant received exactly one of the three
mailings, which focused on schools, change, and corruption, respectively, with no
control group. The mailings were sent out on September 13, and 1,321 participants
were surveyed by telephone between September 17 and September 20th. As predicted
by the Personal Experience Model, the school mailing worked particularly well among
parents.
The respondents to the survey were asked a standard vote question in which the
parties of the candidates, Democrat Steve Beshear and Republican Ernie Fletcher,
were identified. Respondents reported which candidate they currently supported,
indictating whether their support was “definite, probable, or lean.” Given that the

140
respondents were union members, it is not surprising that they generally preferred
the Democrat: 69% supported Beshear (including leaners), 17% supported Fletcher,
and 13% were undecided. The results were similar for the subset who received the
school mailing: 70% Beshear, 15% Fletcher, and 15% undecided (n=448).
Part of the schools mailing is pictured in Figure 5.4. The Personal Experience
Hypothesis suggests that parents would be especially responsive to this piece because
they are able to independently judge politicians on the state of Kentucky schools.
Unfortunately, the survey of respondents did not ask whether the respondent was
a parent. Fortunately, the survey did ask about marital status, however, and a
linked commercial database does have data on children in the household. The 13%
of respondents who have children in the household (as indicated by the commercial
data) was combined with the 18% of respondents who are not single and are under
40 to compose the target group: the 26% of the population likely to have children.
This group was labeled (with some hesitation) “parents.”
Parents who received the school mailing favored Beshear 75%-12% (n=97) as dis-
played in Figure 5.6; parents who received other mailings favored Beshear 68%-22%
(n=223). This result is at the edge of statistical significance. A linear regression
using the 7-point vote preference, and controlling for the prior probability of sup-
porting Beshear (the support score demographic model) the school mailing*parent
finds a significant interaction with p=0.03, two-tailed. An ordinal regression with
the trichotomous Beshear/Fletcher/undecided outcome variable (i.e., less granular-
ity) and no controls produced a p-score for the interaction of 0.15, two-tailed. These
regressions indicate that an independent parent increased his/her probability of sup-

141
porting Beshear by 11 percentage points after receiving the school mailing (from 66%
to 77%).
The population of the experiment, union members, is not representative of the
Kentucky electorate. To maximize the mailing’s effect on Beshear margin on his
chance of winning, Besshear is assumed to have a baseline margin of -3 percentage
points. The mailing has an average treatment effect of 3.9 percentage points, thus
treating 77 percent of union members (under the ATE approach) would push Beshear
over the 50% change of winning mark.
Figure 5.7 displays the significant cost savings provided by the nonparametric
algorithm. Beshear’s probably of winning passes the coin-flip mark with just 10%
of the population treated. Though with such a small sample, this 10% figure has a
lot of noise—that treating 20% of the population decreases the chance of winning
is an indication of this noise. Nevertheless, the algorithm, which selects both the
commercial data on parents and the imputed data in various iterations, provides
clear benefits.
Certainly there are alternative explanations of why parents are more responsive.
First, parents may be more likely to read the mailing because its subject is schools.
This mechanism, however, is consistent with the Personal Experience Model as indi-
viduals might pay more attention to a mailing that they realize: (1) will inform their
political judgments, and (2) that they are more capable of processing the political
arguments of the mailing. Personal experience with political issues is expected to
strengthen both these processes.
A second possible explanation is that parents have a self-interest in school fund-

142
ing. As with the observational data of Chapter 3, this experiment does not allow
one to disentangle these two mechanisms.

5.5.4 Partisan Example: Parents and Education Spending

In March 2008 while the Democratic Presidential primary race was still undecided, a
union tested three anti-McCain mail pieces (along with a control group) in Ohio. The
focus of the mailings with varying degrees of emphasis, was how McCain’s policies
would damage the economy and weaken the working- and middle-class. Figure 5.8
displays part of the most straightforward of the mailings, which were mailed to union
members. These mailings were buttressed with robocalls that reinforced the topic of
the mailing. Participants were surveyed by telephone shortly after the mail pieces
and robocalls were received. The Democrat’s margin was 23 points (Dem 54% -
McCain 31%, n=2,967) among those who received any mailing and 21 points (Dem
53% - McCain 32%, n=987) in the control group—not a significant difference.
The Personal Experience Hypothesis would predict that members who had per-
sonally experience the economic downturn (or were close to those who had) would
react more strongly to the mail and robocall program. The subgroup of interest was
the set of individuals who live in economically distressed areas—defined by several
census characteristics, such as percent unemployed and median household income.
A factor analysis identified the top 30% of respondents who live in disadvantaged
areas (after individuals who could not be matched to census data were eliminated).6
The median household income of the disadvantaged group is about $32,500, which
6
Special thanks to Catalist, LLC for its help in matching voters to census information.

143
was significantly lower than the average of nearly $50,000 in the advantaged group.
This key group’s support for the Democrat increased significantly after receiving
the mailings and robocalls: from 13 points in the control group (Dem 48%, McCain
35%, n=203) to 21 points among anyone who received a mailing (Dem 53%, McCain
32%, n=626). Among those who lived in the 70% most well-off places, this margin
did not move at all between treatment and control, remaining at 19 points (see Figure
2 on the next page).
These data above are suspicious because of the low Dem margin among the dis-
advantaged in the control group. With only 203 disadvantaged control-group respon-
dents, poor randomization may be affecting the results. To reduce this possibility,
vote preference are regressed on the treatment condition and its interaction with the
subgroup of interest, controlling for a demographic-based partisanship score provided
by the union.7 With a linear regression on the 7-point vote question, the treatment
effect for the advantaged subgroup is nearly exactly zero, but for the disadvantaged
subgroup the coefficient is large and nearly statistically significant (p=0.15, two-
tailed). The interaction coefficient can be interpreted as: the average union member
living in a disadvantaged area moved one-third of a survey response unit (e.g., from
”Dem: lean” to ”Dem: probably”) after receiving the mailing.
An ordinal regression on the trichotomous outcome variable (i.e., the only options
are voting for the Democrat, McCain, or undecided) indicates that a member with
the median score would increase his/her probability of voting for the Democrat by 7
percentage points as a result of the treatment (from 38% to 45%).
7
This score was developed before the experiment.

144
 When the nonparametric algorithm is applied to this data, both small- and large-
budget operations benefit. As Figure 5.10 displays on the left panel, treating 10%
of the population increases the Democrat’s probability of winning to 38%, while the
ATE strategy would only produce a win 24% of the time.8 In nearly every iteration,
the algorithm selected the continuous variable of economic distress.
Big gains are also found under loose budget constraints. The ATE is 2.4 percent-
age points, yet treating 80% of the population under the proposed strategy would
produce an effect of 4.5 percentage points. A backlash is present among some of
the population. The algorithm, when the budget constraint estimates that 10% of
the population is negatively treatment responsive (Figure 5.10, right panel), but
that is probably an underestimate since the overall persuasion boost declines to 3.5
percentage points.
As with the Kentucky test, alternative hypotheses are possible. Especially con-
cerning with this test is that disadvantaged members would react strongly to a variety
of issues, not just the economy. Unfortunately, this proposition cannot be tested in
a rigorous manner since there was no control group. The only evidence offered is
that in a mail test later in the 2008 by the same union dealing with Social Security,
economically disadvantaged were less likely to react positively to the treatment than
the economically advantaged.9
8
Again, the margin between the candidates is artificially tightened to examine differences be-
tween the algorithms.
9
This result is not statistically significant at conventional levels. Further, the Social Security
mailing cannot be used as a test of the Personal Experience Hypothesis because the mailing was
delivered to seniors only. Thus the experiment lacks crucial variance in the interacting variable of
interest (seniors v. younger voters and Social Security appeals).

145
5.6 Concluding Remarks

More than forty years ago, Kramer (1966) observed,

In the past two decades, the use of quantitative methods as aids for
decision-making has become common in many fields. [...] By and large,
however, these efforts have not been made by political scientists. [...] This
is unfortunate, for many of the traditional concerns of political scientists
appear to be quite susceptible to this sort of analysis. (p.137)

Yet, political scientists have since generally neglected to use quantitative methods to
directly inform policy makers who must make decisions using available data. This
chapter takes up Kramer’s proposal and show how modern statistical methods can
be used to help practitioners devise strategies and implement optimal policy using
the Personal Experience Hypothesis.
The proposed methodology is agnostic to theories (e.g., the Personal Experience
Model) about why certain subgroups should be more responsive to specific treatment.
Nevertheless, these theories can help practioners form priors, identify sources of po-
tentially useful data, and better explain the results of the nonparemetric method.
Such approaches demonstrate one application of Kramer’s larger goal: arming cam-
paign strategists with the theory and methods to more accurately target voters who
are susceptible to persuasive appeals.

146
 age< 24.5
| age< 19.5
|
age.missing< 0.5
0.62

lg.dens>=6.043
0.72
lg.dens>=5.985
0.66
lg.dens< 5.858
0.42 lg.dens< 5.852
0.38

0.42 0.79
0.42 0.73

Figure 5.2: Final Classification Trees for the Control Group (left panel) and Treat-
ment Group (right panel). The complexity parameters are chosen from 10-fold cross-
validation using the algorithm described in Section 5.3.2 so that the resulting optimal
turnout is maximized on the validation set. In this example, the planner’s budget
allows treatment of at most 10% of the population. At each node, subjects who
meet the node’s criterion are filtered through the left branch of the tree. Covariate
abbreviations: age is the age in years of the subject, age.missing is whether the
age of the participant is unknown, and lg.dens is the log of the subject’s county
population density.

147
 0.03
●

Optimal Strategy
●

Overall Turnout Increase

●

0.02
●

ATE Strategy
●

●
0.01
0.00

0% 20% 40% 60% 80% 100%

Maximum Proportion of Voters Contacted

Figure 5.3: Empirical Evaluation of the Performance of the Proposed Method for the
Text Messaging Experiment. The figure displays the overall turnout increase that re-
sults from two campaign strategies as a function of the maximum proportion of voters
contacted. The first strategy is the average treatment effect or ATE strategy (solid
line), which contacts randomly selected voters. The second strategy is an optimal
approach based on the proposed methodology, which uses covariate characteristics
of voters to determine which voters receive the treatment. Solid circles represent the
estimated optimal turnout using the difference-in-means estimator. The estimator is
applied to the test data which are not used to devise of the optimal strategy.

Figure 5.4: Part of the Kentucky Mailing.

148
 80%
70%
60%
Beshear Margin

nts
Pare

Non−parents
50%
40%
30%

Corruption & Change Schools

Mailing Appeal Type

Figure 5.5: Parents React Strongly to School Mailing. Beshear’s margin by type
of mailing and whether the recipient is believed to be a parent are shown. Point
estimates are represented by filled-in squares and triangles. Dashed lines represent
one standard error; dotted lines represent two standard errors.

149
 80%
70%
60%
Beshear Margin

nts
Pare

Non−parents
50%
40%
30%

Corruption & Change Schools

Mailing Appeal Type

Figure 5.6: Parents React Strongly to School Mailing. Beshear’s margin by type of
mailing and whether the recipient is believed to be a parent. Point estimates are
represented by filled-in squares and triangles. Dashed lines represent one standard
error; dotted lines represent two standard errors.

150
 1.0

1.0
ATE Strategy

Actual Proportion of Voters Contacted

●
Probability of Beshear Winning

0.8

0.8
● Optimal
Strategy
Optimal Strategy
0.6

0.6
● ●
●
●
●
●
●
0.4

0.4
●

ATE Strategy
0.2

0.2
●
●

●
0.0

0.0
●

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Maximum Proportion of Voters Contacted Maximum Proportion of Voters Contacted

Figure 5.7: Empirical evaluation of the performance of the nonparametic method on

the Kentucky data. The results are calculated based on 10-fold cross-validation. The
left panel displays the estimated probability of Beshear winning under two treatment
strategies plotted over the maximum proportion of the electorate treatable under the
budget constraint. The dashed lines represent the “ATE strategy” in which random
voters are contacted. The solid lines represent the optimal strategy based on the
proposed methodology, which uses covariate characteristics of voters to determine
which voters receive the treatment. The optimal strategy outperforms the ATE
approach strategy under tight budget constraints. The right panel plots the actual
proportion of voters contacted by the mailing against the maximum proportion of
voters contacted, which is determined by budget constraint. Only in the situations
where nearly all voters can be contacted, does the algorithm choose not to canvass
some voters—and it chooses these untargeted voters poorly.

Figure 5.8: Part of the Kentucky Mailing.

151
 Figure 2: Members in Disadvantaged Neighborhoods
React Strongly to Economic Mailings
30%
Generic Dem's Margin over McCain

d
tage
dvan
Disa
20%

Advantaged
10%
0%

Control Any Mailing

Treatment Condition

Figure 5.9: The Economically Disadvantaged React Strongly to Economy-focused

Mailings. Generic Democrat’s margin over McCain by treatment condition and
whether the union member lives in a disadvantaged area are shown. Point esti-
mates are represented by filled-in squares and triangles. Dashed lines represent one
standard error; dotted lines represent two standard errors.

152
 1.0

1.0
ATE Strategy

Actual Proportion of Voters Contacted

●
Probability of Obama Winning

0.8

0.8
● Optimal
●
Strategy
●
0.6

0.6
● ●

Optimal Strategy
0.4

0.4
● ●
● ●

ATE Strategy
0.2

0.2

●
●
●
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Maximum Proportion of Voters Contacted Maximum Proportion of Voters Contacted

Figure 5.10: Empirical evaluation of the performance of the nonparametic method on

the anti-McCain experiment data. The results are calculated based on 10-fold cross-
validation. The left panel displays the estimated probability of the generic Demo-
cratic winning under two treatment strategies plotted over the maximum proportion
of the electorate treatable under the budget constraint. The algorithm successfully
finds voters who are very treatment responsive and negatively responsive. See Figure
5.7 for details.

153
Chapter 6

Campaign and Normative

Implications of Microtargeting

6.1 Campaign Microtargeting

Campaign microtargeting is a method for identifying a subset of voters to direct a

tailored message to. The advent of electronic voter files and commercial databases
has made this strategy more efficient and more widely applicable to more groups
of voters (Malchow, 2003). The Personal Experience Model is one theoretical con-
nection between specific voters and issues. However, several alternative group-issue
connections are described in both the academic (Krosnick, 1990; Chong et al., 2001)
and practitioner literature (Sosnick et al., 2006).
Microtargeting is a powerful tool. If voter-issue linkages are identified correctly,
campaigns can have a potent effect on how voters judge candidates in an election.

154
Chapters 3, 4, and 5 provide the empirical evidence to support this claim and demon-
strate how campaigns can leverage personal experience to garner support from voters
who would otherwise vote for the opposing candidate. This chapter focuses has two
sections. First, the situations in which campaigns would not choose to microtar-
get are outlined. In the broad sense, campaigns that need to change the landscape
of a race dramatically would do best not to microtarget. Second, the normative
implications for microtargeting are discussed.
Various methods for identifying voters who will be responsive to narrow issue
appeals are available to campaigns. The easiest method, perhaps, is to use well-
defined categories from readily available voter files. For example, a campaign might
use birth year to target an age group, such as seniors. A second possibility is to
use non-voter-file information (e.g., census data, licensing lists), match this data
to the voter file, and identify specific groups, such as hunters or those with low
incomes. The third, and most complex, method is to survey a group of voters,
asking a question (e.g., “Do you have children under the age of 18?” or “What is the
most important issue to you?”) the answer to which is no available in any database.
After a sufficient number of voters (perhaps on the order of 2,000) have responded to
a question, a campaign analyzes the data using either a parametric (e.g., logit) or a
non-parametric (e.g., CHAID) method to correlate measures available for the entire
electorate the survey response of interest (Malchow, 2003). The campaign can then
assign a probability of being in the specified subgroup to all voters. Similar to the
optimizations described in Chapter 5, the campaign would contact the voters with
the highest probability of being in the target group.

155
 Figure 6.1 displays the results of a hypothetical microtargeting of undecided
voters. Voters in the top decile (by their probability of being undecided) are twice
as likely to be undecided as the population as a whole (20% to 10%).

20%
15%
Percent Undecided

10%
5%
0%

1 2 3 4 5 6 7 8 9 10

Decile of Undecided Score

Figure 6.1: Example of the results of a campaign microtargeting undecideds. The

top decline of voters (10) included twice as many undecided voters as the population
average. Targeting these voters is thus twice as efficient as a random strategy.

None of these methods perfectly identifies a list of voters who will change their
votes (or even shift their votes) if targeted with an appeal. In an attempt to maximize
the probability of a targeted voter changing her vote choice, campaigns often also
use the survey/CHAID method to identify voters who are on (or near) the fence for
their vote choice decisions. Further error is induced when the targeted population is
identified by survey, as electorate-wide covariates may not predict survey responses
well. The use of a training and test set can ensure that microtargeting models do
not over-fit and that they do identify targeted voters better than if the campaign

156
were to deliver messages randomly.
In essence, microtargeting increases the efficiency of campaign issue signals and
the linkage between a candidate’s stances and a shift in voter’s candidate preferences.
If a voter is experienced with an issue, belong’s to an issue public, or has self-interest
in an issue (all microtargetable subgroups), the voter is likely to more readily pay
attention to campaign appeals on that issue. In all of these cases, the voter more
easily processes political arguments on the issue (due to their experience, appetite,
or self-interest) and thus are less likely to ignore messages on the issue because the
candidate’s appeal is either too complex or uninteresting. Microtargeting increases
the pace of voter learning about pivotal issues.
Applying the Personal Experience Model to various forms of voter targeting
demonstrates how microtargeting enhances the connection between candidate stances
and vote choice. Again, assume that campaigns target voters who have experience,
who are in an issue public, or who have self-interest. Because each of these types of
targeted voters has information or opinions that originate independently of political
elites, the voter can “ground-truth” the stances taken by the candidates (see Chapter
2). This independent information allows the voter to judge the politician on the issue
when the voter cue-takes from the politicians’ positions. Since cue-taking polarizes
the electorate, and does not affect vote preferences, an increase in microtargeting
(and hence a reduction in cue-taking) strengthens the relationship between voter
attitudes and candidate evaluation.

157
6.2 Implications for Campaign Strategy

6.2.1 Microtargeting and Message Control

If a campaign can identify a subgroup of voters with experience on an issue, it will

be able to deliver issue-specific messages to these voters. If the voter’s opinions are
congruent with the candidate’s platform, then the campaigns appeal with raise the
voters’ evaluations of the candidate (on average). In some cases, these voters can be
the difference between losing and winning.
However, in some situations, microtargeting is a net loss for the campaign. Mi-
crotargeting and message delivery cost the campaign both in terms of money and
opportunity cost, thus the benefits must be proportional to the resources devoted to
microtargeting.
In three situations, microtargeting is not helpful. First, independent of other
options for resource allocation, microtargeting is not necessarily beneficial for a cam-
paign. For instance if all microtargetable voters disagree with the candidate on the
issue in question, then clearly the campaign should not raise the issue. If micro-
targetable voters do side with the candidate on the issue the campaign should can
identify these like-minded voters and deliver messages only to them. For the rest of
this Chapter, the term “microtargetable” is restricted to this group of voters.
Even if microtargetable, like-minded voters exist, sending these voters tailored
messages is useful only if these individuals are unlikely to vote for the candidate in
the absence of the appeal. Some of these individuals will likely abstain or vote for
the opponent based on their predispositions if the campaign does not target them.

158
However, the proportion of these voters in this circumstance can be small in high-
salience, polarized elections where the dimension of opinions on the issue in question
is similar to the overall ideological dimension of the campaign.1 While potentially
small in number, some portion of the electorate would shift their vote choice only if
microtargeted.
In addition to microtargeting, campaigns can spend resources attempting to con-
trol their overall message agenda. One example is the 2002 midterm elections,
Democrats pushed the issues of health care and corporate responsibility and Re-
publicans attempted to put taxes and national security at the fore of voters’ minds.2
Because of cue-taking, these non-microtargeted issues often have little effect on the
vote preferences of the electorate. For instance, the sudden shift in dialogue in the
2008 presidential election after the third debate to “spreading the wealth” and “Joe
the Plumber” polarized the public on economic issues and did not move the overall
vote.3 The exception to this rule is when the media or public agree that one party
(or candidate) has outperformed the other on an issue. This occurs after the issue
has resolved (Canes-Wrone et al., 2001) and accounts for the importance of economic
voting (Mueller, 1970) and retrospective voting in general (Fiorina, 1981). If a cam-
paign chooses to spend resources to “broadcast” a message on a single issue and that
issue resolves in its favor, then the electoral payoff can be large.
1
A “high salience elections” means few registered voters abstain, so little get-out-the-vote effect
is possible. A polarized electorate means that few voters are undecided or potential defectors.
Similar issue dimensions means that few voters current by supporting the opponent are likely to be
congruent with the candidate for the issue in question.
2
Nagourney, Adam. “Domestic Concerns Take Center Stage In Congress Races.” New York
Times. September 1, 2002.
3
See Gallup Polls on wealth redistribution and aggregate vote choice surveys on pollster.com.

159
6.2.2 Translating the Personal Experience Model into a For-

mal Game

To delineate the situations in which microtargeting is the optimal strategy, I develop

a model that explores the campaign resource allocation decision. In the broad sense,
I presume that a candidate: (1) has decided to run for office, (2) has chosen a plat-
form on which to run, and (3) must decide how to spend the campaign’s limited
resources (i.e., money). The candidate must chose between two methods of cam-
paigning. First, the campaign can microtarget voter-issue pairs for which the voters
have stable opinions that are congruent with the candidate’s platform. Alternatively,
the campaign can broadcast to all voters the candidate’s stance on an issue for which
the voters have unstable opinions but that may resolve in the candidate’s favor.
I also assume that the campaign has a wealth of knowledge about the electorate
to guide its decisions. Through public opinion polls (or other means), the campaign
knows the general predispositions of the electorate. The campaign also has knowledge
about what percentage of the electorate is microtargetable (i.e., has stable opinions)
on each potential campaign issue. In addition, the campaign has accurate beliefs
about the probability that certain issues will resolve in its favor by Election Day
(e.g., an improving economy helps the incumbent party).
The Personal Experience Model demonstrates how specific segments of the elec-
torate are disproportionately responsive to certain issue appeals when forming can-
didate evaluations. Campaigns generally have the ability to identify these voters to
some extent; perfect identification is not necessary. If a campaign delivers targeted
messages to these voters, the campaign can shift some of these microtargeted voters

160
candidate preferences. I also assume that the more money a campaign invests in
microtargeting, the more voters it can reach.
Any voter-issue pairs in which some voters have a stable opinion is labeled as a
”potentially microtargetable” pairing. The campaign whose platform is congruent
with the voter’s opinion in this pairing would prefer that the voter learn about the
candidate’s stances on the issue and have the issue on the top of her head come
Election Day. However, the campaign does not want to emphasize the issue among
voters who have stable opinions that are incongruous with its platform. Hence,
even if wide swaths of the electorate have stable opinion on issues, campaigns could
maximize their efficiency by microtargeting only the voters with congruent opinions.
Segments of the population who lack issue experience (or an alternative issue
connection such as self-interest) are not likely to shift their vote on the issue in
question. Instead, these voters cue take from politicians, their vote choices polarize,
and they do not change their vote choice. This polarization mechanism requires
two streams of considerations (i.e., an appeal from each ideological side)—a criterion
that is not always present. The canonical example of one-sided issue streams is the
early discussion of the Vietnam war (Zaller, 1992). In this case, both liberals and
conservatives alike supported the war as that was the only position represented by
political elites.
Message streams can become one-sided when an issue resolves. For instance,
when the country’s economy tanked in the Fall of 2008, only 5% of the public rated
the economic situation of the country as “excellent” or ”good”.4 Voters can then
4
“How would you rate economic conditions in this country today – as excellent, good, only fair,
or poor?” USA Today/Gallup Poll. Oct. 10-12, 2008. N=1,269 adults nationwide.

161
use the single stream of information from the media to “ground truth” politicians
performance. The evidence that resolved issues affect large swaths of the electorate is
clear when economic and war data are correlated with election results on the national
(Hibbs Jr., 2000) and state levels (Campell, 1992; Cohen and King, 2004).
In the model, an issue can (1) resolve in favor of one candidate, (2) resolve in
favor of the other candidate, or (3) not resolve during the course of the campaign.
Campaigns can choose to focus their appeals on certain issues in an attempt to prime
voters’ minds (i.e., place considerations about this issue on the tops of voters’ heads).
In essence, campaigns can place bets that certain issues will resolve in their favor.
One possible exception to the cue-taking rule is issue ownership Petrocik (1996).
Related to the Broadcast Corollary discussed in Chapter 2, the theory of issue own-
ership stipulates that wide swaths of the electorate favor one side of an issue. If a
party “owns” an issue, it can be considered an “easy issue” (Carmines and Stimson,
1990) in that voters will likely have a stable opinion on the issue independent of
cue-taking.
However, elections in which a large majority of voters side with one candidate
on an “easy” or “owned” issue are usually uncompetitive elections. The campaign
that has this majority of voters on its side most likely has a winning (and perhaps
dominant) strategy to emphasize this issue. Since close campaigns are more interest-
ing, the model instantiations in Section 6.2.5 usually consider cases in which a small
percentage of the electorate is microtargetable.

162
6.2.3 The Model

In an election between two candidates, D and R, the campaigns decide how to allo-
cate resources across two issues (A and B) and whether to microtarget or broadcast
their messages about these issues. In the hypothetical world in which neither cam-
paign makes any effort, the predispositions of the voters dominate. Both campaigns
have common beliefs about their chance of winning in this no-campaign universe
(i.e., a benchmark poll is taken before the campaign to test the mood of the elec-
torate). This belief is represented by a mean percent of the vote for candidate i,
µi , and a common uncertainty, σ 2 (i ∈ D, R). This representation is similar to a
valence advantage in other models (Groseclose, 2001). The vote for candidate i, Vi ,
is distributed normally,
Vi ∼ N (µi , σ 2 ). (6.1)

Candidate i’s utility, Ui , is the campaign’s probability of winning,

µi − 0.5
Ui = Pr(i wins) = Pr(Vi > 0.5) = Φ( ). (6.2)
σ

To simplify the analysis, I assume only two candidates in the race, µD = 1 − µR and
a fixed electorate with 100% turnout. The tiny tails of the vote distribution that
extend beyond the logical range of Vi , [0, 1], are ignored.
Implicitly in the model, ties are broken with a coin flip. If the campaigns garner
exactly the same number of votes on average (i.e., µ = 0.5), then the probability of
winning approaches 0.5 as the uncertainty of the result goes to zero (i.e., σ 2 → 0).
This characteristic of the model is a result of the symmetry of the normal curve

163
about its mean.
Under all circumstances, campaigns want to increase their share of the vote, µi ,
∂Ui
since ∂µi
> 0. They attempt to increase this share by taking stances on issues.
For the moment, let the issues equal advantages for the campaigns. Issue A is an
issue for which a potion of the electorate has experience, although some experienced
voters side with candidate D and some with candidate R. Let the proportion of
all voters who fall into one of those two categories (i.e., who are microtargeted by
the candidates) be mD and mR respectively. For now, assume a level playing field:
mD = mR .
The other issue, B, is not microtargetable but has a chance of resolving by the
election. If this issue resolves in candidate D’s favor, an occurrence with probability
of pD , q percent of voters switch their preference from R to D. A symmetric switch
of q percent of the electorate occurs with probability pR ; i.e., in this case, the issue
resolves in candidate R’s favor. Let pD = pR .
The campaigns do not have the resources to fully utilize the advantages offered
by both approaches. Instead, they must choose a combination of microtargeting and
broadcasting. Assume that both campaigns have equal resources, a budget of 1, and
campaign i spends δi ∈ [0, 1] on microtargeting. A campaign must spend its entire
budget to gain the maximum votes from the strategy-issues combinations above.5
The resulting distribution of the percentage of votes for candidate D, given both
5
If one campaign spends all its resources on broadcasting then, q percent of voters switch. If
both campaigns only broadcast, then 2q voters switch.

164
campaigns disbursements, is

VD (δD , δR ) ∼ N (µi , σ 2 ) + δD mD − δR mR + (2 − δD − δR )BS(pD , pR ) (6.3)

VD (δD , δR ) ∼ N (µi , σ 2 ) + m(δD − δR ) + (2 − δD − δR )BS(p, p) (6.4)

where BS(p1 , p−1 ) is a Bernoulli scheme with probability p1 of outcome 1 and prob-
ability p−1 of outcome -1. The analogous equation shows the vote for candidate R.
The mean and variance of Vi are,

mean(Vi (δi , δ∼i )) = V̄i (δi , δ∼i ) = µ + m(δi − δ∼i ) (6.5)

Var(Vi (δi , δ∼i )) = σ 2 + 2pq 2 (2 − δi − δ∼i )2 (6.6)

If neither campaign spends any resources broadcasting (i.e., δi = 1), each candi-
date’s contribution to the variance (Equation 6.6) goes to 0. As before, the utility
for candidate i is the probability of winning, Ui = Pr(Vi > 0.5). For candidate D,

UD = (1 − 2p)Φ((µ + m(δD − δR ) − 0.5)/σ) (6.7)

+pΦ((µ + m(δD − δR ) + q(2 − δD − δR ) − 0.5)/σ)

+pΦ((µ + m(δD − δR ) − q(2 − δD − δR ) − 0.5)/σ)

This model includes two assumptions that which are rough approximations of the
real world and and not likely to hold in actual campaigns. First, the proportion of
voters who will switch their candidate preference because of microtargeting (m) does
not vary based on the resources spent on broadcasting (1 − δ), or vice versa (with

165
q and δ). In essence, this simplification assumes a uniform distribution of the elec-
torate across the ideological spectrum regardless whether a candidate becomes more
advantaged (i.e., the decision cutpoint on the ideological spectrum shifts), the num-
ber of nearly undecided (i.e., indifferent) voters is constant. Second, if the campaign
microtargets voters who have a high probability of shifting their vote choice, the rate
of successful microtargeting is not affected by spending on broadcasting. In other
words, wasteful spending by campaigns that “doubles up” on voters—contacting
them with both microtargeting and broadcasting when only one of the methods is
necessary to shift the voter’s preference toward the candidate—is assumed not to
occur.

6.2.4 Best Response

The optimal strategy of candidate i, given the strategy of the other candidate (labeled
∼ i) is the level of microtargeting (δi∗ ) that maximizes candidate i’s utility,

δi∗ (δ∼i ) = argmax Ui (δi , δ∼i ) (6.8)

δi

To determine the best response to the opponent’s strategy, δ∼i , the partial deriva-

166
tive of candidate i’s utility is taken with respect to the campaign’s strategy:

∂Ui m
(δi , δ∼i ) = (1 − 2p) φ((µ + m(δi − δ∼i ) − 0.5)/σ) (6.9)
∂δi σ
m−q
+p φ((µ + m(δi − δ∼i ) + q(2 − δi − δ∼i ) − 0.5)/σ)
σ
m+q
+p φ((µ + m(δi − δ∼i ) − q(2 − δi − δ∼i ) − 0.5)/σ)
σ

Under some circumstances, the resource allocation decision is trivial. If the num-
ber of votes that can be shifted with microtargeting (m) is greater than the proportion
of votes possibly shifted if the issue resolves (q), then the campaign should spend all
∂Ui
its money on microtargeting. Formally, m > q → ∂δi
> 0 (see Equation 6.4) and
the candidate’s utility is increasing in δ.
As an intuition for the whether a campaign has an incentive to microtarget or
broadcast, consider the effect of broadcasting on the variance of Vi . Pouring more re-
sources into broadcasting increases the percentage of the population that will switch
under a resolved issue, and since issue B resolves stochastically, increased broad-
casting increases the variance of Vi . For a mathematical intuition of this result, see
∂Var(Vi )
equation 6.6 and note that ∂δi
= −4pq 2 (1 − δi − δ∼i ) is negative, so the variance
of the vote increases with the amount of resources spent on broadcasting.
Similar to the stochastic knapsack problem discussed in Chapter 5, campaigns
want to increase the variance of the election result if they are losing (i.e., the ex-
pected vote percentage is below 50%) and decrease the variance if they are winning
(Carraway et al., 1993). Applying this principle to the microtargeting vs. broad-
casting model, we find an incentive for broadcasting only when a campaign cannot

167
reach 50% (on average) with a combination of predispositions and microtargeting.
To prove this result formally, note that adding or subtracting a value to the point
on a Gaussian probability distribution has the property x < 0 ↔ φ(x + c) > φ(x − c).
Next, Equation 6.9 is rewritten to combine the terms influenced by the potential for
microtargeting, m, and broadcasting, q,

∂Ui m
(δi , δ∼i ) = [(1 − 2p)φ((µ + m(δi − δ∼i ) − 0.5)/σ) (6.10)
∂δi σ
+pφ((µ + m(δi − δ∼i ) + q(2 − δi − δ∼i ) − 0.5)/σ)

+pφ((µ + m(δi − δ∼i ) − q(2 − δi − δ∼i ) − 0.5)/σ)]

q
− [pφ((µ + m(δi − δ∼i ) + q(2 − δi − δ∼i ) − 0.5)/σ)
σ
−pφ((µ + m(δi − δ∼i ) − q(2 − δi − δ∼i ) − 0.5)/σ)].

∂Ui
For ∂δi
to be negative, the campaign must start in a losing position and not be able
to make it up with microtargeting alone: µ + m(δi − δ∼i ) < 0.5. That condition is
necessary, but not sufficient, for the optimal campaign strategy to be 100% broad-
casting. The other necessary condition is that broadcasting must be sufficiently more
potent than microtargeting (q >> m) that Equation 6.11 is negative.
If the campaigns start on equal footing (i.e., µD = µR = 0 as well as mD =
mR = m, pD = pR = p, and qD = qR = q) then the weakly dominating strategy is
to spend all resources on microtargeting; if campaign i spends all its resources on
microtargeting (δ∼i = 1), then any money campaign i spends on broadcasting will be
wasted in the case of issue B not resolving. When the issue does not resolve (which

168
occurs with probability 1 − 2p), campaign i loses with probability Φ( m(1−δ
σ
i)
), which
is greater than 50%. If the issue resolves, then the half the time the issue resolves
in campaign i’s favor nearly balances out with the other half of the time it does not.
But even in this case, the result is a net negative for campaign i since the mean point
is negative and x < 0 → Φ(x + c) < (1 − Φ(x − c)). If candidate ∼ i foolishly does
not spend all his resources on microtargeting, then candidate i can take advantage
of that error and win over half the time by following this logic and spending all his
money on microtargeting.
Thus, in the case of a campaign in a marginal district with symmetric proper-
ties, microtargeting is a weakly dominating strategy. The strategy is not strictly
dominant because of the case where issue B always resolves (p = 0.5). Figure 6.2
illustrates the relationship between the probability of issue B resolving and the use-
fulness of broadcasting relative to microtargeting. The utilities of candidate D for
all possible combinations of microtargeting and broadcasting are shown under three
scenarios, each with increasing probability of issue resolution. As the effect of micro-
targeting shrinks (m) relative to the effectiveness of broadcasting (q), the utility of a
microtargeting-only strategy and the utility of a broadcasting-only strategy approach
the same value: 50%.

6.2.5 Edge Equilibria

Extending this logic of the previous section, in the case where candidates are on equal
footing, only corner equilibria exist. In nearly all circumstances, these equilibria
occur when both campaigns microtarget. In this section, restrictions on symmetry

169
 Low Probability (20%) of High Probability (80%) of Issue Resolution is
Issue Resolution Issue Resolution a Certainty
0.50

0.50

0.50
0.48

0.48

0.48
Probability that Candidate D wins

Probability that Candidate D wins

Probability that Candidate D wins

0.46

0.46

0.46
0.44

0.44

0.44
0.42

0.42

0.42
0.40

0.40

0.40
0.38

0.38

0.38
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Proportion of Candidate D's of Proportion of Candidate D's of Proportion of Candidate D's of
resources spent microtargeting resources spent microtargeting resources spent microtargeting

Figure 6.2: Candidate D’s Utility Under Increasing Probabilities of Issue B resolving.
In the examples depicted, 5% of the electorate is microtargetable by each campaign
(m = 0.05) and 25% of the electorate shifts if the issue resolves (q = 0.25). The
campaigns start on equal footing (µ = 0.5), although uncertainty about the electorate
is high (σ = 0.08). The panels depict candidate D’s utility under the strategy
specified by the x-axis given candidate R spending all resources on microtargeting
δR = 1 and increasing probabilities of issue B resolution (p = 0.1, p = 0.4, and
p = 0.5, respectively).

are relaxed; three scenarios are considered and depicted in Figure 6.3.
The first scenario is similar to the symmetric case considered in the previous
section except that candidate D has an advantage with respect to broadcasting.
More often the issue will resolve in candidate D’s favor (pD = 0.7, pR = 0.3) and
more voters will switch to candidate D in the case of resolution (qD = 0.2, qR = 0.1).
Thus, at high levels of microtargeting by candidate R (approximately, δR > 0.8),
candidate D optimally puts all resources into broadcasting. As shown as an “X”
in panel (a) of Figure 6.3, the equilibrium is candidate D fully broadcasting and

170
candidate R fully microtargeting.
When candidate i has an advantage in the electorate (i.e., µi > 0.5) that can be
maintained when both campaigns microtarget (i.e., µi +mi −m∼i > 0), then campaign
i’s best strategy is often to microtarget only. The microtargeting strategy is optimal
because it lowers the variance of the election result compared to broadcasting. The
disadvantaged campaign’s best response is often a combination of microtargeting
and broadcasting; the exact mix depends on the relative effectiveness of those two
strategies.
An edge equilibrium of this type is depicted in Figure 6.3, panel (b). In this
case candidate R has a pre-campaign edge (µR = 0.54, σ = 0.04) and candidate D
cannot recover this edge via microtargeting (mD = mR = 0.03). In equilibrium, R
fully microtargets to bring his chance of winning up to 72%. Candidate D wants
both a high vote mean and a high variance; microtargeting provides the former and
broadcasting (qD = qR = 0.2, pD = pR = 0.25) provides the latter. Hence, candidate
D spends 28% of his money on microtargeting and the rest on broadcasting.
An exception to the general rule that advantaged candidates microtarget is pre-
sented in panel (c). In this case, candidate R is advantaged both in the electorate’s
predisposition (µD = 0.445, σ = 0.067) and issue resolution (qR = 0.1, qD = 0.1;
pR = 0.7, pD = 0.2). By combining a microtargeting and a broadcasting strategy,
candidate R can increase his probability of winning to 80%. Candidate D’s best
response in this case is to fully microtarget: this pair of strategies is an equilibrium.
In this example, candidate D has an interesting best response curve, which has
three regimes. When candidate R does not microtarget much (δR < 0.2) then can-

171
didate R can increase the mean vote to fairly close to a 50-50 election (mD = 0.044,
mR = 0.008). Also, the emphasis candidate R places on broadcasting in this case
increases the variance of the election, which also raises candidate D’s probability of
winning. When candidate R microtargets to a substantial degree (approx. mD > 0.4)
then the benefit of raising the mean by candidate D via microtargeting is greater
than the benefit of increasing the variance by broadcasting. In the middle of these
two regimes, the need for candidate D to raise the mean is small enough to en-
courage broadcasting. However, candidate D cannot free-ride off candidate R’s full
broadcasting effort, and it is worth candidate D’s while to broadcast himself.
No equilibrium exists in the final example, which is depicted in panel (d). Candi-
date R is advantaged in electorate predispositions and in issue resolution. This cir-
cumstance is interesting because both campaigns may have an incentive to broadcast.
Candidate R can broadcast to increase the mean of his vote distribution. Candidate
D can broadcast to increase the variance of his vote share. (The latter strategy is
helpful to candidate D’s mean of vote share is always below 50%.) However, it is
never the case that both candidates broadcast at the same time. As candidate D
spends more resources on broadcasting only, he contributes to candidate D’s vote-
share variance, so candidate D has an incentive to microtarget. As candidate D
spends more on broadcasting, he contributes to candidate R’s vote-share mean, so
candidate R has an incentive to microtarget and lower his vote-share variance. In
this manner, the candidates never reach equilibrium.

172
6.2.6 No Internal Equilibria

In the microtargeting-broadcasting game, there are no internal equilibria. Nor are

there edge equilibria in which one candidate fully broadcasts. To prove this formally
for the general, non-symmetric case, first I derive the candidate i’s utility and first
derivative in this case.

Ui (δi , δ∼i ) = (1 − pi − p∼i )Φ((µ + mi δi − m∼i δ∼i − 0.5)/σ) (6.11)

+pi Φ((µi + mi δi − m∼i δ∼i + qi (2 − δi − δ∼i ) − 0.5)/σ)

+p∼i Φ((µ + mi δi − m∼i δ∼i − q∼i (2 − δi − δ∼i ) − 0.5)/σ)

∂Ui mi
(δi , δ∼i ) = (1 − pi − p∼i )φ((µ + mi δi − m∼i δ∼i − 0.5)/σ) (6.12)
∂δi σ
mi − qi
+ pi φ((µi + mi δi − m∼i δ∼i + qi (2 − δi − δ∼i ) − 0.5)/σ)
σ
mi + q∼i
+ p∼i Φ((µ + mi δi − m∼i δ∼i − q∼i (2 − δi − δ∼i ) − 0.5)/σ)
σ

The proof that no equilibria exists where both candidates spend resources on
∗
broadcasting follows by contradiction. Assume an equilibrium of (δi∗ , δ∼i ) exists where
∗
δi∗ < 1∧δ∼i < 1. The contribution of microtargeting alone to the candidate’s utility is
∂Ui
always positive: limitqi →0 ∂δi
> 0. For candidate i’s optimal strategy (δi ∗) to include
∂Ui ∗ ∗
some broadcasting, the first derivative is nonpositive: ∂δi
(δi , δ∼i )) ≤ 0. Thus, the

173
broadcasting term of the first derivative is negative:

qi
0 > − pi φ((µi + mi δi − m∼i δ∼i + qi (2 − δi − δ∼i ) − 0.5)/σ) (6.13)
σ
q∼i
+ p∼i Φ((µ + mi δi − m∼i δ∼i − q∼i (2 − δi − δ∼i ) − 0.5)/σ)
σ

But analogous logic for candidate ∼ i yields the result that negation of the right-
hand side of Equation 6.13 must be negative. The negation of a negative cannot be
negative. Contradiction.

6.2.7 Model Extensions: Multiple Issues and Platform De-

cisions

If more than two issues are at play in the election, the model will also inform a
candidate’s decision about which issues to focus his efforts on. A candidate who is
advantaged against his opponent’s optimal strategy seeks to increase the mean of
his vote share and decrease the variance. Candidates who, on average, cannot win a
majority of the vote seek to increase both their vote share and the variance of the
outcome. The model can be extended to addressed multiple issues in a straightfor-
ward manner, though the strategy space would increase to k − 1 dimensions, where
k is the number of issues considered. The logic of Section 6.2.6 would hold within
an issue: no two candidates would broadcast on the same issue, although candidates
might optimally broadcast on two different issues.
Incorporating multiple issues would enable the model to be extended to of plat-
form choice. Similar to the model of Groseclose (2001), candidates would have exoge-

174
nous policy preferences and weights that they would place on winning the election vs.
policy outcome. The choice of which issues to emphasize (either with microtargeting
or broadcasting) would be a function not only of whether the issues can help the
candidate win, but whether the candidate’s position on the issue is consistent with
his policy preferences.

6.2.8 Discussion

The model has three major implications. First, in general, campaigns that are be-
hind microtarget, and campaigns that are ahead broadcast. This implication may
actually be understated by the model, as advantaged campaigns would not want to
microtarget just to increase their vote share, but also to prevent voter defection.
A natural extension of the theory in presented Chapter 2 suggests that delivering
microtargeted appeals before issue resolution will prevent some voters from defect-
ing. Thus, early microtargeting might be even more effective than demonstrated at
lowering the variance of election results.
Second, a losing campaign may be in the situation where it must lower its ex-
pected vote share to increase its chance of winning. This circumstance was epito-
mized by the description provided to John McCain of Sarah Palin during his running-
mate decision process: “high risk, high reward.”6 Since losing campaigns take risks
that on average do not work out for them, more strategic thinking by campaigns
may lead to fewer moderately close elections.
Third, campaigns never purposely talk to the same voters about the same issues.
6
As publicly stated by A.B. Culvahouse, recounting his conversation with John McCain.

175
Opposing campaigns may microtarget the same issue, but they are targeting mutu-
ally exclusive groups: only voters who agree with the campaign on the issue. The
empirical evidence demonstrates that opposing campaigns seek to emphasize distinct
sets of issues (Sellers, 1998), although current events often foil these plans and force
campaigns to talk about the same issues (Sigelman and Buell, 2004).

6.3 Normative Implications

Microtargeting increases the efficiency of individuals learning about candidates’ po-

sitions. Whether or not the increased ability of campaigns to microtarget is helpful
or detrimental to democracy depends on context and situation. Fundamentally, mi-
crotargeting is a tool, and like most tools (e.g., screwdrivers, TNT), it can be used
for positive or negative purposes. Microtargeting has both positive and negative
implications for democracy.

6.3.1 Heuristics and Judging Democracy

A fundamental debate rages in the study of American politics: How much information
does a voter need to make an informed decision in the voting booth? On one side
of the ledger are those who believe that Americans are ill-informed (Delli Carpini
and Keeter, 1996), that they are generally incapable of forming coherent opinions
(Converse, 1964), and that this lack of information is crucial to their decisions (Lau
and Redlawsk, 2001; Bartels, 2005). In the other camp are scholars who believe that
cognitive shortcuts are available (Popkin, 1994), voters seek the information that

176
interests them (Krosnick, 1990), and any errors cancel each other in collective public
opinion (Page and Shapiro, 1992).
Often, when judging democracy, scholars analyze citizens’ vote preferences (Lau
and Redlawsk, 2001; Bartels, 1996). The argument from the “more information
needed” camp is that there is a standard by which citizens should judge candidates.
Lau and Redlawsk (2001) attempt to infer the weights that voters should place on
various issues to determine a “correct” vote. Bartels (1996) assumes that if low-
informed voters become more knowledgable, they will process information similar to
their high-information counterparts—an assumption that is called into question by
research on information processing in political situations (Gilens, 2001).
These assumptions about which candidate preference a voter should have are
troublesome because they impose scholars’ views on how citizens choose among po-
litical alternatives. A safer standard for to judge democracy is to evaluate voters
knowledge of indisputable facts. The debate over heuristics is lively under this nor-
mative standard as well (Gilens, 2001; Bartels, 2002; Caplan, 2008). The downside
is that the outcome of governments (democratic or otherwise) are policy decisions
and citizens’ issue opinions may be farther from that policy positions than their vote
decisions.
To resolve this balancing act, I use two standards for judging democracy. The
primary standard is the amount of accurate information voters have. Often the
accuracy of political opinion is in dispute; even when political elites agree on the
optimal outcome (e.g., lower unemployment) they disagree about how to achieve
that outcomes. Thus, for this first standard, the analysis is constrained to the set

177
of information that is verifiable. Included in this set are candidate positions (which
the candidate defines) and observable facts.
I examine the more subjective standard of policy options and vote choice, albeit
under certain crucial assumptions. Often voters must judge candidates’ platforms
before knowing the effect of the policy stances contained within those platforms.
I assume that certain segments of the population have more knowledge about the
probability of positive resolution of an issue under the various proposed policy op-
tions. If this segment of the population has more sway in the election outcome,
the resultant government’s policy choices are more likely to be optimal. Optimal is
defined as issue resolution that the most people approve of (e.g., good economy or
winning a war).

6.3.2 Positives for Democracy

Microtargeting increases the efficiency of voters learning the candidates’ positions.

If campaigns microtarget, a voter with experience on a political issue is more likely
to receive messages and about that issue. Microtargeting increases the amount of
information at voters’ disposal and is thus a positive for democracy under the first
standard.
An example of this process is seniors learning about Bush’s and Gore’s positions
on Social Security (see Chapter 3 for details). At the beginning of the campaign
(through May 2000), 42% of seniors (65 and older) could correctly identify both
candidates on Social Security, compared to 36% of younger voters: a difference of 6
percentage points. In August, this difference was 8 percentage points; in September,

178
10 points; and in October, 12 points. This interaction between dichotomous age and
time is significant at conventional levels (logit regression, p = 0.05).7
It is unclear how much of this effect was due to microtargeting by the Bush and
Gore campaigns, but clearly seniors were more actively engaged in the Social Security
debate. If the campaigns had been able use the techniques developed just two cycles
later, perhaps more seniors would have learned about the candidates positions. Even
by the end of the campaign, over a third of seniors could not accurately identify both
candidates’ positions on Social Security.
Transitioning to the second standard—correct policy opinion—if experienced vot-
ers have more knowledge about which policy options will resolve favorably, then the
increased information efficiency provided by microtargeting benefits democracy. In
game theory parlance, experienced voters have more knowledge about the “state of
the world” on that particular issue than non-experienced voters.8 In a situation
similar to the “Swing Voter’s Curse” of Feddersen and Pesendorfer (1996), the non-
experienced voters do not know which policy proposal is best on a particular issue.
However, if the pool of voters who judge politicians on the issue for which they
have experience (and thus, more in depth knowledge) is larger than any underly-
ing partisan or incumbency biases, then the candidate with the best policies will be
elected.
As an illustration of why microtargeting helps in this case, consider a challenger
7
For two other issues tested, vouchers and taxes, the interaction effect is in the same direction,
but is half the magnitude and not statistically significant.
8
If, for some reason, experienced voters are less likely to hold the correct policy position than the
general public then the analysis works in the opposite direction and microtargeting is detrimental
to democracy.

179
facing an incumbent who is so incompetent that he is wrong on every issue. If the
challenger does not microtarget and instead broadcasts one issue to all voters, then
only the small subset of voters with experience on that issue will be affected (unless
that issue resolves). The incumbent’s resources (e.g., higher name recognition, more
funds) might be sufficient to overcome a small block of voters casting ballots for the
challenger. However, if the challenger uses modern technology to match voters to the
issues they have experience with, several subsets of voters will judge the incumbent
poorly, potentially increasing the challenger’s probability of success to greater than
50%. As demonstrated in Section 6.3, the challenger will want to microtarget in this
case—this strategy will prevent the incumbent from winning again and implementing
wrong-headed policies.

6.3.3 Negatives for Democracy

There are potential downsides to microtargeting as well. Foremost, microtargeting

may enhance the ability of candidates and parties to have individuals believe in-
correct statements. As shown in the survey experiments in Chapter 4, delivering
congruent cues to partisans increases their attachment to their party’s candidates.
This attachment leads to cue-taking (see Chapters 2 and 4), which means that indi-
viduals uncritically accept the position of their preferred elite. Thus, they may be
more prone to believe false information (Bartels, 2002).
As an example of individuals believing verifiably incorrect information, the 2008
CCES asked potential voters whether they knew Barack Obama’s religion. Despite
several statements by Obama that he was Christian and a highly covered debate in

180
the Spring of 2008 about controversial statements by Obama’s pastor, over a quarter
of the electorate believed (with at least some degree of certainty) that Obama was
a Muslim. Conservative elites (e.g., Fox News) were peddling these falsehoods and
the individuals who believed them were nearly all conservative.9
Another reason why microtargeting might be detrimental to democracy is that
people may have experience with issues that affect only a small part of their lives.
For instance, an environmentalist might be microtargeted on global warming and
vote for the pro-environment candidate even though he does not support that candi-
date’s economic policies. Since global warming works over a decade- or century-long
timeframe, the economy is probably more relevant to the voter. But the voter may
nonetheless voter for the suboptimal candidate based on the issue with which the
voter has experience.
Undoubtedly voters sometimes have policy opinions that are incorrect, even by
their own standards (with the benefit of hindsight). For instance, a third of the public
went from believing the Iraq War was justified to a mistake from 2003 to 2006. Would
a more informed electorate have had a different view of the war in 2003?10 Given
that elites on both sides were advocating their respective positions and Americans
were filtering their consideration intake (e.g., via Zaller’s RAS model), it is not clear
that a more attentive electorate would have had a different opinion. Perhaps a more
attentive electorate would have had more considerations at the tops of their heads,
9
Of the respondents who were able to place themselves somewhere other than 50 on the 0-100
scale of liberal to conservative, 90% were in the conservative half.
10
This question differs from that of whether better intelligence would have changed the minds of
the elites. If more moderate Democrats (e.g., John Kerry, Hillary Clinton) had voted against the
2002 war resolution, a lower percentage of the public would have supported the war.

181
but in the same distributions.
The danger is that voters with strong party or elite affiliation may have tighter
filters. Since microtargeting increases polarization and elite affection (see Chapter 2
for the theory and Chapter 4 for the empirical evidence), advancements in campaigns’
targeting abilities may increase the cue-taking from elites and the congruent choice
of policies that lead to negative outcomes.

6.3.4 Further Discussion: Party Structure

The cue-taking effect that causes large segments of the population to believe false-
hoods also leads to closer party affiliation. Political parties in the United States often
include different “types” of voters (e.g., social conservatives, economic conservatives).
Microtargeting on an issue a voter has a stable opinion about (e.g., abortion) raises
her opinion of her party and leads her to more easily accept the party’s platform on
other issues (e.g., the economy).11 This artifact of microtargeting has both positive
and negative implications for democratic participation. On one hand, voters polarize
when affiliating with parties, thus increasing political participation (Dalton, 2008).
The number of parties is kept to a minimum (since parties can better keep factions
under one roof), which also increases political participation (Blais and Dobrzynska,
1998). On the other hand, parties are less representative of voters’ stable beliefs,
which can produce negative results if the parties abuse the voter’s affiliation by pur-
suing rent-seeking or ideologically extreme policies that are detrimental to the state
of the nation (Sartori, 1976).
11
Partisanship has increased recently (Bartels, 2000), though I would not attribute much of this
change to increased microtargeting as other more powerful forces are at work (McCarty et al., 2008).

182
6.4 Conclusion

The Personal Experience Model provides a theory behind campaigns’ microtargeting

tactics. The more campaigns combine political science research with individual-
level data, the more they can control their own destinies. Instead of hoping that
the Heaven’s resolve an issue in the campaign’s favor a week before the election,
campaigns can plan well in advance who they will microtarget when.
Microtargeting is a powerful tool that helps identify the voters who are most
likely to change their vote preference as a result of a persuasive appeal. These voters
are also more likely to receive information about candidates’ stances that is relevant
to their lives and that makes their voting decisions easier. However, that easier
decision can lead to the voter to be too trusting of elites. Microtargeting thus offers
both positives and negatives for democracy. Which side has more weight, as always,
depends on one’s perspective.

183
 (b) No Starting Advantage;
(a) Candidate R's Advantage Holds
Cand D Issue Res. Adv.
1.0

1.0
X X
Cand. R Strategy: Pct. Micortarget

Cand. R Strategy: Pct. Micortarget

0.8

0.8
● ●
0.6

0.6
0.4

0.4
0.2

0.2
Candidate D Best Resp. Candidate D Best Resp.
Candidate R Best Resp. Candidate R Best Resp.
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Cand D Strategy: Pct. Micortarget Cand D Strategy: Pct. Micortarget

(c) Cand R has advantage; (d) Cand R Starting and Iss.

Cand D Needs Luck to Win Res. Adv.; No Equilibrium
1.0

1.0
Cand. R Strategy: Pct. Micortarget

Cand. R Strategy: Pct. Micortarget

0.8

0.8
0.6

0.6

● ●
0.4

0.4

X
● ●
0.2

0.2

● ●
Candidate D Best Resp. Candidate D Best Resp.
Candidate R Best Resp. Candidate R Best Resp.
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Cand D Strategy: Pct. Micortarget Cand D Strategy: Pct. Micortarget

Figure 6.3: Best Response Plots of Four Scenarios. See Text for specifics.

184
Chapter 7

Conclusion

To be written. Filler below.

Prominent research in political science (Campbell et al., 1960; Bartels, 2002) em-
phasizes the long-lasting attachments voters have to political parties. Yet large por-
tions of the electorate switch parties from election (Key, 1968) to election or split their
ticket within one election (Fiorina, 1996). These defections are often explained by
the preferences of voters relative to the candidates (Downs, 1957; Enelow and Hinich,
1984; Carmines and Stimson, 1990) and the debate over relative effects of partisan
identification versus issue opinion continues in recent research (Ansolabehere et al.,
2008). This dissertation synthesizes the arguments from both camps of thought,
demonstrating how party identification and issue opinions interact jointly form in-
dividuals’ political judgments. To provide observable predictions of the theory, the
model and evidence presented emphasizes the role of personal experience in political
evaluations.

185
 Two recent anecdotes illustrate this connection. From the 2004 to the 2006
elections, the percentage of Hispanics identifying as Democrats, as measured by
exit polls, jumped by 15 percentage points (from 43% to 58%), while non-Hispanics
showed no increase in their inclination to be identified as Democrats. This trend
toward favoring the Democrats continued in the 2008 election, with Obama garnering
67% of the Hispanic vote, as compared to the 53% who supported Kerry. Hispanics’
response to questions in the 2006 exit poll shed light on the underlying reason for the
massive movement: 37% of Latinos indicated that the issue of “illegal immigration”
was “extremely important” in their voting decision (compared to only 29% of the
non-Latino population).While immigration policies might not personally affect these
voters (as they were already citizens), their involvement in the Hispanic community
most likely gave them firsthand experience with the immigration issue.
As Hispanics were flocking to the Democratic party, active duty military personnel
were leaving the Republican party in droves. A 2004 mail survey of subscribers to
Military Times found that 60% of active duty personnel identified as Republicans.
The same survey in 2006 found that this percentage had dropped to 46%. Republican
affiliation among the electorate at large (as measured by exit polls) dropped only 1.5
percentage points over these two years. In 2008, Obama outperformed Kerry among
the active military and veterans despite both Kerry and Obama’s opponent being
veterans. Almost assuredly, the wars in Iraq and Afghanistan had a profound effect
on the political views of the military.
This dissertation argues that changes in an individual’s political allegiance are
likely to occur when the political parties take identifiable positions on issues with

186
which the individual has significant personal experience. When a large group of
people have experience with a prominent issue (e.g., Hispanics and immigration, the
military and the Iraq War), substantial changes can occur in the political landscape.
Under more common circumstances, an individual’s partisan bias weighs heavily on
their political decision-making process as they adopt the positions of ideologically-
similar elites and do not adjust their political worldview. Campaigns can use recent
technological advances to identify (i.e., microtarget) the segments of the population
who have experience with certain issues and who thus may be ripe for a political
change of heart.
In brief, the Personal Experience Model explains a voter-issue linkage that stymies
the cycle of partisan bias that pervades the political arena. The micro-level model
provides deeper understanding of the research on issue publics and self-interest. And
the implications of the model can be leveraged by campaigns to increase the impact
and efficiency of their campaign appeals.

187
Appendix A

Appendix for Chapter 2

A.1 Details for Candidate Evaluation Simulation

The following tables display the initial parameter values of the simulation in Section
2.1.5. Voters’ issue opinions, beliefs of candidate stances, and candidate evaluation
are updated each day of the hypothetical campaign using the formulae of the Personal
Experience Model.

Voters’ Prior Issue Opinions: Mean (Precision)

Voter Traits Health Care Education
Alice the Architect 0.95 (10,000) 0 (6.25) 0 (625)
Ted the Teacher 0.95 (10,000) 0 (6.25) 0.5 (100)
Doris the Doctor 0.95 (10,000) 0.5 (400) 0 (625)

Table A.1: The mean (δ1 ) and precision (τ12 ), in parentheses, of voter’s prior beliefs
on issues and traits.

The voters’ initial issue opinions are displayed in Table A.1. The precision on
traits are extremely high because voters are sure that they want high valence (e.g.,

188
competent, trustworthy, shared-values) politicians. When occupation and policy
intersect (e.g., Ted the teacher and education) the voter’s precision on this issue is
higher than otherwise.

Voter’s Prior Beliefs about Candidate Mandy’s Positions

Voter Traits Health Care Education
Alice the Architect 0.4 (6.25) 0 (6.25) 0 (6.25)
Ted the Teacher 0 (4) 0 (6.25) 0 (6.25)
Doris the Doctor 0.3 (4) 0 (6.25) 0 (6.25)
bottomrule
Voter’s Prior Beliefs about Candidate Nathan’s Positions
Voter Traits Health Care Education
Alice the Architect 0 (4) 0 (6.25) 0 (6.25)
Ted the Teacher 0.4 (6.25) 0 (6.25) 0 (6.25)
Doris the Doctor 0.3 (4) 0.5 (6.25) 0 (6.25)

Table A.2: The mean (µ1 ) and precision (ν12 ), in parentheses, of voter’s prior beliefs
on issues and traits.

Candidate Signals
Candidate Traits Health Care Education
Mandy No signal -0.5 (1.5) 0.5 (1.5)
Nathan No signal 0.5 (1.5) -0.5 (1.5)

Table A.3: The mean (γ1 ) and precision (ψ12 ), in parentheses, of the candidate’s
signals about where they stand on the issues.

The precisions are low because I simulation that the topic of conversation about
each issue lasts 10 days and that one signal is received by the voters each day from
their preferred candidate. Alternatively, I could have simulated a higher precision
and an additional parameter representing the chance that the voter receives a signal
on a day that issue is discussed. (If a voter prefers the candidates equally, there is a

189
50% chance of receiving the signal from one candidate or the other.

Per-Day Change in Issue Weight

Campaign Phase Traits Health Care Education
Phase 1 (day 1) +1.0 0 0
Phase 2 (days 2-10) 0 +0.02 -0.03
Phase 3 (days 11-20) No signal -0.02 +0.03

Table A.4: The weight (δj ) placed on each issue.

At the beginning of the campaign, voters know nothing about the candidates’
issue position and only care about traits. During the second phase of the campaign,
a discussion about health care begins, and voters add considerations about health
care to their candidate evaluations. The third and final phase of the campaign sees
the conversation shift to education to the detriment of health care. Weights are
forced to be nonnegative.

190
Appendix B

Appendices for Chapter 4

B.1 Question Wording in Nationwide Survey

B.1.1 Candidate Descriptions

• Former Governor Richard Miller has been nominated as the Democratic

United States Senate candidate for your state.

Gov. Richard Miller grew up in a working-class family, won a scholarship to

a top college, is now 52, is married and has two children. He served eight
years as the state’s governor, during which time he won re-election by a large
margin. Gov. Miller won praise for crafting an innovative health care policy,
promoting economic growth that outpaced the national average, and working
well with state leaders of the opposite party. He is endorsed by the nurses’
unions and several environmental organizations. Republican leaders, however,
claim he will raise taxes too much.

191
 • Attorney General Mark Jones has been nominated as the Republican United
States Senate candidate for your state.

Attorney General Mark Jones grew up in a small town, graduated top of his
class in law school, and began his career in the local U.S. District Attorney’s
office. After rising quickly in the ranks to head that office, he was elected
Attorney General. As Attorney General, he reduced violent crime in the state
and cracked down on corrupt politicians in the state capital; he easily won re-
election. He is 62, married, has three children and five grandchildren. Jones is
endorsed by the local farmers’ organization and the Fraternal Order of Police.
Democratic leaders, however, claim that his policies are too biased toward big
business.

B.1.2 Candidates’ Issue Signals

• America has always been a nation of immigrants and today’s hard-working

immigrants deserve the chance to achieve the American dream.

• Immigrants who entered the country illegally clearly have no respect for the
rule of law and must be sent back to their country of origin.

• Not only is Iraq a distraction from dangerous terrorists in other countries, but
the U.S. leaving Iraq will finally force the Iraqi government to take responsi-
bility for their county.

• Not only does staying in Iraq provide America security against terrorists, but

192
 America also has a duty to see the Iraq situation through to a peaceful conclu-
sion.

• We should recognize same-sex marriage so that all Americans are treated

equally. America has long outgrown its history of discrimination and we need
to extend marriage rights to all citizens.

• Marriage has always been between one man and one woman, and passing a con-
stitutional amendment affirming that principle ensures that the courts cannot
override the will of the people.

B.1.3 Experience Questions

• Next I’m going to ask about the characteristics of these people with whom you
discuss important matters. Think about the proportion of important discussion
you have with any person that fits the description below.

– About what proportion (percent) of your important discussions is with a

person who currently serves in the U.S. military?

– About what proportion (percent) of your important discussions is with a

person who once served in the U.S. military?

– About what proportion (percent) of your important discussions is with a

person who is an immigrant to the United States?

– About what proportion (percent) of your important discussions is with a

person who is Hispanic?

193
 – About what proportion (percent) of your important discussions is with a
person who is gay or lesbian?

• Including yourself, has anyone living in your house or apartment ever served in
the U.S. military? This includes the National Guard and the Reserves. Check
all that apply.

• How often do you attend religious services, apart from special events like wed-
dings and funerals?

• If you are Christian: Would you say you have been ”born again” or have
had a ”born again” experience that is, a turning point in your life when you
committed yourself to Christ?

• Regardless of whether you now attend any religious services, do you ever think
of yourself as part of a particular church or denomination? Which one?

194
B.2 Auxiliary Regressions

Issue
Inter. Var. Immigration Iraq War Gay Marriage
Issue. Experience 0.21 (0.20) 0.02 (.20) 0.29 (0.22)
Political Attentiveness 0.98 (.76) 1.5 (0.74) 2.9 (0.8)
Issue Certainty 1.4 (1.1) 2.3 (1.1) 0.0 (1.8)

Table B.1: OLS regressions of shift in candidate evaluation. Dependent variable’s

potential range is -100 to +100. Signal incongruity is the distance (magnitude) of
the candidate’s signal (either -10 or +10) to the respondent’s wave one position on
that issue. All variables listed are interacted with signal congruency, meaning the
expected sign is positive. Main effects are not shown; they are insignificant for all
regressions. Political attentiveness and issue certainty (from wave 1) are measured
on a 0 to 1 scale. N is about 270 for each regression.

Issue
Exp. Var. Immigration Iraq War Gay Marriage
Issue. Experience -0.47 (0.34) -0.41 (0.30) -0.80 (0.35)
From favored cand. 0.65 (0.74) -0.45 (0.73) 0.32 (0.69)
Attentiveness -0.26 (0.26) 0.22 (0.25) -0.33 (0.24)
Certainty 0.10 (0.15) 0.09 (0.15) -0.10 (0.13)
Constant 0.47 (1.46) -0.78 (1.50) 2.15 (1.22)

Table B.2: OLS regressions of issue opinion shift in the direction of the delivered
signal. Dependent variable’s potential range is -20 to +20. “From favored cand.” is
a dichotomous variable indicating whether the signal is delivered by the candidate
that the respondent favored in the first wave. N is about 270 for each regression.

B.3 Issue Experience Measures and the Hypothe-

ses

195
 Group Congruity:Exp. Against Bias
Hispanic 0.57 (.59) 0.79 (0.80)
... in workplace (26%) -.32 (.16) -0.2 (0.2)
... in zipcode (31%) 0.0 (0.4) 0.14 (0.63)
Talk to hisp (32%) 0.46 (0.44) 0.35 (0.61)

Serve(d) in Mil. -0.16 (.56) -0.17 (0.75)

HH Served(s) 0.39 (.074) 0.40 (0.4)
Talk to vets/mil (33%) 0.13 (.44) 0.42 (0.59)

Resp is GLBT 1.5 (1.0) 2.3 (1.3)

Talks to GLBTs (30%) 0.6 (0.5) 0.33 (0.69)
Church weekly+ -0.20 (0.5) -0.40 (0.22)
Born again 0.22 (0.46) -0.50 (0.64)

Table B.3: OLS regression coefficients for various experience measures. (See Table
B.1 for regression details.) The first column lists the experience measure and the
percentage of the population included in the measure if there is not a self-evident
boundary. The second column is the regression coefficient (and standard error), with
all respondents who received a signal on that issue included (n is about 270). The
third column is the same regression as column two but only among respondents who
received a signal that went against their partisan predisposition (n is about 135).

196
 Group Shift in Direction of Signal
Hispanic -0.9 pts (n.s)
... in workplace (26%) -1.1 pts (n.q.s)
... in zipcode (31%) no diff.
Talk to hisp (32%) -1.4 (p < .05)

Serve(d) in Mil. -0.1 (n.s.)

HH Served(s) -0.6 pts (n.q.s)
Talk to vets/mil (33%) -0.2 pts (n.s.)

Resp is GLBT -1.1 pts (n.q.s)

Talks to GLBTs -1.1 pts (p < 0.1)
Church weekly+ -2.0 pts (p < 0.02)
Born again -0.4 pts (n.s)

Table B.4: Number of points in the direction of the signal (or cue) respondents
shifted when compared with individuals not in the group listed. The Cue-Taking
hypothesis predicts these values to be negative. When the values are significant at
conventional levels, p-values are listed, n.s. means “not significant” and n.q.s means
“not quite significant at conventional levels.”

197
Thank you very much for taking my survey, which should only take 5 minutes of your
time.

Your participation is completely voluntary and you may withdraw your consent and
discontinue participant of the project at any time. Your refusal to participate will not
result in any penalty. This study has been approved by the University's Institutional
Review Panel for Human Subjects. If you have questions about the project please email
Aaron Strauss <[email protected]>. For answers to any questions you may have
about your rights as a research subject, contact Joseph Broderick, Secretary, Institutional
Review Panel for Human Subjects (609.258.3976).

If you complete the survey you will be entered in a random drawing for a $100 gift
certificate to amazon.com or iTunes (your choice!).

The survey is on multiple pages; answer the first question below and click the "Next"
button below to navigate through the pages.

1. What is your class year at Princeton?

2011 (Freshman)............................................................ 19
2010 (Sophomore) ......................................................... 20
25
2009 (Junior) ................................................................
2008 (Senior) ................................................................
36

[Candidates Rotated Below]

Below are descriptions of two hypothetical candidates running for United States Senate.
For each, please rate that candidate using something we call the feeling scale. Positive
ratings between 1 and 10 mean that you feel favorable and warm toward the person, with
10 being the warmest rating. Negative ratings between -1 and -10 mean that you don't
feel favorable toward the person and that you don't care too much for that person, with -
10 the coolest rating. You would rate the person at 0 if you don't feel particularly warm
or cold toward the person.
2. Former Governor Richard Miller has been nominated as the Democratic United
States Senate candidate for your state.

Gov. Miller grew up in a working-class family, won a scholarship to a top

college, is now 52, married and has two children. He served eight years as the
state’s governor, during which time he won re-election by a large margin. Gov.
Miller won praise for crafting an innovative health care policy, promoting
economic growth that outpaced the national average, and working well with state
leaders of the opposite party. He is endorsed by the state’s farmers organization
as well as the National Organization for Women.
-10: most negative ................................
0 +1: ................................................................
3
-9:................................ 0 +2: ................................................................
6
-8:................................ 0 +3: ................................................................
6
-7:................................ 0 +4: ................................................................
6
-6:................................ 0 +5: ................................................................
15
-5:................................ 0 +6: ................................................................
11
-4:................................ 1 +7: ................................................................
15
-3:................................ 1 +8: ................................................................
14
-2:................................ 1 +9: ................................................................
7
-1:................................ 2 7 +10: most positive................................
7 89
0: neutral ................................
4 Refused ................................ 1

3. Retired Admiral Mark Jones has been nominated as the Republican United
States Senate candidate for your state.

Admiral Jones grew up in a small town, graduated from the Naval Academy, and
climbed the ranks of the Navy quickly. He is 62, married, has three children and
five grandchildren. After retiring from service, he served as Secretary of State of
a Republican presidential administration. During his tenure as Secretary, Admiral
Jones won accolades for achieving peace in several long-standing conflicts.
Admiral Jones is endorsed by several local newspapers and the Fraternal Order of
Police.
-10: most negative ................................
0 +1: ................................................................
4
-9:................................ 0 +2: ................................................................
11
-8:................................ 0 +3: ................................................................
8
-7:................................ 1 +4: ................................................................
11
-6:................................ 0 +5: ................................................................
11
-5:................................ 1 +6: ................................................................
10
-4:................................ 2 +7: ................................................................
7
-3:................................ 3 +8: ................................................................
6
-2:................................ 4 +9: ................................................................
3
-1:................................ 4 16 +10: most positive................................
3 75
0: neutral ................................
9 Refused ................................ 0
 Prefer Democratic Candidate ........................................ 63
Prefer Republican Candidate......................................... 20
Prefer Equally................................................................ 15
Refused on either evaluation ......................................... 1

Next, please consider the following political issues.

Read the following two statements and again rate them on a scale similar to that of the
last two questions. If you agree with Statement A you have ten choices, from "A:1"
which means you slightly agree with Statement A all the way to "A:10" which means you
completely agree with Statement A. Similarly, if you agree with Statement B pick a value
between "B:1" (slightly agree with B) and "B:10" (completely agree with B). Choose "0,
Neutral" if you are unsure or agree with the two statements equally.

[Issue order rotated. A/B messages rotated]

4. With regard to U.S. foreign aid, do you think the federal government should:

A) Cut foreign aid in half because it is ineffective and costs taxpayers over 25
billion dollars a year.

B) Double foreign aid since it accounts for less than one percent of the federal
budget and helps those in need.
A:10, completely B:1, slightly
agree w/ A................................
2 agree w/ B ................................5
A:9 ................................ 0 B:2................................................................
7
A:8 ................................ 1 B:3................................................................
9
A:7 ................................ 0 8
B:4................................................................
A:6 ................................ 2 B:5................................................................
6
A:5 ................................ 2 7
B:6................................................................
A:4 ................................ 3 B:7................................................................
5
A:3 ................................ 3 B:8................................................................
4
A:2 ................................ 3 B:9................................................................
2
A:1, slightly B:10, completely
agree w/ A................................
5 21 agree w/ B ................................6 59
0: neutral ................................
20 Refused ................................ 0
5. On a different issue, do you think the federal government should:

A) Encourage businesses to invest in new, more energy efficient technology,

while letting the market shape the actions of businesses.

B) Implement a cap-and-trade system that would restrict the amount of

greenhouse gases released by businesses to a government-mandated level.
A:10, completely B:1, slightly
agree w/ A................................
5 agree w/ B ................................2
A:9 ................................ 3 B:2................................................................
5
A:8 ................................ 4 5
B:3................................................................
A:7 ................................ 5 B:4................................................................
4
A:6 ................................ 5 B:5................................................................
8
A:5 ................................ 7 6
B:6................................................................
A:4 ................................ 5 B:7................................................................
6
A:3 ................................ 4 B:8................................................................
5
A:2 ................................ 2 B:9................................................................
2
A:1, slightly B:10, completely
4 42
agree w/ A................................ agree w/ B ................................7 49
0: neutral ................................
8 Refused ................................ 1

6. A state's Supreme Court recently mandated that the state government must
legalize same-sex unions, either through civil unions or gay marriage.

In your opinion, should the state:

A) Implement civil unions--a less divisive solution that provides equal rights to
same-sex couples.

B) Recognize gay marriage; anything less would be discriminatory against gays.

A:10, completely B:1, slightly
agree w/ A................................
11 agree w/ B ................................4
A:9 ................................ 1 B:2................................................................
3
A:8 ................................ 2 B:3................................................................
2
A:7 ................................ 2 B:4................................................................
3
A:6 ................................ 3 B:5................................................................
5
A:5 ................................ 4 B:6................................................................
2
A:4 ................................ 3 B:7................................................................
4
A:3 ................................ 4 B:8................................................................
4
A:2 ................................ 3 B:9................................................................
3
A:1, slightly B:10, completely
agree w/ A................................
4 36 agree w/ B ................................
25 55
0: neutral ................................
10 Refused ................................ 0
7. On average, how many times a month have you attended a spiritual or religious
event or service (e.g., church, synagogue, mosque, meditation) during Fall term?
Never .............................................................................
40
Less than once a month ................................................. 23
Once a month................................................................6
Twice a month ............................................................... 6
Four times a month ........................................................ 15
More often than weekly................................................. 8
Refused ..........................................................................
1

8. Over the past twelve months, how many outdoor trips or excursions have you
gone on (e.g., camping, hiking, ski cabin, quiet beach)?
0: ....................................................................................
10
1: ....................................................................................
10
2: ....................................................................................
17
3: ....................................................................................
15
4: ....................................................................................
10
5: ....................................................................................
9
6: ....................................................................................
6
3
7: ....................................................................................
8: ....................................................................................
2
1
9: ....................................................................................
10 or more...................................................................... 15
Refused .......................................................................... 1

9. In the past five years, about how much cumulative time have you spent outside of
the U.S., Canada, Mexico, and western Europe (e.g., South America, Asia,
Africa, eastern Europe)?
21
None ..............................................................................
One week ....................................................................... 6
Two weeks................................................................ 6
Three or four weeks....................................................... 8
One month ................................................................ 7
Two months ................................................................15
Three to size months...................................................... 19
Seven months to a year.................................................. 5
More than one year ........................................................ 12
Refused ..........................................................................
1
10. Do you have friends or family members who are gay, lesbian, bisexual or
transgender?
No ..................................................................................
28
Yes .................................................................................
72
Refused .......................................................................... 0

11. If so, how often do you communicate with these individuals? (Feel free to check
multiple frequencies below if they are applicable to different friends and/or
family.)
Never .............................................................................
1
On occasion ................................................................16
Monthly .........................................................................12
Weekly...........................................................................11
Two or three times a week............................................. 11
Daily ..............................................................................
21

Answered “No” in previous question ............................ 28

Refused in either question .............................................
1

12. Generally speaking, do you think of yourself as a Republican, a Democrat, an

Independent, or another party? And how strong is your affiliation? If you are an
Independent, do you tend to favor (i.e., lean to) Republican candidates or
Democratic candidates?
Strong Republican ......................................................... 5
Not so strong Republican .............................................. 7
Independent, lean Republican ....................................... 10
Independent ................................................................16
Independent, lean Democrat .......................................... 24
Not so strong Democrat................................................. 24
Strong Democrat............................................................ 14
Other party................................................................ 0
Refused ..........................................................................
0
13. When it comes to politics, do you usually think of yourself as:
Extremely Conservative ................................................ 2
Conservative ................................................................10
Slightly conservative ..................................................... 11
Moderate; middle of the road ........................................ 17
Slightly liberal ...............................................................21
32
Liberal............................................................................
Extremely liberal ........................................................... 6
Other party................................................................ 0
0
Refused ..........................................................................

14. Please indicate your sex:

Female ...........................................................................
54
Male...............................................................................
45
Refused ..........................................................................
1

15. Please Enter your age in the box to the right

17-18:............................................................................. 28
19: .................................................................................. 27
20: .................................................................................. 19
21: .................................................................................. 18
22-24:............................................................................. 6
Refused .......................................................................... 1

16. Compared to other Princeton students, how much to do you pay attention to
political news?
A great deal................................................................ 7
Quite a bit ......................................................................
20
39
Some ..............................................................................
Very little.......................................................................
28
Not at all ........................................................................
6
Refused ..........................................................................
0
 Wave Two Candidate Signals

[Respondent receives two signals on distinct issues.]

Signals:

Gay Marriage
• Marriage has always been between one man and one woman. Civil unions provide
the same rights to same-sex couples without alienating the more traditional
segments of the population.
• We should recognize same-sex marriage so that all Americans are treated equally.
America has long outgrown its era of discrimination and we need to extend
marriage rights to all citizens.

Environment
• To combat global warming, we need a cap-and-trade system that will restrict the
amount of dangerous greenhouse gasses released into the atmosphere. Businesses
will always try to cut costs; without tight restrictions, our environment will
continue to suffer.
• A market-based solution to climate change is the best way to ensure the health of
the planet while maintaining a strong economy. Innovative, eco-friendly
businesses will prosper, leading to a healthy environment and continued job
growth.

Foreign Aid
• America has the world’s largest economy by far, and as global leaders we have the
responsibility to assist those less fortunate than us.
• Foreign aid is wasted on corrupt foreign governments. We should spent taxpayers
money in America, where we can better evaluate its effectiveness.
Appendix C

Appendix for Chapter 5

C.1 Computational Appendices

This appendix presents fast and approximate solutions to the planner’s optimization
problem in both partisan and nonpartisan cases.

C.1.1 Nonpartisan Case: the Knapsack Problem

To approximate the solution to the nonpartisan planner’s optimization problem de-

fined in Section 5.3.1, the key is to notice that the above linear optimization problem
is identical to the canonical knapsack problem. In the knapsack problem, one max-
imizes the total value of objects to be placed in a knapsack of fixed sized, with
each object having its own value and size. The analogous case for the nonpartisan
planner is maximizing the number of voters given a budget constraint where each
individual-treatment pairing may be thought of as an object.

206
 Following Dantzig (1957), we approximate the exact solution of this linear pro-
gramming problem by ordering the individual pairs by their maximum vote per dollar
ratio and treat the individuals with the highest such ratio first until the budget is
exhausted. If the ratio is non-positive (i.e., the best non-control treatment for an
individual does not outperform the control), this individual is not treated. In most
cases, this approximation yields solutions very close to the optimal result because the
ratio of the per-use cost of the most expensive treatment (e.g., $15 for a canvassing
shift) is tiny compared to the overall budget (usually at least $10,000). Thus, when
the addition of an expensive and efficient treatment runs just over budget and a
cheaper yet less efficient tactic should be used in its place, inefficiencies at the edge
of the problem are of little importance.

C.1.2 Partisan Case: the Stochastic Knapsack Problem

To derive a fast and approximate solution to the partisan’s optimization problem

defined in Section 5.4.3, the key is to notice that this optimization problem is identical
to the stochastic Knapsack problem, in which one maximizes the probability that the
total value of items in the knapsack equals or exceeds a target value where each
object has a random value and a known size. As in the nonpartisan case, each
individual-treatment pair can be treated as an item.
As an approximate solution to this problem, we use the algorithm that is based
on Geoffrion (1967) where subgroups are ordered by the weighted combination of the
mean and standard error of their posterior vote choice profile, π(ρ). Optimization
is performed over the weight parameter, which can take values between 1 (i.e., only

207
the means of the posteriors matter) and 0 (i.e., only the standard errors matter).
For a discussion of when this approximation fails to yield the optimal result, see
Henig (1990). The intuition behind this algorithm can be developed by considering
the following scenarios. Campaigns with a natural advantage (i.e., would garner a
majority of the vote without treatment) could further increase their probability of
winning by contacting voters who are highly responsive on average and have a low
variance of their treatment response. On the other hand, campaigns that are behind
aim to treat segments of the population that are both highly responsive and that
have high variance. Thus, unlike in the nonpartisan case, the optimal subgroups to
treat change depending on the outcome under the control. The algorithm finds an
approximate solution by limiting its search to the subspace defined by the weight
parameter, which makes optimization feasible when the dimension of δ is large.

208
Appendix D

Appendix for Chapter 6

D.1 Analytical Solution for Variance of Vote Share

The general formula for the variance of a random variable X that is composed of two
normal distributions that occur with probability p and 1 − p respectively is derived.
For an example, Figure D.1 shows the distribution with p = 0.5 and sub-distributions
of N (1, 0.16) and N (−1, 0.16).
To determine the variance of X, let Y be a Bernoulli random variable that de-
termines which guassian produces x. If the two normal distributions that constitute

209
 0.5
0.4
Probability Density

0.3
0.2
0.1
0.0

−3 −2 −1 0 1 2 3

Hypothetical Random Variable X

Figure D.1: Hypothetical distribution of a variable that is Guassian, but the mean
and variance of the Guassian depend on a Bernoulli process.

X are N (µ1 , σ12 ) and N (µ2 , σs2 ), then

Var(X) = Var(E[X|Y ]) + E[Var(X|Y )] (D.1)

Var(E[X|Y ]) = p(µ1 − X̄)2 + (1 − p)(µ2 − X̄)2 (D.2)

E[Var(X|Y )] = pσ12 + (1 − p)σ22 (D.3)

Var(X) = p[(µ1 − X̄)2 + σ12 ] + (1 − p)[(µ2 − X̄)2 + σ22 ], (D.4)

where X̄ = pµ1 + (1 − p)µ2 . Each sub-distribution contributes its own variance

and the squared distance to the mean to the overall variance, in proportion to the
probability that the distribution is activated by the Bernoulli probability.
This derivation is important because the distribution of the vote function is the

210
superposition of four normal curves, each of which occur with a certain probability.
There are four distributions because there are four possible outcomes when both
the Democratic and Republican broadcasted issue have independent probabilities of
either resolving or not resolving.
The mean of the distribution is

V¯D (δD , δR ) = µ + m(δD − δR )

The variance of Vi is a weighted average of the squared distance of the four normal
distributions that constitute Vi to V̄i . Using the derivation above, the variance is,

Var(VD (δD , δR )) = pD (σ 2 + (q(2 − δD − δR ))2 )

+pR (σ 2 + (q(2 − δD − δR ))2 )

+(1 − pD − pR )(σ 2 )

combining terms and assuming pD = pR = p

= σ 2 + 2pq 2 (2 − δD − δR )2

211
Bibliography

Achen, C. and Bartels, L. (2006). It feels like we’re thinking: The rationalizing voter
and electoral democracy. In Annual Meeting of the American Political Science
Association, Philadelphia.

Achen, C. H. and Bartels, L. M. (2005). Partisan hearts and gall bladders: Retro-
spection and realignment in the wake of the great depression. Presented at 2005
the Annual Meeting of the Midwest Political Science Association .

Ansolabehere, S., Iyengar, S., Simon, A., and Valentino, N. (1994). Does attack
advertising demobilize the electorate? The American Political Science Review 88,
4, 829–838.

Ansolabehere, S., Rodden, J., and Snyder, J. M. (2008). The strength of issues:
Using multiple measures to gauge preference stability, ideological constraint, and
issue voting. The American Political Science Review 102, 2, 215–232.

Arceneaux, K. and Kolodny, R. (2007). The messenger or the message? group

endorsements, heuristics, and grassroots campaigning. Unpublished manuscript .

212
Bartels, L. M. (1993). Messages received: The political impact of media exposure.
The American Political Science Review 87, 2, 267–285.

Bartels, L. M. (1996). Uninformed votes: Information effects in presidential elections.

American Journal of Political Science 40, 1, 194–230.

Bartels, L. M. (2000). Partisanship and voting behavior, 1952-1996. American

Journal of Political Science 44, 1, 35–50.

Bartels, L. M. (2002). Beyond the running tally: Partisan bias in political percep-
tions. Political Behavior 24, 2, 117–150.

Bartels, L. M. (2005). Homer gets a tax cut: Inequality and public policy in the
american mind. Perspectives on Politics 3, 1, 15–31.

Berger, J. O. (1985). Statistical Decision Theory and Bayesian Analysis. Springer,

New York, 2nd edn.

Blais, A. and Dobrzynska, A. (1998). Turnout in electoral democracies. European

Journal of Political Research 33, 2, 239–262.

Breiman, L., Friedman, J., Stone, C. J., and Olshen, R. (1984). Classification and
Regression Trees. Chapman & Hall, London.

Brody, R. A. and Page, B. I. (1972). Comment: The assessment of policy voting.

The American Political Science Review 66, 2, 450–458.

Campbell, A., Converse, P., Miller, W., and Stokes, D. (1960). The American Voter.
John Wiley & Sons Inc., New York.

213
Campell, J. E. (1992). Forecasting the presidential vote in the states. American
Journal of Political Science 36, 2, 386–407.

Canes-Wrone, B., Herron, M. C., and Shotts, K. W. (2001). Leadership and pander-
ing: A theory of executive policymaking. American Journal of Political Science
45, 3, 532–550.

Caplan, B. (2008). The Myth of the Rational Voter: Why Democracies Choose Bad
Policies. Princeton Univertsity Press, Princeton, NJ.

Carmines, E. and Stimson, J. (1990). Issue Evolution. Princeton University Press,

Princeton, NJ.

Carraway, R. L., Schmidt, R. L., and Weatherford, L. R. (1993). An algorithm for

maximizing target achievement in the stochastic knapsack problem with normal
returns. Naval Research Logistics 40, 2, 161–173.

Chipman, H. A., George, E. I., and McCulloch, R. E. (2008). BART: Bayesian

additive regression trees. Tech. rep., Department of Statistics, University of Penn-
sylvania. Available at http://arxiv.org/abs/0806.3286v1.

Chong, D., Citrin, J., and Conley, P. (2001). When self-interest matters. Political
Psychology 22, 3, 541–570.

Cohen, J. E. and King, J. D. (2004). Relative unemployment and gubernatorial

popularity. The Journal of Politics 66, 4, 1267–1282.

Converse, P. E. (1964). The nature of belief systems in mass publics. In D. E. Apter,

ed., Ideology and Discontent, 75–169. Free Press, New York.

214
Cox, D. R. (1958). Planning of Experiments. John Wiley & Sons, New York.

Dale, A. and Strauss, A. (2008a). Don’t forget to vote: Text message reminders as
a mobilization tool. Unpublished manuscript .

Dale, A. and Strauss, A. (2008b). Mobilizing the mobiles: How text messaging can
boost turnout. In C. Panagopoulos, ed., Politicking Online: The Transformation
of Election Campaign Communications. Rutgers University Press, New Brunswick,
NJ.

Dalton, R. J. (2008). The quantity and the quality of party systems: Party system
polarization, its measurement, and its consequences. Comparative Political Studies
41, 7, 899–920.

Dantzig, G. B. (1957). Discrete-variable extremum problems. Operations Research

5, 2, 266–277.

Dehejia, R. (2005). Program evaluation as a decision problem. Journal of Econo-

metrics 125, 141–173.

Delli Carpini, M. and Keeter, S. (1996). What Americans Know about Politics and
Why It Matters. Yale University Press, New Haven, CT.

Downs, A. (1957). An Economic Theory of Democracy. Harper & Row, New York.

Eldersveld, S. J. (1956). Experimental propaganda techniques and voting behavior.

American Political Science Review 50, 1, 154–165.

215
Enelow, J. M. and Hinich, M. J. (1984). The Spatial Theory of Voting : An Intro-
duction. Cambridge University Press, New York.

Fazio, R. H. and Zanna, M. P. (1978). On the predictive validity of attitudes: The

roles of direct experience and confidence. Journal of Personality 46, 2, 228–243.

Feddersen, T. J. and Pesendorfer, W. (1996). The swing voter’s curse. The American
Economic Review 86, 3, 408–424.

Ferejohn, J. (1986). Incumbent performance and electoral control. Public Choice 50,
1, 5–25.

Festinger, L. (1957). A theory of cognitive dissonance. Stanford University Press,

Stanford, CA.

Fiorina, M. P. (1981). Retrospective Voting in American National Elections. Yale

University Press, New Haven, CT.

Fiorina, M. P. (1996). Divided Government. 2nd ed. Allyn and Bacon, Needham
Heights, MA.

Geoffrion, A. M. (1967). Solving bicriterion mathematical programs. Operations

Research 15, 1, 39–54.

Gerber, A. and Green, D. P. (1998). Rational learning and partisan attitudes. Amer-
ican Journal of Political Science 42, 3, 794–818.

Gerber, A. S. (2004). Does campaign spending work?: Field experiments provide

evidence and suggest new theory. American Behavioral Scientist 47, 5, 541–574.

216
Gerber, A. S. and Green, D. P. (2000). The effects of canvassing, telephone calls,
and direct mail on voter turnout: A field experiment. American Political Science
Review 94, 653–663.

Gerber, A. S., Green, D. P., and Shachar, R. (2003). Voting may be habit-forming:
Evidence from a randomized field experiment. American Journal of Political Sci-
ence 47, 3, 540–550.

Gershkoff, A. R. (2006). How issue interest can resue the american public. Doctoral
Disseration, Princeton University, Department of Politics .

Gilens, M. (2001). Political ignorance and collective policy preferences. American

Political Science Review 95, 2, 379–396.

Gilens, M. and Murakawa, N. (2002). Elite cues and political decision-making. In

M. X. Delli Carpini, L. Huddy, and R. Y. Shapiro, eds., Research in Micropol-
itics: Political Decision-Making, Deliberation, and Participation (Vol. 6), X–Y.
JAI Press, New York.

Gosnell, H. F. (1927). Getting-Out-the-Vote: An experiment in the stimulation of

voting. University of Chicago Press, Chicago.

Grant, D. A., Hake, H. W., and Hornseth, J. P. (1951). Acquisition and extinction of
a verbal conditioned response with differing percentages of reinforcement. Journal
of Experimental Psychology 41, 1.

Green, D. P. and Gerber, A. S. (2008). Get Out the Vote: How to Increase Voter
Turnout. Brookings Institution, Washington D.C., 2nd edn.

217
Groseclose, T. (2001). A model of candidate location when one candidate has a
valence advantage. American Journal of Political Science 45, 4, 862–886.

Gunter, L., Zhu, J., and Murphy, S. (2007). Variable selection for optimal decision
making. Poceedings of the 11th Conference on Artificial Intelligence in Medicine
AIME 2007, LNCS/LNAI 4594, 149–154.

Hartmann, G. W. (1936). A field experiment on the comparative effectiveness of

“emotional” and “rational” political leaflets in determining election results. Jour-
nal of Abnormal Psychology 31, 1, 99–114.

Henig, M. I. (1990). Risk criteria in a stochastic knapsack problem. Operations

Research 38, 5, 820–825.

Hibbs Jr., D. A. (2000). Bread and peace voting in u.s. presidential elections. Public
Choice 104, 1, 149–180.

Hill, J. L. and McCulloch, R. E. (2008). Bayesian nonparametric modeling for causal

inference. Tech. rep., Graduate School of Business, University of Chicago.

Hotelling, H. (1929). Stability in competition. The Economic Journal 39, 153, 41–57.

Iyengar, S. and Kinder, D. (1989). News that matters : Television and American
opinion. University Of Chicago Press, Chicago, IL.

Kelleher, C. A. and Wolak, J. (2006). Priming presidential approval: The condition-

ality of issue effects. Political Behavior 28, 3, 193–210.

218
Key, V. (1968). The Responsible Electorate: Rationality in Presidential Voting, 1936-
1960. Harvard University Press.

Knutson, K. M., Wood, J. N., Spampinato, M. V., and Grafman, J. (2006). Politics
on the brain: An fmri investigation. Social neuroscience 1, 1, 25–40.

Kramer, G. (1966). A decision-theoretic analysis of a problem in political campaign-

ing. In J. L. Bernd, ed., Mathematical Applications in Political Science, Vol. II,
137–160. Southern Methodist University, Dallas, TX.

Krosnick, J. A. (1990). Government policy and citizen passion: A study of issue

publics in contemporary america. Political Behavior 12, 1, 59–92.

Lau, R. R. and Redlawsk, D. P. (2001). Advantages and disadvantages of cognitive

heuristics in political decision making. American Journal of Political Science 45,
4, 951–971.

Lenz, G. (2006). What Politics is About. Ph.D. thesis, Princeton University.

Levitt, S. D. and Wolfram, C. D. (1997). Decomposing the sources of incumbency

advantage in the u. s. house. Legislative Studies Quarterly 22, 1, 45–60.

Lodge, M., McGraw, K. M., and Stroh, P. (1989). An impression-driven model of

candidate evaluation. The American Political Science Review 83, 2, 399–419.

Lord, C. G., Ross, L., and Lepper, M. R. (1979). Biased assimilation and attitude
polarization: The effects of prior theories on subsequently considered evidence.
Journal of Personality and Social Psychology 37, 11, 2098–2109.

219
Lupia, A. (1994). Shortcuts versus encyclopedias: Information and voting behavior
in california insurance reform elections. The American Political Science Review
88, 1, 63–76.

Lupia, A. and McCubbins, M. D. (1998). The Democratic Dilemma: Can Citizens

Learn What They Need To Know? Cambridge University Press, New York.

Malchow, H. (2003). The New Political Targeting. Campaigns & Elections, Wash-
ington, D.C.

Malhotra, N. and Kuo, A. G. (2009). Emotions as moderators of information cue

use. American Politics Research 37, 2, 301–326.

Manski, C. F. (2005). Social Choice with Partial Knowledge of Treatment Response.

Princeton University Press.

Markus, G. B. and Converse, P. E. (1979). A dynamic simultaneous equation model

of electoral choice. The American Political Science Review 73, 4, 1055–1070.

McCarty, N., Poole, K. T., and Rosenthal, H. (2008). Polarized America: The Dance
of Ideology and Unequal Riches. MIT Press, Cambridge, MA.

McFadyen, R. G. (1998). Attitudes toward the unemployed. Human Relations 51,

2, 179–199.

Mueller, J. E. (1970). Presidential popularity from truman to johnson. The American

Political Science Review 64, 1, 18–34.

220
Mutz, D. C. (1992). Mass media and the depolitization of personal experience.
American Journal of Political Science 36, 2, 483–508.

Nickerson, D. and Arceneaux, K. (2006). Who is mobilized to vote? a meta-analysis

of 7 field experiments. In Annual Meeting of the Midwest Political Science Asso-
ciation.

Nickerson, D. W. (2007). Quality is job one: Professional and volunteer voter mobi-
lization calls. American Journal of Political Science 51, 2, 269–282.

Nickerson, D. W. (2008). Is voting contagious?: Evidence from two field experiments.

American Political Science Review 102, 1, 49–57.

Page, B. and Shapiro, R. (1992). The Rational Public: Fifty Years of Trends in
American’s Policy Preferences. University Of Chicago Press.

Parker-Stephen, E. (2007). When perceptions polarize: How motives and information

shape partisan inference. In Annual Meeting of the Midwest Political Science
Association, Chicago.

Petrocik, J. R. (1996). Issue ownership in presidential elections, with a 1980 case

study. American Journal of Political Science 40, 3, 825–850.

Petty, R. E. and Cacioppo, J. T. (1981). Attitude and Persuasion: Classic and

Contemporary Approaches. Wm. C. Brown, Dubuque, IA.

Pocock, S. J., Assmann, S. E., Enos, L. E., and Kasten, L. E. (2002). Subgroup
analysis, covariate adjustment and baseline comparisons in clinical trial reporting:
current practice and problems. Statistics in Medicine 21, 2917–2930.

221
Popkin, S. (1994). The Reasoning Voter: Communication and Persuasion in Presi-
dential Campaigns. University of Chicago Press, Chicago, IL.

Regan, D. T. and Fazio, R. H. (1977). On the consistency between attitudes and

behavior: Look to the method of attitude formation. Journal of Experimental
Social Psychology 13, 1, 28–45.

Richard Johnston, Michael G. Hagen, K. H. J. (2004). The 2000 Presidential Election

and the Foundations of Party Politics. Cambridge University Press.

Ripley, B. D. (1996). Pattern Recognition and Neural Networks. Cambridge Univer-

sity Press, Cambridge.

Rubin, D. B. (1990). Comments on “On the application of probability theory to

agricultural experiments. Essay on principles. Section 9” by J. Splawa-Neyman
translated from the Polish and edited by D. M. Dabrowska and T. P. Speed.
Statistical Science 5, 472–480.

Sanders, M. S. (2000). Uncertainty and turnout. Political Analysis 9, 1, 45–57.

Sartori, G. (1976). Parties and party systems. Cambridge University Press, New
York.

Savage, L. J. (1951). The theory of statistical decision. Journal of the American

Statistical Association 46, 253, 55–67.

Sawyer, A. G. (1973). The effects of repetition of refutational and supportive adver-

tising appeals. Journal of Marketing Research 10, 1, 23–33.

222
Sears, D. O. and Funk, C. L. (1990). Self-interest in americans’ political opinions. In
J. J. Mansbridge, ed., Beyond self-interest, 147–170. University Of Chicago Press,
Chicago.

Sellers, P. J. (1998). Strategy and background in congressional campaigns. The

American Political Science Review 92, 1, 159–171.

Sigelman, L. and Buell, Emmett H., J. (2004). Avoidance or engagement? issue con-
vergence in u.s. presidential campaigns, 1960-2000. American Journal of Political
Science 48, 4, 650–661.

Sosnick, D. B., Dowd, M. J., and Fournier, R. (2006). Applebees America: How Suc-
cessful Political, Business, and Religious Leaders Connect with the New American
Community. Simon & Schuster, New York.

Sotirovic, M. (2001). Effects of media use on complexity and extremity of attitudes

toward the death penalty and prisoners rehabilitation. Media Psychology 3, 1,
1–24.

Stoye, J. (2009). Minimax regret treatment choice with finite samples. Journal of
Econometrics Forthcoming.

Taber, C. S. and Lodge, M. (2006). Motivated skepticism in the evaluation of political

beliefs. American Journal of Political Sciencey 50, 3, 755–799.

van Knippenberg, D. and Daamen, D. (1996). Providing information in public opin-

ion surveys: Motivation and ability effects in the information-and-choice question-
naire. International Journal of Public Opinion Research 8, 1, 70–82.

223
Verplanken, B. (1991). Persuasive communication of risk information: A test of cue
versus message processing effects in a field experiment. Personality and Social
Psychology Bulletin 17, 2, 188–93.

Wald, A. (1950). Statistical Decision Functions. Wiley, New York.

Washington, E. (2008). Female socialization: How daughters affect their legislator

fathers voting on womens issues. American Economic Review 98, 311–332.

Weston, D. (2007). The Political Brain: The Role of Emotion in Deciding the Fate
of the Nation. Public Affairs, Cambridge, MA.

Wood, W. (1982). Retrieval of attitude-relevant information from memory: Effects

on susceptibility to persuasion and on intrinsic motivation. Personality & social
psychology bulletin 42, 5, 798–810.

Zaller, J. R. (1992). The Nature and Origins of Mass Opinion. Cambridge University
Press, New York.

224