See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330842670

Issue Importance and Voting Behavior

Preprint · October 2019

CITATIONS READS
0 5,381

2 authors:

Wilson Law Patrick Flavin

Baylor University Baylor University
18 PUBLICATIONS 169 CITATIONS 64 PUBLICATIONS 1,264 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Wilson Law on 07 October 2019.

The user has requested enhancement of the downloaded file.

 Issue Importance and Voting Behavior
Patrick Flavin
Department of Political Science
Baylor University
One Bear Place #97276, Waco, TX 76798-7276
[email protected]
(254) 710-7418

Wilson Law
(Corresponding Author)
Department of Economics
Baylor University
One Bear Place #98003, Waco, TX 76798-8003
[email protected]
(254) 710-4390
https://orcid.org/0000-0001-9408-7001

Abstract
This paper investigates whether the reported subjective importance of policy positions in surveys is influential
in citizens’ vote choices. We derive a novel estimation model from the canonical proximity voting literature
and, in contrast to previous studies on this question, directly estimate the weights of policies from vote choices.
Using this new method and data from the American National Election Studies for 1996-2008, we find evidence
that citizens’ voting behavior is consistent with their reported policy positions and subjective issue importance.
A one standard deviation change in the policies that are extremely important to the voters is associated with a
7.6 percentage point increase of voting for the more ideologically proximate candidate, about a quarter of the
size of the effect from identifying as a Strong Democrat or Strong Republican. These findings suggest that
policy positions are influential in citizens’ vote choices even after controlling for party identification and
demographic characteristics and, contrary to previous studies, the impact of policy proximity is noticeably
heightened for issues that citizens deem important. Methodologically, the technique we introduce for
estimating policy weights can be applied to future studies of public opinion and voting behavior.
Keywords: Voter Preference, Spatial Voting Model, Discrete Choice Model, Issue Salience, Identification

In a democratic society where candidates propose policy positions to win votes, it is important to understand

how sensitive (or not) citizens’ voting behaviors are to these positions. More specifically, scholars are

interested in learning how vital each individual policy is to citizens when they consider whom to vote for in

elections. 1 This research agenda often employs the spatial or proximity voting model (Hotelling, 1929; Black,

1
Other political behavior traditions focus on drivers such as party identification (e.g., Converse, 1964) and
group membership and identities (e.g., Lazarsfeld, Berelson and Gaudet, 1968; Achen and Bartels, 2017).
1
1948; Downs, 1957; Davis, Hinich, and Ordeshook, 1970) that posits that voters hold preferred positions for

different policies and then select candidates whose positions are closest to them. When multiple issues are

being considered, the relative importance of these policies is essential in addition to the distances to the

candidates. To date, the general consensus among the scholars (e.g., Niemi and Bartels, 1985; Grynaviski and

Corrigan, 2006; Leeper and Robison, 2018) suggests that policy importance levels reported by respondents

have a minimal bearing on voters’ evaluations of candidates and their vote choice. 2 This lack of an effect on

issue importance in previous studies is somewhat surprising, we argue, for several reasons. First, citizens are

more likely to frequently think of issues that are important to them, less likely to oscillate in their opinions on

those issues over time, and more likely to form presidential candidate preferences on the basis of those

opinions (Krosnick, 1990; Lavine, Borgida and Sullivan, 2000; Lecheler, de Vreese and Slothuus, 2009).

Second, citizens are more likely to seek out additional information about political issues they deem important

and are more likely to learn where candidates stand on those issues (Converse, 1964; Holbrook et al., 2005;

Iyengar et al., 2008; Bolsen and Leeper, 2013; Henderson, 2014). Third, citizens are more likely to engage in

political participation when issues that are important to them are at stake (Price et al., 2006; Holbrook et al.,

2016). In short, strong empirical support from related studies leads us to expect that issue importance enhances

proximity voting.

Our primary methodological contribution in this paper is to develop an estimation model from the

canonical proximity voting model by recognizing that policy positions and levels of importance (weights)

reported in most data sets (e.g., that of American National Election Studies (ANES)) are ordinal (not cardinal)

values. In previous studies, a common approach is to assume implicitly the reported weights are linear

transformation of the true weights. We suspect this approach introduces measurement errors that ultimately

attenuate the estimated effects of policy weights. Heterogeneous voter preference may present additional

challenges to precisely estimating the effects. In our model, we use the reported importance levels as grouping

2
Some research, such as Krosnick (1988b) and Krosnick (1990), demonstrate that different policies are
important to different citizens. However, the magnitude of the effects are minimal, and their methodology is
subject to the critique by Grynaviski and Corrigan (2006) and this paper.
2
variables and estimate the average weight for each level of subjective importance directly. Further, our model

allows us to evaluate survey data aggregated over multiple years. By aggregating more than 6,000 individual

observations from the ANES presidential election surveys for 1996-2008, we obtain more precise estimates

of the policy weights. Importantly, we find that half of the voters in our data hold some positions closer to the

candidates they did not choose to vote. These conflicted voters help us identify the issue importance weights

they use to reconcile their differences between the candidates.

Our primary substantive contribution is to the extensive literature on proximity voting (in particular,

regarding the salience of policy issues to voters) by directly estimating the policy weights from data. We find

that a one standard deviation change in the policies that are extremely important to voters is associated with

about 7.6 percentage point increase from voting for the candidates. The size of the effect is about a quarter of

that from identifying as a Strong Democrat or a Strong Republican. The finding that citizens engage in

proximity voting for issues they deem important contrasts with the “democratic realists” (Carpini and Keeter,

1996; Achen and Bartels, 2017) who generally contend that citizens cannot possibly come to informed

judgments about the ideological placement of themselves and candidates for the dozens of issues at play in

any given election. Moreover, we contribute to the literature that focuses on party identification, group

membership, and identities (e.g., Converse, 1964; Lazarsfeld, Berelson and Gaudet, 1968; Achen and Bartels,

2017) by showing that policy preferences are influential for vote choice even after accounting for party

identification and demographic characteristics. Taken together, this paper has important implications for our

understanding of citizens’ levels of democratic competence and suggests a more important role for political

issues in explaining voting behavior in American elections.

3
1. MODEL AND IDENTIFICATION

In this section, we begin with the canonical proximity voting model. We review the commonly adopted

estimation procedure under a binary vote choice setting. 3 We then propose our estimation model and discuss

the identification and interpretation of the parameters.

We index individuals in the survey by 𝑖𝑖 , and the candidates by 𝑘𝑘 ∈ {𝐷𝐷, 𝑅𝑅} , where D denotes the

Democratic candidate and R denotes the Republican candidate. Conditional on voting, a voter chooses the

candidate 𝑐𝑐𝑖𝑖 who maximizes utility 𝑢𝑢𝑖𝑖𝑖𝑖 . Formally,

𝑐𝑐𝑖𝑖 = arg max 𝑢𝑢𝑖𝑖𝑖𝑖

k∈{D,R}

We assume voters choose the candidate based on the candidate’s valence and the net benefit of voting (𝜌𝜌𝑖𝑖𝑖𝑖 )

and the ideological distance:

𝑢𝑢𝑖𝑖𝑖𝑖 = 𝜌𝜌𝑖𝑖𝑖𝑖 − 𝑑𝑑(𝑟𝑟𝑖𝑖 , 𝑦𝑦𝑖𝑖𝑖𝑖 ; 𝜔𝜔𝑖𝑖 ) + 𝜖𝜖𝑖𝑖𝑖𝑖

Where 𝑑𝑑(⋅) is a distance function that returns the distance between the positions of the respondent 𝑟𝑟𝑖𝑖 and

those of the candidate 𝑦𝑦𝑖𝑖𝑖𝑖 with issue importance (or policy weights) 𝜔𝜔𝑖𝑖 . Our interest is to learn about 𝜔𝜔𝑖𝑖 .

Note that candidate 𝑦𝑦𝑖𝑖𝑖𝑖 is indexed by 𝑖𝑖 because respondents use the perceived candidates’ positions in the

decision-making process. Lastly, 𝜖𝜖𝑖𝑖𝑖𝑖 represents the idiosyncratic shock to the voter’s utility that researchers

do not observe at the time of the decision. The two commonly used distance functions are Euclidean and city

block norms (Grynaviski and Corrigan, 2006). With Euclidean norm, the weighted distance is defined as:

𝑃𝑃
2
𝑑𝑑𝑖𝑖𝑖𝑖 ≡ 𝑑𝑑(𝑟𝑟𝑖𝑖 , 𝑦𝑦𝑖𝑖𝑖𝑖 ; 𝜔𝜔𝑖𝑖 ) = � 𝜔𝜔𝑖𝑖𝑖𝑖 �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 �
𝑗𝑗=1

where each policy is indexed by 𝑗𝑗 = 1,2, … , 𝑃𝑃. With city block norm, the corresponding distance function is

defined as:

3
The exposition here can be applied to linear models and discrete choice models with more than two choices.
4
 𝑃𝑃

𝑑𝑑𝑖𝑖𝑖𝑖 ≡ 𝑑𝑑(𝑟𝑟𝑖𝑖 , 𝑦𝑦𝑖𝑖𝑖𝑖 ; 𝜔𝜔𝑖𝑖 ) = � 𝜔𝜔𝑖𝑖𝑖𝑖 �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 �

𝑗𝑗=1

In a binary vote choice model, only the difference between the two utility functions matters. The probability of

choosing the Republican candidate is:

𝑃𝑃(𝑐𝑐𝑖𝑖 = 𝑅𝑅) = 𝑃𝑃(𝑢𝑢𝑖𝑖𝑖𝑖 − 𝑢𝑢𝑖𝑖𝑖𝑖 > 0)

= 𝑃𝑃(𝜖𝜖𝑖𝑖𝑖𝑖 − 𝜖𝜖𝑖𝑖𝑖𝑖 < 𝜌𝜌𝑖𝑖𝑖𝑖 − 𝜌𝜌𝑖𝑖𝑖𝑖 + 𝑑𝑑𝑖𝑖𝑖𝑖 − 𝑑𝑑𝑖𝑖𝑖𝑖 )

= 𝑃𝑃(𝜖𝜖𝑖𝑖𝑖𝑖 − 𝜖𝜖𝑖𝑖𝑖𝑖 < 𝑣𝑣𝑖𝑖 )

where 𝑣𝑣𝑖𝑖 is a linear function of the difference in the closeness to the two candidates and the corresponding

policy weights. The probability of voting for the Republican is commonly estimated with a probit or logit

model. With city block norm, the linear part of the model becomes

𝑃𝑃

𝑣𝑣𝑖𝑖 = 𝜌𝜌𝑖𝑖𝑖𝑖 − 𝜌𝜌𝑖𝑖𝑖𝑖 + � 𝜔𝜔𝑖𝑖𝑖𝑖 ��𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � − �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 ��

𝑗𝑗=1

𝑃𝑃

= 𝜌𝜌𝑖𝑖𝑖𝑖 − 𝜌𝜌𝑖𝑖𝑖𝑖 + � 𝜔𝜔𝑖𝑖𝑖𝑖 𝑥𝑥𝑖𝑖𝑖𝑖 (1)

𝑗𝑗=1

Where 𝑥𝑥𝑖𝑖𝑖𝑖 = �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � − �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � is the difference in distances between the voters and the candidates.

Recall that our goal is to learn about 𝜔𝜔𝑖𝑖𝑖𝑖 . Since everyone has different weight for different policies, and we

typically observe only one decision for each voter, it is impossible to recover 𝜔𝜔𝑖𝑖𝑖𝑖 . Can we use the reported

policy weight 𝑤𝑤𝑖𝑖𝑖𝑖 ? A common approach is to assume implicitly the reported weight (𝑤𝑤𝑖𝑖𝑖𝑖 ) is a linear

transformation of the true weight (𝜔𝜔𝑖𝑖𝑖𝑖 ). The true weight is likely continuous and the reported weight is

typically on some integer scale such as 1 to 5. Mapping a continuous variable onto an integer scale likely

introduces some measurement errors. With an additional assumption that the policy weights by all individuals

5
can be rescaled by constants 𝑎𝑎𝑗𝑗 and 𝑏𝑏𝑗𝑗 , we can rewrite the relationship as: 𝜔𝜔𝑖𝑖𝑖𝑖 = 𝑎𝑎𝑗𝑗 + 𝑏𝑏𝑗𝑗 𝑤𝑤𝑖𝑖𝑖𝑖 + 𝜉𝜉𝑖𝑖𝑖𝑖 where 𝜉𝜉𝑖𝑖𝑖𝑖

is the measurement error that is zero in expectation. We can interpret 𝑎𝑎𝑗𝑗 as the average importance of policy

𝑗𝑗. 𝑏𝑏𝑗𝑗 is a scaling factor of the reported importance and thus has no structural interpretation. Substituting this

relationship in the theoretical model above (Equation 1), we obtain: 4

𝑃𝑃

𝑣𝑣𝑖𝑖 = 𝜌𝜌𝑖𝑖𝑖𝑖 − 𝜌𝜌𝑖𝑖𝑖𝑖 + � 𝑎𝑎𝑗𝑗 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝑏𝑏𝑗𝑗 𝑤𝑤𝑖𝑖𝑖𝑖 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝜉𝜉𝑖𝑖𝑖𝑖 𝑥𝑥𝑖𝑖𝑖𝑖 (2)
𝑗𝑗=1

In a comprehensive study, Leeper and Robison (2018) follow Krosnick (1988b) to estimate the following model:5

𝑃𝑃
′ 𝑅𝑅𝑅𝑅𝑅𝑅 𝑅𝑅𝑅𝑅𝑅𝑅
𝑣𝑣𝑖𝑖 = 𝛼𝛼 𝒛𝒛𝑖𝑖 + � 𝛽𝛽1𝑗𝑗 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝛽𝛽2𝑗𝑗 𝑤𝑤𝑖𝑖𝑖𝑖 + 𝛽𝛽3𝑗𝑗 𝑤𝑤𝑖𝑖𝑖𝑖 𝑥𝑥𝑖𝑖𝑖𝑖
𝑗𝑗=1

𝑅𝑅𝑅𝑅𝑅𝑅
where 𝒛𝒛𝑖𝑖 is a vector of voter characteristics. They use Relative Issue Distance, 𝑥𝑥𝑖𝑖𝑖𝑖 = �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦�𝐷𝐷𝐷𝐷 � −

�𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦�𝑅𝑅𝑅𝑅 � to measure the difference in distances to the average candidate positions ( 𝑦𝑦�𝐷𝐷𝐷𝐷 , 𝑦𝑦�𝑅𝑅𝑅𝑅 ) of all

respondents. 𝑤𝑤𝑖𝑖𝑖𝑖 is the reported subjective importance of policy 𝑗𝑗 in which higher values represent a more

important issue to 𝑖𝑖. 𝛼𝛼 and 𝛽𝛽 are the parameters to be estimated. In this model, as evidence from Equation (2),

𝛽𝛽1𝑗𝑗 represents the average weight of policy 𝑗𝑗 to all individuals. A large and significant estimate suggests the

policy is important to vote choices. 𝛽𝛽2𝑗𝑗 does not have an apparent structural interpretation because voters

should not make vote choice based on the level of importance in an issue in isolation. Recall that higher 𝑣𝑣𝑖𝑖

implies a higher likelihood to vote Republican. Suppose 𝛽𝛽2𝑗𝑗 is positive, it means that voters will more likely

vote Republican if they think issue 𝑗𝑗 is more important to them independent of their policy positions. If voters

are closer to the Democratic candidate, and the issue is important to them, we should expect a lower 𝑣𝑣𝑖𝑖 and

hence a negative 𝛽𝛽2𝑗𝑗 . Therefore, 𝛽𝛽2𝑗𝑗 does not have a clear interpretation and potentially reflects the average

positions of the sample. Current literature often uses the estimate of 𝛽𝛽3𝑗𝑗 to evaluate whether reported

4
See Grynaviski and Corrigan (2006) and Malhotra and Tahk (2011) for a more thorough discussion of the scale invariant form
in models where the reported distances are linear transformations of the true distances as well.
5
Bélanger and Meguid (2008) and Fournier et al. (2003) use a similar approach.
6
importance is meaningful in vote choices. As we reported above, the current findings are mixed. When a

statistically significant effect is observed, it is often small. To explain these puzzling results, Krosnick (1988b)

observes that in issues that respondents deem unimportant (and likely uninformed as a result), they are more

likely to report the candidate’s positions to be closer to theirs. This phenomenon might diminish the difference

in the estimated effects between the important and unimportant issues. We suspect measurement errors and

heterogeneity may obfuscate these findings as well. Recall that weights on an integer scale are measured with

errors. From Equation (2), 𝜉𝜉𝑖𝑖𝑖𝑖 𝑥𝑥𝑖𝑖𝑖𝑖 is included in the error term and thus correlates with 𝑥𝑥𝑖𝑖𝑖𝑖 which is likely to

bias the estimates of 𝛽𝛽3𝑗𝑗 toward zero. In addition, the scaling variable 𝑏𝑏𝑗𝑗 may vary by individuals and present

a challenge in estimating the mean of them precisely.

In this paper, we propose a new approach. We begin with the proximity voting model with city block

norm as in Equation 1, and using demographic characteristics to approximate the difference in candidate

preference and net benefit of voting (𝜌𝜌𝑖𝑖𝑖𝑖 − 𝜌𝜌𝑖𝑖𝑖𝑖 ) as in Leeper and Robison (2018), we have:

𝑃𝑃
′
𝑣𝑣𝑖𝑖 = 𝛼𝛼 𝒛𝒛𝑖𝑖 + � 𝜔𝜔𝑖𝑖𝑖𝑖 𝑥𝑥𝑖𝑖𝑖𝑖 (3)
𝑗𝑗=1

where 𝜔𝜔 𝑖𝑖𝑖𝑖 denotes the true unobserved weight 𝑖𝑖 puts on policy 𝑗𝑗. 𝒛𝒛𝑖𝑖 is a vector of voter characteristics and an

election fixed effect, and 𝑥𝑥𝑖𝑖𝑖𝑖 is the difference in distances between respondents and candidates. The individual

characteristics control for other unobserved policy positions and provide a frame of reference to interpret the

coefficients of the issue proximity. We further assume the difference in error terms is independent and

identical Type I extreme value random variable so that the formulation results in a familiar logit model.

From here, we can maintain the spatial and proximity voting tradition by treating the reported differences

in policy positions as linearly transformed cardinal measures of distances with measurement errors (as in

Equation 1). Or, we can follow Leeper and Robison (2018), Krosnick (1988b) and others to use Relative Issue

Distance which uses the average candidate positions in place of the respondents’ subjective candidate positions.

Regardless of how we measure the distances, we show in the appendix that the reported policy weights remain

7
as important as our main results. However, we do not favor using the differences between the respondents’

own positions and the average candidate positions because proximity models posit a psychological mechanism

such that individuals make decisions based on their perceived differences (Grynaviski and Corrigan, 2006;

Merrill, 1995). We also do not favor using the differences between perceived positions because the data are

usually ordinal. Policy positions are commonly reported as ordinal category values, but their differences are

often used as cardinal ideological distances. These differences might not be comparable between individuals

and policies. For example, the difference between favoring a policy “a great deal” and “moderately” is one

unit, but we cannot be certain that this relative distance is the same between individuals. We propose an

alternative measure for the difference in proximity by using discrete values to indicate which candidates are
𝑐𝑐
closer to the respondents. Let 𝑥𝑥𝑖𝑖𝑖𝑖 be the relative closeness of 𝑖𝑖 to the candidates in policy 𝑗𝑗,

⎧ 1, if �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � < �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � (closer to Republican)

⎪
𝑐𝑐
𝑥𝑥𝑖𝑖𝑖𝑖 = 0, if �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � = �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � (equal distance to both)
⎨
⎪−1, if �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � > �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � (closer to Democrat)
⎩

This measurement captures the most reliable information from the responses because we only need to assume

that individuals use equal distance in each policy assessment, they correctly assess the relative distances to

the candidates and use these distances to determine their vote choices. This measurement is not sensitive to

scaling issues common in models with Euclidean norms. 6 Our measurement has the flavor of directional

voting (e.g., Rabinowitz and Macdonald, 1989), but we do not consider whether the respondents are on the

same “side” of the issues as the candidates.

𝑐𝑐
With 𝑥𝑥𝑖𝑖𝑖𝑖 decided in our choice model (Equation 3), we consider the estimation of 𝜔𝜔𝑖𝑖𝑖𝑖 . In a typical election

survey, we only observe an individual’s voting decision once, 𝜔𝜔𝑖𝑖𝑖𝑖 , therefore, cannot be recovered from data

because everyone most likely has different weights for different policies. 7 Although we cannot recover

6
If differences in positions are treated as cardinal, the choice model needs to be invariant to the scale chosen by the
surveys (e.g., 1-7 vs. 1-100). Grynaviski and Corrigan (2006) and Malhotra and Tahk (2011) provide excellent guidance
on developing various scale invariant models.
7
With enough data, we can potentially estimate the average weights for each policy (𝐸𝐸�𝜔𝜔𝑗𝑗 �).
8
individual weights, we can recover the averages of the weights for each reported level of importance in the

surveys ranging from “the policy is not important at all” to “the policy is extremely important.” More formally,

we can identify the average weights 𝛿𝛿ℓ ≡ 𝐸𝐸�𝜔𝜔𝑖𝑖𝑖𝑖 �𝑤𝑤𝑖𝑖𝑖𝑖 = ℓ� , where ℓ = 1,2, … , 𝐿𝐿 is the importance level

reported in the surveys. In our data set, 𝐿𝐿 = 5. We can rewrite Equation 3 to be:

𝑐𝑐 𝑐𝑐
𝑣𝑣𝑖𝑖 = 𝛿𝛿1 � 𝑥𝑥𝑖𝑖𝑖𝑖 + ⋯ + 𝛿𝛿5 � 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝛼𝛼 ′ 𝒛𝒛𝑖𝑖
𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 =1 𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 =5

where 𝛿𝛿 can be estimated from data. 8 An equivalent model is to use the reported weight variables as dummy

variables and interacts them with the relative closeness (𝑥𝑥𝑖𝑖𝑖𝑖𝑐𝑐 ): 𝑣𝑣𝑖𝑖 = 𝛼𝛼 ′ 𝒛𝒛𝑖𝑖 + ∑𝑃𝑃𝑗𝑗=1 ∑5ℓ=1 𝛿𝛿ℓ 1�𝑤𝑤𝑖𝑖𝑖𝑖 = ℓ�𝑥𝑥𝑖𝑖𝑖𝑖𝑐𝑐 where 1(⋅)

is an indicator function. 9 With sufficient data, we can further estimate the average weight by policy where 𝛿𝛿𝑗𝑗ℓ ≡

𝐸𝐸�𝜔𝜔𝑖𝑖𝑖𝑖 �𝑤𝑤𝑖𝑖𝑖𝑖 = ℓ�, but we choose to focus on the simplest case first. An advantage of this model is that we can

aggregate multiple years of survey data in a single estimation to obtain more precise estimates. When we

aggregate multiple years of data, we normalize the sum of relative positions by the corresponding total number
1
of questions: 𝑞𝑞𝑖𝑖 = ∑ 𝑥𝑥 𝑐𝑐
𝑄𝑄𝑡𝑡 𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 =ℓ 𝑖𝑖𝑖𝑖
where 𝑄𝑄𝑡𝑡 is the number of questions asked in year t. 10 𝑞𝑞𝑖𝑖 ∈ [−1, 1] is the

net proportion of policies that is closer to Democrat (-1) or Republican (1). The choice model to be estimated

becomes:

𝑣𝑣𝑖𝑖 = 𝛿𝛿1 𝑞𝑞𝑖𝑖1 + 𝛿𝛿2 𝑞𝑞𝑖𝑖2 + 𝛿𝛿3 𝑞𝑞𝑖𝑖3 + 𝛿𝛿4 𝑞𝑞𝑖𝑖4 + 𝛿𝛿5 𝑞𝑞𝑖𝑖5 + 𝛼𝛼 ′ 𝒛𝒛𝑖𝑖 .

When the data set is small, we can simplify the model to have only two levels:

𝑐𝑐 𝑐𝑐
𝑣𝑣𝑖𝑖 = 𝛿𝛿𝐿𝐿 � 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝛿𝛿𝐻𝐻 � 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝛼𝛼 ′ 𝒛𝒛𝑖𝑖
𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 <𝜓𝜓 𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 ≥𝜓𝜓

8
More precisely, each 𝛿𝛿ℓ represents both the scale and weight of the policies in the voting decision. Let 𝛿𝛿ℓ = 𝜑𝜑𝜔𝜔ℓ , where
𝜑𝜑 > 0 is the scale and 𝜔𝜔ℓ is the weight with the restrictions that 𝜔𝜔ℓ > 0, ∀ℓ and ∑ℓ 𝜔𝜔ℓ = 1. We can obtain the scale
𝜑𝜑 = ∑ℓ 𝛿𝛿ℓ and 𝜔𝜔ℓ = 𝛿𝛿ℓ /𝜑𝜑 .
9
1(𝑎𝑎) = 1 when 𝑎𝑎 is true, and 0 otherwise.
10
For sake of readability, we omit the 𝑡𝑡 subscript in the discussion.

9
With 𝜓𝜓 being some cutoff level to separate the more important issues from the less important issues:

𝛿𝛿𝐿𝐿 = 𝐸𝐸�𝜔𝜔𝑖𝑖𝑖𝑖 �𝑤𝑤𝑖𝑖𝑖𝑖 < 𝜓𝜓� and 𝛿𝛿𝐻𝐻 = 𝐸𝐸�𝜔𝜔𝑖𝑖𝑖𝑖 �𝑤𝑤𝑖𝑖𝑖𝑖 ≥ 𝜓𝜓�. To normalize the sum of relative positions, we apply the

same procedures above to create:

1 1
𝑞𝑞𝑖𝑖𝑖𝑖 = � 𝑥𝑥𝑐𝑐𝑖𝑖𝑖𝑖 , 𝑞𝑞𝑖𝑖𝑖𝑖 = � 𝑥𝑥𝑐𝑐𝑖𝑖𝑖𝑖
𝑄𝑄𝑡𝑡 𝑄𝑄𝑡𝑡
𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 <𝜓𝜓 𝑗𝑗:𝑤𝑤𝑖𝑖𝑖𝑖 ≥𝜓𝜓

before estimating the parameters with the following choice utility:

𝑣𝑣𝑖𝑖 = 𝛿𝛿𝐿𝐿 𝑞𝑞𝑖𝑖𝑖𝑖 + 𝛿𝛿𝐻𝐻 𝑞𝑞𝑖𝑖𝑖𝑖 + 𝛼𝛼 ′ 𝒛𝒛𝑖𝑖

We can now test several hypotheses with data. If policy positions do not matter in vote choices, we should

expect all the 𝛿𝛿’s to be small and statistically insignificant. If policy positions matter, but reported issue importance

does not matter, all the 𝛿𝛿’s should be statistically significant but of similar magnitude. If voters are consistent with

their stated importance of policy preferences, we expect:

𝛿𝛿𝐿𝐿 < 𝛿𝛿𝐻𝐻 , and

𝛿𝛿1 < 𝛿𝛿2 < 𝛿𝛿3 < 𝛿𝛿4 < 𝛿𝛿5

The policy weights, 𝛿𝛿’s, are identified by the variations in 𝑥𝑥𝑖𝑖𝑖𝑖 at different important levels and voting

decisions. One worry is the potential endogeneity between candidate preference and perceived positions in

which the respondents report their positions to be closer to the candidates they favor (Converse and Markus,

1979; Page and Jones, 1979) rather than deducing their favored candidates from ideological positions as theory

suggested. 11 Relatedly, voters may project their own ideological position onto their favored candidate such

11
One common solution to this problem is to use Relative Issue Distance (e.g., Krosnick, 1988b; Visser, Krosnick and
Simmons, 2003; Leeper and Robison, 2018) by taking the average of all the respondents’ perceived candidate positions
and computing the distances between the respondents’ positions and the average positions. This method implicitly
assumes that there is a fixed location for each candidate for all respondents. If true, the averaging procedure might
reduce the measurement error and move closer to an accurate distance. However, the respondents need to respond to
this true location, which is unknown to them. Therefore, Merrill (1995) and Grynaviski and Corrigan (2006) think the
voting behavior should depend on voters’ self placement and own placement of the candidates. If candidates present
different positions to different groups of voters, or different voters receive different messages, the candidate placement
will differ by respondents. The averaging procedure will introduce measurement error and bias the estimates downward
(Lewis and King, 1999; Tomz and Van Houweling, 2008). Another concern is that the independent variables are
10
that their subjective placement markedly differs from a more objective or expert placement of that candidate

(Wilson and Gronke, 2000; Ansolabehere and Jones, 2010). The implication is that our estimates can be

potentially biased upward. We believe this concern is mitigated in several ways. In our data, about half of the

respondents have at least one issue that is closer to the candidate who they did not choose to vote for. It

suggests that a significant number of voters are conflicted, and their positions are not entirely a projection

from their vote choices. In fact, these conflicted voters help identify the weights of the issues. If issue

proximity is completely consistent with vote choices in every issue, it is not possible to identify the weights.

Our conclusions remain the same when we limit our analysis only to the conflicted voters. Moreover, there is

little reason to expect that citizens are more likely to project their preferences onto candidates for issues they

deem more important. Krosnick (1988a) and Krosnick (1990) find that issues that are important to citizens

are highly resistant to change and highly stable over time. If citizens are persuaded by their preferred

candidates, it is most likely for unimportant issues. In our empirical analysis, we include party identification

and a common set of demographic characteristics as controls. In the subsequent analysis, we show that party

identification significantly influences vote choices, and we believe it also absorbs a significant amount of

projection if present. In summary, we believe that any endogeneity that is not accounted for, it is likely to

manifest itself in the less important issues. Lastly, another objective of this paper is to resolve the puzzle that

voters’ decisions do not appear to depend on the importance level they reported. Current literature has not

demonstrated that it significantly affects vote choice.

2. DATA

The data comes from pre-election surveys conducted by the American National Election Studies (ANES) in the

presidential election years for 1996-2008. 12 We limit our sample to those who planned to vote for the two major-

correlated with each other through the averages of other observations. It invalidates the independent assumptions in the
error terms.
12
We omit data from 1980 and 1984 because some policy position questions were asked in the pre-election surveys and some in the
post-election surveys. As policy positions likely change as campaign progresses, we feel it inappropriate to combine the two sets
of questions. Also, in the 1980 and 1984 surveys, the scale of issue importance items is different than those in the later years. It
makes it difficult to aggregate them in the same study.

11
party candidates in the presidential elections. The summary statistics of the vote choices, party identification,

and demographic characteristics are reported in Tables 1 and 2.

[Tables 1 and 2 approximately here]

In the surveys, individuals were asked to report their positions in different policies and those of the

presidential candidates in mostly seven-point ordinal values. Table 3 lists the policy questions, and the years

they were asked. In addition to policy positions, the respondents report the level of the importance of each

issue to them on a five-point scale ranging from “not important at all” to “extremely important.”

[Table 3 approximately here]

[Figure 1 approximately here]

To investigate the relationship between policy positions and vote choices, we tally the proportions of

policies 13 that are closer to the Democrat candidate, the Republican candidate, or Neither for each respondent.

If individuals are not sure about their positions or that of the candidates, we assume they are closer to neither

because respondents cannot determine which candidate is closer to that particular policy. Figure 1 presents

the distribution of the proportions of policy positions that are closer to each candidate or neither conditional

on vote choice. Not surprisingly, contingent on voting Democrat, the density is higher in the higher proportion

of policies that is closer to the Democratic candidates. Conditional on voting Republican, the distribution is

analogous. Table 4 reports the summary statistics of the proportions for all levels of importance and by levels

of importance to the respondents. On average, about half of the issues are closer to that of the candidates the

respondents are planning to vote for. About 12 to 13% of the policy positions are closer to the candidates they

choose not to vote for. Importantly, in our sample, close to half of the respondents have at least one policy

position that is closer to the candidates they choose not to vote for. When we subdivide the statistics by the

level of importance, we continue to see that a higher proportion of the policy positions are closer to that of the

13
Proportions are more comparable between years than the number of policies because each year has different numbers of policy
questions.

12
preferred candidates. There are higher proportions of the policy positions in the more important categories.

Respondents are less likely to be indifferent in the policies they deem more important than those they consider

less important. The proportions of positions closer to the favored candidates are much higher than the

opponents in the more important policies. These sample statistics suggest that a large percentage of voters are

conflicted with their choices. There is some preliminary evidence that voters reconcile these conflicting

positions by evaluating and aggregating how important is each policy to them. In the results section, we

present more evidence of this evaluative process.

[Tables 4 and 5 approximately here]

𝑐𝑐
To estimate our voting model, we create a variable 𝑥𝑥𝑖𝑖𝑖𝑖 to indicate individual 𝑖𝑖’s relative positions to that

𝑐𝑐
of the candidates in policy 𝑗𝑗. 𝑥𝑥𝑖𝑖𝑖𝑖 = 1 when the respondent’s position is closer to that of the Republican

candidate than the Democratic candidate, −1 if the Democratic candidate is closer, and 0 if they are of equal
𝑐𝑐
distance or the respondents are uncertain about the positions. Summing up 𝑥𝑥𝑖𝑖𝑖𝑖 for all policies for each

1
individual and dividing it by the number of questions asked each year, we have a measurement 𝑞𝑞𝑖𝑖𝑖𝑖 = ∑ 𝑥𝑥 𝑐𝑐
𝑄𝑄𝑡𝑡 𝑗𝑗 𝑖𝑖𝑖𝑖𝑖𝑖

comparable between years. This normalized sum has a range of -1, where all the issues are closer to the

Democratic candidate, to 1, where all the issues are closer to the Republican candidate. Figure 2 presents the

distribution of the normalized policy closeness conditional on vote choice. Reassuringly, respondents tend to

vote for the candidates with which the net proportion of issues is closer to them. In other words, the

respondents are sensitive to policy positions. But why did some respondents choose to vote for the candidate

who has more issues farther away from that voter than their opponent? It is evident that the voters weigh the

importance of the policies, and choose with a weighted sum. Figure 3 presents the distributions of the net

positions by the level of importance. The distribution patterns are similar to that of the total values except that

zero is more prevalent in the less important issues. They are consistent with the statistics reported in Table 4.

It suggests that the respondents are either not knowing the positions of the less important policies, they are

indifferent between the candidates, or the number of issues closer to the two candidates is equal. Table 5

13
presents the summary statistics of the distributions in Figures 2 and 3. One notable feature is that the standard

deviation is higher for more important issues.

[Figures 2 and 3 approximately here]

3. RESULTS

We estimate our model with the aggregated data for 1996-2008. The estimates of weights and party

identification are reported in Table 6. 14 In the first column, we present the estimated average policy weights

for the five levels of importance. All coefficients have the expected positive signs. They are statistically

significant at any conventional level (p-value<0.01), and the magnitude is large relative to the party

identification. Put differently, voters’ policy preferences are predictive of their vote choice even after

controlling for party identification. The size of the coefficients decreases with the level of importance except

for “Not important at all.” The second column presents the results of the statistical tests on the differences of

the coefficients. We cannot statistically distinguish the difference between that of the Very Important issues

and the Moderately Important issues, but the coefficients are mostly in the right order. The coefficient of the

policies that are Not Important At All is surprisingly large. As Krosnick (1990) points out, voters are more

likely to be persuaded by their preferred candidates in the unimportant issues. It is likely that voters are not

informed on the issues not important to them. When they were asked to report the policy positions, they place

themselves closer to their favored candidates. We believe this is a result of candidate preference projection

instead of their true issue importance. In the next four columns, we present four different alternative

specifications that vary with the inclusion of party identification, demographic characteristics, and election

fixed effects. Election fixed effects reflect the average advantage Republican candidates have over Democratic

candidates for different election years. When the fixed effects are present, issue importance and party

identification have larger coefficients, but the increase is minimal. Demographic characteristics have a very

small impact on the estimates as well. Party identification has the most substantial influence on the issue

importance estimates. When the party identification is omitted, the issue importance increases almost

14
The coefficients of the demographic characteristics and election fixed effects are reported in the Appendix.
14
uniformly, and the order of the weights remains unchanged. It shows that party preference, and most likely

candidate preference as well, correlates with policy preference, but it is far from explaining all the variation,

nor does it change the relative levels of issue importance in vote choices. In the appendix, we present the

results from using two different measurements of issue proximity used in previous studies: (1) the Subjective

Issue Distance technique (i.e., the linear distance between respondents’ policy positions and their subjective

placement of the candidates’ positions and (2) the Relative Issue Distance technique (i.e., the linear distance

between respondents’ policy positions and the average placement of candidates). We find that the estimates

for the least important issues are sensitive to the measurements, but for the more important policies, the results

follow the same pattern as our main results. Therefore, if we set aside the least important issues, voters’

behaviors are consistent with proximity voting, and their reported issue importance has a strong influence on

their decisions regardless of how we measure issue proximity.

[Tables 6 and 7 approximately here]

In Table 7, we report the estimated weights for two importance levels with different cutoffs. All of them

are estimated with very high precision, and the more important issues carry a heavier weight than the less

important issues. The differences between the two levels are statistically significant at any conventional level.

The party identification parameters all have the expected signs, are precisely estimated, and are very stable

across all models.

[Table 8 approximately here]

Table 8 reports the marginal effects at the mean of the data. All of the results are precisely estimated.

Using the sample statistics in Table 5, we can derive that a one standard deviation change in the extremely

important position (0.23) is associated with about 7.6 percentage points (0.329 × 0.23 ≈ 0.076) increase in

voting for a Republican candidate. The size of this effect is about a quarter of that from identifying as a Strong

Democrat or a Strong Republican relative to an Independent. For very important issues, a one standard

deviation change is associated with a 6.8 percentage point (0.260 × 0.26 ≈ 0.068) increase. In contrast, for the

issues not important at all, there is about one percentage point (0.186 × 0.06 ≈ 0.01) increase due to a smaller

15
standard deviation and a more modest marginal effect than those of the more important issues. However, we

believe this marginal effect reflects the candidate preference rather than the ideological preference. The

marginal effects for the two-level models continue to be significant as expected.

Overall, our data suggests that voters have diverse policy preferences, which are in stark contrast with the

polarized positions of the elected officials. Voters reconcile the conflicting policy preferences between

candidates by weighing policies that are important to them against those that are less important. We also find

that policy preferences have a sizable influence in vote choices relative to party identification.

4. CONCLUSION

One crucial way citizens might encourage greater congruence between public opinion and the policy decisions

made by elected officials is by determining their own opinion on a political issue, comparing that opinion to

the positions of the candidates in a campaign on that issue, and then voting for the candidate who is most

proximate to their position. While this process of (spatial or proximity) issue voting would ideally promote

greater political responsiveness, “democratic realists” might rightfully contend that citizens cannot possibly

come to informed judgments about the ideological placement of themselves and candidates for the dozens of

issues at play in any given election (e.g., Carpini and Keeter, 1996; Achen and Bartels, 2017). Instead, we

might more reasonably expect this process to play out among “issue publics”–groups of citizens who report

that government implementing policy following their preferences on a particular issue is especially important

to them (e.g., Converse, 1964; Krosnick, 1990). From a theoretical standpoint, it is those citizens who report

an issue being subjectively important to them that we would expect to seek out information about candidates’

ideological positions on the issue and then cast their vote for the more ideologically proximate candidate (e.g.,

Krosnick, 1988a; Hutchings, 2001; Visser, Krosnick and Simmons, 2003; Holbrook et al., 2005; Iyengar et

al., 2008; Bolsen and Leeper, 2013; Henderson, 2014).

Despite the intuitive appeal of this issue public proposition, empirical studies to date have not found much-

supporting evidence that citizens who deem an issue important are any more likely to engage in proximity

issue voting (e.g., Niemi and Bartels, 1985; Grynaviski and Corrigan, 2006). Indeed, a recent study puts its
16
conclusion rather bluntly: “there is scant evidence that subjective issue importance consistently moderates the

relationship between issue positions and vote choice” (Leeper and Robison, 2018, 2). In contrast, we show in

this paper that there is strong statistical evidence to suggest that citizens are choosing among candidates based,

at least in part, on the reported subjective importance of their policy positions. Importantly, the results we

uncover are robust to different ways of measuring proximity including those used in previous studies.

Moreover, we find that policy preferences are essential for vote choice even after accounting for party

identification and a standard set of demographic characteristics. Our data also show that voter policy

preferences are highly heterogeneous, in contrast with the polarized positions among elected officials in recent

decades.

With the methodology introduced in this paper that allows for the estimation of policy importance, future

researchers can perform a variety of counterfactual exercises. For example, one application is to assess how

sensitive vote choices are to a change in a candidate’s policy position on a specific issue or to determine which

policy might induce the most substantial shift in voter support. In addition, one concern with using the

observational data is the potential bias that results from voters projecting their candidate preference in the

least important issues. To tackle this challenge, using data from experimental studies such as those in Leeper

and Robison (2018) and Jenke and Munger (Forthcoming) with our method may obtain more accurate

estimates, but at the cost of losing the context of an actual election.

More broadly, our findings highlight the importance of including issue salience or subjective importance

in future studies of voting behavior. Unlike politicians, citizens do not necessarily hold a position in every

policy, nor do they inevitably find every issue to be relevant to them. Accordingly, spatial voting models that

reduce the dimension of voters’ positions (e.g., Jessee, 2009; Jessee, 2016) may need to more carefully

consider and account for the salience of each policy being analyzed.

17
Table 1: Distribution of Vote Choice and Party Identification (1996-2008)
Vote Choice Percent Party Identification Percent
Democrat 58.38 Strong Democrat 24.55
Republican 41.62 Not very strong Democrat 17.14
Leaning Democrat 12.77
Independent 6.58
Leaning Republican 10.01
Not very strong Republican 13.22
Strong Republican 15.73
Total 100.00 Total 100.00

Table 2: Demographic Characteristics (1996-2008)

Variable Mean Std. Dev Min Max
Age* 48.88 16.54 17 97
Proportion
Male 0.43
White 0.77
Some College 0.32
College Grad 0.34
Low Income** 0.18
High Income** 0.23
Observations 6273
Note:

* One individual in this sample was 17 years old at the time of

interview but would turn 18 by Election Day.

** Low and high income are defined as the bottom and top quintiles
respectively in household income in the sample for each year.

18
 Table 3: Policy Questions on ANES from 1996 to 2008

Policy 1996 2000 2004 2008

TS1 TS2 Panel
Increase/decrease Spending and Services X X X X
Increase/decrease Defense Spending X X X X
Government Provision of Job/Standard of Living X X
Aid to Blacks/Minority Groups X X X X
Abortion X X
Women’s Equal Role in Society X X
Protect the Environment vs. Protect Jobs X X X
Protect the Environment vs. Burdening Business X
Lower Emission Standards X
Gun Control X X
Government vs. Private Health Insurance X
Universal Health Care X X
Citizenship Process: Illegal Immigrants X X
Same-Sex Marriage X
Raising Taxes: >200,000 X
Pay for Low-Income Seniors Prescription Drugs X
Habeas Corpus Rights for Terrorists X
Court Orders for Wiretaps X
3-Year Work Stay for Undocumented Immigrants X
Deadline for Iraq Troop Withdrawal X
Note: 2008 TS1 and TS2 are two sets of policy questions randomly assigned to the 2008
Time Series study. 2008 Panel is a separate panel study. We use the survey results from
October.

Figure 1: Distribution of Policy Positions Conditional on Vote Choice

19
 Table 4: Summary Statistics of Policy Positions Conditional on Vote Choice

Vote Democrat Vote Republican

Fraction of Positions Closer to Mean Std Dev Mean Std Dev
All levels of importance
Democrat 0.48 0.30 0.13 0.17
Neither 0.40 0.31 0.39 0.29
Republican 0.12 0.16 0.49 0.29
Observations 3662 2611

Vote Democrat Vote Republican

Fraction of Positions Closer to Mean Std Dev Mean Std Dev
5-Extremely important
Democrat 0.14 0.21 0.02 0.07
Neither 0.07 0.14 0.06 0.14
Republican 0.02 0.07 0.13 0.19
4-Very important
Democrat 0.18 0.20 0.04 0.09
Neither 0.12 0.18 0.11 0.17
Republican 0.04 0.09 0.18 0.20
3-Moderately important
Democrat 0.12 0.16 0.04 0.09
Neither 0.13 0.18 0.14 0.18
Republican 0.04 0.09 0.13 0.17
2-Slightly important
Democrat 0.03 0.09 0.01 0.05
Neither 0.05 0.11 0.05 0.11
Republican 0.01 0.05 0.04 0.09
1-Not important at all
Democrat 0.01 0.05 0.00 0.03
Neither 0.03 0.09 0.02 0.08
Republican 0.00 0.03 0.01 0.05
Observations 3662 2611

Table 5: Summary Statistics of Net Policy Positions

Mean Std Dev Min (Closest to Dem) Max (Closest to Rep)

All levels of importance -0.06 0.52 -1 1
5-Extremely important -0.03 0.23 -1 1
4-Very important -0.02 0.26 -1 1
3-Moderately important -0.01 0.20 -1 1
2-Slightly important -0.00 0.10 -0.78 0.67
1-Not important at all -0.00 0.06 -0.67 1
Observations 6273

20
Figure 2: Net Policy Positions Conditional on Vote Choice

21
Figure 3: Net Policy Positions Conditional on Vote Choice and Level of Importance

22
 Table 6: Issue Importance in Five Levels
Dependent Variable: Vote Republican
(1) Main Results Alternative Specifications
Variable Coefficient Test (2) (3) (4) (5)

Extremely important (𝛿𝛿̂5 ) 4.946*** 𝐻𝐻0 : 𝛿𝛿5 ≤ 𝛿𝛿4 4.821*** 4.852*** 6.203*** 6.324***
(0.347) p-value < 0.01 (0.339) (0.335) (0.292) (0.283)
Very important (𝛿𝛿̂4 ) 3.911*** 𝐻𝐻0 : 𝛿𝛿4 ≤ 𝛿𝛿3 3.744*** 3.806*** 4.971*** 5.055***
(0.261) p-value = 0.127 (0.254) (0.252) (0.215) (0.209)
Moderately important (𝛿𝛿̂3 ) 3.452*** 𝐻𝐻0 : 𝛿𝛿3 ≤ 𝛿𝛿2 3.194*** 3.234*** 4.181*** 4.233***
(0.297) p-value < 0.001 (0.289) (0.291) (0.232) (0.228)
Slightly important (𝛿𝛿̂2 ) 1.552*** 𝐻𝐻0 : 𝛿𝛿2 ≤ 𝛿𝛿1 1.492*** 1.542*** 2.790*** 2.892***
(0.493) p-value = 0.869 (0.476) (0.474) (0.378) (0.374)
Not important at all (𝛿𝛿̂1 ) 2.797*** 2.659*** 2.505*** 4.106*** 4.122***
(0.993) (0.969) (0.949) (0.762) (0.733)

Party Identification
Strong Democrat -2.608*** -2.588*** -2.639***
(0.206) (0.202) (0.197)
Not very strong Democrat -1.507*** -1.529*** -1.526***
(0.163) (0.159) (0.156)
Leaning Democrat -1.632*** -1.583*** -1.553***
(0.179) (0.174) (0.172)
Leaning Republican 1.167*** 1.140*** 1.217***
(0.177) (0.171) (0.167)
Not very strong Republican 1.526*** 1.403*** 1.569***
(0.171) (0.165) (0.161)
Strong Republican 2.628*** 2.624*** 2.766***
(0.234) (0.230) (0.225)

Demographics Yes Yes No Yes No

Election Fixed Effects Yes No No No No
Observations 6,273 6,273 6,273 6,273 6,273
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

23
 Table 7: Issue Importance in Two Levels

Dependent Variable: Vote Republican

Variables Cutoff 𝜓𝜓 = 3 𝜓𝜓 = 4 𝜓𝜓 = 5

More Important (𝛿𝛿̂𝐻𝐻 ) 4.037*** 4.298*** 4.946***

(0.166) (0.204) (0.345)
Less Important (𝛿𝛿̂𝐿𝐿 ) 1.802*** 2.944*** 3.411***
(0.440) (0.241) (0.172)
𝐻𝐻0: 𝛿𝛿𝐻𝐻 ≤ 𝛿𝛿𝐿𝐿. p-value <0.0001 <0.0001 <0.0001

Party Identification
Strong Democrat -2.599*** -2.610*** -2.625***
(0.205) (0.205) (0.205)
Not very strong Democrat -1.508*** -1.504*** -1.501***
(0.163) (0.162) (0.162)
Leaning Democrat -1.636*** -1.638*** -1.649***
(0.179) (0.178) (0.178)
Leaning Republican 1.180*** 1.186*** 1.172***
(0.177) (0.177) (0.177)
Not very strong Republican 1.540*** 1.520*** 1.511***
(0.171) (0.171) (0.171)
Strong Republican 2.634*** 2.604*** 2.602***
(0.233) (0.233) (0.232)

Demographics Yes Yes Yes

Election Fixed Effects Yes Yes Yes
Observations 6,273 6,273 6,273
Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

24
 Table 8: Marginal Effects

Dependent Variable: Vote Republican

5 Levels 2 Levels
Variable Cutoff 𝜓𝜓 = 3 𝜓𝜓 = 4 𝜓𝜓 = 5

Extremely important 0.329***

(0.0220)
Very important 0.260***
(0.0164)
Moderately important 0.230***
(0.0191)
Slightly important 0.103***
(0.0328)
Not important at all 0.186***
(0.0660)
More Important 0.270*** 0.288*** 0.331***
(0.00936) (0.0121) (0.0221)
Less Important 0.120*** 0.197*** 0.229***
(0.0293) (0.0155) (0.0103)

Party Identification
Strong Democrat -0.293*** -0.294*** -0.296*** -0.297***
(0.0218) (0.0219) (0.0219) (0.0218)
Not very strong Democrat -0.187*** -0.188*** -0.188*** -0.187***
(0.0204) (0.0205) (0.0205) (0.0205)
Leaning Democrat -0.200*** -0.202*** -0.203*** -0.204***
(0.0218) (0.0219) (0.0218) (0.0218)
Leaning Republican 0.156*** 0.159*** 0.160*** 0.157***
(0.0238) (0.0240) (0.0239) (0.0239)
Not very strong Republican 0.202*** 0.206*** 0.203*** 0.201***
(0.0228) (0.0230) (0.0229) (0.0229)
Strong Republican 0.330*** 0.335*** 0.331*** 0.330***
(0.0275) (0.0277) (0.0278) (0.0277)

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

25
REFERENCES

Achen, Christopher H., and Larry M. Bartels. 2017. Democracy for Realists: Why Elections Do Not
Produce Responsive Government. Princeton Studies in Political Behavior, Princeton University Press.

Ansolabehere, Stephen, and Philip Edward Jones. 2010. “Constituents’ Responses to Congressional
Roll-Call Voting.” American Journal of Political Science, 54(3): 583–597.

Black, Duncan. 1948. “On the Rationale of Group Decision-making.” Journal of Political Economy, 56(1):
23–34.

Bélanger, Éric, and Bonnie Meguid. 2008. “Issue salience, issue ownership, and issue-based vote choice.”
Electoral Studies, 27: 477–491.

Bolsen, Toby, and Thomas J. Leeper. 2013. “Self-Interest and Attention to News Among Issue Publics.”
Political Communication, 30(3): 329–348.

Carpini, Michael X. Delli, and Scott Keeter. 1996. What Americans Know about Politics and Why It
Matters. Yale University Press.

Converse, Philip E. 1964. “The Nature of Belief Systems in Mass Publics.” In Ideology and Discontent,
ed. David E. Apter, 206–261. New York: The Free Press of Glencoe.

Converse, Philip E., and Gregory B. Markus. 1979. “Plus ça change...: The New CPS Election Study
Panel.” American Political Science Review, 73(1): 32–49.

Davis, Otto A., Melvin J. Hinich, and Peter C. Ordeshook. 1970. “An Expository Development of a
Mathematical Model of the Electoral Process.” American Political Science Review, 64(2): 426–448.

Downs, Anthony. 1957. An Economic Theory of Democracy. Harper & Row.

Fournier, Patrick, André Blais, Richard Nadeau, Elisabeth Gidengil, and Neil Nevitte. 2003. “Issue
Importance and Performance Voting.” Political Behavior, 25(1): 51–67.

Grynaviski, Jeffrey D., and Bryce E. Corrigan. 2006. “Specification Issues in Proximity Models of
Candidate Evaluation (with Issue Importance).” Political Analysis, 14(4): 393–420.

Henderson, Michael. 2014. “Issue Publics, Campaigns, and Political Knowledge.” Political Behavior,
36(3): 631–657.

Holbrook, Allyson L., David Sterrett, Timothy P. Johnson, and Maria Krysan. 2016. “Racial
Disparities in Political Participation Across Issues: The Role of Issue-Specific Motivators.” Political
Behavior, 38(1): 1–32.

Holbrook, Allyson L., Matthew K. Berent, Jon A. Krosnick, Penny S. Visser, and David S. Boninger.
2005. “Attitude Importance and the Accumulation of Attitude-Relevant Knowledge in Memory.” Journal of
Personality and Social Psychology, 88: 749–769.

26
Hotelling, Harold. 1929. “Stability in Competition.” The Economic Journal, 39(153): 41–57.

Hutchings, Vincent L. 2001. “Political Context, Issue Salience, and Selective Attentiveness: Constituent
Knowledge of the Clarence Thomas Confirmation Vote.” Journal of Politics, 63(3): 846–868.

Iyengar, Shanto, Kyu S. Hahn, Jon A. Krosnick, and John Walker. 2008. “Selective Exposure to
Campaign Communication: The Role of Anticipated Agreement and Issue Public Membership.” The Journal
of Politics, 70(1): 186–200.

Jenke, Libby, and Michael Munger. Forthcoming. “The Significance of Salience in Candidate Evaluation: Is
Mouse Tracking a Better Measure?” Public Choice.

Jessee, Stephen. 2016. “(How) Can We Estimate the Ideology of Citizens and Political Elites on the Same
Scale?” American Journal of Political Science, 60(4): 1108–1124.

Jessee, Stephen A. 2009. “Spatial Voting in the 2004 Presidential Election.” American Political Science
Review, 103(1): 59–81.

Krosnick, Jon A. 1988a. “Attitude importance and attitude change.” Journal of Experimental Social
Psychology, 24(3): 240 – 255.

Krosnick, Jon A. 1988b. “The role of attitude importance in social evaluation: A study of policy preferences,
presidential candidate evaluations, and voting behavior.” Journal of Personality and Social Psychology,
55(2): 196–210.

Krosnick, Jon A. 1990. “Government Policy and Citizen Passion: A Study of Issue Publics in Contemporary
America.” Political Behavior, 12(1): 59–92.

Lavine, Howard, Eugene Borgida, and John L. Sullivan. 2000. “On the Relationship Between Attitude
Involvement and Attitude Accessibility: Toward a Cognitive-Motivational Model of Political Information
Processing.” Political Psychology, 21(1): 81–106.

Lazarsfeld, Paul Felix, Bernard Berelson, and Hazel Gaudet. 1968. The People’s Choice: How the Voter
Makes Up His Mind in a Presidential Campaign. A Columbia paperback, 83. 3 ed., Columbia University
Press.

Lecheler, Sophie, Claes de Vreese, and Rune Slothuus. 2009. “Issue Importance as a Moderator of Framing
Effects.” Communication Research, 36(3): 400–425.

Leeper, Thomas J., and Joshua Robison. 2018. “More Important, but for What Exactly? The Insignificant
Role of Subjective Issue Importance in Vote Decisions.” Political Behavior.

Lewis, Jeffrey B., and Gary King. 1999. “No Evidence on Directional vs. Proximity Voting.” Political
Analysis, 8(1): 21–33.

27
Malhotra, Neil, and Alexander Tahk. 2011. “Specification Issues in Assessing the Moderating Role of
Issue Importance: A Comment on Grynaviski and Corrigan (2006).” Political Analysis, 19(3): 342–350.

Merrill, Samuel. 1995. “Discriminating between the directional and proximity spatial models of electoral
competition.” Electoral Studies, 14(3): 273–287.

Niemi, Richard G., and Larry M. Bartels. 1985. “New Measures of Issue Salience: An Evaluation.” The
Journal of Politics, 47(4): 1212–1220.

Page, Benjamin I., and Calvin C. Jones. 1979. “Reciprocal Effects of Policy Preferences, Party Loyalties
and the Vote.” The American Political Science Review, 73(4): 1071–1089.

Price, Vincent, Clarissa David, Brian Goldthorpe, Marci McCoy Roth, and Joseph N. Cappella. 2006.
“Locating the Issue Public: The Multi-Dimensional Nature of Engagement with Health Care Reform.”
Political Behavior, 28(1): 33–63.

Rabinowitz, George, and Stuart Elaine Macdonald. 1989. “A Directional Theory of Issue Voting.” The
American Political Science Review, 83(1): 93–121.

The American National Election Studies, 1996-2008. 2018. These materials are based on work supported by
the National Science Foundation under grant numbers SES 1444721, 2014-2017, the University of
Michigan, and Stanford University.

Tomz, Michael, and Robert P. Van Houweling. 2008. “Candidate Positioning and Voter Choice.” Ameri-
can Political Science Review, 102(3): 303–318.

Visser, Penny S., Jon A. Krosnick, and Joseph P. Simmons. 2003. “Distinguishing the cognitive and
behavioral consequences of attitude importance and certainty: A new approach to testing the common- factor
hypothesis.” Journal of Experimental Social Psychology, 39(2): 118–141.

Wilson, J. Matthew, and Paul Gronke. 2000. “Concordance and Projection in Citizen Perceptions of
Congressional Roll-Call Voting.” Legislative Studies Quarterly, 25(3): 445–467.

28
APPENDIX

A. Demographic and Fixed Effects of Primary Analysis

In Tables 9 and 10, we report the coefficients of the demographic characteristics and election fixed effects

from our primary analysis. The election fixed effects, when combined with the intercept, represent the average

Republican advantage over the Democratic candidates.

[Tables 9 and 10 approximately here]

B. Alternative Distance Measures

In this section, we present the results from re-estimating our model with two different distance measures:

Relative Issue Distance and Subjective Issue Distance. Relative Issue Distance is computed as followed:

𝑅𝑅𝑅𝑅𝑅𝑅
𝑥𝑥𝑖𝑖𝑖𝑖 = �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦�𝐷𝐷𝐷𝐷 � − �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦�𝑅𝑅𝑅𝑅 �

where 𝑟𝑟𝑖𝑖𝑖𝑖 is the respondent 𝑖𝑖’s position for policy 𝑗𝑗. 𝑦𝑦�𝑘𝑘𝑘𝑘 is the average perceived position of candidate 𝑘𝑘 =

𝐷𝐷, 𝑅𝑅 for all respondents in policy 𝑗𝑗. Leeper and Robison (2018), Krosnick (1988b), and others use this measure

in their analyses. Subjective Issue Distance replaces the average candidate positions with the respondents’

subjective perceived positions:

𝑥𝑥𝑖𝑖𝑖𝑖 = �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 � − �𝑟𝑟𝑖𝑖𝑖𝑖 − 𝑦𝑦𝑖𝑖𝑖𝑖𝑖𝑖 �.

Subjective Issue Distance reflects more directly the respondents’ perceived positions than Relative Issue

Distance does. It also captures more variation than the relative closeness measure that we propose. However,

as we argue in the paper, the additional variation does not necessarily carry additional information as

differences between ordinal values are not necessarily comparable between individuals and issues. Therefore,

we do not favor either of these measures in our study. Squared Euclidean distance is another standard measure

of distance. Lewis and King (1999) find that they “rarely differ empirically by very much.” We include the
29
following analyses to show that our methods do not depend on our proposed measurements. With both

measures, we rescale the distances and normalize them by the number of questions for the survey years so

that all measurements are between -1 (closest to the Democrat) and 1 (closest to the Republican) with 0 being

an equal distance to both candidates.

In Table 11, we report the estimates of the five levels of issue importance using Relative Issue Distance

and their corresponding marginal effects. The pattern on the estimated importance is similar to that in our

main results. The standard errors are larger, but the coefficients are still precisely estimated. One key

difference is the coefficient of the issues that are Not Important At All is much larger. It might be one of the

reasons that current studies do not find a significant difference between different reported issue importance.

With a much larger estimate for issues that are not important at all, it makes it difficult to distinguish the

average weights between more and less important issues as presented in Table 12. We do not find the weights

to be statistically different when the cutoff is at the Moderately Important (𝜓𝜓 = 3) issues, but statistically

different at all other cutoffs.

[Tables 11 and 12 approximately here]

In Tables 13 and 14, we present the results of the same analyses above with Subjective Issue Distance.

The pattern of the coefficients follows our main results. The coefficient of the Not Important At All is much

smaller relative to that of using other distance measures. It is approximately the same as that of the Slightly

Important issues. Therefore, when we estimate the weights for two levels of importance, we find the

differences to be statistically different at any conventional level for any cutoff.

[Tables 13 and 14 approximately here]

In conclusion, we find that the estimates for the least important issues are sensitive to the measurement of

distance. However, for the more important issues, we find the same pattern as our main results.

30
 Table 9: Main Results - Demographics and Fixed Effects

Dependent Variable: Vote Republican

Main Results Alternative Specifications
Variable (1) (2) (3) (4) (5)

Policy Weights and Party Identification Coefficients in Table 6.

Male -0.0319 -0.0310 0.0858

(0.101) (0.0988) (0.0787)
Age 0.00379 0.00338 -0.00115
(0.00299) (0.00293) (0.00228)
Non-White -1.015*** -0.868*** -1.489***
(0.133) (0.130) (0.107)
Some College 0.130 0.204* 0.462***
(0.120) (0.117) (0.0943)
College Grad 0.0169 0.0919 0.538***
(0.132) (0.128) (0.101)
Low Income -0.0991 -0.0668 -0.215**
(0.131) (0.128) (0.105)
High Income 0.0575 0.0213 0.185*
(0.129) (0.127) (0.0982)
Year = 2000 1.371***
(0.154)
Year = 2004 1.632***
(0.172)
Year = 2008 1.060***
(0.143)
Constant -1.169*** -0.166 -0.151 -0.336** -0.340***
(0.238) (0.208) (0.116) (0.146) (0.0368)

Observations 6,273 6,273 6,273 6,273 6,273

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Note: Model (1) represents the full model with election fixed effects. (2) and (3) contain party
identification, but vary by demographic controls. (4) and (5) have no party identification, but
vary by demographic controls.

31
 Table 10: Main Results (Two-Level) - Demographic and Fixed Effects
Dependent Variable: Vote Republican
Variable Cutoff 𝜓𝜓 = 3 𝜓𝜓 = 4 𝜓𝜓 = 5

Policy Weights and Party Identification Coefficients in Table 7.

Male -0.00763 -0.0213 -0.0261

(0.101) (0.101) (0.101)
Age 0.00382 0.00391 0.00366
(0.00298) (0.00298) (0.00298)
Non-White -1.011*** -1.009*** -1.020***
(0.132) (0.132) (0.132)
Some College 0.118 0.123 0.128
(0.120) (0.120) (0.120)
College Grad -0.00140 0.0240 0.0280
(0.132) (0.131) (0.131)
Low Income -0.0898 -0.0776 -0.0868
(0.130) (0.130) (0.130)
High Income 0.0639 0.0667 0.0630
(0.129) (0.129) (0.129)
Year = 2000 1.373*** 1.360*** 1.358***
(0.154) (0.153) (0.153)
Year = 2004 1.653*** 1.620*** 1.620***
(0.172) (0.171) (0.171)
Year = 2008 1.065*** 1.045*** 1.039***
(0.143) (0.143) (0.143)
Constant -1.184*** -1.175*** -1.146***
(0.238) (0.237) (0.237)

Observations 6,273 6,273 6,273

Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

32
 Table 11: Issue Importance in Five Levels with Relative Issue Distance
Variable Vote Republican Test Margins

Extremely important (𝛿𝛿̂5 ) 6.588*** 𝐻𝐻0 : 𝛿𝛿5 ≤ 𝛿𝛿4 0.552***

(0.748) p-value = 0.0627 (0.0621)
Very important (𝛿𝛿̂4 ) 5.105*** 𝐻𝐻0 : 𝛿𝛿4 ≤ 𝛿𝛿3 0.428***
(0.609) p-value = 0.0127 (0.0505)
Moderately important (𝛿𝛿̂3 ) 3.009*** 𝐻𝐻0 : 𝛿𝛿3 ≤ 𝛿𝛿2 0.252***
(0.695) p-value = 0.430 (0.0581)
Slightly important (𝛿𝛿̂2 ) 2.755** 𝐻𝐻0 : 𝛿𝛿2 ≤ 𝛿𝛿1 0.231**
(1.249) p-value = 0.967 (0.105)
Not important at all (𝛿𝛿̂1 ) 6.680*** 0.560***
(1.773) (0.148)

Party Identification
Strong Democrat -2.973*** -0.387***
(0.188) (0.0233)
Not very strong Democrat -1.603*** -0.273***
(0.145) (0.0251)
Leaning Democrat -1.790*** -0.295***
(0.160) (0.0257)
Learning Republican 1.480*** 0.287***
(0.154) (0.0289)
Not very strong Republican 1.703*** 0.323***
(0.151) (0.0277)
Strong Republican 3.142*** 0.479***
(0.212) (0.0255)

Demographics Yes
Election Fixed Effects Yes
Observations 6,273
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

33
 Table 12: Issue Importance in Two Levels with Relative Issue Distance
Dependent Variable: Vote Republican
Variable Cutoff 𝜓𝜓 = 3 Margins 𝜓𝜓 = 4 Margins 𝜓𝜓 = 5 Margins

More Important (𝛿𝛿̂𝐻𝐻 ) 4.858*** 0.409*** 5.698*** 0.479*** 6.584*** 0.553***

(0.380) (0.0313) (0.469) (0.0386) (0.747) (0.0622)
Less Important (𝛿𝛿̂𝐿𝐿 ) 4.042*** 0.340*** 3.330*** 0.280*** 4.168*** 0.350***
(1.040) (0.0874) (0.575) (0.0481) (0.413) (0.0342)
𝐻𝐻0 : 𝛿𝛿𝐻𝐻 ≤ 𝛿𝛿𝐿𝐿 . p-value 0.231 0.000848 0.00252

Party Identification
Strong Democrat -2.982*** -0.389*** -2.968*** -0.387*** -2.971*** -0.388***
(0.188) (0.0234) (0.188) (0.0233) (0.188) (0.0233)
Not very strong Democrat -1.611*** -0.276*** -1.603*** -0.274*** -1.600*** -0.273***
(0.144) (0.0252) (0.145) (0.0251) (0.144) (0.0251)
Leaning Democrat -1.795*** -0.298*** -1.785*** -0.295*** -1.793*** -0.296***
(0.160) (0.0258) (0.160) (0.0258) (0.160) (0.0257)
Learning Republican 1.486*** 0.291*** 1.481*** 0.288*** 1.481*** 0.288***
(0.154) (0.0290) (0.154) (0.0289) (0.154) (0.0289)
Not very strong Republican 1.694*** 0.324*** 1.697*** 0.323*** 1.698*** 0.323***
(0.150) (0.0279) (0.150) (0.0278) (0.150) (0.0278)
Strong Republican 3.153*** 0.482*** 3.141*** 0.480*** 3.149*** 0.481***
(0.212) (0.0255) (0.212) (0.0255) (0.212) (0.0255)

Demographics Yes Yes Yes

Election Fixed Effects Yes Yes Yes
Observations 6,273 6,273 6,273
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

34
 Table 13: Issue Importance in Five Levels with Subjective Issue Distance
Variable Vote Republican Test Margins

Extremely important (𝛿𝛿̂5 ) 9.198*** 𝐻𝐻0 : 𝛿𝛿5 ≤ 𝛿𝛿4 0.620***

(0.647) p-value = 0.0359 (0.0416)
Very important (𝛿𝛿̂4 ) 7.706*** 𝐻𝐻0 : 𝛿𝛿4 ≤ 𝛿𝛿3 0.520***
(0.514) p-value = 0.00821 (0.0326)
Moderately important (𝛿𝛿̂3 ) 5.794*** 𝐻𝐻0 : 𝛿𝛿3 ≤ 𝛿𝛿2 0.391***
(0.590) p-value = 0.139 (0.0388)
Slightly important (𝛿𝛿̂2 ) 4.469*** 𝐻𝐻0 : 𝛿𝛿2 ≤ 𝛿𝛿1 0.301***
(1.057) p-value = 0.540 (0.0710)
Not important at all (𝛿𝛿̂1 ) 4.666*** 0.315***
(1.671) (0.113)

Party Identification
Strong Democrat -2.549*** -0.282***
(0.208) (0.0215)
Not very strong Democrat -1.433*** -0.177***
(0.162) (0.0203)
Leaning Democrat -1.524*** -0.187***
(0.178) (0.0217)
Learning Republican 1.347*** 0.179***
(0.177) (0.0233)
Not very strong Republican 1.590*** 0.209***
(0.170) (0.0224)
Strong Republican 2.756*** 0.339***
(0.237) (0.0266)

Demographics Yes
Election Fixed Effects Yes
Observations 6,273 6,273
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

35
 Table 14: Issue Importance in Two Levels with Subjective Issue Distance
Dependent Variable: Vote Republican
Variable Cutoff 𝜓𝜓 = 3 Margins 𝜓𝜓 = 4 Margins 𝜓𝜓 = 5 Margins

More Important (𝛿𝛿̂𝐻𝐻 ) 7.560*** 0.512*** 8.309*** 0.561*** 9.197*** 0.623***

(0.323) (0.0187) (0.398) (0.0238) (0.642) (0.0414)
Less Important (𝛿𝛿̂𝐿𝐿 ) 4.504*** 0.305*** 5.397*** 0.364*** 6.542*** 0.443***
(0.892) (0.0601) (0.482) (0.0316) (0.343) (0.0211)
𝐻𝐻0 : 𝛿𝛿𝐻𝐻 ≤ 𝛿𝛿𝐿𝐿 . p-value <0.001 <0.0001 <0.001

Party Identification
Strong Democrat -2.558*** -0.285*** -2.553*** -0.283*** -2.571*** -0.286***
(0.208) (0.0216) (0.208) (0.0215) (0.208) (0.0215)
Not very strong Democrat -1.439*** -0.179*** -1.437*** -0.178*** -1.441*** -0.179***
(0.162) (0.0204) (0.162) (0.0203) (0.162) (0.0204)
Leaning Democrat -1.527*** -0.189*** -1.521*** -0.187*** -1.542*** -0.190***
(0.177) (0.0218) (0.177) (0.0217) (0.178) (0.0218)
Learning Republican 1.330*** 0.178*** 1.354*** 0.180*** 1.334*** 0.178***
(0.176) (0.0235) (0.177) (0.0234) (0.176) (0.0234)
Not very strong Republican 1.590*** 0.211*** 1.591*** 0.210*** 1.572*** 0.208***
(0.170) (0.0226) (0.170) (0.0224) (0.169) (0.0225)
Strong Republican 2.743*** 0.342*** 2.748*** 0.340*** 2.734*** 0.339***
(0.235) (0.0269) (0.236) (0.0268) (0.235) (0.0267)

Demographics Yes Yes Yes

Election Fixed Effects Yes Yes Yes
Observations 6,273 6,273 6,273
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

36

View publication stats