“But when we get right straight down to... all the various ramifications of the public service for our own individual lives, what difference does it make whether the men who do the work are Republicans or Democrats; whether they are high tariff or low tariff? We want the men who will do the work well and honestly.”

— Major Henry T. Lee advocating the adoption of nonpartisan municipal elections to a meeting of Good Government organizations, Los Angeles, California, 1909

Introduction

At the turn of the 20th century, Progressives advanced a set of reforms designed to erode the strength of parties and political machines in local politics (see e.g., Welch and Bledsoe 1988). Along with the institutions of city manager and at-large elections, nonpartisan electoral rules were a crucial component of the Progressives’ strategy. Reformers argued that the raison d’être of municipal government is the provision of essential services, which requires technical expertise rather than partisan fealty. Advocates of the reform movement espoused the benefits of business-like efficiency and a universalist approach to governance in contrast to the waste, corruption, and particularistic benefits associated with machine politics (Welch and Bledsoe 1988; Bridges 1997). The movement was largely successful: the vast majority of municipal governments in the US still feature such reform institutions (Moulder 2008).

Nonpartisan elections—among the most prevalent and durable of the Progressive-era municipal reforms—operate on the logic that voters will choose different candidates depending on the presence or absence of partisan information.Footnote 1 Advocates of this institutional reform clearly thought that victorious candidates would not only be different, but be better along some dimension, be it ideological leaning or fitness for the job. The first question we seek to answer is, were the Progressive-era reformers correct? Do different types of candidates win when party labels are removed? Secondly, how do winning candidates differ under the two regimes? In the absence of party labels, do voters rely on identity politics, or do they give greater weight to other aspects of candidate biography such as previous government or private sector experience?

Previous scholarship indicates that the answer to the first question is likely to be yes. Partisan ballots provide voters a powerful, low-cost information shortcut (Popkin 1991; Rahn 1993). Knowing nothing more than party labels, voters can infer candidates’ ideology and issue positions with some degree of certainty. Removing these labels may induce voters to rely on alternative heuristics, such as race or ethnicity (Pomper 1966; Bullock 1984; Bullock and Campbell 1984; Squire and Smith 1988), to infer candidates’ partisanship or ideology. The institution of nonpartisan elections may change the kinds of candidates who win election simply by altering the relative cost and probative value of different heuristics.

Granting that nonpartisan elections do change who wins, what sorts of candidates are likely to be advantaged by the absence of party cues? Nonpartisan ballots may induce voters to give greater weight to other characteristics beyond partisanship and ideology such as markers of competence or quality (Schaffner et al. 2001; Lim and Snyder 2015). One important dimension of candidate quality is political experience. Incumbents and prior office holders have electoral advantages in partisan legislative elections at both the national and state levels (Jacobson and Kernell 1983; Jewell and Breaux 1988; Jacobson 1997; Lee 2008). Several studies find that incumbency advantages extend further down ballot to mayors (Ferreira and Gyourko 2009; Ferreira and Gyourko 2014) and city council members (Trounstine 2011). These studies indicate that voters may rely on an incumbency heuristic when evaluating candidates; a remaining open question is whether this reliance itself depends on electoral rules.

Studying the effects of nonpartisan elections is difficult for two main reasons. The first is common to settings in which the researcher does not manipulate the causal variable of interest directly. We cannot be sure whether the observed differences in election outcomes for partisan and nonpartisan elections are due to the causal effect of the electoral institution itself or some other feature of the electoral context. It may be that localities that opt to institute nonpartisan elections place a higher value on leaders’ political experience than those with partisan elections due to the tastes and preferences of the local electorate, not the electoral rules. Statistical fixes for this problem such as multiple regression or matching only help if we are willing to assume that after conditioning on a set of observable characteristics of elections, the electoral institution is “as-if” randomly assigned. The plausibility of such an assumption varies from context to context. Even in localities that hold nonpartisan mayoral elections but partisan congressional elections, one may still be worried that the electoral rules are nonrandomly applied.

The second challenge is that elections that are nonpartisan in theory are not necessarily nonpartisan in practice (Adrian 1959). Candidates’ party affiliations may be widely known to voters or revealed during the campaign. Even in nominally nonpartisan elections, partisanship remains a systematic predictor of voters’ preferences when information about candidates’ party or ideology is readily available (Squire and Smith 1988; Schaffner et al. 2001). It is not clear how to measure and account for these complexities. For example, if we were to find that in nonpartisan elections, experienced candidates are more likely to be elected, this result might actually reflect the subtle dynamics of candidates’ decisions to compete in elections where candidates’ partisanship is not advertised but nevertheless common knowledge. In such a scenario, liberal candidates in conservative districts might be systematically less experienced because the high-quality, liberal would-be candidates, knowing they have a poor chance of victory, pursue careers outside of politics or run in other constituencies.

Our experimental design cleanly sidesteps both problems. In two separate implementations of the same design, we invite subjects to participate in a series of “elections” that are conducted as conjoint survey experiments. In each election, subjects see the profiles of two candidates and must choose between them. Unlike standard conjoint experiments that only randomize the levels of a fixed set of attributes, we randomize in addition whether the partisanship attribute itself is shown to respondents. This design enables us to determine the effect of partisan information, not just on candidate choice, but also on how subjects use the other attributes to evaluate candidates. Unlike some nonpartisan elections outside the survey environment, our nonpartisan elections are unambiguously devoid of partisan information.

Survey experiments are sometimes described as being high on internal validity but lower on external validity (Mutz 2011). We are sympathetic to this critique and think it is important to distinguish between two kinds of external validity. The first concerns the extent to which an experiment conducted on one sample would generalize to the same experiment being conducted on a different sample. We directly show that our experiment exhibits this kind of external validity by conducting it on both Amazon’s Mechanical Turk and on a nationally representative sample administered by YouGov.Footnote 2 The second, and in our view more important, sort of external validity concerns the extent to which the causal processes at work in the survey environment map on to the political phenomena we wish to investigate in the real world. Here we rely on an analogy between the survey environment and the voting booth. Our survey respondents have to choose between two hypothetical candidates on the basis of the five or six pieces of information we provide. Especially for down-ballot offices, voters may have to choose between two candidates on the basis of information provided on the ballot itself. In addition to candidates’ names (from which gender, race, and ethnicity can be imperfectly inferred), some ballots provide incumbency, occupation, or partisanship information (McDermott 1998, 2005). In this way, electoral choice is not so different from a survey response. The survey experimental context is of course very different from voting – the stakes are much lower and responding to survey questions is far less meaningful than casting ballots. In this case, however, we are willing to trade a decrease in verisimilitude for an increase in our ability to directly manipulate the information environment.

To preview our results, we see a clear difference across electoral institutions in how subjects use information about candidate experience in both samples. In nonpartisan elections, we find that the effect of candidate experience on vote choice is approximately 10 percentage points higher than in partisan elections. We find that withholding partisan information has different effects for different subjects. In particular, Republicans rely more on career experience whereas Democrats turn to political experience when candidates’ partisanship is unknown.

Voting and Information Shortcuts

Since the early voting studies of the Columbia and Michigan schools, political scientists have consistently documented uneven and generally low levels of political knowledge and interest among voters (e.g., Lazarsfeld et al. 1944; Campbell et al. 1960; Converse 1964; Delli Carpini and Keeter 1996). Although most voters tend to know very little about political candidates and their policy positions, information shortcuts or heuristics can guide political decision-making (e.g., Downs 1957; Popkin 1991). Indeed, some evidence suggests that heuristics can enable low-information voters to make nearly the same choices they would make if they were fully informed (Althaus 2003; Lupia 1994). Voters may rely on any number of shortcuts, including retrospective evaluations, endorsements, or candidates’ personal characteristics, but party identification tends to be the most potent heuristic (Rahn 1993). Party identification is a “shortcut or default value, a substitute for more complete information about parties and candidates” (Popkin 1991, p. 14). A party label generally provides a reliable proxy for candidates’ ideology and issue positions. For voters, party identification also appears to be a stable and enduring attachment, akin to other social identities such as ethnicity, religion, or class (Campbell et al. 1960; Green et al. 2002).

Voters may evaluate candidates using a likability heuristic that relies on their affect toward politically salient groups (Brady and Sniderman 1985). If voters use this information shortcut, their perceptions of candidates’ ideological positions would reflect their own beliefs weighted by their feelings toward opposing groups. For example, Brady and Sniderman (1985) find that on average, conservative survey respondents dislike liberals more intensely than liberals dislike conservatives, and they argue conservative respondents consequently overestimate the ideological distance between the two groups. Another possibility is that voters rely on a representativeness or goodness-of-fit heuristic by making inferences about candidates based on how well they represent a given group or type (Tversky and Kahneman 1974; Popkin 1991). Carnes and Sadin (2015), for instance, argue that a representativeness heuristic leads subjects to mistakenly infer that candidates from working class families are more liberal on economic policy than candidates from affluent backgrounds.

Among studies of nonpartisan elections, a common finding is that voters rely on party cues when they can and look to other information shortcuts, such as race or incumbency, when necessary. Prior research suggests that characteristics such as race (Brady and Sniderman 1985; McDermott 1998), gender (Huddy and Terkildsen 1993; McDermott 1998), and class (Sadin 2014) also influence perceptions of candidates. Women and African-American candidates are seen as more liberal and more Democratic than white men (Huddy and Terkildsen 1993; McDermott 1998). Using occupation as a proxy for social class in a survey experiment, Sadin (2014) finds that respondents rate upper class candidates as more competent relative to either working-class candidates or candidates whose social class is unknown.

Information Shortcuts in Nonpartisan Elections

Our expectation is that in nonpartisan elections voters will rely less heavily on partisan heuristics to choose between candidates. The empirical record to date generally supports the expectation that partisanship and vote choice should be less strongly associated in nonpartisan elections. For example, Pomper (1966) analyzes ward-level election results in Newark, New Jersey and finds vote shares for candidates of the same party are highly correlated in partisan state legislative elections but not in nonpartisan municipal elections. In a recent study, Lim and Snyder (2015) find strong correlations (0.88–0.99) between the Democratic vote share for state judges and the Democratic “normal vote” in partisan elections. A different pattern emerges in nonpartisan elections, where judges’ vote shares are less strongly correlated with their co-partisans’. Schaffner et al. (2001) also find a systematic relationship between partisanship and Democratic vote share in partisan contests, but partisanship is not a statistically significant predictor of the vote in most of the nonpartisan elections they analyze.

Even when nonpartisan rules make candidates’ party affiliations difficult or costly to uncover, some evidence suggests that voters may try to infer party from other information. In a study of nonpartisan judicial retention elections in California, Squire and Smith (1988) leverage a pre-election survey that provided a random subset of respondents with the name of the governor who appointed each judge. Treatment group respondents were more likely to support retaining judges appointed by copartisan governors. Recent experimental evidence also indicates that voters may infer candidates’ party affiliations from issue positions. Bonneau and Cann (2015) provide descriptions of hypothetical candidates for state supreme court, with a random subset of subjects receiving party cues. Descriptions of the Republican candidate, for example, highlight support for the death penalty and a commitment to traditional family values while Democrats are described as advocates of same-sex marriage who believe the courts should take an active role in promoting equality. The experimental results show a strong link between partisanship and vote choice even in the absence of an explicit party cue.

Nonpartisan elections are not devoid of ideological or partisan content, but because voters do not have access to the partisan shortcut, such information is relatively more costly to acquire. We therefore predict that in nonpartisan elections, voters will be more likely to resort to other cues. These cues may include indicators of fitness for the job, including private sector and political experience (Schaffner et al. 2001; Lim and Snyder 2015).

A first glance at this prediction comes from the historical record of 1010 US mayoral elections held between 1945 and 2007. This dataset covers 225 unique municipalities and records background characteristics of winners and runners-up, contextual information such as electoral rules, and electoral outcomes. We constructed this dataset principally by examining the newspaper accounts of candidates and electoral outcomes.Footnote 3

Table 1 shows the political experience of the winning candidate in each election, subset according to whether the election was nominally nonpartisan or partisan. The statistically significant \(\chi ^2\) statistic indicates that the political experience of winning candidates differs across partisan and nonpartisan elections. In partisan elections, 30% of winning candidates have no previous political experience whereas in nonpartisan elections, the share of inexperienced winning candidates drops by 9 percentage points to 21%. That fewer inexperienced candidates win in nonpartisan elections fits with our predictions.

The 9 percentage point difference might reflect the effect of nonpartisan ballots on who gets elected, but it could just as easily reflect other differences between cities that do and do not hold partisan elections. For example, larger cities are more likely to hold partisan contests, and in larger cities, candidates with nonpolitical experience (such as attorneys and business executives) may be more likely to run for election. Our own data bear this out. In partisan elections, 63% of candidates are attorneys or business executives while in nonpartisan elections the corresponding figure is 48%. Another potential confounder is that partisan organizations provide financial and institutional support, creating opportunities for politically inexperienced candidates. Observational analyses of the effects of election type are further complicated the concern mentioned above that some elections are nonpartisan in name only.

While we do see that winners in nonpartisan elections appear to have more political experience than winners in partisan elections, an alternative analysis of the historical record paints a murkier picture. Subsetting our dataset to only those elections in which candidates have different levels of political experience (779 elections), we see that the candidate with more experience wins about 62% of the time in partisan elections and 61% of the time in nonpartisan elections. This difference is not statistically significant (\(p = 0.791\)).

Table 1 Political Experience of Winning Candidate in Partisan and Nonpartisan US Mayoral Elections
Full size table

Conjoint Candidate Choice Survey Experiments

In an effort to combat the challenges outlined above, we have adopted the conjoint survey design, ideal for studying multidimensional preferences (Hainmueller et al. 2014). Within political science, conjoint experiments have been applied to the study of immigration preferences (Hainmueller and Hopkins 2015; Bansak et al. 2017), complex policy preferences (Bechtel et al. 2015), and (as in our case) candidate preference (Hainmueller et al. 2014; Carlson 2015; Franchino and Zucchini 2015). The conjoint design will also allow us to evaluate the separate impacts of a large set of causal factors on subjects’ preferences over candidates. While these experiments are artificial in the sense that they present subjects with an abstract choice, Hainmueller et al. (2015) show that conjoint experiments can produce externally valid estimates by comparing their experimental results to real-world outcomes.

In our studies, subjects judge five successive elections in which five (or six) attributes of two competing candidates are displayed: their race, gender, political experience, career experience, age, and in some cases, political party. The attributes of each candidate are fully randomized so that every possible candidate profile is equally likely. Figures 1 and 2 show examples of an election in which partisan information is available (1) and an election in which it is withheld (2). The possible levels of each attribute are displayed in Table 2. Some levels were added to the Political Experience and Career Experience attributes in the YouGov version of the study in order more fully account for the range of plausible biographies.

Fig. 1
figure 1

Experimental Stimuli: A Partisan Election

Full size image
Fig. 2
figure 2

Experimental Stimuli: A Nonpartisan Election

Full size image
Table 2 Attributes
Full size table

A great deal of the methodological literature on conjoint analysis is concerned with the selection of attributes and levels. Attributes should be independent of one another and levels should describe a wide range of possibilities (Green and Srinivasan 1978). A recurring question is how many attributes to include. The consensus seems to be that six or seven attributes is the limit. Above this limit, survey researchers caution that subjects may resort to cognitive shortcuts when evaluating profiles, causing two problems for inference. First, subjects may over-weight the first few attributes presented to them. Second, they may over-weight particularly salient attributes. We address the first problem by following the advice of (Hainmueller et al. 2014, p. 7) to randomize the order of the attributes.

The second problem is, in our view, a feature, not a bug, of our design. Candidates’ party is likely the most salient detail when subjects are choosing between profiles. By randomizing whether or not subjects are shown the party label, we can directly test whether the injection of partisanship into an election changes the impacts of the other attributes. Further, this design feature reflects the real-world variation in electoral institutions and is therefore our main experimental manipulation.

We conducted our experiment on a Mechanical Turk (MTurk) convenience sample and on a nationally representative sample constructed by YouGov. The demographic profile of the MTurk sample is quite different from that of the YouGov sample. On average, the MTurk sample is whiter, more male, more liberal, more Democratic, better educated, and younger.Footnote 4 In addition to these measured characteristics, the samples may differ on unobserved dimensions. Indeed, many social scientists are skeptical of MTurk samples because of these unmeasured dimensions (Goodman et al. 2013). Others (Berinsky et al. 2012; Mullinix et al. 2015; Coppock 2017) are optimistic that experimental results on MTurk can generalize to other populations but stress the need for careful consideration of the individual level moderators that might invalidate generalizing from one context to another. In our case, we believe that the most important moderator is respondents’ partisanship. Fortunately, MTurk offers sufficient numbers of both Democrats and Republicans to obtain relatively precise estimates for each group, even if MTurk partisans are not representative of partisans nationally.

We will limit our exploration of treatment effect heterogeneity to partisan differences only, for two reasons. First, because we randomized whether or not candidates’ partisanship is displayed to subjects, it is appropriate to test whether the effects of candidates’ partisanship are moderated by subjects’ own party affiliation. Second, we are concerned about the multiple comparisons complications we would encounter with additional subgroup analyses.

Analysis

Our main dependent variable is candidate choice, which is asked “Which of these two candidates do you prefer?” A second dependent variable, candidate competence, is asked “On a scale from 0 to 100, how competent do you think these candidates would be as mayor?” We will use this dependent variable to explore a possible mechanism by which candidate attributes and electoral contexts affect vote choice.

We will analyze the effects of our experimental manipulations on these dependent variables using two models, shown in Eqs. 1 and 2. The coefficient vectors \(\beta _1, \beta _2, ...\) and \(\alpha _1, \alpha _2 ...\) are each of length \(k-1\), where k refers to the total number of levels within an attribute. Individual-level idiosyncrasies in candidate preferences are captured by the error terms \(\epsilon\) and \(\eta\). The required assumption that the errors are independent of each other and of candidate attributes is justified by the experimental design. We will estimate Eq. 1 among the subset of elections that do not include party and Eq. 2 among the elections that do include party.

$$Y= \beta _0 + \beta _1Race + \beta _2Age + \beta _3Gender + \beta _4Political\_Exp + \beta _5Career\_Exp + \epsilon$$
(1)
$$Y= \alpha _0 + \alpha _1Race + \alpha _2Age + \alpha _3Gender + \alpha _4Political\_Exp + \alpha _5Career\_Exp + \alpha _6\_Party + \eta$$
(2)

Our experiment is motivated by the extent to which the party heuristic overwhelms the other factors contributing to candidate choice. Accordingly, we are especially interested in the differences between \(\beta _1, \beta _2, ...\) and \(\alpha _1, \alpha _2 ...\). We will estimate Eqs. 1 and 2 by ordinary least squares (OLS) with standard errors clustered by respondent. (Hainmueller et al. 2014, p. 15) show that this approach is asymptotically equivalent to their average marginal component effect (AMCE) estimator.Footnote 5 We will further condition the estimation on respondents’ own party identification, focusing on effects among Democrats versus Republicans including leaners.

We will test for the equality of the corresponding coefficients in Eqs. 1 and 2 by interacting the attributes with an indicator for election type in the full sample. We will test for the equality of coefficients between the Democrats and Republicans by interacting the treatment variables with an indicator for partisanship.

A short note on presentation: all together, these analyses will render a very large number of coefficient estimates. For this reason, we will present our results graphically using coefficient plots, in which attribute levels are placed on the vertical axis and point estimates with 95% confidence intervals are placed on the horizontal axis. For those who prefer tables, the corresponding regression output for each figure is presented in the appendix. We recognize that this presentation mode obscures some details while highlighting others—we have endeavored to maintain both clarity and transparency in our presentation choices.

Results

We will present three sets of results. First, we will examine the effects of candidate attributes, split by election type. Second, we will split our samples by respondent partisanship in order to examine the possibly heterogeneous effects of candidate attributes and election types. Third, we will examine a possible mechanism (perceptions of competence) by which election type may affect the attributes that voters favor.

Effects of Partisan Elections on Candidate Choice

Figure 3 presents the results of the MTurk study. In the first column, the estimates of Equation 1 are shown. The strongest effects are observed for the political experience attribute. Relative to a candidate with no political experience, respondents prefer candidates who are City Council Members, State Legislators, Mayors, or Representatives in Congress by a margin of 25–30 percentage points. Candidates who previously held a mayoral office were rewarded most for their political experience. By contrast, we observe relatively muted effects for the job experience, race, age, and gender attributes, although our respondents do express a mild preference for candidates who are female and nonwhite. Our respondents’ preferences for candidates varied non-monotonically with age: 45-year-olds are preferred to 35-year-olds and 55- and 65-year olds.

Fig. 3
figure 3

Mechanical Turk Main Analysis Dependent Variable: Candidate Preference

Full size image

In partisan elections, we observe a similar pattern, though the effects for the political experience variables are more muted. On average, our sample prefers independents to partisan candidates of either stripe, though this average masks some heterogeneity by respondent party identification, as we will explore in the next section.

The final column of Fig. 3 shows the difference between partisan and nonpartisan elections across the attributes they have in common. For job experience, race, age, and gender, the presence or absence of party labels makes no difference. However, we do observe statistically significantly different weight being given to the political experience variables, depending on election type. In nonpartisan elections, the effects of candidates’ political experience are approximately 10 percentage points larger than in partisan elections.

Figure 4 presents the identical analyses using the YouGov data. Overall, we observe a very similar pattern of results. In nonpartisan elections, political experience is heavily rewarded. We added the “School Board President” level to test the alternative explanation that respondents prefer any experience to “No Political Experience.” Indeed, respondents do prefer school board presidents to political neophytes, but higher offices are nevertheless preferred to school board presidents as well. In the YouGov sample, we observe a similar interaction between election type and the effects of political experience. Political experience matters more in nonpartisan elections.

Fig. 4
figure 4

YouGov Main Analysis Dependent Variable: Candidate Preference

Full size image

We added the “Stay-at-Home Dad/Mom” and “Electrician” levels to the job experience attribute.Footnote 6 Both of these careers were viewed negatively in both partisan and nonpartisan elections. We observe similarly small effects of gender and age in the YouGov sample as we did in the MTurk sample.

Heterogeneous Effects by Respondent Partisanship

In Figs. 5 and 6, we reproduce the main analyses, splitting the samples based on respondents’ own partisanship. As shown in the top center panel of each figure, Democrats dislike Republican candidates and Republicans dislike Democratic candidates. The differences in these preferences are large and statistically significant. Intriguingly, in both the MTurk and YouGov samples, partisans dislike the out-party (relative to an independent candidate) more than they like the in-party. Given the ambiguity surrounding an unknown independent candidate’s policy positions or ideology, respondents may optimistically perceive independents as sharing their own preferences (Tomz and Van Houweling 2009).

Fig. 5
figure 5

Mechanical Turk Heterogeneous Effects Analysis: Candidate Preference

Full size image
Fig. 6
figure 6

YouGov Heterogeneous Effects Analysis: Candidate Preference

Full size image

When we disaggregate by respondent partisan identification, we do see some small patterns with respect to candidate gender and race emerge. Republican respondents marginally prefer white candidates while Democrats marginally prefer nonwhite candidates. Republicans do not appear to have a gender preference, while Democrats are 10 percentage points more likely to choose a female candidate than a male candidate. These race and gender patterns do not differ much by election type.

Turning next to political experience, Republicans and Democrats both reward more highly-experienced candidates in both partisan and nonpartisan elections. However, our main theoretical prediction—that political experience will matter more in nonpartisan elections than in partisan elections—is only borne out among Democratic respondents, not Republican respondents. This pattern is clearest in the MTurk sample, though it does obtain in the YouGov sample as well. It may be that, in the absence of party labels, Republicans and Democrats turn to different markers of competence. In the YouGov sample, Republican respondents give greater weight to occupational experience in nonpartisan elections, while Democratic respondents give greater weight to political experience. These findings resonate with those of Sadin (2014) who finds that candidates’ occupations influence perceptions of their ideology.

Mechanism: Candidate Competence

Thus far, our findings have shown that the absence of party labels changes the types of candidates that respondents prefer. Doubtless many pathways from nonpartisan elections to vote choice could be responsible for these effects. Existing theory and evidence highlights at least one possibility: perceptions of candidate competence. Lacking a clear party cue, respondents try to infer candidate competence from the information available to them. In this section, we focus on the plausibility of this competence mechanism, but we acknowledge that nonpartisan elections likely influence vote choice through many causal pathways of which competence is only one. We briefly consider two more pathways (perceptions of candidate ideology and satisficing) at the end of this section.

In order to assess the possiblity that the nonpartisan treatment operates by changing peceptions of candidate competence, we asked respondents to rate the competence of both candidates on a scale ranging from 0 to 100. This measure will help us to substantiate a pillar of our main theoretical claim: in nonpartisan elections compared to partisan elections, voters will give relatively more weight to nonideological dimensions when evaluating candidates. While we cannot conduct a formal mediation analysis here because the required assumption of sequential ignorability (Imai et al. 2011) is difficult to justify in this context, this mechanism is rendered more plausible if we observe the same pattern of treatment effects on the competence dependent variable as we did for vote choice.

Figures 7 and 8 repeat the analyses presented in Figs. 5 and 6 using the competence dependent variable. Most importantly, both Republicans and Democrats rate candidates as more competent when they have more political experience. As shown in the difference column, Democratic respondents (but not Republican respondents) rate such candidates as even more competent in the absence of party labels. We observe small effects of race, gender, and age on competence ratings, although across both datasets, Republican respondents appear to rate white candidates as marginally more competent than nonwhite candidates, while the opposite pattern holds for Democratic respondents. We observe larger differences in competence ratings by occupation, with Republican respondents rating police officers, small business owners, and business executives more highly than educators, while Democrats hold the opposite views on such candidates. Neither party’s respondents rated stay-at-home parents as more competent than educators.

Fig. 7
figure 7

Mechanical Turk Heterogeneous Effects Analysis: Competence

Full size image
Fig. 8
figure 8

YouGov Heterogeneous Effects Analysis: Competence

Full size image

These figures lend support to the idea that in nonpartisan elections, voters prefer more experienced candidates because they give greater weight to nonideological dimensions. The same candidate types that respondents view as more competent are the ones that they tend to elect at higher rates in nonpartisan elections. We do, however, interpret these results with caution as there may be other (unmeasured) pathways beyond competence by which voters prefer some types more in nonpartisan elections. For example, it is plausible that party labels remind subjects of rancorous party politics, which in turn makes them marginally more likely to select outsider candidates. While we do not think this explanation is particularly likely, we cannot rule it (or other similar explanations) out as a possible mechanism by which nonpartisan elections affect candidate choice.

Beyond the competence mechanism, the effects of nonpartisan elections may operate through perceptions of candidate ideology. In an effort to address this possibility, we asked respondents how likely candidates would be to achieve certain ideologically-inflected policy goals. The results (presented in detail in Appendix C.2) suggest that respondents, regardless of party, use occupation as a shortcut for ideology in nonpartisan elections. For example, both Democrats and Republicans view small business owners and business executives as more likely to implement conservative policies, and there is some suggestive evidence that this effect is stronger in nonpartisan elections. A link between candidate occupation and perceived ideology could help explain the heterogeneous effects of occupation, but our analyses also signal that something more than ideology influences respondents in a nonpartisan setting. Indeed, Democrats prefer experienced candidates in nonpartisan elections, but they do not see these candidates as more likely to implement liberal policies.

Finally, a more pedestrian mechanism that could account for our findings is satisficing (Bansak et al. 2017). Because subjects evaluate candidates on five attributes in the nonpartisan elections and on six attributes in the partisan elections, they may mechanically assign more weight to remaining traits when party is omitted. To address this possibility, we exploit a situation in which the number of attributes stays fixed but partisanship varies. Such a scenario arises when subjects evaluate a pair of candidates that share the same partisanship versus when they evaluate a pair who are from different parties. Appendix C.3 reports the results of this analysis. At least in the Mechanical Turk sample, we find that subjects give greater weight to political experience in same-party elections compared with cross-party elections.

Discussion

Drawing on both observational and experimental data, we have shown how a specific electoral institution—nonpartisan balloting—can influence candidate selection. We relied on a theory of candidate choice that posits that in the absence of the party label shortcut, voters have more difficulty inferring the ideology of candidates and as a result rely more heavily on other characteristics.

The implications of this theory of candidate choice were borne out in two survey experiments conducted on both convenience and nationally representative samples. The institutional context matters for the evaluation of candidates based on their attributes. The effect of previous political experience was shown to be statistically significantly larger in nonpartisan elections. This finding directly supports our major theoretical prediction. Our results are also consistent with earlier studies that find candidate quality, particularly incumbency, is more consequential when party does not appear on the ballot (e.g., Schaffner et al. 2001; Lim and Snyder 2015).

We conducted our experiment twice, once on a convenience sample and again on a nationally representative sample. In the appendix, we explore the correspondence across samples more deeply, finding that the correlation of effect estimates is very strong at 0.95. Thus, our findings contribute to a small but growing literature on the correspondence of survey experimental estimates across samples [e.g., Mullinix et al. (2015); Coppock (2017)].

The conjoint experimental design allows us to avoid many of the challenges inherent in studying nonpartisan elections, in particular the problem that cities with partisan and nonpartisan elections may differ in systematic ways. The survey experimental design ensures the clear delineation of partisan and nonpartisan contests. However, these studies were not without limitations. First, we are unable to account for local political contexts. Factors such as retrospective evaluations (Oliver et al. 2012) or inter-group conflict (Kaufmann 2004) might alter the salience and effects of certain cues; we did not control in any way the other features of the electoral context that our subjects may have been imagining. Second, hypothetical candidate choice is related to, but distinct from, actual vote choice. However, it is unclear which way the “biases” from this difference would cut. Considering the thin information environment, one might make the claim that the effect of the electoral institution on the weight given to nonpartisan attributes is understated in these experiments.

These results have important implications for the institutional features of elections beyond local contests. Some hold the normative position that members of the judiciary should be selected for their competence not their ideology. Our results suggest that nonpartisan elections may be a powerful institutional tool for achieving this goal. Voters in primary elections seek to know the ideological positions of candidates but cannot rely on a party cue. In this constrained information environment, voters may give special consideration to candidates’ backgrounds and resumes.

In particular, we manipulated the presence of one shortcut and measured its effects on other shortcuts. Voters seek to make the best decisions possible, given available knowledge and a constrained budget for acquiring new information. When one cue—for example party labels—is no longer available, voters turn to other sources of information.