Membership in a stigmatized religious minority and political support: nonreligious individuals running for office in the United States

Abstract

Atheists can expect discrimination when running for office. We know less about political appraisal of other types of nonreligious candidates or how the influence of nonreligion compares to other factors. Using a conjoint experiment, I examine how the impact of nonreligion on vote choice depends on (1) the label describing nonreligion; (2) the electoral scenario in which voters face the candidate; and (3) voters' partisanship and religiosity. I find that atheists and nonbelievers are at a substantial disadvantage but secular candidates suffer a smaller penalty. While nonreligion reduces political support, it is not the most important influence, plays a smaller role in lower than in higher level elections, and is generally not a factor for Democratic and nonreligious voters. In contrast, it is a major liability for Republican, Independent, and religious voters, especially when Republicans vote in nominating contests and when they face atheists or nonbelievers as opposed to secular candidates.

When the U.S. Constitution was adopted, religious qualifications for elective office were common in the states even though they typically did not require adherence to a particular faith. At the same time, members of certain religions and nonbelievers were legally excluded from holding office. To this day, atheists are banned from holding public office in seven state constitutions (Lee, 2021). While these bans may not be enforceable because they defy the U.S. Constitution, they communicate the message that belief in God is necessary for good moral citizenship. They also place the burden of challenging them on the shoulders of any duly elected nonbelievers in the event that they are prevented from holding office (Lee, 2021).

Although the U.S. Constitution and most state Constitutions ban religious qualifications for office, candidates for elective office from both sides of the political spectrum are expected to display their religiousness. They benefit from confirming their religious credentials because they then look more trustworthy and earn greater political support. In contrast, “not demonstrating religiousness [can be] a significant roadblock for winning public office in the United States” (Djupe and Calfano, 2014; Clifford and Gaskins, 2016, 1088; cf., Campbell et al., 2021). This potentially leaves nonreligious candidates in a bind because prejudice toward nonbelievers, a fast-growing religious minority in the United States,¹ is higher than toward almost any other religious minority in the country. Significant reluctance to support an atheist candidate for elective office is one expression of this prejudice (Clifford and Gaskins, 2016; McCarthy, 2019).

Given considerable evidence of the pervasive nature of anti-atheist-prejudice, it is not surprising that real-life candidates have been doing their best to avoid being labeled atheists—whether or not they happen to identify as such. For example, the media have described members of Congress Jared Huffman (D-CA) and Kyrsten Sinema (D-AZ) as atheists but both Huffman and Sinema insist that they are, respectively, a “humanist” and “unaffiliated” (Garrand, 2019; Smith, 2019). Similarly, former presidential candidate and current U.S. Senator Bernie Sanders (I-VT) has resisted the media's and his political opponents' efforts to frame him as an atheist.

Even though actual candidates distance themselves from the atheist label, the bulk of existing research on political evaluation of candidates who eschew traditional religious affiliations has focused on willingness to support atheists. We know relatively little about the ways in which voters may respond to candidates who are not atheists but are not religious either (cf., Campbell et al., 2021). Relatedly, we do not know a lot about whether the impact of nonreligion conceptualized more broadly depends on the nature of the electoral context and voters' predispositions. Finally, existing scholarship does not report how the magnitude of the influence of nonreligion on candidate evaluation compares to other electorally influential cues.

Using a conjoint experimental design and Amazon Mechanical Turk (MTurk) participants, I build on the existing literature by showing how different nonreligious labels interact with political context and voters' attributes to shape political support. I find that nonreligious candidates are generally disadvantaged compared to religious candidates though candidates described as secular suffer less in political support than those who are described as atheists or nonbelievers. Additionally, I find that the deficit in political support that nonreligious candidates suffer tends to be smaller in elections to lower than higher level office and in general than in nominating elections. Reactions to nonreligious candidates are also contingent on respondents' partisanship and religiosity such that Democrats and nonreligious respondents are generally unbiased toward them whereas Republicans and religious respondents penalize them in their evaluations, even when Republicans are faced with a choice between a Democrat and a fellow Republican. While nonreligion is not the most important predictor of vote choice overall, it can be a deciding factor in close elections which are becoming increasingly common in this polarized age.

Previous research on voters’ responses to nonreligious candidates for political office

Existing research on nonreligion largely focuses on public antipathy toward atheists. Atheists are disliked because they are perceived as immoral, untrustworthy, and threatening religion in the public sphere (Djupe et al., 2014; Cook et al., 2015) and are also stereotyped as Democrats or ideological liberals (Campbell et al., 2021). Atheist candidates for elective office are perceived to lack desirable attributes of good politicians and to be less competent on different issues of public policy (Madrid et al., 2022). Empirical tests of these explanations have been largely based on observational data (e.g., Yancey, 2010; Gervais et al., 2011; cf., Djupe et al., 2014; Franks and Scherr, 2014; Clifford and Gaskins, 2016; Franks, 2017; Grove et al., 2020; Campbell et al., 2021; Madrid et al., 2022).

In contrast, growing experimental research on political support for nonreligious candidates is relatively modest (e.g., Castle et al., 2017; Campbell et al., 2021; Madrid et al., 2022). Focusing mostly on atheists, this research converges on several conclusions. First, negative perceptions of atheists' aptitude for elective office and, especially, perceptions that they are immoral and untrustworthy are primarily responsible for the stigma that they suffer (e.g., Djupe et al., 2014; Franks and Scherr, 2014; Cook et al., 2015; Clifford and Gaskins, 2016; Franks, 2017; Grove et al., 2020; Madrid et al., 2022). Second, bias against atheists is only somewhat malleable with exposure to counter-stereotypic information (Djupe et al., 2014; Franks, 2017; see also, Grove et al., 2020, for related findings). Third, atheist candidates are evaluated more negatively by voters who are religious, conservative, or Republicans (Clifford and Gaskins, 2016; Campbell et al., 2021; Madrid et al., 2022). Yet, secularity does not necessarily preclude electability because candidates who do “not currently identify with a religion” or are “not particularly religious” earn more support than those who are described as “atheists” or “not believing in God” (Castle et al., 2017; Campbell et al., 2021, 21).

In sum, anti-atheist prejudice in the United States is pervasive, exceedingly difficult to modify, and fueled by perceptions of atheists as immoral and untrustworthy. It is thus no wonder why actual candidates who are not religious—whether or not they privately happen to identify as atheists—eschew the atheist label in public life. Research on reactions to such candidates has been on the rise but we still know relatively little about (1) voters' responses to candidates who are not atheists but are not religious either, (2) how the impact of candidate religion may depend on the level of office for which s/he is competing, (3) whether nonreligion plays the same role in nominating and general elections, and (4) how the magnitude of the causal influence of candidate religion compares to the impact of other electorally important cues (cf., Castle et al., 2017; Campbell et al., 2021). It is to the discussion of these questions that I turn next.

Does the label used to describe candidate religion influence the amount of political support s/he receives?

Since atheists are one of the less numerous subgroups in the nonreligious population and atheist candidates are rare in electoral politics, it is important to consider whether and how the label used to encapsulate candidate nonreligion influences voters' responses (Smith and Cragun, 2019).² Atheists likely face more animosity than other nonreligious candidates because “the word ‘atheism’ has a hard edge to it” (Campbell et al., 2021, 187). Put differently, atheists are likely disliked more because they are perceived to be more radical or doctrinaire and, therefore, likely to pose a greater threat to respondents' values and religion in the public sphere. In contrast, candidates embracing other nonreligious identities—like those who find religion irrelevant to their lives and thus describe themselves as nonreligious, unaffiliated, or secular—are likely perceived as posing less threat to respondents' values and religion in the public sphere and are thus likely to suffer fewer adverse socio-political consequences.³ In short, recent research favors a prediction that candidates with a “softer secular edge” should be evaluated more positively than atheists—although there is also some evidence that candidates described as nonbelievers may draw even more animosity than atheists (Campbell et al., 2021).

While it is outside of the scope of research I summarize in this paper to examine whether different nonreligious identities are associated with different amounts of perceived threat, I build on Campbell et al.'s (2021) work to test how the label used to describe a candidate's nonreligion influences political appraisal. I expect that:

H₁: Overall, nonreligious candidates—especially when they identify as atheists or nonbelievers as opposed to secular—will earn less political support than religious candidates—or those who are described as Christian or religious.

How does the impact of religion compare to other attributes that shape political appraisal?

Research to date convincingly establishes that religion is a significant factor in electoral choice. That said, we still do not have empirical evidence on how the magnitude of its impact compares to the influence of other electorally important cues. I set out to fill this lacuna by comparing and contrasting the magnitude of the causal influence of religion to party identification, a factor of paramount importance to vote choice in most elections to political office in the United States, and other attributes that may influence voters' impressions. The latter include candidate gender, race or ethnicity, education, age, previous occupation, experience in public office, family status, salient personal characteristics, and policy area of expertise (Ono and Burden, 2019).⁴ This portion of my investigation is more exploratory in the sense that I do not—a priori—have a specific hierarchy of effects in mind with the exception of anticipating that party identification and religion will be especially important influences on vote choice.

How does the influence of candidate religion depend on the level of office for which s/he is competing?

Elections in the U.S. federal system take place at national, state, or local levels. Contests for seats in lower level offices promise to have less general impact because they culminate in elections of governing bodies that are responsible for smaller geographical areas than elections to national office. In contrast, elections to higher level offices may be perceived as more consequential because they lead to decisions that have broader impact and are likely seen as more prestigious than elections to state or local office. To the extent that otherwise reluctant voters might be inclined to support a nonreligious candidate, this suggests that they should be more willing to do it when the stakes are lower or when a nonreligious candidate is running for a lower- than a higher-level office. I hypothesize that:

H₂: Nonreligious candidates will suffer a bigger penalty in support, especially when they are described as atheists or nonbelievers as opposed to secular, when they are competing for the office of the U.S. President than for a seat in respondents' state legislature.

How do voters’ partisanship and religiosity condition the impact of candidate religion?

While previous research demonstrates that candidate religion is an important factor in electoral choice, it shows that its effects are also contingent on voters' attributes (e.g., Castle et al., 2017; Campbell et al., 2021). For one thing, voters' partisanship moderates the influence of religion on political support because partisanship is a central force in most elections to political office and has a well-established association with voters' religion (Campbell et al., 2021). But, “voters’ own partisanship is not the whole story” because their religiosity or secularism also shapes how they respond to candidate religion “in similar ways within the two parties” (Campbell et al., 2021, 206). Guided by this research, I examine heterogeneous treatment effects as a function of voters' partisanship and religiosity.⁵

Partisanship

Identification with the Democratic party comes with lower and less intense religious commitment and more open-mindedness toward religious difference than identification with the Republican party (e.g., Hewstone et al., 2002; Castle et al., 2017; Campbell et al., 2021). The Democratic party is also home to a large contingent of individuals who reject conventional religious anchors, in part because the nonreligious have been increasingly migrating into the Democratic party's fold in order to align their worldviews with this party's platform (Margolis, 2022). At the same time, counter to the stereotype that Democrats are not religious, a large number of religious activists contribute to the Democratic party membership's diversity (Campbell et al., 2021).⁶ In contrast, the Republican party is dominated by individuals who are higher in religious devotion even though, counter to its stereotype, “a small but vocal group of Secularist libertarians” calls the Republican party their home (e.g., Layman, 1997, 2001; Kaufmann, 2004; Campbell et al., 2021, 20).⁷ We know relatively less about the religious composition of pure Independents. The percentage of religiously unaffiliated Independents in the 2021 General Social Survey falls between the percentages of religiously unaffiliated Democrats and Republicans (SDA: Survey Documentation and Analysis, No date). Other national surveys demonstrate that Independents account for a similar percentage of the “religious nones” as Republicans (Pew Research Center, 2023). On balance, I hypothesize that:

H₃: Democrats will evaluate nonreligious and religious candidates similarly whereas Republicans and Independents will discriminate against nonreligious compared to religious candidates, especially when the former are atheists or nonbelievers as opposed to secular.

Religiosity

My prediction regarding the moderating role of respondents' religiosity follows straightforwardly from the existing literature (e.g., Campbell et al., 2021). I expect that:

H₄: Religious respondents will be less likely to support nonreligious candidates over their religious counterparts, especially when the former are described as atheists or nonbelievers as opposed to secular whereas candidate religion will not have a significant bearing on nonreligious respondents' choices.

How does the influence of candidate religion depend on whether voters face a candidate in a nominating or a general election contest?

The influence of candidate religion should also depend on the informational environment in which a candidate is running—or whether voters are making their choices in nominating or general elections. Candidates for a party's nomination typically share their partisanship with the voters. Because vote choice in nominating elections cannot be made on the basis of a shared party identification, biases related to a candidate's membership in a stigmatized minority should have more room to influence it. In contrast, given that party identification plays an outsized role in general election voting (e.g., Bafumi and Shapiro, 2009), voters facing nominees of different political parties may be inclined to “forgive” a candidate for her or his nonreligion when that candidate belongs to the same partisan tribe. Simultaneously, willingness to overlook or downplay the absence of a religious identity should be further contingent on respondents' partisanship. I hypothesize that:

H₅: Democrats will not be influenced by candidate religion in either informational scenario whereas Republicans will be more likely to discriminate against nonreligious over religious candidates, especially when the former are atheists or nonbelievers, in nominating than in general elections. In the absence of a partisan anchor, Independents will rely on candidate religion as a heuristic and will discriminate against nonreligious candidates in both informational scenarios.

A summary of all the hypotheses can be found in Table 1.

Table 1. Summary of hypotheses

Method

To recap, I examine how voters' responses to nonreligious candidates are contingent on the label used to describe their views, the level of office for which they are competing, and whether they are running in simulated nominating or general elections. Additionally, I investigate how the impact of nonreligion depends on my participants' partisanship and religiosity as well as how the magnitude of its causal influence compares to the impact of other electoral cues. While previous research on nonreligion and electoral choice examined evaluations of individual candidates, I mimic electoral choice more effectively by having respondents choose between two candidates. A conjoint survey experiment is ideally suited for addressing these questions.

Conjoint survey experiments have grown in popularity in political science as a method for analyzing multidimensional preferences such as those that voters form when they evaluate politicians inevitably possessing some attributes that they like and others that they do not (Bansak et al., 2021). Conjoint designs make it possible to simultaneously vary a large number of factors and to calculate the magnitude of the causal influence of each of these factors, in addition to their statistical significance, while controlling for all other manipulated factors. These designs are also more efficient than factorial experiments because they require far fewer participants, boast higher external validity, and may suppress social desirability biases better than factorial experiments (Bansak et al., 2021).

Before I proceed, a note on terminology. Since I seek to establish how individuals who eschew religious affinities fare in electoral politics compared to those who embrace them, I use “nonreligious” when talking about the former and “religious” when referring to the latter. My use of both terms is inspired by how scholars of religion and politics have conceptualized religious identities as having three components—“belonging to a religious group, engaging in religious behavior, and holding religious beliefs” (Castle et al., 2017, 146). Historically, belonging to a religious group was linked with electoral choice and, to some extent, it continues to influence it. However, in the late 20th century, engaging in religious behavior or religiosity has become an increasingly important identity that shapes socio-political behavior (Castle et al., 2017, 146; see also Smidt et al., 2017; Campbell et al., 2021). Thus, in my use of the term, the religious category includes belonging to a religious group like Christians or Jews or Muslims as well as religious behaviors such as attending a place of worship, praying, or giving money to religious organizations (Layman et al., 2021). Similarly, in my use of the term nonreligious, I include individuals who may consider themselves nonreligious because they do not identify with a religious group and, by default, do not practice their religion as well as others like atheists, agnostics, humanists, or seculars who not only lack an identification with a religious group but may embrace beliefs and practice behaviors that are consistent with their specific identity. For the sake of simplification, I use the term religion when I have both religious and nonreligious identities in mind.

I programmed the experiment using the Conjoint Survey Design Tool and subsequently embedded it into a Qualtrics survey (Strezhnev et al., 2016). On the model of a typical conjoint experiment in political science, participants were presented with a series of tables featuring profiles of 10 political candidates who might be competing in electoral campaigns in which they are eligible to vote (Bansak et al., 2021). The instructions introducing participants to the task emphasized that the exercise was purely hypothetical. Therefore, they were asked to make a choice even if they were not entirely sure of their preference. Each table included a set of 12 randomly varied attributes of each candidate in the pair. Candidates in each pair were labeled “Candidate A” and “Candidate B” and each pair was presented on a separate screen. Participants were asked to familiarize themselves with each candidate for at least 25 seconds and then indicate which of the two candidates they would vote for in the election specified in the instructions. While all participants made 10 choices, the 10 pairs of candidates were divided into two blocks of five pairs each. One block introduced the candidates as running for a seat in the participant's state legislature whereas the preface to the other block indicated that the candidates were running for U.S. President. Participants were reminded before evaluating each pair of candidate profiles for what office the candidates were competing. The order of the presentation of the two blocks was randomized across participants.

Each candidate profile included a randomly assigned religion that had five levels. Three levels represented nonreligious identities: (1) “atheist”—to represent the most stigmatized nonreligious group, (2) “nonbeliever”—to represent an alternative framing of an atheist identity, and (3) “secular”—to represent a nonreligious identity that represents not only a rejection of traditional religious ties but may come with a distinct ideology (Layman et al., 2021). Religious permutations of the candidates were either: (1) “Christian”—to represent a dominant religious tradition in the United States or (2) “religious”—a shorthand term to communicate the zeal with which the candidate is practicing her or his religion and one which corresponds to religious cues that real-life political candidates communicate about themselves (Clifford and Gaskins, 2016).

Other attributes in the design are typically found in political candidates' campaign materials and may influence voters' choices—demographics, previous occupation and experience in public office, salient personal characteristics, areas of policy expertise, and party affiliation (Ono and Burden, 2019). All attributes included in each conjoint table and their levels are listed in Table 2. A sample conjoint table is provided in Table 3.

Table 2. Manipulated attributes and their levels

Table 3. Sample conjoint table

Four more points need to be mentioned about the information that was displayed on each participant's screen. First, the order of all attributes' presentation was randomized across respondents but fixed for each respondent in order to minimize the cognitive burdens that they faced. For example, candidate religion may have appeared in any of the 12 rows of each candidate's profile but it always appeared in the same spot for each research participant. Second, in order to ensure plausibility, some combinations of education and previous occupation were restricted from appearing in candidate profiles. This is known as conditionally independent randomization. Specifically, candidates whose previous occupation was described as “educator” were prevented from having less than at least some college at the same time and candidates described as “lawyer” were only allowed to have a graduate degree. Conditionally independent randomization does not detract from one's ability to make causal inferences based on a conjoint design but it does require some extra analytical footwork. I discuss the latter in the “Results” section. Completely independent randomization was used for all remaining attributes in the design—meaning that any candidate profile had an equal probability of including any level of these remaining attributes. Third, based on the most recent U.S. census, I used weighted randomization with regard to the candidates' race or ethnicity. As a result, a majority (61%) of the candidates generated in the experiment were white, 13% were Black, 20% were Hispanic, and 6% were Asian. Finally, I imposed a handful of attribute order constraints in order to facilitate the tasks of reviewing the candidate profiles. Specifically, the following sets of attributes were listed consecutively within any given profile: (1) information about previous occupation and experience in public office and (2) salient personal characteristics, party affiliation, and policy area of expertise.⁸

The data were collected on July 19, 2021 on the MTurk data collection platform hosted on Amazon Web Services. While samples that can be recruited on this platform for a fee are not representative (see below for more information about sample characteristics), they are widely used in social sciences and are deemed appropriate for experimental—or even correlational—research (Berinsky et al., 2012; Levay et al., 2016).⁹ For a robustness check, I recalculate all the results I report below with sample weights based on the most recent national census but find no difference in the findings.

Participants

One thousand three hundred and sixty-seven MTurk members were invited to participate in the experiment in exchange for a $1.50 incentive payment. Data from seven respondents were excluded from all analyses because they had either terminated their participation immediately or had skipped over the experimental portion of the design. This resulted in usable data from 1,360 participants. Because each participant evaluated 10 candidate pairs, the resulting sample available for analysis consists of 13,600 candidate pairs and 27,200 unique candidate profiles. Such a sample is sufficiently large to estimate the unique effect of each attribute in the design.¹⁰

My participants' profile is similar to MTurk samples employed in other research (e.g., Levay et al., 2016). A majority of participants (66.4%) were white and male (61.7%). While they ranged in age from 18 to 85, an “average” participant was 40 years old. An “average” participant also reported having a bachelor's degree. Half the sample reported an affiliation with the Democratic party and a smaller number embraced an ideologically liberal self-identification (37.9%). Half the participants were Catholic and almost 20% reported no religious affiliation or identified as an atheist, a nonbeliever, or secular. About a third of the sample reported attending religious services every week or almost every week and, at the other end of the scale, a similar number reported never attending them.

Data analysis

A key statistic of interest in the analysis of data from conjoint experiments is the average marginal component effect or AMCE (Hainmueller et al., 2014). In a choice experiment I report in this paper, the AMCE corresponds to the average change in the likelihood that a profile with a particular attribute will be selected instead of the baseline attribute value while controlling for the effects of all other attributes (Hainmueller et al., 2014; Bansak et al., 2021).

The dependent variable is measured with a question about vote choice and is coded “0” when a profile of a particular candidate is not selected and “1” when a profile is selected. On the model of the approach described in Hainmueller et al. (2014), I calculate AMCEs for each attribute using the cjoint package in R, version 2.1.0 (Strezhnev et al., 2016). The standard errors in this analysis are “clustered by the respondent to account for the dependence of observations across respondents” (Ono and Burden, 2019, 595).

Results

The impact of candidate religion on vote choice

I start by estimating a series of models of vote choice with dummy variables coding for candidate religion, excluding different baselines in each model. These models also include a full set of dummy variables representing the levels of each manipulated attribute—save for the excluded baselines—in order to determine how the magnitude of the influence of candidate religion compares to other attributes that may shape electoral choice. The models also include an interaction between candidate education and previous occupation because, as I noted above, some combinations of education and previous occupation were prevented from occurring (Hainmueller et al., 2014). Because each predictor in this and subsequent analyses is coded on a 0–1 scale, the magnitude of the effect of any level of any attribute is directly comparable to others and easy to interpret. Specifically, the size of each coefficient associated with any level of any manipulated attribute represents either a percent advantage (positive coefficient) or a percent disadvantage (negative coefficient) of possessing that particular attribute compared to the excluded baseline.

I test the hypothesis that nonreligious candidates will earn less political support than religious candidates (H₁) by comparing the AMCEs for atheist, nonbeliever, and secular candidates with the AMCEs for Christian and religious candidates. To simplify, only the significant effects of candidate religion are shown in Figure 1. Full results, including standard errors, are available in the Appendix.

Figure 1. Influence of candidate religion on vote choice. Significant differences only; full results, including standard errors, can be found in the Appendix.

Figure 1 shows percentage change in support, based on the AMCEs, that each nonreligious candidate enjoys or suffers compared to the excluded baseline that is highlighted in gray at the bottom of each set of comparisons. In line with H₁, the first set of the three bars on the left side of this figure demonstrates that atheist, nonbeliever, and secular candidates lose about 4, 5, and 3%, respectively, in political support compared to Christian candidates. Similarly, the middle set of the two bars shows that atheist and nonbeliever candidates suffer in political support compared to religious candidates by 3 and 4%, respectively. Secular candidates also suffer a smaller deficit in political support than atheist or nonbeliever candidates compared to candidates described as Christian (3%), do not suffer a political penalty compared to religious candidates, and are preferred over nonbeliever candidates, by an average margin of about 2%. Additional significance testing demonstrates that evaluations of atheist candidates, compared to nonbeliever or secular candidates, are statistically indistinguishable. On the other hand, nonbeliever candidates suffer significantly more discrimination compared to their secular counterparts (p < 0.001).

This initial estimation shows that vote choice is also affected by candidate age, experience in public office, family status, party affiliation, and education (results are available in the Appendix). Compared to a 35-year-old candidate, the excluded baseline, 59-year-old candidates lose 2% in political support and 71-year-old candidates lose 8% in political support. Compared to candidates with 12 years of experience in public office, those with 8 years of experience lose 3% in political support, those with 4 years of experience lose 7%, and candidates with no experience suffer a costly 12% loss of political support. Compared to married candidates with two children, single candidates are at a significant disadvantage, losing 3% on average when they are also divorced and 2% when they were never married. Compared to Democratic candidates, Republican candidates are on average 6% less likely to win the favor of the research participants qua voters—not surprisingly in light of the sample's composition. Finally, compared to candidates with college degrees, candidates with a high school degree or some college are at a political disadvantage of about 3% in each case. In sum, younger, more experienced, married, well-educated, and Democratic candidates are preferred to their older, less experienced, single, poorly-educated, and Republican counterparts.

The impact of candidate religion on vote choice in state legislative versus presidential elections

I next separately calculate the impact of candidate religion on vote choice for candidates competing for a seat in a respondent's state legislature and those competing for U.S. President in order to test the hypothesis that candidate religion would have more influence in elections to higher than lower office. The significant effects within each electoral context are shown in Figure 2. Full results are available in the Appendix.

Figure 2. Influence of candidate religion on vote choice in state legislative versus presidential elections. Significant differences only; full results, including standard errors, can be found in the Appendix.

What is not clear from the effects summarized in Figure 2 is whether they are significantly different across the two electoral scenarios. To shed light on this question, I rely on the dependent samples t-tests. In partial support of H₂, atheists running for President are at a significantly greater disadvantage compared to Christians than when they are running for a state legislature (p < 0.01). At the same time, atheists are evaluated similarly across the two electoral scenarios compared to religious candidates (p < 0.49). As expected, the penalty that nonbelievers suffer significantly depends on the electoral scenario when they are compared to a religious baseline (p < 0.01) but nonbelievers are evaluated similarly in legislative and presidential contexts compared to Christian candidates (p < 0.4). Unexpectedly, nonbelievers suffer a significantly greater penalty when running for a state legislature than for President. Finally, in support of H₂, secular candidates do not suffer a disadvantage compared to Christian candidates when they are running for a state legislature but lose an average of 4% in support when competing for the U.S. presidency (p < 0.002). Secular candidates also enjoy a significant advantage in state legislative races compared to nonbelievers but are evaluated similarly to nonbelievers in presidential contests (p < 0.001).

Voters' partisanship and religiosity as moderating factors

I have so far demonstrated that, all else equal, nonreligious candidates for office are penalized for their religious identity relative to their Christian or religious counterparts, especially when they are running for higher level office. Since vote choice is strongly anchored in partisan loyalties and partisanship is correlated with religiosity, I next consider heterogeneous treatment effects as a function of respondents' partisanship and religiosity. The significant heterogeneous treatment effects are shown in Figure 3.

Figure 3. Influence of candidate religion on Republicans' and religious respondents' vote choice. Significant differences only; full results, including standard errors, can be found in the Appendix.

Partisanship

I compare and contrast the influence of candidate religion among Democrats, Independents, and Republicans. Since existing scholarship suggests that Independent leaners are closeted partisans (e.g., Klar and Krupnikov, 2016), I classify all Democrats, whether strong or weak, and Independents leaning Democratic as Democrats. Similarly, I classify all Republicans, whether strong or weak, and Independents leaning Republican as Republicans. This coding results in a relatively small number of pure Independents.¹¹ Therefore, I perform a robustness check by calculating the results for pure Independents as one group and Independents combined with leaners as another.

In line with H₃, Democrats' choices are not affected by candidate religion regardless of whether Democratic leaners are counted as Democrats or Independents. To some extent, the results for Independents depend on whether one examines pure Independents' choices or those of Independents combined with all leaners. In partial support of H₃, pure Independents only discriminate against atheists compared to religious candidates by about 6%. When Independents are combined with leaners, they penalize atheists compared to Christian, religious, and secular candidates by an average margin of 5% in each case. Additional significance testing demonstrates that Independents who are combined with leaners evaluate atheist and secular candidates similarly to those who are described as nonbelievers. In contrast, Independents combined with leaners are significantly less likely to support an atheist compared to a secular candidates (p < 0.01).

As expected, Republicans' choices are stable regardless of whether Republican leaners are combined with Republican identifiers. Specifically, Republicans respond more negatively to all nonreligious candidates, whether compared to their Christian or religious counterparts. Compared to Christian candidates, Republicans are less likely to support atheists by an average margin of 14%, nonbelievers by an average margin of 12%, and secular candidates by an average margin of 9%. Averaging across the three nonreligious identities, Republicans discriminate against nonreligious compared to Christian candidates by an average margin of 11.67%. Similarly, Republicans discriminate against atheists, nonbelievers, and secular candidates in comparison to religious candidates by average margins of 12, 10, and 7%, respectively. Averaging across these differences, Republicans penalize nonreligious candidates compared to their religious counterparts by an average margin of 9.67%. Additional testing demonstrates that Republicans' responses to atheist and secular candidates, respectively compared to nonbelievers, are statistically indistinguishable. However, as predicted by H₃, Republicans are significantly less hostile to secular candidates than those described as atheists (p < 0.04).

Religiosity

To examine heterogeneous treatment effects as a function of religiosity, I classify respondents as religious or nonreligious based on the frequency of their attendance at a place of worship. Respondents who reported that they attended a place of worship every week, almost every week, or once or twice a month were coded as religious. The remaining respondents, classified as nonreligious, indicated that they either never attended a place of worship or attended it only a few times a year.

In line with H₄, religious respondents generally discriminate against nonreligious compared to religious candidates. More specifically, religious respondents are less likely to support atheist and nonbeliever candidates compared to their Christian counterparts by 10% margins and secular candidates by 7%. Religious respondents also penalize atheist, nonbeliever, and secular candidates compared to candidates described as religious though by relatively smaller margins (7, 8, and 5%, respectively). Averaging across all these differences, religious respondents discriminate against nonreligious candidates by an average margin of 9% compared to Christian candidates and an average margin of 6.67% compared to candidates described as religious. Additional significance testing demonstrates that religious respondents' bias against nonreligious candidates does not significantly depend on how their nonreligion is described. In support of H₄, nonreligious respondents' choices are not influenced by any permutation of candidate religion.

A summary of all significant heterogeneous treatment effects is shown in Figure 3. Full results are available in the Appendix.

Candidate religion in nominating versus general elections

The preceding section demonstrates that the impact of candidate religion on vote choice is moderated by voters' partisanship. One question that is still outstanding is whether and how candidate religion and respondents' partisanship further interact with the informational environment in which respondents are making their choices. To illuminate this question, I separately model the effects of candidate religion for Democrats, Republicans, and Independents: (1) when both candidates are Democrats; (2) when both candidates are Republicans; and (3) when one candidate is a Democrat and another a Republican. The first two scenarios mimic the partisan matchup in Democratic and Republican nominating contests, respectively, and the third imitates general elections. Because candidate party affiliation was fully randomized in the experiment, Democrats and Republicans also vote in scenarios in which both candidates represent the political party with which they do not identify. While such scenarios are much less likely, they are plausible. For example, California Republicans were choosing between two Democrats in the 2016 general election when the current U.S. Vice President Kamala Harris was running for the U.S. Senate. I show the significant results from this portion of the analysis in Figures 4 and 5. I do not include results for pure Independents because candidate religion did not influence their choices when they were examined in different informational scenarios. However, full results are available in the Appendix.

Figure 4. Influence of candidate religion on Democrats' choice in different informational environments. Democratic leaners combined with Democratic identifiers. Significant differences only; full results, including standard errors, can be found in the Appendix.

Figure 5. Influence of candidate religion on Republicans' choice in different informational environments. Republican leaners combined with Republican identifiers. Significant differences only. Full results, including standard errors, can be found in the Appendix.

Democrats do not use candidate religion as a heuristic when choosing between two Democrats. On the other hand, when choosing between a Democrat and a Republican, Democrats discriminate against nonbeliever candidates compared to their religious and atheist counterparts, by 5 and 4% respectively. In the electoral scenarios featuring two Republicans, Democrats prefer secular and atheist candidates over their religious counterparts by 6% margins.

Republicans show greater aversion to nonreligious candidates when choosing between two Republicans than when choosing between a Democrat and a Republican (Figure 5). More specifically, in scenarios mimicking Republican nominating contests, Republicans penalize atheist compared to Christian and religious candidates by an average of 19.5%, nonbeliever candidates by 18.5%, and secular candidates by 14%. In contrast, in scenarios mimicking general election contests, Republicans discriminate against atheist and nonbeliever candidates compared to Christian and religious candidates by an average of 10% and against secular candidates by an average of 6% compared to Christian candidates. Republicans also discriminate against nonbeliever compared secular candidates by an average of 5%.

Finally, Republicans respond negatively to nonreligious candidates when they are choosing between two Democrats, discriminating against atheists over Christian and religious candidates by 11% on average and nonbeliever and secular candidates compared to their Christian counterparts by 8 and 10%, respectively. Additional significance testing demonstrates that, with the exception of their significantly greater aversion to atheist compared to secular candidates when choosing between a Democrat and a Republican (p < 0.05), Republicans' responses to all other pairs of nonreligious candidates are similar in all informational scenarios.

Summary and discussion

To recap, the results from the conjoint experiment I report in this paper are substantially in line with the hypotheses concerning the links between candidate religion and political appraisal. In the aggregate, these results identify a substantial political disadvantage that candidates who happen to be atheists, nonbelievers, or secular suffer in comparison to Christian or religious candidates. While reactions to atheists were not significantly different from reactions to nonbelievers, candidates described as secular were significantly preferred to nonbelievers.

Countering this interpretation, some might argue that a 3–5% aggregate penalty that nonreligious candidates suffer is substantively modest. We can put this penalty in perspective by juxtaposing it with the magnitude of the biases linked with other attributes in the design. The overall deficit associated with nonreligion is smaller than the penalty that comes with old age or little to no political experience. But, it is comparable to or larger than the impact of other significant attributes in the experiment. Similarly, evidence from a recent conjoint experiment on the role of gender in political appraisal demonstrates that the aggregate penalty that female candidates suffer is considerably smaller (1.3%) than the price that nonreligious candidates pay in my experiment (Ono and Burden, 2019). In contrast, the aggregate disadvantage that comes with a sexual minority or, especially, transgender status is considerably higher than the disadvantage caused by nonreligion (6.7% in the former case and 11% in the latter) (Magni and Reynolds, 2021). Overall, whether the disadvantage a member of an underrepresented group suffers in electoral politics amounts to 1% or more is arguably substantively important because many elections are decided by very small margins in this polarized age.

As importantly, the aggregate results camouflage important nuance in the role that religion plays in political appraisal because its influence varies with the level of office for which a candidate is competing, participants' partisanship and religiosity, and the informational environment. Specifically, nonreligion tends to detract from political support to a greater extent in presidential than state legislative elections—although the results did not follow this pattern in reactions to candidates described as nonbelievers. As expected, Democrats' choices are largely not influenced by candidate religion whereas Republicans exhibit a strong preference for religious and, especially, Christian candidates over those who are atheists, nonbelievers, or secular. With one exception—of finding secular candidates significantly more palatable than atheists—Republicans' reactions to other combinations of nonreligious candidates are similar. Pure Independents are only more averse to supporting an atheist compared to a religious candidate. When pure Independents are combined with Independent leaners, they are also more averse to an atheist compared to Christian and secular candidates. Also, as expected, nonreligious respondents' choices are not influenced by candidate religion whereas religious respondents generally favor both Christian and religious candidates over those who are atheists, nonbelievers, or secular.

Finally, the magnitude of the impact of candidate religion on vote choice depends on the combination of participants' partisanship and the informational environment in which members of different partisan tribes are forming their impressions of the candidates. Democrats are generally less likely to take candidate religion into account when making their choices except that they discriminate against nonbelievers compared to both atheist and religious candidates when choosing between a Democrat and a Republican. On the other hand, Democrats are more likely to vote for atheist or secular candidates compared to their religious counterparts when both candidates are Republicans—likely because the concept of religious Republicans conjures up the unfavorable images associated with the religious right and its embrace of extreme positions on social policies that Democrats reject. In contrast, Republicans are especially unlikely to vote for nonreligious candidates in primary election-like scenarios though they still discriminate against them by large margins in other informational environments. While Republicans are decidedly unlikely to support nonreligious candidates, they find secular candidates significantly more acceptable than those described as atheists.

These findings have a host of implications for the political careers of nonreligious candidates. As expected, they demonstrate that the success or failure of nonreligious candidates for office may depend on how they frame their identities and the level and type of office for which they are competing. These candidates are likely to pay a greater political price if they identify as atheists or nonbelievers and a smaller price when they embrace a secular label instead, especially when they are competing for high level political office.

What is more, the discrimination that nonreligious candidates may experience has a partisan dimension. Castle et al. (2017) recently pondered whether nonreligious Republicans and religious Democrats can “find a way to get elected despite the electoral disadvantage that their mix of identities creates” (158). In keeping with Castle et al.'s findings, my results are far more optimistic about the chances of religious Democrats than those of secular Republicans. Yet, my findings—like those reported by Campbell et al. (2021)—suggest that rejecting traditional religious ties “is not necessarily fatal” for political candidates (21). While atheists and nonbelievers draw a good deal of animus, candidates who are simply “not religious”—in Campbell, Layman, and Green's research—or “secular” in my study seem more palatable even to research participants strongly predisposed to reject religious difference.

While the findings I summarize in this paper significantly add to the existing scholarship on nonreligion in political appraisal, questions for future research remain. For example, given the heterogeneity of the nonreligious population, more can be learned about the effects of nonreligion on political appraisal by manipulating it in other ways. In addition, future studies could vary the informational environment in which participants make their choices more systematically by independently manipulating the type (legislative versus executive) and level (national versus state versus local) of office for which candidates are competing as well as test the role that nonreligion plays in nonpartisan elections.

For now, I conclude by pondering the implications of the sample I have used to test the hypotheses regarding the links between candidate religion and political evaluation. MTurk samples, while widely used in experimental research in social sciences, differ in some ways from national samples (Berinsky et al., 2012; Levay et al., 2016). Compared to participants in national samples, participants in my sample are younger, more well-educated, more Democratic, and more likely to be men. While MTurk samples have been described as more secular than the general population, I have not found this to be the case in my sample, using the 2021 GSS data as a reference (SDA: Survey Documentation and Analysis, No date). Overall, then, the profile of my sample may have made it more difficult to detect significant effects of religion because younger, more well-educated, Democratic identifiers, and male participants are less likely to discriminate against nonreligious individuals. That said, the results I report are robust when weighted with the sample weights based on the most recent U.S. census.