Electoral Studies 37 (2015) 28e40

Contents lists available at ScienceDirect

Electoral Studies
journal homepage: www.elsevier.com/locate/electstud

Changing motivations, time of the voting decision, and

short-term volatility e The dynamics of voter heterogeneity
Jan Eric Blumenstiel*, Thomas Plischke 1
University of Mannheim, A5, 6, D-68131 Mannheim, Germany

a r t i c l e i n f o a b s t r a c t

Article history: The notion of “voter heterogeneity” e that different voters decide according to different
Received 1 February 2014 sets of choice criteria e has become widely accepted. But while in a given election some
Received in revised form 3 November 2014 voters are candidate-oriented and others are issue-oriented, little is known about the
Accepted 5 November 2014
temporal dimension of heterogeneity: are candidate voters at t1 also candidate voters at
Available online 15 November 2014
t2? We argue that individual voters' motivations can change over time and expect that the
extent of inter- as well as intra-individual heterogeneity impacts the process of individual
Keywords:
decision-making. Using panel data collected in the run-up to the German Federal Election
Vote decision
Subjective reasons
2009 we demonstrate that certain choice motivations and intra-individual changes in
Heterogeneity these motivations lead to later vote decisions and contribute to the explanation of vote
Time of vote decision switching.
Volatility © 2014 Elsevier Ltd. All rights reserved.

1. Introduction sources of variations in voters' voting calculi, but research

In deciding how to vote, voters are not all alike but
rather apply different decision-making strategies and differ
as to the motivation of their choice. Some voters, e.g., might
be policy-driven, others may put a special emphasis on
candidates, and still others may simply vote habitually.
Such differences in the influence of decision criteria on
their vote choice can cause two voters with identical policy
preference to elect different parties (Bartle, 2005; Rivers,
1988). However, ignoring this variation in assuming ho-
mogenous voting calculi has until recently been the stan-
dard in electoral choice. In the meantime, several studies
(e.g. Bartle, 2005; Blumenstiel, 2014; Blumenstiel and
Rattinger, 2012; Peterson, 2005; Rivers, 1988; Roy, 2011)
have both demonstrated considerable heterogeneity in
electoral decision-making and have started to analyze the
* Corresponding author. Tel.: þ49 0 621 181 2820.
E-mail addresses: jan_eric.blumenstiel@uni-mannheim.de
Blumenstiel), thomas.plischke@uni-mannheim.de (T. Plischke).
1 An example of this type of heterogeneity would be that
Tel.: þ49 0 621 181 3653.
on the topic is still limited.
Among other things, the consequences of heterogeneity
and its longitudinal dynamics are still largely unknown.
Regarding possible consequences of voter heterogeneity,
we will focus on the timing of the vote decision and on
short-term fluctuations of the vote intention. As to the
longitudinal dynamics, previous research almost exclu-
sively focused on cross-sectional data, thus the temporal
(in)stability of voter heterogeneity has been largely dis-
regarded. Therefore, we differentiate between two types of
heterogeneity to uncover the temporal dimension of voter
heterogeneity. The first type concerns differences in the
importance voters assign to various considerations in their
decision at one point in time. An example for this type of
heterogeneity would be that voter A places more weight on
her issue orientations than voter B, who primarily bases her
decision on her impression of the candidates. In the
following, these differences between voters will be referred
to as “inter-personal heterogeneity”. The second type
(J.E. concerns the individual stability of consideration weights.

http://dx.doi.org/10.1016/j.electstud.2014.11.003
0261-3794/© 2014 Elsevier Ltd. All rights reserved.
 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
voter A consistently makes issue-oriented decisions in two
consecutive elections, whereas the decision calculus of
voter B changes over time. In the following, the instability
of consideration weights will be referred to as “intra-per-
sonal heterogeneity”.
Against this background, we will focus on two research
questions: First, do inter- and intra-personal heterogeneity
affect the time when voters make their decision? Second,
do both types of heterogeneity influence the likelihood of
vote switching during an electoral campaign? To analyze
these questions, we use data from a seven-wave online
campaign panel conducted as part of the German Longi-
tudinal Election Study (GLES).
The article proceeds as follows. In section 2, we define
our research question in more detail by discussing previous
research on inter- and intra-individual heterogeneity. Sec-
tion 3 introduces our measure of voter heterogeneity which
is based on voters' self-reports of their most important
reasons for their voting decision. In section 4, we postulate
more specific hypotheses about how inter- and intra-
individual heterogeneity may affect the voter decision
process in terms of the timing of the decision and short-
term volatility. After describing data and the coding pro-
cess of the self-reported motivations in section 5, we pre-
sent the results of our analyses in section 6. In the
conclusion, we discuss the consequences of heterogeneity
for voter decision making and also their wider ramifica-
tions, but also refer to some limitations of our study and
questions to be addressed in future research.
2. Inter- and intra-personal heterogeneity
Voter heterogeneity by no means is a new concept (see
Blumenstiel, 2014 for a summary of previous findings). As
Roy (2011) notes, individual differences in decision-making
have already been stressed by Campbell and colleagues
(1960). It has been argued in The American Voter, for
instance, that for voters who are well informed about the
parties' candidates and policy positions, party identifica-
tion should be less important for their decision than for
those who are, apart from feeling attached to a party, little
involved in politics (Campbell et al., 1960: 136). However,
more often than not the insight that voters are not all alike
as to their choice motivation has been widely disregarded
in the discipline, possibly for a lack of both adequate
research methods and theoretical grounds. What has been
denoted as the ‘homogeneity assumption’ became the
default (Lewis-Beck et al., 2008: 81). So it was up to Douglas
Rivers in his seminal article (1988: 737) to bring the argu-
ment that disregarding voter heterogeneity might be
misleading back to the agenda:
“In fact, if two voters have identical policy preferences
and demographic characteristics, then any of the stan-
dard methods of analyzing voting behavior would pre-
dict that the two would cast identical votes. Yet it is easy
to think of situations where such a prediction would be
unwarranted. If issues have different levels of salience to
voters, then identical policy preferences do not neces-
sarily imply identical (or even similar) voting patterns
for voters.”
29
In assuming homogenous consideration weights for all
voters, previous studies have thus assumed a simplistic and
direct relationship between attitudes and behavioral in-
tentions, where in fact the influence of each attitude on a
behavioral intention should be weighted by the personal
importance of this attitude to an individual (Fishbein and
Ajzen, 1975). As Bartle (2005) illustrates, including voter
heterogeneity as an additional independent variable to the
explanation of voting decisions can help to understand why
voters with identical political preferences decide for
different parties.
If there are hence enough reasons to assume that the
electorate has never been homogenous, the argument has
been fostered by the finding of declining social determi-
nation of voting decisions. According to Dalton (2008:
184f.), partisan dealignment is accompanied by increasing
electoral volatility, a tendency of voters to decide later in
the campaign and growing importance of short-term fac-
tors such as issues and candidates, so that party choice for
most voters now actually is a decision rather than a pre-
determined routine. In short, the electorate is likely to have
become even more heterogenous over time. Furthermore,
while the authors of The American Voter clearly had been
aware of voter heterogeneity, their failure to include indi-
vidual weights of the considerations in their model has
been criticized (Lewis-Beck et al., 2008: 81).
Including the time horizon into the study of individual
variations in the voting motivation, a further distinction
can be made between inter-personal and intra-personal
heterogeneity. If the former is being studied, the aim is to
find and to explain variations between voters in the
importance they assign to various considerations in their
decision at one point in time. In studying this type of het-
erogeneity, one could for example try to identify those
voters who primarily base their decisions on their candi-
date orientations and analyze how these voters differ from
other groups of voters. Most previous research devoted to
the study of voter heterogeneity focused on this type of
heterogeneity, i.e. on differences between voters (cf. e.g.
Rivers, 1988; Bartle, 2005; Clarke et al., 2008; Roy, 2011;
Blais et al., 1998; Blumenstiel and Rattinger, 2012). If
however intra-personal heterogeneity is to be studied, the
focus shifts to the temporal stability of individual consid-
eration weights. Regarding this type of heterogeneity, one
could for example ask “Are candidate voters at t1 also likely
candidate voters at t2?”. Depending on the time period
studied, t1 and t2 could either be consecutive elections or,
as in this article, different points in time during one elec-
toral campaign.
Implicitly the notion of e.g. candidate-oriented voters is
suggestive of an assumed stability of voting motivations at
least in short-term perspective. However, only very few
studies have actually addressed the temporal aspect of
voter heterogeneity. As far as we are aware, Peterson
(2005) first explicitly mentioned the individual-level dy-
namics of voter heterogeneity and later argued that
changes in voters' uncertainty about the candidates during
the campaign can affect the weights of attitudes in the
voting decision (Peterson, 2009). Using longitudinal panel
data, Blumenstiel (2014) demonstrated that voters' attitude
weights are moderately associated between two
30 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
consecutive elections so that, first, those voters who for
instance primarily rely on issue orientations at election 1
also tend to do so at election 2 but, second, there is also
considerable intra-individual change with hitherto un-
known consequences.
Further theoretical and empirical underpinning for the
notion of intra-personal heterogeneity and changes of
motivation comes from research about the “priming” of
political attitudes and considerations. According to this
strand of research, individuals do not retrieve all of the
information stored in long-term memory in order to make
fast but satisfying decisions in daily life. What people think
about is strongly determined by the media or by the
campaign, which are drawing attention to some aspects of
politics while neglecting others. In this way, they also
determine (“prime”) what information comes to mind e
and hence the evaluative standards applied e when a
judgment or a decision is called for. This effect has been
demonstrated in various controlled experiments (cf.
Iyengar and Kinder, 1987), but also in large-scale election
studies, where priming effects have been observed in re-
action to media exposure (e.g. Druckman, 2004) and inter-
personal communication (e.g. Mendelsohn, 1996). Thus,
changes in motivation to vote for a party are quite frequent
during an electoral campaign.
While there is thus enough theoretical and empirical
evidence for the existence of inter- and intra-personal vari-
ability of decision criteria, the consequences of the two as-
pects of heterogeneity for voter decision-making are still
only partly understood (Blumenstiel, 2014). This is in part
due to the methodological difficulty of how to measure voter
heterogeneity (Bartle, 2005). Moreover, in order to examine
the consequences of individual changes in choice motiva-
tions, it is not only sufficient to measure them adequately,
but also to do so repeatedly for the same individuals.
In the following, we use data from an innovative
campaign panel to study both aspects of heterogeneity
simultaneously. Our main research questions focus on the
consequences of voter heterogeneity on two important
aspects of the decision-making process: First, do inter- and
intra-personal heterogeneity affect the time when voters
make their decision? And second, can heterogeneity
contribute to the explanation of vote switching during an
electoral campaign? Before postulating more specific hy-
potheses in section four, we will first describe our measure
of voter heterogeneity in the following section.
3. Measuring voter heterogeneity
One important complication in studying voter hetero-
geneity is that the weights voters assign to various atti-
tudes in their decision cannot be easily measured (Bartle,
2005). The approximation we use in this paper to mea-
sure which factors are crucial to a voter's decision are self-
reported reasons for vote choice provided to an open-
ended question.2 This approach is efficient and elegant in
2
The exact question wording of the open-ended question was: “Why
would you [did you] vote for this party? Please state your most important
reasons”.
that it requires only a few seconds of questionnaire time
and does actually provide each voter with the possibility to
indicate her subjective rationale. Yet, the validity of these
self-reports might be doubtful since it has been argued in
cognitive psychology that people have only limited intro-
spective access to higher-order cognitive processes. Most
prominently, this point was made in an often-cited article
by Nisbett and Wilson (1977). Following their argument,
respondents are quite willing to answer questions about
the reasons of their behavior. But as they cannot be aware
of their own motivations, their responses are mostly
rationalizations based on a-priori causal theories (e.g.,
grounded on cultural norms, media perceptions, or what
they think is socially desirable) rather than valid answers
based on true introspection so that their answers are not
closely related to the unknown actual reasons. If this was
true, self-reports could not possibly be a fruitful way of
identifying voter heterogeneity.
The argument did not remain uncontested, though.
What Nisbett and Wilson have presented is, based on a
comprehensive literature review, a plea against a privileged
introspective accessibility of causal processes in higher-
order cognitive processes in certain experimental situations.
In the shortened reception their argument is often mis-
perceived as general evidence against the usefulness of any
introspective self-reports. This interpretation clearly dis-
regards both the restrictions that are implied in the original
argument as well as the views expressed in a number of
articles critically discussing the findings of Nisbett and
Wilson (cf. e.g. Ericsson and Simon, 1980; White, 1988).
While denying the introspective accessibility of mental
processes, Nisbett and Wilson do not negate that people
can accurately report on “contents” such as plans, emo-
tions, evaluations etc. (1977: 255). Put simply, subjects have
no privileged access to answering the question of how a
decision has emerged in terms of causal chains and inter-
mediate states, but can easily answer the question of what
(i.e. the criteria, facts, feelings or opinions) has influenced a
decision (White, 1988). For the identification of heteroge-
nous subgroups of voters, mental processes can be ignored
and only the latter type of information is relevant. If pro-
cesses are to be examined, experimental approaches such
as process-tracing seem to be more appropriate anyway.
Beyond, since the evidence presented in the Nisbett/
Wilson article is entirely based on experimental choice
situations, the external validity of the findings can be
questioned. Some of the studies they cite are based on
hypothetical choice situations. For instance, in one famous
example probands where asked to choose among a number
of completely identical pairs of socks. Regarding the finding
presented by Ericsson and Simon (1980) that motivation
and salience strongly influence what is stored in long-term
memory, these experiments cannot be directly compared
to the voting decision. The latter is a real, voluntarily, and
quite rarely performed action that thus should be highly
salient to the voters e at least compared to being forced to
choose between identical socks.
Furthermore, while self-reports surely are selective
(Ericsson and Simon, 1980), there is reason to assume that
respondents do not report randomly from their memory,
but rather state their most salient attitudes in their
 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
verbatim answers. As we do not seek to model decision
processes that led to the final decision but rather are
interested in what considerations are top of the head of the
voters' head right at the moment an intention is uttered,
this saliency filter may actually be useful as a mechanism to
reduce complexity without losing too much information.
Using self-reports to uncover the subjectively most
important considerations is therefore in fact consistent
with the findings of Nisbett and Wilson.
As to previous findings, whereas open-ended questions
had been used in early classical studies in electoral science
such as Campbell et al. (1960), they have been largely
ignored in electoral science ever since. Furthermore, the
major findings of the few existing studies are divergent.
While Blais et al. (1998) conclude that open-ended ques-
tions can indeed help to identify subgroups for which a
given consideration is more important, other studies draw
the opposite conclusion.
For Germany, only limited evidence exists. Blumenstiel
and Rattinger (2012) use a very similar coding scheme to
the one employed here and show that the identified groups
vary as expected in terms of political attitudes and pre-
dispositions. For instance, party voters strongly identify
with their preferred party, rarely change their vote decision
between two Bundestag elections, and make their de-
cisions early in the campaign whereas issue voters are most
knowledgeable about politics; strategic voters most often
split their ticket. Most importantly, they find significant
interactions between reasons mentioned in the self-reports
and the regression weights of corresponding predictors of
the Michigan model of voting behavior (Campbell et al.,
1960), indicating that these aspects are indeed more
closely associated with their decision even in a multivariate
test. The cumulated theoretical and empirical evidence let
us come to share the conclusion of Bartle (2005) that using
voters' self-reports might be a promising, if not the most
promising, way to measure motivation heterogeneity.
To include the self-reports into the analysis, voters must
be classified to subgroups with different decision motiva-
tions subject to their verbatim answers. Based on the
theoretical framework of the Michigan model of voting
behavior, we differentiate between seven motivations for
vote choice (see Table 1). The first motive establishes the
group of “party attachment voters” who referred to a
formal membership or a long-term attachment to a party,
mentioned the party they voted for as the advocate of a
social group they belong to, or stated to habitually vote for
this party. The “performance voters” in group two relate to
what is called “attitudes toward parties as managers of
government” in the American Voter (1960: 526) meaning
that they either mention the party's record in government
or characteristics which are desirable for a governing party
in general. The “issue voters” in group three either
mentioned a specific issue or the general party platform of
their chosen party. A fourth group of voters reported can-
didates as the most important decision criterion.
Beyond these groups, a further three groups were
identified which are not directly related to the theoretical
framework of the American Voter. Strategic voting and
31
Table 1
Coding scheme for self-reported reasons of vote choice.
Reported reason Definition
1) Party Attachment 
attachment (e.g. „I am committed to
that party“)
 formal membership
 general representation of personal/group
interests
 habit („I always vote for that party“)
2) Performance of (potential) performance in government
parties in managing (“did a good job”, “would do a better job”
affairs of etc.) or qualities that are desirable for
government governing (credible, not corrupt, keep their
promises, reform-oriented etc.)
3) Issues specific issues or reference to the party
platform in general
4) Candidates reference to chancellor candidates/
front-runners
5) Strategy support (or prevent) a certain coalition,
change in government, balance of
powers etc.
6) Gut Feeling „the lesser evil“, by gut feeling etc.
7) Protest protest, disenchantment with established
parties or politics
protest voting are motives that are commonly mentioned
in the literature (e.g., McGregor, 2012) and are also
mentioned by a number of voters in their verbatim an-
swers. Finally, a last group of voters can be referred to as
“gut feeling voters”. Rather than giving any meaningful
reason for their choice, they either literally state to have
decided by gut feeling or that they just voted for what they
think was the lesser evil.
We rely on the presented coding scheme to measure
both types of heterogeneity. First, the classification into the
seven mentioned groups is treated as a measure of inter-
personal heterogeneity. We assume that voters who are
categorized into the group of issue-voters place consider-
ably more weight in their decision on policy-orientation
compared to other voters, whereas individuals in the
group of candidate voters are assumed to place above-
average weight on candidate orientations. Second, the in-
dividual stability of the classification into the seven groups
is treated as a measure of intra-personal heterogeneity. For
each pair of two consecutive panel waves, we created a
dummy variable which indicates whether the same code
was assigned to an individual in both waves. In this case her
voting calculus is assumed to be stable between both
waves. In combining these dummy variables for all possible
pairs of consecutive panel waves, we obtain a continuous
measure of intra-personal heterogeneity.
It is important to note that reducing differences be-
tween voters in their voting calculus to their single most
important motivation clearly is a simplification, since some
voters are likely to consider more than one decision crite-
rion. Yet this does not necessarily imply an over-
simplification. Given the limited capacity of human short-
term memory, it is in fact unrealistic to assume that most
voters rely on a great number of different considerations
(Gigerenzer et al., 1999; Plischke, 2014). Cognitively more
plausible decision rules explicitly include one reason
decision-making models (Gigerenzer et al., 1999).
32 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40

4. Hypotheses H2. The more the type of motivation varies during the
campaign, the more unstable is the voting intention
From our theoretical considerations presented above,
Related to our first hypothesis, voter heterogeneity
we can derive the hypotheses which will guide the
should also be positively associated to short-term volatility
remainder of this article. As outlined above, our main
of party preferences between panel waves. Yet in contrast
research interest is devoted to study the effects of inter-
to the time of the vote decision, the inter-personal
and intra-personal heterogeneity on two important aspects
dimension of heterogeneity should be less important in
of the decision-making process, namely the time of the
this context. While there might be empirical differences
vote decision and short-term electoral volatility.
between the voter groups in terms of short-term volatility,
H1a. The type of motivation provided in the self-reported there is no straightforward theoretical explanation why for
reasons is related to time of the vote decision instance candidate voters should more frequently change
their preferred party than issue voters. However, irre-
Variation in the timing of final vote decision has long
spective of which specific reasons are mentioned by a voter,
been considered a consequence of voter heterogeneity (see
the more distinct reasons she perceives as being most
e.g. Box-Steffensmeier and Kimball, 1999; Campbell et al.,
important during the seven waves of the campaign panel,
1960; Chaffee and Choe, 1980). As Chaffee and Choe
the more often she should reassess and change her vote
(1980) argue, voters who are attached to a particular
intention, i.e. the number of preferred parties across all
party should have been already decided before the
panel waves should increase.
campaign starts. Similarly, it can be argued that perfor-
mance voters are guided by retrospective evaluations of the
parties' performance in the past and can therefore decide
5. Data and coding of self-reported reasons
early in the campaign. By contrast, issue-voters and
candidate-voters need information about the specifics of
This paper utilizes data from a seven-wave online panel
the particular campaign, e.g. the attributes of the main
study which was conducted in the run-up to the German
contenders and the issue stances of each party on the
Federal Election 2009, as a component of the German
currently most important issues. Hence these voters should
Longitudinal Election Study (Rattinger et al., 2012). In this
make up their mind after they have gathered enough in-
survey, 4.552 interviewees eligible to vote were selected
formation, at some point during the campaign. Finally,
according to pre-defined quota. The first panel wave fielded
there are voters who are hardly concerned about politics
from July 10 to July 20, 2009, approximately three months
and elections. If these apolitical voters participate at all,
before the federal election; subsequent panel waves started
they tend to make very late decisions which are based on
every two weeks. After the election on September 27, the
rather idiosyncratic considerations or a “gut feeling” (for a
seventh and last panel wave took place from September 29
similar argument see McAllister, 2002). However, some of
to October 7. All in all, 3.301 respondents took part in more
those who decide very late can also be expected to be
than three panel waves and 1.462 respondents completed
rather sophisticated “strategic voters” who are guided in
all seven.
their vote choice by the latest polls and coalition signals
In every pre-election panel wave, respondents were
which are often sent late in the campaign (McGregor, 2012).
asked to state their vote intention. If they named a party,
Finally, protest voters also make their decisions compara-
the follow-up question read: “Why would you vote for this
tively late (McGregor, 2012).
party?” In the post-election panel wave, the question was
H1b. The more the type of motivation varies during the slightly modified and posed only to self-confessed voters:
campaign, the later a voter decides “Why did you vote for this party?” Respondents were free
to make a statement in a provided text box. However, since
In addition to these arguments for inter-personal het-
writing down subjective reasoning via the keyboard is a
erogeneity as a source of variation in the timing of vote
cognitively more challenging and certainly a more time-
decision, we expect that changes in the category of self-
consuming task than simply answering Likert-scale ques-
reported reasons during the campaign delay the time of
tions, on average 18 percent of the respondents refused to
the decision. While we expect that some of the voters'
answer. Also, voters who were undecided in a particular
motivations are stable over the course of the campaign, we
panel wave did not receive this question, so that we obvi-
will most certainly also observe over-time variability.
ously do not have information on their most important
Changing their most important decision criterion should
reason to vote at these points of time. In the following
motivate voters to reconsider their vote intention, thereby
analyses, we only included respondents who provided their
increasing the likelihood of changing party preferences or
subjective reasons in at least four panel waves. Note that
e at least e decreasing the certainty of an intention. Both
this reinforces already existent biases in the composition of
mechanisms should delay the point in time from which the
our sample: the remaining respondents are older, more
decision remains unchanged and is perceived as being final.
partisan, more interested in politics, more politically
Even if eventually inconsequential for their party choice,
knowledgeable, and more educated, among other things.
most voters will perceive this reassessment as the time of
Yet without this restriction the variability of the motiva-
their final decision and should therefore report later
tions would have been confounded with the individual
decisions.
participation history.
 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
All coding of the open-ended self-reports was done by
five coders. Prior to actual coding, a reliability test of the
coding scheme was conducted, yielding a sufficiently high
intercoder reliability of .8 (Holsti's CR). However, we were
unwilling to readily accept coding mistakes as this might
have influenced our findings. Instead we applied an exten-
sive three-step coding process of all material with multiple
coders. In the first step, two coders (not including the re-
searchers) coded all verbatims with an overall percent
agreement of 78.2. In the second step, two additional coders
independently coded all cases with deviating values from
the first step. Afterwards, the modal value of the four cod-
ings was assigned as final code. For the 5.7 percent of the
cases for which the mode was not uniquely defined, a fifth
coder made a final decision between the two conflicting
codes. While this procedure was very time-consuming, it
ensures that the coding as closely follows the coding
scheme as possible, which is crucial if coded self-reports are
to be compared at the individual level between panel waves.
Simply relying on a limited reliability test could have
resulted in deviating codings for some individuals between
two waves although they provided very similar answers and
did actually not change their decision motivation at all.
For the analyses presented below, we only make use of
the first codings of the verbatims which we consider to
represent the most important reason. For about a fifth of the
self-reports, a second code was assigned. We re-estimated
all analyses including the information from the second
vote reasons and found virtually no differences except for
minimally reduced associations and model fit values which
we interpret as indication of the information contained in
the second reasons to be somewhat more dubious.
The relative frequencies of the seven categories across
the panel waves are displayed in Fig. 1. Generally, the dis-
tribution is relatively stable during the campaign. If
40%
35%
30%
25%
20%
15%
10%
5%
0%
1 2 3 4
Party Performance
Strategy Gut Feeling
Fig. 1. Self-reported reasons for vote intention by wave (%).
33
anything, party-related reasons are less frequently reported
at the end of the campaign, whereas strategic reasons are
more frequently mentioned in waves 6 and 7, which is
consistent with coalition signals often not being expressed
until late in the campaign. Issue- and party related reasons
are the relatively most frequent stated reasons in all panel
waves. Candidates are referred to in less than five percent
of the self-reports in all stages of the campaign, only about
as often as gut feeling or protest are mentioned.
Before testing the hypotheses postulated above with the
described data, we first have to present some empirical
evidence in support of the coded self-reports as measures
of inter- and intra-personal heterogeneity. Although we
cannot ultimately prove the validity of the self-reports, we
can test whether the necessary conditions are met that
individual self-reports are related to the strength of the
attitudeeintention relationship and display some degree of
stability over time. If self-reports were mere ration-
alizations, they would reflect aspects respondents recently
read or heard about in the media but not their true con-
siderations. Hence, the answers between any two panel
waves would be randomly distributed.
As to the stability of the self-reports on the individual
level, the association of the coded reasons between any two
panel waves varies between .4 and .5 as measured by
Cramer's V. To better illustrate the extent of stability in the
seven categories, we compared the likelihood of being
classified into a given category dependent on the classifi-
cation in the previous wave (see Fig. 1 in the online
appendix). Generally, respondents are much more likely
to be coded into a specific category if they were already
coded into the same category in the previous wave. For
example, about 40 percent of those who were classified as
candidate voters by their self-report in wave 1 are classified
into the same category based on their self-reported reasons
5 6 7
Issues Candidates
Protest Other
34 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40

in wave 2, compared to only two percent of those who were Table 2

classified to any other category in wave 1. While the dif- Time of vote decision by self-reported reason.

ference in percentage points between the two groups var- Self-reported reason Decided within the last N
ies, the same pattern applies for all categories of the coding in post-election weeks of campaign (%)
scheme. If respondents had merely provided random an- panel wave

swers in their self-reported reasons, the classification into a 1) Party 29 (5) 411
given group would instead be independent of the classifi- 2) Performance 26 (5e7) 224
3) Issues 37 479
cation in the previous wave. Consequently, the codings of
4) Candidates 43 65
voters' self-reported reasons show a considerable amount 5) Strategy 48 (1e2) 168
of stability, even though of course quite a few respondents 6) Gut feeling 51 (2) 63
change their most important motivation between any two 7) Protest 51 (2) 63
Total 36 1473
panel waves.
While this pattern points to rather high stability of the Numbers in parentheses are results from Scheffe
-tests. Interpretation: The
self-reports (and the reliability of the coding), this does not share of voters who have reported a late vote decision is significantly
(p < .05) smaller in group 1 (party voters) compared to group 5 (strategic
necessarily imply that the coded verbatims are also valid
voters).
indicators of voters' motivations. To demonstrate that the
self-reports capture real differences in voters' attitude Table 3
weights we estimated a series of bivariate and multivariate Time of vote decision by number of self-reported reasons.

conditional logit models of party choice (see Table A.1 in
the Appendix). All models were separately estimated for
groups of voters with different self-reported reasons. The
dependent variable of all models is the voting decision as
stated in the post-election wave 7. As independent variable,
only the criterion mentioned in the self-reported reason is
included. For example, model A only includes party iden-
tification as an independent variable. As Table A.1 in the
Appendix demonstrates, party identification alone
correctly explains 53.7 percent of the choices made by re-
spondents who were classified as party voters, compared to
only 42.9 percent of the choices made by respondents
coded to any other group (i.e. who did not mention party
identification as their most important decision criterion).
Generally, the percentage of voting decisions being
correctly explained by any given predictor of vote choice is
between six and ten percentage points higher if the cor-
responding reason was mentioned in the open-ended
questions. In sum, these findings corroborate the validity
of the coded self-reports as measures of heterogeneity.
6. Results
We have hypothesized in hypotheses 1a and 1b that
voter heterogeneity contributes to the understanding of the
time of the final vote decision in two ways. First, the type of
motivation to vote for a party should affect the timing,
because at which stage in the campaign the relevant in-
formation becomes available differs between these moti-
vations (H1a). Second, as changes in voters' primary
motivation over the course of the campaign should cause
them to reconsider their party preference, greater vari-
ability in choice criteria is expected to delay the time of
final vote decisions (H1b).
With regard to hypothesis 1a, one simple way to assess
the strength of the assumed relationship is to calculate the
share of voters who reported “late” decision (defined as
having made the final decision in the last weeks of the
campaign) in dependence of their self-reported reason to
vote for a party, as measured in the post-election panel
wave. The data in Table 2 support the hypothesis that
different types of voters tend to decide in different stages of
the campaign. The earliest decisions are indeed made by
Number of distinct Decided within the last N
self-reported reasons weeks of campaign (%)
1 29 298
2 32 575
3 39 509
4þ 54 190
Total 36 1572
g ¼ .19***
* ¼ p < .05, ** ¼ p < .01, *** ¼ p < .001.
party-voters and performance-voters; less than 30 percent
of voters in both groups made their decision in the final
weeks of the campaign. Both groups have in common that
the information underlying their most important decision
criterion had been readily available in the earliest stages of
the campaign. Contrary to that, issue voters and candidate
voters more heavily rely on prospective evaluations, and
consequently yield a higher share of late decisions (37% and
43%, respectively). As expected, the latest decisions were
reported by strategic voters (48%), gut-feeling-voters, and
protest voters (51%).
In order to examine the second assumed mechanism of
how voter heterogeneity affects the time of vote decision
stated in H1b, Table 3 reports the share of late deciders in
dependence of how many distinct self-reported reasons
were stated over the course of the campaign.3 As hypoth-
esized, respondents with a higher number of different
motivations decided later: There is a monotonic positive
and significant relationship between the number of distinct
reasons and the share of late decisions. Only 29 percent of
the respondents whose vote motivation remained stable
over all panel waves made late decisions, while 54 percent
of respondents with four or more different reasons decided
late.
3
The indicator for the number of distinct motivations ranges from 1 (0
is not a possible value as only respondents are included in the analysis
who provided self-reports in at least four waves) to a theoretical
maximum of 7 (for voters who provided a different motivation in each
wave). On average, the respondents included in the analysis provided 2.4
different motivations. About 20 percent of the respondents provided the
same motivation in all waves in which they participated, 1 percent of the
respondents provided five different reasons, no respondent provided
more than five different motivations.
 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40 35

Table 4
OLS-regression for time of vote decision.

Independent Variables Model 1 Model 2 Model 3 Model 4

Age .009*** (4.09) .008*** (3.90) .010*** (4.58) .009*** (4.27)

PID strength .22*** (10.82) .21*** (9.96) .21*** (10.15) .20*** (9.48)
Political Interest .15 (1.20) .22 (1.63) .19 (1.38) .26 (1.88)
Political Knowledge .01 (1.62) .01 (1.46) .01 (1.81) .01 (1.61)
Ambivalence 1.40*** (6.90) 1.34*** (6.64) 1.31*** (6.51) 1.28*** (6.38)
Self-reported reason (Party ¼ reference):
Performance .02 (.25) .06 (.58)
Issues .17* (2.14) .18* (2.30)
Candidates .37* (2.38) .28 (1.79)
Strategy .45*** (4.15) .36** (3.31)
Gut feeling .55** (3.50) .51** (3.25)
Protest .16 (.99) .08 (.53)
No. of distinct self-reported reasons .18*** (5.58) .16*** (4.96)
Adj. R2 .137 .144 .155 .166
Log-Likelihood 2279.800 2263.829 2264.346 2251.542
LR-difference (compared to model 1) 31.941*** 30.908*** 56.517***
N 1457 1457 1457 1457

* ¼ p < .05, ** ¼ p < .01, *** ¼ p < .001. Coefficients are unstandardized regression coefficients, t-values in parentheses.

While these bivariate statistics provide initial support
for hypotheses 1a and 1b, the crucial question is whether
the type and intra-individual diversity of choice motiva-
tions contribute to the explanation of the timing of vote
choice over and above to established predictors. From a
recent meta-analysis on the predictors of the time of vote
decision (Plischke, 2014) it is known that the existence and
strength of party identification is the most important single
variable in explaining time of vote decision. Furthermore,
early decisions tend to be made by politically interested
and knowledgeable voters as well as by older ones (see e.g.
Schmitt-Beck and Partheymüller, 2012). Our set of control
variables also includes an indicator named “ambivalence”
as previous research has shown that many citizens are in
ambivalent choice situations and that ambivalence is an
important predictor of both the time of the voting decision
(Mutz, 2002) and response variability over time (Alvarez
and Brehm, 1995). The indicator for voters in ambivalent
choice situations employed here is based on the distance in
party feeling thermometers between the best and the
second best rated party. To create an indicator of ambiva-
lence, this distance is inverted and rescaled to range from
0 to 1, so that it takes on higher values the closer the ratings
for the two best rated parties are (see description of
variables in the Appendix).
In order to test whether the inclusion of heterogenous
voter motivations contributes to the explanation of the
time voters make decisions, we apply a stepwise modeling
strategy. In the first step, time-of-vote decision is regressed
on the full set of conventional predictors (see Table 4,
Model 1).4 These voter characteristics explain 13.7 percent
4 contribute to the explanation of time of vote decision. To
Since the dependent variable is measured with an ordinal five-point
scale, estimating an ordered logit model instead of an OLS regression
might be considered more appropriate. However, there is some debate in
the literature whether estimating OLS models for binary and ordinal
dependent variables is preferable, not least for the more intuitive inter-
pretation. As results from OLS- and the ordered logistic model are very
similar, we decided to present the OLS model for easier interpretation
(see for comparison the results from the ordered logistic model in Table 1
in the online appendix).
of the variance in time of vote decision. Not surprisingly,
the most important predictor is partisanship: Voters who
are not or only weakly attached to a party tend to decide
later. Also, ambivalence and younger age contribute to the
explanation of late deciding. In contrast, the effects of po-
litical interest and political knowledge are not significant at
the five-percent level.
In the next step, the set of explanatory variables was
expanded by the motivations for the voting decision as
provided in the post-election wave. The different reasons
enter the regression equation in the form of six dummy
variables, with party voters serving as reference group.
Hence, the positive regression coefficients in column 3
indicate that all other groups except for performance voters
make later decisions compared to the party voters. With
one notable exception, the differences in time of vote de-
cision between voter types are very similar to the bivariate
statistics in Table 2: Issue voters, as a group, make up their
mind a little later than party voters, whereas the compar-
atively latest decisions were being made by respondents
who stated a gut feeling, strategic considerations or the
candidates as their most important reason for party choice.
The only difference affects protest voters: Once the influ-
ence of partisanship is controlled for, protest voters do not
make significant later decisions than performance or party
voters. Hence, it is not the protest motive that leads to later
vote decisions, but the fact that protest voters tend to be
independents (but see McGregor, 2012). All in all, the in-
clusion of the self-reported reasons lead to a rather small,
but significant improvement of model fit, as a likelihood-
ratio test confirms.
In Model 3, we further tested whether changes in voters'
motivations over the course of the campaign also
this end, we added the recorded number of distinct moti-
vations as an independent variable to the set of established
predictors (see Table 4, Model 3). Again, the results support
our hypothesis: The more different self-reported reasons
are stated during the campaign, the later voters decided.
The effect is both statistically significant and substantially
relevant: Voters who stated at least four different reasons
36 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
on average scored .54 scale points higher on the time-of-
decision scale than voters who displayed a single, and
thus stable, motivation over the course of the campaign.
One might object that there are some obvious inter-
relationships between our indicators for inter- and intra-
personal heterogeneity. For example, voters who eventu-
ally reported to have voted strategically could have been
party or issue voters in earlier stages of the campaign. Put
differently, there might be a positive correlation between
those motivations with significant effects in Model 2 and
the number of changes in the prime motivation to vote for a
party. In this case, the significant effects in Model 2 would
reflect a spurious relationship which arises from neglecting
the count variable introduced in Model 3.
Indeed, strategic voters and candidate voters have
switched their most important motivation somewhat more
often (on average 2.9 and 2.8 times, respectively; numbers
are not displayed in tables) than party voters (2.3 times),
performance voters (2.5 times), and issue voters (2.2
times). But while some effect sizes are smaller in Model 4
compared to model 2, the main pattern remains un-
changed. We can therefore conclude that both inter- and
intra-individual variation in voters' motivations are signif-
icant and substantially meaningful predictors over and
above established determinants of time of vote decision.
We have argued that variability in voters' motivations
should lead to greater instability of their voting intentions.
Maybe the best analogy is that of a regression equation
where a change in predicted outcomes does not only occur
due to changes in the values of the independent variables,
but also by changes in the impact of each independent
variable, expressed by a change in the regression co-
efficients. Translated to the explanation of swing vote, a
change of vote intention need not necessarily be a conse-
quence of changing attitudes and cognitions, but rather can
be a consequence of the weights ascribed to them. This
argument seems straightforward, but has hitherto not
received much attention (but see Fishbein and Ajzen, 1975).
Contrary to that, as explicated above the content of self-
reported motivations should not be related to vote
switching.
As expected, there are indeed no significant differences
in the number of distinct parties for which vote intentions
were formulated5 and the type of self-reported motivation
in the post-election panel wave (see Table 4 in the online
appendix). While there is a slight tendency that strategic
voters, gut feeling voters, and protest voters switched more
often than other voters, these differences are small and not
statistically significant. However, as stated in H2, the
number of preferred parties monotonically increases with
the number of distinct self-reported reasons (Table 5).
5 6
The number of preferred parties theoretically ranges from 1 (as only
respondents who stated a vote intention for one of the parties received
the open-ended question about the reasons for vote choice) to 6 (if each
of the five major parties CDU/CSU, SPD, FDP, GRUENE, and DIE LINKE was
preferred in at least one wave as well as one of the smaller parties not
represented in the parliament). Empirically, respondents in the analysis
preferred on average 1.4 parties with about two thirds preferred only one
single party and less than one percent preferring four different parties. No
respondent preferred more than four different parties.
Table 5
Number of distinct vote intentions by number of self-reported reasons.
Number of distinct Preferred parties (mean) N
self-reported reasons
1 1.23 273
2 1.37 534
3 1.48 482
4þ 1.81 188
Total 1.44 1477
g ¼ .36***
* ¼ p < .05, ** ¼ p < .01, *** ¼ p < .001.
Respondents with stable motivations on average preferred
1.23 parties during the whole length of the campaign,
whereas voters with four or more different motivations
over the course of the campaign preferred 1.81 parties. This
association is significant with gamma amounting to .36.
In the next step, we tested the relationship stated in H2
in a multivariate model.6 The dependent variable is the
number of preferred parties over the course of the
campaign. As control variables, we identified an array of
voter characteristics which are commonly applied to
explain swing voting. The model includes time-constant
variables like strength of party identification, political in-
terest, ambivalence and political knowledge (cf. e.g.;
Schoen, 2003), each of them measured in the post-election
panel wave. We also included variables expressing attitude
change: the number of changes in party identification (cf.
Clarke and McCutcheon, 2009 on the instability of party
identifications), the number of changes of the party iden-
tified as the most competent one in solving the most
important problem, and changes in party ratings, an index
measuring the amount of intra-individual volatility in the
parties' feeling thermometers. The control variables
explain a sizeable portion of swing voting (adj. R2 ¼ .14, see
Table 6, Model 1). The most important predictor is strength
of party attachment with strong partisans less frequently
switching parties. Additionally, the three indicators for
attitude change also contribute to the explanatory power of
the model. Frequent switches of the party perceived as
most competent or in the party most highly acclaimed
corresponded with a high number of preferred parties.
While these relationships are not very surprising, they
contribute to make the test of our hypothesis as challenging
as possible.
Consistent with our expectations, entering dummy
variables for the different motivations to vote for a party
does not significantly improve the model fit (Model 2).
Contrarily, adding the number of different reasons to the
model yields a highly significant regression coefficient,
Since the dependent variable of the model is a count variable, esti-
mating a Poisson regression model instead of an OLS regression might be
considered more appropriate. However, as the range of the dependent
variable does not include zero, the dependent variable must be trans-
formed for doing so. Also the interpretation of coefficients is more intu-
itive in the OLS model. As the results from the OLS model and the Poisson
regression are very similar, we decided to present the OLS model for
easier interpretation here (see for comparison the results from the
Poisson model in Table 2 in the online appendix).
 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40 37

Table 6 Methodologically speaking, what is needed is a panel-

OLS-regression for number of preferred parties. model explaining intra-individual change in vote in-
Independent Variables Model 1 Model 2 Model 3 Model 4 tentions over time by accounting for intra-individual
PID strength .07*** .07*** .06*** .06***
changes in the motivations over time.
(5.04) (4.98) (4.62) (4.65) The appropriate method for this purpose is the so-called
Political Interest .07 .07 .10 .10 “fixed-effects” (FE) panel regression (see Allison, 2009).
(.72) (.71) (1.07) (1.10) This model has two noteworthy properties: First, for the
Ambivalence .40** .39** .33* .34*
estimation of effect parameters, this method solely relies
(2.66) (2.61) (2.22) (2.31)
Political Knowledge .06 .06 .05 .05 on differences within individuals over time, and not on
(1.83) (1.82) (1.63) (1.6) comparisons between persons, as is the case in conven-
No. of changes in PID .17*** .17*** .16*** .16*** tional cross-sectional regression models. Hence, this setup
(5.32) (5.23) (5.14) (5.08)
enables testing the proposition that changes in the
No. of changes party .08*** .08*** .07*** .07***
competence (4.42) (4.39) (4.22) (4.15)
dependent variable coincide in time with changes in the
Change in party feeling .07*** .07*** .06*** .06*** independent variables. It does not make any difference
thermometers (4.77) (4.57) (4.00) (3.77) whether respondents took part in only two panel waves or
Self-reported reason (Party ¼ reference): in all seven ones. All that is needed for a respondent to be
Performance .13 .15*
included in the model are at least two points of observa-
(1.89) (2.33)
Issues .08 .06 tions. Second, since FE-models explain differences with
(1.37) (1.12) differences, the model implicitly controls for all time-
Candidates .12 .21* constant individual characteristics without explicitly
(1.10) (1.97) including them in the model. This property of FE panel
Strategy .05 .13
(.67) (1.82)
models greatly reduce the potential for omitted variables
Gut feeling .01 .05 bias which usually is a concern in cross-sectional regres-
(.07) (.44) sion models.
Protest .06 .13 Since the units of analysis in FE-Models are in-
(.56) (1.17)
dividuals at different points of time, it is necessary to
No. of distinct self- .12*** .14***
reported reasons (5.86) (6.20) transform the dataset into the “stacked format”, so that
Adj. R2 .140 .138 .174 .177 individuals are represented in the dataset as often as
Log-Likelihood 708.887 706.563 691.873 687.418 they have participated in the panel study. For every
LR-difference (compared 4.648 34.029*** 42.397*** individual and each panel wave starting from the sec-
to model 1)
N 803 803 803 803
ond interview, we have the information whether the
individual switched her vote intention compared with
* ¼ p < .05, ** ¼ p < .01, *** ¼ p < .001. Coefficients are unstandardized
the previous interview (1) or whether the vote inten-
regression coefficients, t-values in parentheses.
tion has remained constant (0). This is the dependent
variable of the model. Similarly, we used a dummy
indicating that a one-number increase in the number of variable to measure change in the voters' motivations,
motivations leads, on average, to a .12 increase in the taking on the value 1 when the most important reason
number of preferred parties. In fact, this effect is the single changed or 0 when it remained stable compared to the
most predictive variable in the model (according to the t- previous interview. Stable voter characteristics do not
values). In a bivariate model, it explains 5.3 percent of the have to be included since they are automatically
variance of the dependent variable (results not shown controlled for (Allison, 2009).
here). Entering it on the top the whole set of predictor Yet it is important to consider indicators for attitude
variables (Model 4), it increases the adjusted R2-statistic by change as control variables. This is complicated by the fact
additional 3.7 points. that few of our control variables included in Table 6 were
However, it might be criticized that the multivariate surveyed in each panel wave. We therefore estimated a
model in Table 6 provides only indirect support for the very parsimonious model, in which only one indicator for
claim that changes in vote intentions are indeed a conse- attitude change is included as independent variable using
quence of changing motivations. The analysis reveals that the party feeling thermometers. These variables are suit-
respondents who frequently change their reasons for party able because they can be considered summary judgments
choice also frequently change their voting intentions. An which comprise more specific attitudes. For each panel
even more convincing test of our hypothesis would be to wave, the current rank of each party was calculated based
show that both changes coincide7 in time. on the feeling thermometers. If the party the individual
intended to vote for in the previous wave lost at least one
rank compared to the previous panel wave, a dummy var-
7
If the relationship in fact is a causal relationship, it would imply a iable takes on the value “1”. Such a rank change indicates
time lag, and no simultaneity, as the word “coincide” suggests. However, attitude change in that the previously preferred party has
with regard of the relationship between calculus change and vote change, lost some ground against its competitors. We would
we would expect the time lag to be very small, perhaps only a few sec- therefore expect a change of vote intention, too. The
onds. Since we do not have individual observations for different seconds,
but only for different weeks, we would expect to observe simultaneous
dummy variable takes on the values “0” if the rank of the
changes in both dependent and independent variables, even though the previously preferred party was constant or even got better,
changes should actually occur with a very small time lag. making a vote switch less likely.
38 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
Table 7
Fixed-Effects OLS-regression for change of vote intention.
Independent variables Model
Change in feeling thermometer (“attitude change”) .11*** (12.26)
Change of self-reported reason (“motivation change”) .04*** (4.76)
R2 (within) .028
N (observations) 9192
N (individuals) 2643
* ¼ p < .05, ** ¼ p < .01, *** ¼ p < .001. Coefficients are unstandardized
regression coefficients, t-values in parentheses. Estimated parameters for
wave-specific constants are not displayed.
Table 7 contains the results of the FE-model.8 In total,
data of 2643 respondents were used. On average, these
respondents took part in 3.48 panel waves, so that in total
9152 observations were used in this model (2643 multi-
plied by 3.48). The R2-statistic refers to the proportion of
intra-individual variance in vote switching that can be
explained by attitude change and motivation change, and
amounts to 2.8 percent. As changes within persons are
more difficult to explain, low R2 values are a typical trait of
FE-models.
Regarding the effect coefficients, the results displayed in
Table 7 support our previous findings. Both independent
variables e the indicator for attitude change as well as the
indicator for intra-individual heterogeneity e significantly
contribute to the explanation of vote switching. The like-
lihood to change a vote intention increases on average by 11
percentage points if the attitude toward the previously
preferred party deteriorates relative to the competing
parties. The impact of a motivation change is somewhat
smaller e the likelihood for a vote switch increases by 4
percentage points e but substantially improves the model
fit. Moreover, it has to be considered that changes in
motivation occur more frequently: On average, 46 percent
of individuals changed their motivation to vote between
two panel waves, while a rank loss of the previously most
preferred party took place only in 19 percent of all obser-
vations. This further illustrates the contribution of intra-
individual heterogeneity to the explanation of the voting
decision process.
7. Conclusion
As a popular adage in electoral science goes, “voters are
not all alike”. Previous studies have shown that e contrary
to the formerly conventional “homogeneity assumption” e
voters differ in their motivations to vote for a party and that
these differences influence both their party choice and
their decision making process (cf. e.g. Bartle, 2005; Roy,
2011). While previous research has almost exclusively
focused on inter-personal heterogeneity, we have argued
8
Since the dependent variable is binary, the most appropriate method
would be a logistic FE-regression. However, estimates from logistic FE-
regression are difficult to interpret, because predicted probabilities
cannot be calculated. As results from OLS- and logistic FE-models are very
similar, we decided to present the OLS model for easier interpretation
here (see for comparison the results of the logistic FE panel regression in
Table 3 in the online appendix).
that the reality is even more complex, and that an impor-
tant aspect in the study of voter decision-making has
hitherto been overlooked. Not only are voters not all alike,
but they are also not staying the same. Voters sometimes
change their motivations over time e something we have
called “intra-personal heterogeneity” e and shifting moti-
vations in turn contribute to the explanation of important
aspects of the voting decision process.
Rather than analyzing the determinants of voter het-
erogeneity, we used a proxy measure of voters' most
important motivations based on their self-reports to
analyze consequences of inter- and intrapersonal hetero-
geneity. More precisely, our research question focused on
the effects of the type of choice motivation and intra-
individual changes in these motivations over time on the
time of the vote decision and on short-term volatility in
vote intentions.
Regarding the timing of vote decisions, we have
demonstrated that inter-as well as intra-personal hetero-
geneity explain when voters make their decisions. In
accordance with previous findings, late decisions are made
by candidate voters, strategic voters and voters who refer to
a “gut feeling” as their most important reason for party
choice. However, what has not been recognized thus far,
late decisions are also being made by voters who frequently
shift their choice motivations over the course of the
campaign.
Regarding vote switching, the dominant account thus
far has been attitude change: If voters develop distaste for
a party they had previously endorsed, it is likely that they
change their vote intention. However, changes in party
preferences can also occur if there is change in which
attitudes primarily motivate an individual's vote inten-
tion. Our analyses have demonstrated that this mecha-
nism contributes to the explanation of short-term
volatility even if attitude change is controlled for, and
future work on vote switching may benefit from incor-
porating these insights.
In times of possibly increasing voter heterogeneity,
these mechanisms could become ever more important.
However, one might speculate about whether these two
forms of voter heterogeneity depend on differences in
electoral and party systems. With regard to inter-personal
heterogeneity, it is possible that candidate voting might
be more prevalent in countries where candidates get
elected, opposed to party-centered countries such as
Germany, where party-related motivations dominate.
Additionally, the extent of strategic motivations in the
electorate depends on the opportunities provided by the
electoral system to behave strategically. One could argue
that Germany, with its five-percent threshold for parties
to enter parliament and its tradition of coalition govern-
ments, offers more opportunities for strategic behavior
than the United States. As these motivations are related to
different time of voting decisions, voter heterogeneity
could explain some of the differences in time of voting
decision between different countries (see e.g., McAllister,
2002).
Likewise, the relationship between intra-personal
heterogeneity and vote switching might depend on the
particular party system. In Germany, switching votes takes
 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40
places when voters have several parties in their choice set,
and typically these choice sets are comprised by parties of
similar ideologies and issue stances. Multi-party systems,
in other words, makes vote switching more likely, because
voters have more ideologically related alternatives. In such
settings, when voters have difficulties to decide between
parties toward which they have similar attitudes, the
particular motivation can make the difference for one or
the other party, thus affecting the outcome of the election.
While the presented theoretical mechanism should be a
universal one, in two party systems, and especially in
polarized party systems, changes in motivations could be
somewhat less consequential.
Finally, some limitations as to the generalizability of our
data and the validity of our measures have to be
mentioned. With regard to the data, remember that re-
spondents were recruited from an online-access panel and
therefore are not representative for the German electorate.
It might be possible that the mechanisms underlying the
respondents' decision-making processes are different from
voters who do not participate in online access-panels.
Moreover, several authors have doubted the validity of
self-reports, doubting voters' ability to reconstruct their
most important reasons for voting the way they did,
depreciating them as “rationalizations” (cf. e.g., Nisbett and
Wilson, 1977). However, we have brought forward theo-
retical and empirical evidence for our claim that voters'
self-reports are meaningful. Of course, it would be desir-
able to develop measures of voter heterogeneity that are
even more precise. Nonetheless, we are convinced that a
neglect of inter- and intrapersonal heterogeneity would
seriously hurt our understanding of voter decision-making
processes.
Acknowledgments
The authors wish to thank two anonymous reviewers
for their valuable comments and suggestions and Kon-
stantin Gavras for skilled research assistance. The GLES is
funded by the German Research Foundation (DFG).
Appendix A. Supplementary data
Supplementary data related to this article can be found
at http://dx.doi.org/10.1016/j.electstud.2014.11.003.
Appendix
Variables in the analysis.
Time of the vote decision. “When did you decide how
you were going to vote in the federal election?” (1) A long
time before the election, (2) A few months before the
election, (3) In the last few weeks before the election, (4)
In the last few days before the election, (5) Not until
election day.
Voting intention (second vote). “You have two votes in the
federal election. The first vote is for a candidate from your
constituency and the second is for a party. Please tell me
39
which party you will vote with your second vote.”(1) CDU/
CSU, (2) SPD, (3) FDP, (4) GRUENE, (7) DIE LINKE, (8) Other
party.
Change of voting intention. Pairwise comparisons of
voting intentions between any two consecutive panel
waves (excluding don't know and no answer): (0) voting
intention unchanged compared to previous wave; (1)
voting intention changed compared to previous wave;
index ranges from 0 to 6 changes.
Party identification. “Many people in Germany feel close
to a particular political party for a longer period of time
even if they occasionally vote for another party. What about
you? In general terms, do you feel close to a particular
political party? And if so, which one?”
Party identification, strength. “All in all, how strongly or
weakly attached are you to this party?”(0) no party iden-
tification, (1) Very weakly, (2) Fairly weakly, (3) Moderately,
(4) Fairly strongly, (5) Very strongly.
Changes in party identification. Pairwise comparisons of
party identification between any two consecutive panel
waves (excluding don't know and no answer): (0) party
identification unchanged to previous wave; (1) party
identification changed compared to previous wave; index
ranges from 0 to 6 changes.
Political interest. “In general terms: How interested in
politics are you?” (1) Not interested at all, (2) Slightly
interested, (3) Moderately interested, (4) Very interested,
(5) Extremely interested.
Political Knowledge. Index of two factual knowledge
questions: A) Importance of the first and second vote, B)
Electoral threshold; index ranges from (0) no question
correct, (1) one question correct, (2) two questions correct.
Feeling thermometer. “How do you feel about the polit-
ical parties in general? Please use a scale ranging from 5
to þ5.” (CDU/CSU, SPD, FDP, GRUENE, DIE LINKE), (5) I
have a very negative view of this party to, (5) I have a very
positive view of this party; recoded for the analyses to
range from 0 to 1.
Change of feeling thermometers. Pairwise comparisons
of feeling thermometers (rescaled to range from 0 to 1)
for the five major parties between any two consecutive
panel waves ranging from 0 (no difference) to 1
(maximum absolute difference). Index: sum of pairwise
comparison, (theoretically) ranging from 0 to 30 (five
parties times six wave transitions), the empirical
maximum is 11.7.
Ambivalence. Distance between the best-rated and the
second best-rated party in the feeling thermometer Reco-
ded to range from 0 (highest distance, i.e. lowest ambiva-
lence) to 1 (zero distance, i.e. highest ambivalence).
Party competence: party best to solve most important
issue. “And which political party do you think is best able to
solve this most import issue?”
Change of the party competence. Pairwise comparisons
between any two consecutive panel waves (excluding don't
know and no answer): (0) party best able to solve most
important issue unchanged; (1) change in party best able to
solve most important issue; index: 0e6 changes.
40 J.E. Blumenstiel, T. Plischke / Electoral Studies 37 (2015) 28e40

Table A.1
Explanatory power of conditional logit models for vote choice by self-reported reasons

Model Independent Variable(s) Voter group McFadden R2 Correctly predicted (%) N (obs.)

A Party identification Party voters .508 53.7 2350

Others .341 42.9 6705
B Party performance Performance voters .443 49.5 1138
Others .307 41.4 4622
C Party competence Issue voters .456 76.3 2025
Others .364 70.1 2425
D Candidate thermometer Candidate voters .459 58.2 415
Others .379 52.2 6099
E Party identification, party performance, Gut feeling, protest, strategic voters .422 60.2 882
party competence, candidate thermometer Others .694 83.5 3007

A voter's decision is classified as correctly predicted if the party with the highest predicted probability is chosen.
Interpretation: The results stem from five conditional logit models for the vote choice as reported in wave 7. Each model has been estimated separately for
two groups of respondents with the first group comprising those respondents whose type of motivation corresponds to the independent variable(s) included
in the model and the second group comprising all other respondents. If the self-reports measure real differences in attitude weights between voters, the
explanatory power for each model should be higher in the first group except for model E. As to the latter, the explanatory power of the full model should be
lower for the first group because the rationale of the decisions of gut feeling, strategic and protest voters is beyond the theoretical frame of the Michigan
model.

References
Allison, P.D., 2009. Fixed Effects Regression Models. SAGE Publications,
London.
Alvarez, R.M., Brehm, J., 1995. American ambivalence towards abortion
policy: development of a Heteroskedastic Probit model of competing
values. Am. J. Polit. Sci. 39 (4), 1055e1082.
Bartle, J., 2005. Homogeneous models and heterogeneous voters. Polit.
Stud. 53 (4), 653e675.
Blais, A., Martin, P., Nadeau, R., 1998. Can people explain their own vote?
Introspective questions as indicators of salience in the 1995 Quebec
referendum on sovereignty. Qual. Quant. 32 (4), 355e366.
Blumenstiel, J.E., 2014. Voter fragmentation and the differentiation of vote
functions. In: Weßels, B., Rattinger, H., Rossteutscher, S., Schmitt-
Beck, R. (Eds.), Voters on the Move or on the Run? Oxford University
Press, Oxford, pp. 17e39.
Blumenstiel, J.E., Rattinger, H., 2012. Warum haben Sie das getan? Sub-
jektive Gründe der Wahlentscheidung bei der Bundestagswahl 2009.
In: Schmitt-Beck, R. (Ed.), W€ ahlen in Deutschland. PVS Sonderheft,
vol. 45. Nomos, Baden-Baden, pp. 251e275.
Box-Steffensmeier, J.M., Kimball, D., 1999. The timing of voting decisions
in presidential campaigns. In: Paper Prepared for Presentation at the
1999 Meeting of the Midwest Political Science Association.
Campbell, A., Converse, P.E., Miller, W.E., Stokes, D.E., 1960. The American
Voter. John Wiley, New York, London.
Chaffee, S.H., Choe, S.Y., 1980. Time of decision and Media use during the
Ford-Carter campaign. Public Opin. Q. 44 (1), 53e69.
Clarke, H.D., McCutcheon, A.L., 2009. The dynamics of party identification
reconsidered. Public Opin. Q. 73 (4), 704e728.
Clarke, H.D., Kornberg, A., Scotto, T.J., 2008. Making Political Choices:
Canada and the United States. University of Toronto Press, Toronto.
Dalton, R.J., 2008. Citizen Politics: Public Opinion and Political Parties in
Advanced Industrial Democracies, fourth ed. CQ Press, Washington
DC.
Druckman, J.N., 2004. Priming the vote: campaign effects in a I.S. Senate
election. Polit. Psychol. 25 (4), 577e594.
Ericsson, K.A., Simon, H.A., 1980. Verbal reports as data. Psychol. Rev. 87
(3), 215e251.
Fishbein, M., Ajzen, I., 1975. Belief, Attitude, Intention, and Behavior: an
Introduction to Theory and Research. Addison-Wesley, Reading.
Gigerenzer, G., Todd, P.M., the ABC Research Group, 1999. Simple Heu-
ristics that Make Us Smart. Oxford University Press, Oxford.
Iyengar, S., Kinder, D.R., 1987. News that Matters: Television and American
Opinion. University of Chicago Press.
Lewis-Beck, M.S., Jacoby, W.G., Norpoth, H., Weisberg, H.F., 2008. The
American Voter Revisited. University of Michigan Press, Ann Arbor.
McAllister, I., 2002. Calculating or capricious? The new politics of late
deciding voters. In: Farrell, D.M., Schmitt-Beck, R. (Eds.), Do Political
Campaigns Matter? Campaign Effects in Elections and Referendums.
Routledge, London, pp. 22e40.
McGregor, R.M., 2012. Voter sincerity and the time-of-voting-decision.
Elect. Stud. 31 (4), 715e725.
Mendelsohn, M., 1996. The Media and interpersonal communications: the
priming of issues, leaders, and party identification. J. Polit. 58 (1),
112e125.
Mutz, D.C., 2002. The consequences of cross-cutting networks for political
participation. Am. J. Polit. Sci. 46 (4), 838e855.
Nisbett, R.E., Wilson, T.D., 1977. Telling more than we can know e verbal
reports on mental processes. Psychol. Rev. 84 (3), 231e259.
Peterson, D.A.M., 2005. Heterogeneity and certainty in candidate evalu-
ations. Polit. Behav. 27 (1), 1e24.
Peterson, D.A.M., 2009. Campaign learning and vote determinants. Am. J.
Polit. Sci. 53 (2), 445e460.
Plischke, T., 2014. Wann Wa €hler entscheiden. Abla €ufe von Entschei-
dungsprozessen und der Zeitpunkt der Wahlentscheidung. Nomos
Verlag, Baden-Baden.
Rattinger, H., Roßteutscher, S., Schmitt-Beck, R., Weßels, B., 2012. German
Longitudinal Election Study e Short-term Campaign Panel, 07/10/
2009-10/07/2009. GESIS, Ko €ln. http://dx.doi.org/10.4232/1.11131.
ZA5305, Version 3.0.0.
Rivers, D., 1988. Heterogeneity in models of electoral choice. Am. J. Polit.
Sci. 32 (3), 737e757.
Roy, J.J., 2011. Information heterogeneity, complexity and the vote cal-
culus. Journal of elections. Public Opin. Parties 21 (1), 29e56.
Schmitt-Beck, R., Partheymüller, J., 2012. Why voters decide late: a
simultaneous test of old and new hypotheses at the 2005 and 2009
german federal elections. Ger. Polit. 21 (3), 299e316.
Schoen, H., 2003. W€ ahlerwandel und Wechselwahl: Eine vergleichende
Untersuchung. Westdeutscher Verlag, Wiesbaden.
White, P.A., 1988. Knowing more about what we can tell: 'introspective
access' and causal report accuracy 10 years later. Br. J. Psychol. 79,
13e45.