Sticking with Your Vote : Cognitive Dissonance and Political Attitudes

Sticking with Your Vote:
Cognitive Dissonance and Political Attitudes

Sendhil Mullainathan (Harvard University and NBER)
Ebonya Washington (Yale University and NBER)*
December 2007
A b s t r a c t

Cognitive dissonance theory predicts that the act of voting for a candidate
leads to a more favorable opinion of the candidate in the future. We find support
for the empirical relevance of cognitive dissonance to political attitudes. We
examine the presidential opinion ratings of voting age eligibles and ineligibles
two years after the president's election. We find that eligibles show 2-3 times
greater polarization of opinions than comparable ineligibles. We find smaller
effects when we compare polarization in opinions of senators elected during high
turnout presidential campaign years with senators elected during non-presidential
campaign years.



*We are extremely grateful to Thomas Wang for superb research assistance and to LaRue Allen, Stephen Ansolabehere,
Jonathan Bendor, Don Green, Greg Huber, Lawrence Katz and Todd Rogers for very helpful suggestions. We would also
like to thank seminar participants at the Toulouse Summer School in Psychology and Economics, Yale, University of
Pittsburgh and Berkeley for comments. e-mail: mullain@fas.harvard.edu and ebonya.washington@yale.edu

---

---

1 Introduction
While models of voting behavior vary considerably, one common assumption of
models of both turnout and vote choice is that voting behavior is an expression of preferences
or beliefs. Whether voters are motivated by a desire to shift the outcome of the election to
their desired outcome, as in instrumental models, or by the desire to express strong feelings
on the part of themselves or of their group, as in intensity and popularity models respectively,
voting models assume that preferences are a variable in the voting decision equation. (See for
example Aldrich, 1993; Coate and Conlin, 2004; Grossman and Helpman, 2001 and Matsuaka
and Palda, 1999 for summaries of voter turnout models.) Psychologists on the other hand have
highlighted that causation may also run in the opposite direction: actions themselves may
drive preferences and beliefs. Numerous experiments have led to the conclusion that
behavioral change may precede attitudinal change (Bandura, 1989). One explanation for the
impact of behaviors on beliefs is cognitive dissonance (Festinger, 1957) which refers to one's
internal need for consistency. If an individual performs an activity that is antithetical to his
beliefs, the individual may unconsciously change his beliefs to alleviate the discomfort of
having inconsistent attitudes and actions.1
For example, in a classic experiment (Festinger and Carlsmith, 1959), subjects were
asked to perform for an hour the boring task of placing knobs on pegs, turning them, and then
taking the pegs off again. After the task was completed, each experimental subject was told
that the research assistant had not shown up and that the scientists needed the subject's help in
recruiting more participants. Subjects were told they would receive either $1 or $20 ($7 or
$135 in 2006 dollars) for their assistance. After each subject convinced the new recruit, really
a confederate of the experiment, that the task was fun, the subject was asked for a rating of

1 This impact of behavior on attitudes is most commonly known as cognitive dissonance, which is also how
we will refer to it. Psychological research on the other hand has shown that several other mechanisms besides
the one emphasized by a narrowly defined cognitive dissonance theory could produce a similar effect. For
example, self perception theory (Bem, 1967) provides a second explanation for the impact of behaviors on
beliefs: Individuals infer their opinions from their own actions. The publication of the 1967 article sparked
great debate: “But as evidence began to accumulate that dissonance was indeed an unpleasant state of arousal,
self-perception theory began to wane as an explanation for dissonance phenomena.” (Hogg and Cooper, 2003).
We continue to use the phrase cognitive dissonance because of its use in common parlance, and not to signify

1

---

how much s/he truly enjoyed the experiment. Those who were paid $1 rated the task much
more favorably than those who were paid $20. The theory of cognitive dissonance explains
the result by assuming that most people believe themselves to be truthful unless they have strong
incentives to behave otherwise. Those who were told they would receive $20 had a strong incentive
to lie. Those who were told they would receive $1 had no such incentive. Therefore those in the $1
group felt the discomfort (dissonance) of having inconsistent actions (lying to new recruit) and
beliefs (I am a truthful person and the task was really boring). Unable to change the past action of
telling the new recruit that the task was enjoyable, those in the $1 group had no other option but to
change their belief that the task was boring to believe that the task was, in fact, enjoyable.2
In the intervening years, empiricists have discovered that dissonance can be aroused
even when behaviors are in line with attitudes. Aronson et. al (1991) had college students
create videos to encourage high school students to practice safe sex. Some of these students
were then asked to think about their own failures to use condoms in the past. For this group of
students, the discrepancy between their past behavior and the message they were currently
preaching appears to have aroused dissonance. Members of this group stated greater intentions
to use condoms in the future.3 Dissonance is not limited to cases in which subjects are trying
to persuade others. Since the 1959 study, empiricists have explored the relevance of cognitive
dissonance to a large variety of contexts including socialization of children, curing snake
phobias, interpersonal attraction, proselytizing, gambling and water conservation (Aronson,
1999).
Applying cognitive dissonance to the context of voting, two years after an election a
citizen who voted for a candidate may hold a favorable opinion of that politician in part to
avoid the internal discomfort of having voted for a person for whom the individual has a poor

a position on which specific psychological mechanisms may be at work.
2 Because the idea of being a truthful person was a longer held and probably more fundamental part of a
person’s self-perception, it was likely easier for subjects to change opinions of the experiment than of their
own veracity.
3 The treatment group not only had greater intentions, they also had greater behavioral changes in condom use.
(Stone et .al, 1994).

2

---

opinion.4 Suggestive evidence that dissonance applies in the voting context comes most
recently from Besley and Joslyn (2001).5 Using the National Election Study, the authors
demonstrate that those who report voting in the presidential election show greater polarization
in their ratings of the two candidates immediately after the election than immediately before
(as compared to those who do not report having voted). Clearly, the endogeneity of the voting
decision (and the reporting of the voting decision) leaves us unable to treat these results as
evidence of a causal link between voting and increased polarization. Suppose that in
December 2004 we compare two Republicans, both of whom favored Bush in the 2004
election, but only one of whom actually turned out to vote. Let's call them Persons V (voter)
and N (non-voter). It would not be surprising to observe, even under the traditional purview,
that Person V had a more positive opinion of Bush than Person N. If probability of turnout is
increasing in intensity of preference and preferences linger, such a correlation would arise
nearly mechanically. Thus, in our test of the relevance of cognitive dissonance to voting
behavior, our identification strategy must rely on variables that impact turnout but that are
independent of voter preferences.
The age restriction on voting is one such variable. Consider two individuals in 1996
who both support Clinton and who are eighteen and seventeen years old respectively. Let's
call them Persons O (older) and Y (younger). The eighteen year old is able to express his
support of Clinton at the polls, whereas the seventeen year old is not. Suppose now we
examine their views of Clinton in 1998 when his approval ratings had fallen. Traditional
models predict no systematic differences in the trends in views of these two persons. On the
other hand, cognitive dissonance theory would lead us to believe that Person Y (now 19)
would show a greater fall in his Clinton approval rating than Person 0 (now 20). Person 0

4 These processes may be reinforced by social networks. Individuals may find it hard to renounce a position
that they have not only voted in accordance with, but also announced publicly, previously. In fact in a survey
of New York City voters in the week prior to the 2005 mayoral election, we found that simply asking voters
whether they planned to vote (to which they nearly unanimously replied yes) increased turnout by four
percentage points in a group whose mean turnout was extremely high (88 percent). Respondents were
randomly selected to the commit/not commit to vote groups.
5 Ginsberg and Weissberg (1978), Sticker (1964) and Thomsen (1938) also examine opinion change in views
of candidates from before to after an election.

3

---

having made a concrete act of commitment to Clinton, through voting, would find it harder to
walk away from that position. In contrast, if the two had been Dole supporters in 1996, the
psychological model would predict a different pattern of change, from before to after the
election. Cognitive dissonance theory would now predict that Person 0 would show the larger
decline in Clinton approval, jumping on information that supported his previous action,
whereas Person Y would show a smaller decline viewing new information in a more
moderated way. This logic implies a fairly simple prediction: voting eligibles should show
greater post-election polarization than voting ineligibles.
We test this prediction using data from US elections from 1976 to 1996. Our dependent
variables are individuals' ratings of presidential performance, gathered two years after each
presidential election in the National Election Study. Our sample consists of young people who were
eligible to vote in the previous election (20 and 21 year olds) as well as those who were ineligible
(18 and 19 year olds). We compare the polarization (by party) of these two groups in their attitudes
towards the president. For a variety of attitudinal measures, we find a great deal of increased
polarization. Eligible youth are nearly two times as polarized as ineligible ones. When we scale our
results by turnout, we find that voters are three times as polarized as nonvoters. We examine a
second source of exogenous variation in voter turnout. Senatorial elections vary greatly in turnout
depending on whether they are held in presidential or interim election years. For voters of all ages,
we compare attitudes towards senators based on the year in which the senator was most recently
(re)elected. We again find differential polarization by party. Attitudes towards senators elected in
presidential years are roughly 25 percent more polarized than towards senators elected in non-
presidential years.6 Scaling by the turnout differential between the two groups, we find that the act
of voting more than doubles polarization.
Before concluding that these results demonstrate the applicability of cognitive
dissonance theory to voter turnout, we first consider the relevance of three potentially
confounding factors. First, our results may be driven by age-induced polarization. Perhaps,

6 Turnout percentages are calculated using the National Election Survey’s voter verification survey conducted
in 1976, 1978, 1980, 1984, 1986, 1988 and 1990.

4

---

older Americans have more divergent views.7 The senatorial results demonstrate that aging
cannot explain all of our results. We further test the aging hypothesis directly in two ways.
First, we compare 20 and 21 year olds to 22 and 23 year olds, thus comparing different age
groups all of whom were eligible to vote in the previous presidential election. In contrast to
the age induced polarization hypothesis, we find greater polarization among the younger
group than the older. However the difference is not statistically significant. In a second
placebo test we focus on presidential election years, comparing opinions of 18 and 19 year olds
to those of 20 and 21 year olds. Neither of these groups was eligible to vote for the sitting
president. Again, we find no significant difference in polarization between groups. These
results suggest that our findings are not due merely to aging.
Second, we consider the possibility that our results are biased by the fact that party is
measured at the same time as the attitudinal surveys are conducted. Perhaps a respondent's
positive (negative) feelings about the president lead her to report herself as a member (not a
member) of the president's party. We address the potential endogeneity of party report in two
ways. First, we substitute predicted vote choice (based on demographic characteristics) for
the party variable. Results are qualitatively unchanged. Our second test circumvents the need
for a party variable. We show that eligibles are unconditionally more likely to hold views in
the tails of the presidential feeling distribution.
Finally, we consider that increased polarization may be driven by information rather than
dissonance. Perhaps those who vote collect more political information during or after the electoral
campaign. If they interpret this information in a manner that is favorable to their candidate, as
confirmatory bias would suggest (Lord, Ross and Lepper, 1979) they will show increased
polarization. To test this hypothesis, we examine knowledge, exposure and interest in politics
for eligible and ineligible youth. We find no significant differences in levels of knowledge,
exposure or interest, suggesting that interest and information are not the mechanism driving
the link between voting and increased polarization.

7 This would be consistent with (Converse, 1969) evidence that partisanship is increasing in length of
membership in party.

5

---

As a whole, these results suggest the practical importance of cognitive dissonance
theory for the political arena.8 This finding has implications for the political capital of
politicians. By definition in a two party race more than half of voters vote for the winner.
Therefore elected officials receive a boost in their approval ratings due to dissonance.
In the remainder of the paper we present our results more formally. In Section 2, we
discuss the data and methodology. Results that exploit the voting age regulations are
presented in Section 3. Results that exploit the variation in turnout between presidential and
non-presidential years are shown in Section 4. Section 5 concludes by discussing implications
of our results for voting behavior.
2 E m p i r i c a l M e t h o d o l o g y a n d D a t a
The core hypothesis we would like to test is that voting for a particular candidate
today increases one's opinion of that candidate in the future. The key difficulty in testing this
theory is that causality also surely runs in the other direction as well. Individuals who have
stronger feelings about a particular candidate are more likely to vote. Moreover, conditional
on voting, individuals who have more positive feelings about Candidate A are more likely to
vote for Candidate A. This reverse causality means that it would be unsurprising to find a
positive correlation between voting for a candidate and feelings for that candidate in the
future.
To address this concern we need an exogenous factor that drives voters to vote that is
unrelated to their preferences at the time. We identify two such factors. The first is age. Only
individuals who have reached the age of 18 on Election Day may cast a ballot. The second is the
timing of the election. Turnout is higher in presidential elections than in interim elections. Thus
there are exogenous shifts in turnout for congressional elections.
2 . 1 Age R e s t r i c t i o n s
During the time frame of our study (1976-1996) the minimum voting age in the United
States was 18. Only individuals who reached their 18th birthday by the date of the election

8 Akerlof and Dickens (1982) discuss the potential applications of dissonance theory to social security,
innovation and advertising. Rabin (1994) considers the implications of cognitive dissonance for efforts to

6

---

were eligible to vote. Those who were 17 or under on that day were ineligible. This
discontinuity allows us to compare the opinions of ineligible and eligible voters two years
after the election. Dissonance theory predicts that eligible voters will show more polarization
in their opinion of the candidate than ineligible voters.
An empirical problem in implementing this strategy is that we must impute which
candidate an ineligible voter would have voted for had they voted. We, therefore, examine
polarization by party affiliation for eligible and ineligible voters. Such a procedure is sensible
since party identification strongly predicts vote choice (Keith et. al, 1992 and Miller and
Shanks, 1996).
Empirically, we estimate an equation of the form:
(1) Opinion of Presidentit = a + b E l i g i b l ei(t-1) + c P a r t y i + d ( E l i g i b l e i(t-1) * P a r t y i ) +
?

it.
Here the variable "Opinion of President" is a rating of the respondent’s feelings toward
some aspect of the president's leadership. The variable "Eligible" indicates that the respondent
was eligible to vote (18 or older on Election Day) and the variable "Party" is a dummy for
whether the political party of the respondent and the president coincide. Thus we would
expect the coefficient c to be positive since politicians are typically viewed more favorably by
members of their own party. The interaction term "Eligible*Party" is our independent variable
of interest. Cognitive dissonance theory says that those who were old enough to vote and did
in fact vote for the elected official in period t-1 should have a higher opinion of the official in
period t. A positive and significant coefficient d would support this theory.
Our data are drawn from the National Election Study (NES), a survey of the political
behaviors and opinions of a cross-section of voting age Americans. Respondents’ views are
collected in face-to-face interviews conducted in the fall of even numbered years.
As presidential elections occur every four years, the NES provides data from both
presidential and interim election years. We focus on non-presidential election years,

promote social change.

7

---

comparing those who were eligible to vote in the prior election to those who were ineligible.9
Our sample period is from 1978 to 2000;10 therefore we have six non-presidential election
years of data.
The great advantage of the NES is its multitude of questions on individuals' opinions
of elected officials. Our main dependent variable is a thermometer question which asks
respondents to rate their feelings regarding the president on a scale from 0 to 100. Other
questions are more specific and more crudely measured. The NES asks respondents to rate on
a scale from 1 to 4 the extent to which they agree that the president is inspiring, is
knowledgeable, is moral, is a good leader, is caring, earns your approval in general and earns
your approval in his handling of the economy. The NES asks respondents whether they agree
with the statement that the president makes them afraid, angry, hopeful and proud and
whether they approve of the president.
Since our identification strategy exploits the voting eligibility age threshold, we focus
on young people. For greater comparability between eligible and ineligible voters we limit our
sample to individuals whose age places them within two years of voting eligibility. Hence we
limit our sample to individuals aged 18-21 at time t, the non-presidential year. This sample
consists of two groups. Those who are 18-19 were 16-17 at the time of the election and were
ineligible to vote. Those who are 20-21 were 18-19 at the time of the election and were
eligible to vote.11
This reduces our sample to 554, of whom 248 were ineligible and 306 were eligible to
vote. Table 1 presents the means for the full sample (column 1) and by age group (columns 2

9 The cross-sectional nature of the data means that we measure party affiliation in the current period and do
not know a person’s party affiliation two years prior. This could potentially produce some endogeneity issues
if party affiliation is changing in the two year period, an issue we address in our discussion of potential
confounds. One would have thought that since we are looking at changes in opinions, the short panels in the
NES would be useful. Unfortunately, because those who are ineligible to vote are too young to be interviewed
in the first wave of the NES panels, these data do not enable us to improve our identification strategy.
10 The 1974 election data is excluded because while individuals voted for Nixon in 1972, he had left office by
November 1974, the survey date. The 1968 election is excluded because of the undersampling of 18-21 year
olds in 1970. The undersampling of young people prior to 1972 in states in which they did not have the
franchise prevents us from exploiting the variation in voting age restrictions across states as an additional
source of variation. Elections prior to 1968 are excluded because the NES did not collect the opinion
thermometer we use as our key dependent variable.
11 The small sample of young voters in the NES dictates the window. Larger samples would have allowed us

8

---