Determinants of Defection:
The Effect of Information Signals on Strategic Voting
Abstract
While theoretical work on strategic voting emphasizes the importance of elite messages in
persuading minor party supporters to abandon their first preference, few empirical studies have
examined this relationship. I argue that while poll results certainly increase the likelihood of
changing one’s vote, explicit information signals can increase this probability even more.
Furthermore, these effects will be moderated by the presence of a counter message and the
sponsor of the explicit information signal. These hypotheses are tested with data generated from
two experiments.
Minor party candidates are becoming a permanent fixture in elections in the United States.
While minor party candidates have had a difficult time winning higher levels of office, their
effects on the outcomes of elections have been felt across many contests, most notably the 2000
U.S. Presidential election. After the election, many scholars and pundits argued that Nader’s
presence in the race cost Al Gore the election, presumably the second preference of Nader
supporters (Burden 2003). The 2000 election is not an anomaly: in 2002, Libertarian candidates
may have cost Republican candidates the election in Senate races in South Dakota and Nevada
and in gubernatorial races in Wisconsin and Oregon (Miller 2002), and in 2004, the Democratic
candidate may have lost the senate seat in Alaska given the presence of the Green Party candidate
(Tsong 2004) and the gubernatorial race in Washington due to the Libertarian party candidate
(Modie 2004).
The examples above illustrate the possibility that the presence of minor party candidates
and the behaviour of their supporters may cost the Condorcet winner, or the alternative that beats
all others in pair wise comparison, the election. However, if minor party supporters had switched
their vote to their second preference in these contexts, then the Condorcet winner may have
carried the election. With so many examples of when strategic voting could have made a
difference, I examine the conditions under which minor party supporters abandon their first
choice and vote for their second preference. Extant scholarship, primarily experimental based,
has offered factors such as electoral history, poll results, ballot position, and signals from
campaign spending (e.g., Felsenthal et al. 1988; Forsythe et al. 1993; Rapoport et al. 1991; Rietz,
Myerson and Weber 1998). However, such studies generally provide poll results in the context of
perfect information, which is rarely, if ever, present in elections. Furthermore, studies have not
examined how more explicit campaign messages might impact the likelihood of defection, or how
the presence of certain counter-messages might dampen such cues.
There are three questions that this paper addresses. First, do information signals about
1
the competitive context of the race increase the likelihood that minor party supporters switch their
vote to their second preference? 1 Second, does the presence of explicit information signals
further increase this probability? Finally, what factors (message source and presence of countercoordination messages) condition an individual’s receptivity to explicit information signals?
These questions are addressed using two experimental designs in which subjects were
presented with a context in which their first preference had little chance of winning, while their
least preferred candidate was leading the race. However, if a certain percentage of subjects
switched their vote, then the second preferred candidate could win. An experiment is beneficial
to examining the factors that help subjects switch their vote in that the effects of different types of
information can be isolated; thus, we know if subjects were exposed only to poll results or to poll
results and an explicit information signal, such as “a vote for your first preference is a vote for
your least preferred candidate.” Finally, experiments make it easier to test the conditional effects
of the sponsor of the message and the presence of counter messages, which is often difficult to
measure with data from campaigns.
Answers to these questions have broad implications for the quality of representation in
the American political system. As mentioned before, if the behaviour of minor party supporters
results in the Condorcet winner losing real elections, it indicates a clear problem in the quality of
representation. If we know about the factors that lead minor party supporters to switch their vote
to their second preference, then we might be more likely to avoid these failures. Presumably,
with knowledge of these factors, the major party candidate who is being harmed by the presence
of the minor party, or interest groups with concerns similar to minor party supporters, will know
how to appeal to this segment of the electorate.
Past Work on Information Signals and Strategic Voting
Strategic voting occurs when voters cast ballots for their second preference as opposed to
their first preference because they do not want to cast a “wasted” vote. Such a context may occur
2
when voters perceive their first preference as having little to no chance of winning and when the
race is close between their second and third preference.
McKelvey and Ordeshook (1972) developed an expected utility model which
demonstrates that if an individual’s first preference has little chance of winning, the race is close
between her second and third preference, and the difference in utility between them is sufficiently
large, then she might be inclined to vote for her second preference. Many scholars have applied
this model (or a variant of it) to multi-candidate elections across countries and across time, and
have demonstrated strong support for it (e.g., Abramson et al. 1992, 2004; Bartels 1988; Black
1978; Blais and Nadeau 1996; Ordeshook and Zeng 1997).
There is however one problem—the probability that a voter casts a decisive vote will
always be close to zero in a large electorate. Riker and Ordeshook (1968) acknowledged this
problem, and argued that what matters is the voter’s subjective sense of being decisive between
pairs of candidates, which will be inflated on its own or due to messages being sent by elites (also
see Aldrich 1993). In terms of the latter, elites send cues that the probability of an individual’s
vote being decisive is higher than it is objectively.
Game theoretic models have resolved the “irrationality” of strategic voting by
demonstrating that strategic voting occurs when voters take into account possible actions by other
voters. In one class of games characterized by complete and perfect information and sequential
voting (e.g. Farquharson 1969; Niemi and Frank 1981; Niou 2001; and Rapoport, Felsenthal and
Maoz 1991), elites incite strategic voting by coordinating the behaviour of voters with the same
preferences. In another class of games characterized by simultaneous games of incomplete
information, elites play an indirect role in the publishing of poll results (e.g., Cox 1997; Fey
1997; Myerson and Weber 1993; Palfrey 1989).
While all of the theoretical models posit some role for elites, few empirical studies of
multicandidate elections have looked at the role elites play in inciting strategic voting. Some
3
exceptions include studies in the U.K., where scholars have presented qualitative evidence of
efforts by the parties to woo minor party supporters (Evans, Curtice and Norris 1998), and have
tested indirect indicators of elite messages (Fieldhouse, Pattie and Johnston 1996). Some work in
the U.S. has also looked at the effects of elite cues to vote strategically in the form of candidate
visits in the 1992, 1996, and 2000 elections (Burden 2005). The results of the study were mixed
with respect to the effect of candidate visits.
The vast majority of studies on the effectiveness of signals to vote strategically have been
conducted by a group of scholars who use experimental methods (e.g., Felsenthal, Rapoport, and
Maoz 1988; Forsythe et al. 1993, 1996; Rapoport, Felsenthal and Maoz 1991; Rietz, Myerson,
and Weber 1998). 2 All of the studies use a context in which the Condorcet loser is the plurality
winner. For example, one bloc of voters (A) has one candidate as their first preference, and this
bloc is the largest in the electorate. At the same time, the first preference of voters in Bloc A is
the least preferred candidate of voters in two other blocs (say B and C), which are individually
smaller in size to bloc A, but collectively larger. In order to create an incentive to vote
strategically, subjects are paid based on the outcome of the race. Thus, preferences are solely
determined by the monetary compensation associated with each candidate winning the election.
Typically, the electorate is small and the candidates are abstract, with letters for names.
Thus, in order to ensure that the Condorcet loser does not win the election, voters in blocs
B and C need to coordinate on a candidate. These scholars have found several factors that aid in
coordination in plurality elections. 3 First, they have found that subjects “learn” to vote
strategically through repeated elections. That is, when subjects learn that a Condorcet loser wins
in Round 1, they are more likely to switch their vote in Round 2 (Felsenthal, Rapoport, and Maoz
1988; Rapoport, Felsenthal, and Maoz 1991; Forsythe et al. 1993, 1996). For example, Forsythe
et al. (1993) found that in Round 1, the Condorcet loser won 2/3rds of the time, while in the next
round, the Condorcet loser won less than half of the time. Second, Forsythe et al. (1993) also
4
tested the impact of pre-election polls and the results indicated that the Condorcet loser only won
33.3 percent of the time, compared to 87.5 percent of the time without polls. In later work,
Forsythe et al. (1996) found that across several rounds of play, the Condorcet loser won 26
percent of the time without polls, and only 19.8 percent of the time with polls. 4 Rietz, Myerson
and Weber (1998) tested for the effects of polls and added a campaign contribution signal, in
which subjects were allowed to contribute money toward ads for candidates, which popped up on
each subjects’ screen. They found that the poll and campaign contribution treatments had similar
effects, with the Condorcet loser winning only 33.33 percent of the time, a percentage identical to
the poll condition in the Forsythe et al. study (1993).
The overall level of strategic voting in these studies is quite similar, ranging from 32.7
percent in Rietz, Forsythe, and Myerson (1991) to 36.45 percent in Forsythe et al. (1993). In a
recent review of the experimental work, Rietz (forthcoming) argues that without any type of
information, voters are often unable to coordinate on their second preference. His research
cluster has found that the signals that help subjects to coordinate their votes vary in effectiveness,
with repeated elections being the least effective. Non-binding pre-election polls and campaign
contributions have similar effects and are much more effective than the former. Finally, the most
effective method is institutional: a majority requirement with a runoff system.
While these experiments have greatly expanded our understanding of the conditions
under which citizens can coordinate on their second preference, there are some limitations and
some areas for extension. First, while the studies are very high in internal validity, they are lower
in external validity with respect to the electoral context. For example, existing studies are very
abstract in nature, with letters for candidates instead of names, and rely on a small electorate,
which is not usually the case in any U.S. election. Second, and related to the issue of external
validity, the payoffs are generally structured so that subjects are only paid based on who wins the
election. This feature is important since it allows the researchers to assign preferences to subjects
5
(based on the money received from the outcome of the election). There are two potential
critiques to this set-up. First, this framework does not model any expressive benefit a minor party
supporter might receive from voting for their first preference or a long-term goal of building a
political party. 5 That is, only instrumental benefits are incorporated. Second, some might argue
that preferences derived from monetary compensation are not a good reflection of preferences in
real elections.
There are also many areas for extensions. Existing studies have yet to test more explicit
information signals that are often used in multi-candidate elections, such as “voting for Ralph
Nader helps elect George W. Bush” 6 in the 2000 election or “a vote for Anderson is a vote for
Reagan” 7 in the 1980 election. Second, we do not know whether the source of an explicit signal
has any impact on an individual’s receptivity to it. Finally, studies have also not considered the
role of counter information on the part of the trailing candidate. In an election, we might expect
that the minor party candidate would respond to coordination cues being sent by other elites.
This question is far from idle—the minor party candidates with enough support to tip the outcome
of the election may also have the resources necessary to counter any information signals to switch
one’s vote that are sent by a major party.
I overcome the artificiality of existing studies by placing my experiments in the context
of a mock mayoral election. In one study, I also compare a purely instrumental payoff structure
to one that incorporates both instrumental and expressive benefits. In a second study, subjects are
all paid a flat fee for participation in the study. Instead of assigning preferences based on
monetary compensation, subjects develop their own preferences based on candidates’ positions
on an issue. Furthermore, in both studies, I examining explicit coordination signals, sponsored by
different elites, and counter-coordination signals. Given that I only run one election at a time, I
am unable to test for the effects of repeated elections. Since the results are similar across extant
studies, the loss of information from being able to conduct repeated elections is justified.
6
Types of Coordination Cues
Past experimental studies of strategic voting find that without information on the
competitive context of the race, few minor party supporters switch their vote, resulting in the
Condorcet loser winning the election (e.g., Felsenthal et al. 1988; Rapoport et al. 1991; Forsythe
et al. 1993, 1996). Minor party supporters might be led to abandon their first preference from
tracking polls alone. Having learned that their first preference has no chance of winning, while
the race is close between the major party candidates, they might be inclined to switch their vote,
in the hope of ensuring at least their second best outcome. However, minor party supporters
generally do not know whether others like them have the same second preference. 8 Given this
uncertainty, additional information might be required to induce strategic voting, which could
come in the form of messages being sent by elites, such as the other major party candidate, the
media, or interest groups. 9 These elites might send a cue that other minor party supporters have a
similar preference profile and that if they all coordinate on candidate B, they can ensure their
second best outcome. 10
To summarize, we should find higher levels of strategic voting among those who are
exposed to messages about the competitive context of the race, and this effect should be more
pronounced among those who are exposed to explicit strategic voting messages, compared to
those who are only exposed to tracking polls.
Conditional Factors
One factor that might affect the likelihood of voting strategically is the response by the
minor party candidate. As Cox (1997) points out, minor party candidates might try to counter the
instrumental arguments of the major party candidate by focusing on the long-term goals of the
minor party. Furthermore, they might emphasize one’s duty to vote for their first preference.
Studies on persuasion find that individuals are receptive to one-sided persuasive messages;
whereas they will be less likely to incorporate a message if an argument against that message is
7
presented (e.g., Lau Smith and Fiske 1991; Petty and Cacioppo 1986; Zaller 1992). Thus, the
presence of a counter-strategic voting message might render the explicit message ineffective.
The impact of explicit messages to change one’s vote might also vary depending on the
sponsor of the cue. Many studies of persuasion and priming find that individuals are more likely
to incorporate messages when they trust the sender of the message, especially in low information
and motivation environments (e.g., Chaiken 1980; Miller and Krosnick 2000; Petty and Cacioppo
1986). Given this literature, we might expect that minor party supporters will be more likely to
incorporate and use messages that are sponsored by interest groups that share their preferences,
compared to messages sponsored by the major party candidate that is trying to appeal to them.
They might view some interest groups as having credibility over the issues they care about
(Ansolabehere and Iyengar 1994; Iyengar and Valentino 1999; Petrocik 1996), while they distrust
the motives of the major party candidate. 11
To briefly summarize, the hypotheses that are tested with the data generated from the
experiments are as follows:
H1:
Exposure to polling results will increase the likelihood of voting for one’s second
preference.
H2:
Exposure to explicit information signals will further increase the likelihood of voting for
one’s second preference.
H3:
Exposure to counter-information signals will depress the likelihood of voting for one’s
second preference.
H4:
Signals sponsored by “credible” interest groups will be more effective than candidate
sponsored signals.
Study 1
Method
Participants
8
289 student subjects were recruited for the study from universities in California, North
Carolina and Pennsylvania. 12 13 Participants were recruited with flyers and were compensated
monetarily for participating in the study. Given that the format of the study was the same across
the universities, the data was pooled in the analyses that follow. With this pooled data, the
demographic breakdown of the student sample is as follows: 55.7 percent male, 71.3 percent
white, 40.8 percent Democrat, 22.8 percent Republican; and, a mean political information score
of 3.5 out of 6 points.
Context of the Study
Subjects took part in a mock three-candidate mayoral election in Allentown,
Pennsylvania. They were led to believe that it was a special mayoral election that took place a
few months before the study. The subjects were assigned preferences over the three candidates,
and all subjects were given the same preference profile, to ensure an identical pool of minor party
supporters. 14 However, the subjects read that there were six different types of voters in the
electorate, representing the six possible strict preference orderings over the three candidates. The
competitive context of the race was also held constant (among the treated groups): David
Howells (the least preferred candidate) was the plurality winner, while Lou Ervin (the first
preference) had no chance of winning. Thomas Burke (the second preference) narrowly trailed
Howells.
Two different forms of compensation were used in the study. In one version, subjects
were compensated based on the outcome of the race, receiving $15 if Ervin won (their first
preference), $10 if Burke won (their second preference), and $5 if Howells won (their third
preference). In a second version, an expressive component was built in the payoff structure.
Thus, subjects were compensated based on the outcome of the election, to capture instrumental
considerations of strategic voting, and for whom they voted, to capture an expressive benefit of
voting for one’s favourite candidate (e.g., Carter and Guerette 1992; Fischer 1996; Hirschman
9
1970; Schuessler 2000). A subject was compensated the highest amount if he or she voted for
Ervin and he won ($15). If Burke won the election, and the subject voted for Ervin, then she
received 10 dollars, while if she voted for Burke, she received 9 dollars. If Howells won, and she
voted for Ervin, she received 5 dollars, while if she voted for Burke, she received 3 dollars.
Given the competitive context, the highest actual payoff, for either compensation scheme, was 10
dollars. Overall, we should find that those in the former compensation set-up are more likely to
switch their vote. Finally, to simulate behaviour in a large electorate, subjects were told that the
outcome would be determined by adding the votes they cast with the results in Allentown and the
votes cast in other classes that took part in the study.
Procedure
Upon arrival, subjects were told that they would be participating in a mock election for a
mayoral race that took place in Allentown, Pennsylvania. They were told that the purpose of the
study was to see what types of information people consider in elections. The subjects then filled
out a short pre-survey, which included a battery of demographic and political information
questions. Subjects were then given a packet, which contained their preference profile,
compensation table, and information about the campaign. 15 Subjects then voted and completed a
post-survey, which included questions about their perceptions of who would win, and their levels
of efficacy and trust. After the post-test, the subjects were debriefed and compensated.
Treatment
To test the hypotheses related to the type of message and the conditional effects of the
sponsor and counter messages, subjects were exposed to one of four types of information (no
information, poll information, poll and explicit information signal, and the latter plus a counter
signal from the trailing candidate), with the explicit message delivered by either candidate Burke
or an interest group. Combining these components, subjects were randomly assigned to one of
10
the six groups (control, poll, candidate signal (CIS), interest group signal (IGS), candidate
counter signal (CCIS), and interest group counter signal (CIGS).
The treatment was embedded in the articles section of the campaign packet. Subjects in
the control group read a neutral article about projected levels of turnout in the election and an
article about the importance of revitalizing downtown Allentown, while those in the poll
treatment read the first neutral article, and an article with a tracking poll in which they learned
that Ervin was trailing in the polls with 15 percent support, while Burke had 40 percent support,
and Howells was leading with 45 percent support.
Those in the two explicit signal treatments read the first neutral article, the poll article,
and an article about an advertisement sponsored by either Thomas Burke (CCC), or Good
Schools Pennsylvania (ICC), in which the ad says that “a vote for Ervin is a vote for Howells”
and that if all Ervin supporters switch their vote, they can ensure that Howells does not win.
Finally, the two counter groups read the neutral article, the poll results, the explicit cue,
from Burke or Good Schools Pennsylvania, and an article about Lou Ervin’s response to the ad.
The article quotes Ervin as arguing that “the ad encourages his supporters to go against one of the
most cherished acts in a democracy…voting for anyone other than one’s first preference goes
against one’s duty as a citizen.”
To refresh, we should find that those who received polling information, or all treated
groups, exhibit higher levels of voting for their second preference, Burke, compared to those in
the control group (H1). Furthermore, we should find that those in the explicit signal conditions
have a higher level of voting for Burke compared to those in the poll condition (H2), with the
interest group sponsored cue being more effective than the candidate sponsored one (H4).
Finally, if we compare the counter conditions to the explicit signal conditions, we should find
lower levels of voting for Burke in the former (H3).
Results
11
Before turning to the differences across treatments, there is a simple manipulation check
that was conducted. If subjects in the treated groups paid attention to the articles, we should find
that the percentage of subjects who think their first preference will win will be significantly lower
among the treated groups compared to the control group, since the treated groups read that Ervin
had 15 percent support in the polls. As expected, all of the treated groups were less likely to
think that their first preference would win compared to the control group, according to a series of
difference in proportions tests. The overall percentage who voted for their second preference was
50.5 percent. Figure 1 presents the percentage of subjects who voted for Burke by treatment
group. As expected, we find support for H1, in that only about 6 percent of those in the control
group voted for Candidate Burke, while the percentage ranged from 47.9 percent to 79.2 percent
across the treated groups, and these differences were significant for each comparison group
according to difference in proportions tests.
[Figure 1 about here]
As expected, a higher percentage of subjects voted for Burke in the candidate information
signal (CIS), 64 percent, and the interest group signal, 79.2 percent, compared to those in the poll
group (47.9 percent), and these differences were significant. Furthermore, the percentage is also
significantly higher among those exposed to the interest group cue verses the candidate cue.
Finally, while the addition of a counter coordination cue to the candidate cue depressed the
percentage voting for Burke, these differences were not statistically significant. However, those
who received the counter signal to the interest group signal were significantly less likely to vote
for Burke. Overall, it appears that there is some initial support for the arguments that explicit
information signals play an important roll in enhancing the likelihood of switching one’s vote,
above poll results only, especially when sent by an interest group. However, there is mixed
support for the argument that counter messages should depress the likelihood of switching one’s
vote to their second preference.
12
While the difference in proportions test indicate some support for the hypotheses, they
only indicate whether there were significant differences across treated and control groups. Since
each treated group received poll results, the previous results can not tell us the independent effect
of explicit information signals or counter information signals. In order to test for these
independent effects, a multivariate test is needed. Thus, to separate the effects of poll results, we
need to create a dummy variable in which all of those exposed to poll results (all treated groups)
are coded as one and zero otherwise (the control group). Next, we need to create dummy
variables for those exposed to the explicit information signals. First, I created a general explicit
information signal dummy variable, which combines those who received the candidate or interest
group signal (CIS, IGS, CCIS and CIGS), and then created separate measures for the candidate
information signal and interest group signal. Finally, I created a dummy variable for all of those
who received a counter information signal (CCIS and CIGS), and separated those who received a
counter message by whether the explicit cue was sponsored by the candidate or interest group.
The dependent variable is whether the subject voted for Burke, their second preference, which is
coded as 1 if they did and zero otherwise. Given that the dependent variable is dichotomous, I
ran probit analysis. In addition to the treatment dummy variables (the control serves as the
baseline), a dummy variable was included for whether the subject was compensated based on the
outcome only, and for whether the respondent is male. 16 The latter variable, male, is included
because there was a slight over-representation of males in some groups compared to others,
according to difference in proportion tests.
Table 1 presents the results. Model 1 presents the results for a model that includes the
control variables, as well as the dummy variables for poll results, explicit information signals, and
counter signals. The first column presents the coefficients and standard errors, while the second
column presents the first differences. As expected, the poll treatment is significant and in the
expected direction; thus, learning of the competitive context of the race made subjects more likely
13
to vote for Burke than the control group, in support of H1. Furthermore, the explicit information
signal dummy variable is also significant and positive, in support of H2. Thus, subjects receiving
an explicit signal became even more likely to switch their vote to their second preference.
Finally, as expected (H4), the counter information signal had a negative effect, decreasing the
likelihood of switching one’s vote. With respect to the control variables, the compensate variable
is positive and significant, as is the male dummy variable. Thus, men and those who were paid
based on the outcome only were more likely to switch to their second preference. The model also
performs well with 66.78 percent correctly predicted.
[Table 1 about here]
Since the probit coefficients are not directly interpretable, I calculate the first differences
using Clarify (King, Tomz, and Wittenberg 2000; Tomz, Wittenberg, and King 2001). All of the
treatment dummies were held at the minimum, along with the control variables. Thus, the
baseline subject is a female student who received the 2nd form of compensation. Exposure to the
poll results increased the likelihood of voting for Burke by 32.5 percent. To get the effect of the
explicit information signal, the poll results were set at their max, since this group was exposed to
the poll results. The change in the probability of voting for Burke among this group was 24.5
percent. Thus, having been exposed to the explicit signal further increased the probability of
voting for one’s second preference. The effect for those in the counter group was a decrease in
the probability of voting for Burke by 21.3 percent. Thus, the addition of a counter message
rendered the explicit signal much less effective.
One feature of the first Model is that it does not differentiate the sponsor of the explicit
signal; thus, in Model 2, separate dummy variables are included for the sponsor of each message.
Here, we find that all of the manipulation dummy variables are significant, with the exception of
the candidate counter signal. In order to see whether messages sponsored by interest groups are
more effective than candidate signals, we need to look at the first differences. As expected, we
14
find that the change in the probability of voting for Burke is only 15.5 percent among those
exposed to the candidate information signal, but much higher, 33.1 percent, among those exposed
to the interest group signal. In this model, the poll results increase the probability of switching
one’s vote by 32.9 percent, while the counter interest group signal decreases the probability by
30.3 percent. Overall, the results support the four hypotheses.
Study 2
For study 2, the context of the study was the same, a mock-mayoral election in Allentown
Pennsylvania with Lou Ervin, Thomas Burke, and David Howells. The treatment articles were
almost identical and were only slightly modified to shorten the content. 17 However, a few
differences were introduced in order to test the robustness of the results. First, a non-student
sample was used. Second, a different form of compensation is employed. Some may argue that
the compensation in the first study is not realistic, given that voters do not receive monetary
compensation based on the outcome of elections. To address this concern in this second study, all
subjects were paid the same amount, regardless of the outcome of the election. This should prove
a more difficult test for the treatments. Finally, rather than assign preferences to subjects, they
were asked to rank order their preference over the three candidates, after being presented with
their positions on a school voucher program. A computer program ensured that the first
preference of subjects was presented as the trailing candidate. The next sections detail the
context of the study, the procedures, the treatment, and the results.
Participants and Design
Study 2 was a computer-based study that ran from late August to early September 2005.
Research subjects were adults recruited from three populations: residents of two Northern
California towns and employees of a large university in Northern California. 18 The 417 adult
participants were compensated $30 for their time. The average reported age of the respondents
was 46; 63.1 percent were female; the modal subject (33 percent) indicated having a college
15
degree; 78.7 percent identified as white; and, with respect to partisanship, 24.0 percent identified
as Republican, 46.8 percent as Democrat, and the remainder as Independent or Other.
Subjects were randomly assigned to one of the six treatment groups. Tests of whether
subjects were evenly distributed across groups, according to demographic and attitudinal
variables (party identification, ideology, political sophistication, gender, education, home
ownership, income and race) revealed that there was a slight over-representation of females and
home owners in some of the groups.
Procedures
Participants reported to the experimental lab in order to participate in a study about current
events and opinions. Once seated, individually, in front of a computer terminal, a research assistant
initiated the program and the participant was randomly assigned to either the control group or one
of the six treatment groups. The study began with questions on basic demographics, opinions on
school vouchers, and party identification. The subjects were then presented with the candidates’
positions on a scale of favour to oppose with respect to a voucher program. After being presented
with the candidates positions, the subjects rank ordered their preferences over the three candidates.
They then proceeded to read the newspaper articles associated with their assigned control or
treatment group. The computer program assigned their first preference as the trailing candidate and
their least preferred candidate as the leading candidate in the treatment articles. After reading the
articles, the subjects voted in the election and were asked similar questions to the previous study.
Results
The same manipulation check question was asked in Study 2. As in study 1, all of the
treated groups were less likely to think that their first preference would win compared to the
control group, according to a series of difference in proportions tests. The overall percentage
who voted for their second preference was 24.5 percent. Figure 2 presents the percentage of
subjects who voted for their second preference by treatment group. As expected, we find support
16
for H1, in that only about 11.6 percent of those in the control group voted for their second
preference, while the percentage ranged from 18.9 percent to 41.4 percent across the treated
groups. According to difference in proportions tests, the control group was significantly less
likely to vote for their second preference compared to all of the treated groups, with the exception
of the poll treatment group, which was just outside of traditional significance levels (p<.131). As
expected, the proportion voting for their second preference is lower for this study, since
compensation was not at all based on the outcome of the election.
[Figure 2 about here]
As expected, a higher percentage of subjects voted for their second preference in the
candidate information signal (CIS), 29.9 percent, and the interest group signal, 41.4 percent,
compared to those in the poll group (18.8 percent), and these differences were significant. As
with the previous study, the percentage voting for their second preference is significantly higher
among those exposed to the interest group cue verses those exposed to the candidate cue. The
results are also similar with respect to those in the counter information groups. There are no
significant differences between the candidate information signal and counter candidate
information signal groups. However, those exposed to the counter interest group article voted for
their second preference at a lower proportion, 21.9 percent, compared to those exposed to the
interest group signal.
Since there were some demographic differences across treated and control groups, and to
test the independent effects of the different types of signals, we turn to multivariate tests. The
same models were used as in the previous analysis, with the exception of the control variables.
Home ownership and gender are included as controls since subjects were not evenly distributed
according to these characteristics.
The results for Model 1, which does not differentiate by the sponsor of the message, are
presented in the first two columns of Table 2. As with Study 1, the explicit information signal
17
and counter signal are significant, with the former increasing and the latter decreasing the
likelihood of voting for one’s second preference. However, receiving poll information is just
outside of traditional significant levels (p<.123, one-tailed). It is somewhat surprising that poll
results have weaker effects in this study. However, given that individuals chose their own
candidates and were not paid based on the outcome, it may have taken more of a push to get
subjects to switch their vote to their second preference. According to the substantive effects,
subjects exposed to poll information became 7.9 percent more likely to switch their vote, with the
effect more pronounced among those exposed to explicit signals to switch their vote, 14.8
percent. Finally, messages to stick with one’s first preference again decreased the probability of
voting for one’s second preference, by 10.9 percent.
[Table 2 about here]
Model 2 presents the results taking into account the sponsor of the message. The results
for the poll information dummy variable are the same as in Model 1. As with Study 1, both the
candidate and interest group explicit information signals are significant and positive, while the
counter interest group signal is significant and negative. Furthermore, the effect of the interest
group signal is again larger than the effect of the candidate information signal. The probability of
voting for one’s second preferences increases by 10.6 percent among those exposed to the
candidate signal, and by 21.2 percent among those exposed to the interest group one. Finally, the
counter interest group signal again almost washes away the effect of the message to switch one’s
vote, decreasing the probability of voting for one’s second preference by 18.3 percent. 19
Discussion and Conclusions
Overall, the results of the study have important implications for the quality of
representation in the U.S. As was discussed before, sometimes failures in coordination among
minor party supporters can result in the Condorcet winner losing real elections, a clear problem in
the quality of representation. The findings of this study shed light on some factors that might
18
help or hinder minor party supporters from switching to their second preference, in a context in
which they have the same second preference. First, as expected, and consistent with previous
work, all subjects were more likely to switch their vote when given some information about the
competitive context of the race. Second, explicit information signals added to poll results were
often more effective than poll results alone in helping minor party supporters switch to their
second preference. The effect for explicit information signals was more pronounced when the
message was sponsored by an interest group than by a candidate. However, the presence of a
counter cue often dampened the effect of explicit information signals.
One striking finding was that subjects exposed to explicit coordination cues in addition to
poll results were more likely to switch their vote compared to those only exposed to poll results.
This result obtained with both the student sample and adult non-student sample. These findings
support the argument that sometimes more information is needed for minor party supporters to
switch to their second preference. However, the sponsor of the message may condition this
relationship, with the effects being more pronounced in both studies with the interest group
message.
Another important feature of the results is that they were consistent across the two
studies, even though the samples consisted of very different populations, students verses nonstudent adults, and the differences in compensation. The students were paid based on the
outcome of the election and were assigned preferences, while the adults received a flat fee for
participation and were allowed to determine their own preference rankings. Granted, the levels of
defection for each treated group were higher in the student study, but the pattern of the results
was the same across the two.
In relation to strategies in the empirical world, the results suggest that explicit signals can
play an important role in increasing the likelihood of defection among minor party supporters,
with the effect likely most pronounced if the message is sent by an interest group aligned with the
19
minor party. In the 2000 election, only the National Abortion Rights Action League aired an
explicit advertisement warning that a vote for Nader is a vote for Bush, and it only aired 47 times
(Campaign Media Analysis Group). These results suggest that if other interest groups that Nader
supporters perceived as credible had made similar appeals, more Nader supporters may have
voted for Gore.
Another important finding was that the presence of a counter message by the minor party
candidate, Lou Ervin, did depress the effect of the explicit information signal in both studies,
particularly among those who received the interest group information signal. In appealing to
strategic voting as a violation of one’s civic duty, the explicit information signals to switch one’s
vote were rendered less effective. This finding can have implications for understanding the
circumstances under which minor parties might prove successful in the American political
system. While minor parties will likely have few resources to mount a counter-campaign in
presidential elections compared to the major party candidates, they might have more resources in
other contests, which would enable them to counter the claims by the major parties.
While the study enhances our knowledge of the effectiveness of different types of
information signals, it also leads to more questions for future research. First, how might counter
cues from the other major party fare? Second, how would more implicit information signals fare,
compared to more explicit signals? Most of the appeals to Nader supporters sponsored by the
DNC and Gore campaign in the 2000 election were not explicit, but implicit in nature,
highlighting Bush’s weaknesses on the environment. Finally, does the likelihood of defection
depend on the nature of the issue? For example, are individuals more likely to defect on
economic issues verses moral issues?
20
References
Abramson, Paul R., John H. Aldrich, Phil Paolino, David W. Rhode. 1992. “’Sophisticated
Voting in the 1988 Presidential Primaries.” American Political Science Review 86: 55-69.
Abramson, Paul R., John H. Aldrich, Matthew Diamond, Abraham Diskin, Renan Levine, and
Thomas J. Scotto. 2004. “Strategic Abandonment or Sincerely Second Best? The 1999
Israeli Prime Ministerial Election.” Journal of Politics 66: 706-728.
Aldrich, John H. 1993. “Rational Choice and Turnout.” American Journal of Political Science 37:
246-278.
Ansolabehere, Stephen and Shanto Iyengar. 1994. “Riding the Wave and Claiming Ownership
over Issues: The Joint Effects of Advertising and News Coverage in Campaigns.” Public
Opinion Quarterly 58: 335-357.
Bartels, Larry M. 1988. Presidential Primaries and the Dynamics of Public Choice. Princeton,
NJ: Princeton University Press.
Black, J.H. 1978. “The Multicandidate Calculus of Voting: Applications to Canadian Federal
Elections.” American Journal of Political Science 22: 609-638.
Blais, Andre and Richard Nadeau. 1996. “Measuring Strategic Voting: A Two-Step Procedure.”
Electoral Studies 15: 39-52.
Burden, Barry C. 2005. “Minor Parties and Strategic Voting in Recent U.S. Presidential
Elections.” Electoral Studies 24: 603-618.
Burden, Barry C. 2003. “Minor Parties in the 2000 Presidential Election.” In Models of Voting in
Presidential Elections: The 2000 Election, ed. Herbert F. Weisberg and Clyde Wilcox.
Stanford, CA: Stanford University Press.
Carter, John R. and Stephen D. Guerette. 1992. “An Experimental Study of Expressive Voting.”
Public Choice 73: 251-260.
21
Chaiken, Shelly. 1980. “Heuristic versus Systematic Information Processing and the Use of
Source versus Message Cues in Persuasion.” Journal of Personality and Social
Psychology 39: 752-66.
Cooper, Russell, Douglas V. DeJong, Robert Forsythe, and Thomas Ross. 1989. “Communication
in Battle of the Sexes Game: Some Experimental Results.” The Rand Journal of
Economics 20: 568-587.
Cooper, Russell, Douglas V. DeJong, Robert Forsythe, and Thomas Ross. 1992. “Communication
in Coordination Games.” The Quarterly Journal of Economics 107: 739-771.
Cox, Gary W. 1997. Making Votes Count. Strategic Coordination in the World’s Electoral
Systems. Cambridge: Cambridge University Press.
Evans, Geoffrey, John Curtice and Pippa Norris. 1998. “New Labour, New Tactical Voting? The
Causes and Consequences of Tactical Voting in the 1997 British General Election.” In
Denver, David, Justin Fisher, Philip Cowley and Charles Pattie, eds. Elections and
Parties Review. London: Frank Cass.
Farquharson, Robin. 1969. Theory of Voting. New Haven, CT: Yale University Press.
Felsenthal, D.S., A Rapoport A, and Z Maoz. 1988. “Tacit Cooperation in Three Alternative Noncooperative Voting Games: A new Model of Sophisticated Behaviour under the Plurality
Procedure.” Electoral Studies 7: 143-161.
Fey, Mark. 1997. “Stability and Coordination in Duverger’s Law: A Formal Model of Preelection Polls and Strategic Voting.” American Political Science Review 91: 135-147.
Fieldhouse, E.A., C.J. Pattie and R.J. Johnston. 1996. “Tactical Voting and Party Constituency
Campaigning at the 1992 General Election in England.” British Journal of Political
Science 26: 403-439.
Fischer, A.J. 1996. “A Further Experimental Study of Expressive Voting.” Public Choice 88:
171-184.
22
Fisher, Stephen D. and David P. Myatt. 2001. “Strategic Voting Experiments.” Paper presented at
the Annual Meeting of the American Political Science Association, San Francisco, CA.
Forsythe, Robert, Roger B. Myerson, Thomas A. Rietz, and Robert J. Weber. 1993. “An
Experiment on Coordination in Multicandidate Elections: The Importance of Polls and
Election Histories.” Social Choice and Welfare 10: 223-247.
Forsythe, Robert, Roger B. Myerson, Thomas A. Rietz, and Robert J. Weber. 1996. “An
Experimental Study of Voting Rules and Polls in Three-Candidate Elections.”
International Journal of Game Theory 25: 355-383.
Gerber, E.R., R.B. Morton and T.A. Rietz. 1998. “Minority Representation in Multi-member
Districts.” American Political Science Review 92: 127-144.
Gomez, Brad T., and J. Matthew Wilson. 2001. “Political sophistication and economic voting in
the American electorate: A theory of heterogeneous attribution.” American Journal of
Political Science 4: 899-914.
Hirschman, Albert O. 1970. Exit, Voice, and Loyalty: Responses to Decline in Firms,
Organizations, and States. Cambridge, MA: Harvard University Press.
Iyengar, S. and N. Valentino. 1999. “Who Says What? Source Credibility as a Mediator of
Campaign Advertising.” In Lupia, A., M. McCubbins, and S. Popkin, Eds. Elements of
Reason. New York: Cambridge University Press.
King, Gary, Michael Tomz, and Jason Wittenberg. 2000. “Making the Most of Statistical
Analyses: Improving Interpretation and Presentation.” American Journal of Political
Science Vol. 44(2): 341-355.
Lau, Richard R., Richard A. Smith and Susan T. Fiske. 1991. “Political Beliefs, Policy
Interpretations, and Political Persuasion.” Journal of Politics 53: 644-675.
Lupia, Arthur and Mathew McCubbins. 1998. The Democratic Dilemma. Cambridge; Cambridge
University Press.
23
McKelvey, Richard D. and Peter C. Ordeshook. 1972. “A General Theory of the Calculus of
Voting.” In Mathematical Applications in Political Science, Vol. 6, ed. James F. Herndon
and Joseph L. Bernd. Charlottesville: University of Virginia Press.
Miller, Joanne M. and Jon A. Krosnick. 2000. “News Media Impact on the Ingredients of
Presidential Evaluations: Politically Knowledgeable Citizens are Guided by a Trusted
Source.” American Journal of Political Science 44: 301-315.
Miller, John J. 2002, November 20. “What did the GOP do to deserve this?” Raleigh News and
Observer, Section A, page 19.
Modie, Neil. 2004, December 14. “Libertarian May Have Helped Rossi, Experts Agree.” The
Seattle Post-Intelligencer, Section B, page 1. Retrieved June 29, 2005, from
http://web.lexis-nexis.com.
Myerson, Roger. and Robert. Weber. 1993. “A Theory of Voting Equilibria.” American Political
Science Review 87: 102-114.
Niemi, Richard G. and Arthur Q. Frank. 1981. “Sophisticated Voting Under the Plurality
Procedure.” In Political Equilibrium, ed. Peter C. Ordeshook and Kenneth A. Shepsle.
Kluwer. Nijhoff Publishing.
Niou, Emerson. 2001. “Strategic Voting Under Plurality and Runoff Rules.” Journal of
Theoretical Politics 13: 209-221.
Ordeshook, Peter C. and Langche Zeng. 1997. “Rational Voters and Strategic Voting: Evidence
from the 1986, 1980, and 1992 Elections.” Journal of Theoretical Politics 9: 167-187.
Palfrey, T.R. 1989. “A Mathematical Proof of Duverger’s Law.” In P.C. Ordeshook, ed. Models
of Strategic Choice in Politics. Ann Arbor, MI: University of Michigan.
Petrocik, John R. 1996. “Issue Ownership in Presidential Elections with a 1980 Case Study.”
American Journal of Political Science 40: 825-850.
24
Petty, R.E. and J.T. Cacioppo. 1986. Communication and Persuasion: Central and Peripheral
Routes to Attitude Change. New York: Springer-Verlag New York, Inc.
Plott, CR. 1991. “A Comparative Analysis of Direct Democracy, Two Candidate Elections and
Three-candidate Elections in an Experimental Environment. In Palfrey, T.R., Ed.
Laboratory Research in Political Economy. Ann Arbor, MI: The University of Michigan
Press.
Rapoport, Amnon, Dan S. Felsenthal, and Zeev Maoz. 1991. “Sincere versus Strategic Voting
Behaviour in Small Groups.” In Palfrey, T.R., ed. Laboratory Research in Political
Science. Ann Arbor: The University of Michigan Press.
Riker, William H. and Peter C. Ordeshook. 1968. “A Theory of the Calculus of Voting.”
American Political Science Review 62: 24-42.
Rietz, Thomas. Forthcoming. “Three-way Experimental Election Results: Strategic Voting,
Coordinated Outcomes and Duverger’s Law.” In Plott, C.R. and VL Smith, Eds. The
Handbook of Experimental Economics Results. Amsterdam: Elsevier Science.
Rietz, T.A., R.B. Myerson, and R.J. Weber. 1998. “Campaign Finance Levels as Coordinating
Signals in Three-way Experimental Elections.” Economics and Politics 10: 185-217.
Schuessler, Alexander A. 2000. A Logic of Expressive Choice. Princeton and London:
Princeton University Press.
Tomz, Michael, Jason Wittenberg, and Gary King. 2001. “Clarify: Software for Interpreting and
Presenting Statistical Results.” Version 2.0 Cambridge, MA: Harvard University, June 1.
Http://gking.harvard.edu
Tsong, Nicole. 2004, October 31. “Third-party Candidate Reach Out for Recognition.”
Anchorage Daily News, Section A, page 8. Retrieved June 29, 2005, from
http://web.lexis-nexis.com.
25
Zaller, John. 1992. The Nature and Origins of Mass Opinion. Cambridge: Cambridge University
Press.
26
Figure 1: Percentage Who Voted for Burke, Study 1
90
80
70
Percentage
60
50
Series1
40
30
20
10
0
Control
Poll
CIS
CCIS
Treatment
27
IGS
CIGS
Figure 2: Percentage Voting for their Second Preference, Study 2
45
40
35
30
25
Series1
20
15
10
5
0
Control
Poll
CIS
CCIS
Treatment Group
28
IGS
CIGS
Table 1: Probit on Vote Burke, Study 1
Model 1
Poll
Explicit Information Signal
Candidate Information Signal
Interest Group Information
Signal
Counter Signal
Candidate Counter Signal
Coefficient
(Standard Error)
1.550**
(0.340)
0.632**
(0.229)
------------
Δ
------------
-------
-0.548**
(0.190)
------------
-21.3%
32.5%
24.5%
-------
-------
Model 2
Coefficient
Δ
(Standard Error)
1.550**
32.9%
(0.340)
-----------------0.412*
(0.261)
0.887**
(0.277)
------------
15.5%
33.1%
-------
-0.325
-12.8%
(0.261)
Interest Group Counter Signal
------------------0.805**
-30.3%
(0.279)
Compensate
0.389**
14.4%
0.393**
14.9%
(0.161)
(0.162)
Male
0.271**
10.5%
0.270*
10.1%
(0.163)
(0.163)
Constant
-1.920**
-1.923**
(0.318)
(0.318)
N
289
289
Pseudo R-squared
.179
.187
Percent Correctly Predicted
66.78%
66.78%
**p<.05 * p<.10, one-tailed. Marginal effects (Δ) indicate the percentage change in the predicted
probability of y given a discrete change in x from its minimum to maximum.
29
Table 2: Probit on Vote Second Preference, Study 2
Model 1
Poll
Explicit Information Signal
Candidate Information Signal
Interest Group Information
Signal
Counter Signal
Candidate Counter Signal
Coefficient
(Standard Error)
0.331
(0.285)
0.484**
(0.233)
------------
Δ
------------
-------
-0.331**
(0.170)
------------
-10.9%
7.9%
14.8%
-------
-------
Model 2
Coefficient
Δ
(Standard Error)
0.331
7.9%
(0.285)
-----------------0.343*
(0.259)
0.637**
(0.262)
------------
10.6%
21.2%
-------
-0.134
-4.3%
(0.240)
Interest Group Counter Signal
------------------0.540**
-18.3%
(0.244)
Female
0.247*
5.5%
0.250*
5.4%
(0.156)
(0.157)
Home
-0.270**
-4.4%
-0.258**
-4.3%
(0.163)
(0.153)
Constant
-1.191**
-1.201**
(0.239)
(0.240)
N
370
370
Pseudo R-squared
.050
.055
Percent Correctly Predicted
75.41%
75.14%
**p<.05 * p<.10, one-tailed. Marginal effects (Δ) indicate the percentage change in the predicted
probability of y given a discrete change in x from its minimum to maximum.
30
1
Strategic voting will be used throughout to indicate voting for one’s second preference as opposed to
one’s first preference in a context in which one does not want to cast a wasted vote.
2
Other experiments focus on testing across competing theoretical models (e.g., Fisher and Myatt 2001;
Plott 1991; Rapoport, Felsenthal and Maoz 1991).
3
Other work has looked at the effect of different institutional settings. Rapoport, Felsenthal, and Maoz
(1991) find that voters have a more difficult time coordinating under approval voting, while Gerber,
Morton and Rietz (1998) find that minority candidates are more likely to win under cumulative voting, as
opposed to straight voting.
4
These percentages were even smaller under approval and Borda voting methods.
5
While strategic voting experiments have not tested for this tradeoff between instrumental and expressive
motivations, other scholars have done experimental work in this domain (Carter and Guerette 1992; Fischer
1996). Both studies found evidence that the lower the probability of being decisive, the more likely the
voter would act on expressive motivations. Fischer (1996) also found some evidence that when votes are
not secret, people are more likely to vote expressively.
6
Reported in an article in the Boston Herald. Battenfeld, Joe and Andrew Miga. 2000, October 26.
“Nader lashes out at Gore-Says tight prez race is veep’s own fault.” The Boston Herald, Section News,
page 1. Retrieved October 19, 2005, from https://web.lexis.nexis.com.
7
Reported in an article in the New York Times. Harris, Myron. 1980, September 29. “One Democrat’s
Reason for Choosing Anderson.” The New York Times, Section A, Page 18. Retrieved October 19, 2005,
from https://web.lexis.nexis.com.
8
Most of the existing experimental work gives voters perfect information on the distribution of preferences
among the groups of voters and the size of these groups.
9
Some voters might also require a linkage mechanism due to lower levels of political sophistication. For
example, in a study of economic voting, Gomez and Wilson (2001) find that less sophisticated voters were
unable to link their personal economic situation to their vote choice. We might also find that in the context
31
of strategic voting, less sophisticated voters also need more information linking their individual behavior to
aggregate outcomes.
10
Some support for this claim can also be found in studies of communication signals between players in
coordination games. Some studies demonstrate that communication can help overcome coordination
failures in a variety of games, such as one way communication in battle of the sexes games (Cooper et al.
1989) and two way communication in simple coordination games (Cooper et al. 1992). While these studies
only deal with two players, we can imagine a similar process occurring if messages are sent by elites
among a broader group of players.
11
Lupia and McCubbins (1998) argue that shared interests are not a necessary condition for persuasion.
Rather, they argue that individuals will use a cue if they perceive the speaker as knowledgeable and
providing reliable information, which occurs if they have incentives to provide truthful information. Even
with this argument, individuals will be less likely to perceive the candidate as having an incentive to
provide truthful information, while this will be more likely with some interest groups.
12
The vast majority of subjects were from California, 144, and North Carolina, 119, with only 26 subjects
from Pennsylvania. The studies in North Carolina and Pennsylvania took place in the fall of 2002, while
the studies in California took place in the spring of 2005. The time difference had to do with waiting to
securing more funding to increase the n of the study.
13
Student subjects are a convenient population and previous studies in this research domain have obtained
compelling results with similar subject pools (e.g., Felsenthal et al. 1988; Forsythe et al. 1993; Rapoport et
al. 1991).
14
15
Most strategic voting studies assign preferences to subjects to hold utility constant.
The campaign articles included background information about the city, candidate biographies, and
campaign information. The section on candidate biographies did not give issue positions, to avoid affecting
the utility component, and all of the candidates had similar education levels and past public service.
16
Dummies for the location of the study were tested but were not significant; thus, they are not included in
the analysis.
32
17
However, subjects only read the articles and not the other campaign packet information from the
previous study.
18
750 individuals from each of the three populations were recruited using mailings. A sample of
individuals from the local populations was purchased from a marketing company; employees from the
campus population were sampled using a campus directory and drawing the first and last names for each
letter of the alphabet until the sample was complete (only individuals working on the main campus were
included; researchers, deans, professors, lecturers, Fellows, executive administrative assistants, directors,
and associate directors were excluded from the process).
19
For both studies, the models were run with the sample split at the median level of political sophistication.
The pattern of effects was similar for both groups, with the results more substantial and more in line with
expectations for high sophisticates.
33