Corruption, anti-corruption, and voter preference for women candidates

Recent experimental studies find that voters tend to prefer women political candidates, i.e., that there is a ‘pro-woman bias’ among the electorate (Aguilar et al., Reference Aguilar2015; Schwarz and Coppock, Reference Schwarz and Coppock2022).Footnote ² An adjacent literature suggest that citizens hold gendered stereotypes (Bauer, Reference Bauer2017; Ellemers, Reference Ellemers2018; Sweet-Cushman, Reference Sweet-Cushman2022). Citizens sometimes hold stereotypes about women being more honest, compassionate, and better at dealing with education and health care, or other ‘women’s issues’, and less competent when it comes to issues such as the military, policing, and the economy (e.g., Alexander and Andersen, Reference Alexander and Andersen1993; Huddy and Terkildsen, Reference Huddy and Terkildsen1993; Dolan, Reference Dolan2010). Traditional traits and competences associated with female candidates have also been viewed as incongruent with office holding. However, as the proportion of women in politics has increased, research has started to question whether stereotypes consistently harm female candidates (Dolan, Reference Dolan2017) and suggest that parties and other elites instead seek to consciously exploit stereotypes about women being less corrupt in order to gain electoral advantage and to restore trust in the face of a corruption scandal (Funk et al., Reference Funk2021; Armstrong et al., Reference Armstrong2022).

Some studies focus directly on investigating whether voters perceive women candidates as being less involved in misconduct and corruption scandals (Barnes and Beaulieu, Reference Barnes and Beaulieu2014; Eggers et al., Reference Eggers2018; Barnes et al., Reference Barnes2018; Wiesehomeier and Verge, Reference Wiesehomeier and Verge2020; Batista Pereira, Reference Batista Pereira2021; Le Foulon and Reyes-Housholder, Reference Le Foulon and Reyes-Housholder2021). For instance, Batista Pereira (Reference Batista Pereira2021) finds that Mexican voters support women accused of corruption less than their male counterparts, while no gender differences were found in the Brazilian sample. Relatedly, Eggers et al. (Reference Eggers2018) find that male voters do not punish corrupt women more harshly in general but that women voters do so. In contrast, Le Foulon and Reyes-Housholder (Reference Le Foulon and Reyes-Housholder2021) find that voters prefer allegedly corrupt women candidates over male candidates in Uruguay, but not in the more corrupt contexts of Argentina and Chile. Interestingly, this study does not find that women candidates who have fought corruption have an advantage. Still, studies in the USA confirm that voters expect women in office to be less corrupt (Barnes and Beaulieu, Reference Barnes and Beaulieu2019) and that women candidates reduce suspicion of fraud (Barnes and Beaulieu, Reference Barnes and Beaulieu2014).

However, most studies to date investigate the extent to which voters support a woman candidate allegedly implicated in a corruption scandal. Fewer studies investigate if women are seen as political cleaners by voters, i.e., that they would also be perceived as fighting corruption levels more broadly, post-hoc. Effectively fighting corruption is, in part, different from refraining from being implicated in corruption scandals, and a candidate may therefore benefit from partly different qualities.

Furthermore, most experimental studies to date build on single-country studies (Barnes and Beaulieu, Reference Barnes and Beaulieu2014; Barnes et al., Reference Barnes2018; Eggers et al., Reference Eggers2018; Wiesehomeier and Verge, Reference Wiesehomeier and Verge2020), which make the findings less comparable across different contexts. Although some recent studies compare data from several countries (e.g., Batista Periera, 2021; Le Foulon and Reyes-Housholder, Reference Le Foulon and Reyes-Housholder2021; Schwindt-Bayer et al., Reference Schwindt-Bayer, Esarey, Schumacher, Stensöta and Wängnerud2018), experimental studies on gender and corruption follow the general approach of the wider field of candidate choice experiments (Incerti, Reference Incerti2020) and are largely single-country based, and sometimes based on convenience samples. This means that we know little about the extent to which stereotypes about women and corruption are generalizable across diverse contexts.

However, there is growing evidence on the link between a larger share of women in office and lower levels of corruption (e.g., Alexander, Reference Alexander and Bågenholm2021), and recent evidence that women in office may actually cause such a change (Esarey and Schwindt-Bayer, Reference Esarey and Schwindt-Bayer2019; Bauhr and Charron, Reference Bauhr and Charron2021). Furthermore, research indicates that some voters believe that women candidates will be less likely to be implicated in corruption scandals (Barnes and Beaulieu, Reference Barnes and Beaulieu2014; Reference Barnes and Beaulieu2019; Eggers et al., Reference Eggers2018; Wiesehomeier and Verge, Reference Wiesehomeier and Verge2020; Batista Pereira, Reference Batista Pereira2021; Le Foulon and Reyes-Housholder, Reference Le Foulon and Reyes-Housholder2021), Drawing together these insights, we hypothesize that voters also see women as being more likely to combat corruption and that this effect may at least partly travel across contexts.

Following the broader literature on corruption and electoral accountability (Davis et al., Reference Davis2004; Birch, Reference Birch2010; Sundström and Stockemer, Reference Sundström and Stockemer2015; Bauhr and Charron, Reference Bauhr and Charron2018), we propose that voter preferences can be expressed in at least two different ways, both as an active choice (voice) and as a more passive abstention (exit) from choosing any of the available candidates (cf. Hirschman, Reference Hirschman1970). How citizens express discontent with candidates has wider implications for the functioning of democratic institutions. Ideally, citizens should be able to distinguish between two candidates and decide which candidate they prefer. This would allow for the best available candidate to gain support and hopefully influence outcomes in the preferred direction. Disillusionment with the ability of any of the available candidates to deal with corruption is instead associated with resignation, alienation, and disengagement (Bauhr and Grimes, Reference Bauhr and Grimes2014; Chong et al., Reference Chong2015).

Much in line with previous work, we thus seek to investigate these two alternative responses, since they both have implications for citizens’ demand for accountability. We hypothesize that candidate gender is used by voters as a consequential heuristic for both of these electoral responses. We expect participants to prefer a female candidate over a male candidate in order to fight corruption, but also that two male candidates on the ballot will make participants more likely to perceive that neither of the available candidates will successfully fight corruption and therefore refrain from choosing any of the available candidates. This leads to the following hypotheses:

How can we understand women candidate advantage? In the following, we depart from our baseline assumption that voters simply see women as more honest and trustworthy when it comes to fighting corruption and investigate the explanatory power of three alternative theoretical frameworks: risk aversion, outsider status, and a social welfare policy profile. These are all well-established and popular explanations for the link between the share of women in office and lower levels of corruption. Drawing on documented gender differences in risk aversion (Bord and O’Connor, Reference Bord and O’Connor1997; Slovic, Reference Slovic1999; Watson and McNaughton, Reference Watson and McNaughton2007; Charness and Gneezy, Reference Charness and Gneezy2012), studies advance that the link between more women in office and lower levels of corruption may at least partly be attributed to women being more risk averse, and therefore less likely to engage in corruption, particularly if there is a risk of corruption being detected and punished (Esarey and Schwindt-Bayer, Reference Esarey and Schwindt-Bayer2018).Footnote ³ Alternatively, women candidates may have fewer opportunities to be implicated in corruption scandals since they are ‘outsiders’, who, on average, are less embedded in the collusive and often male dominated networks where corrupt transactions are made (Goetz, Reference Goetz2007; Bjarnegård, Reference Bjarnegård2013; Bauhr and Charron, Reference Bauhr and Charron2021).

Finally, building on studies that show a link between women’s descriptive representation and policy outcomes (Schwindt-Bayer, Reference Schwindt-Bayer2006; Wängnerud, Reference Wängnerud2009; Clayton, Reference Clayton2021), recent studies suggest that women’s policy priority on delivering social welfare services also contributes towards reducing corruption (Alexander and Ravlik, Reference Alexander and Ravlik2015; Bauhr et al., Reference Bauhr2019; Reference Bauhr2024; Gao and Mahutga, Reference Gao and Mahutga2023). This could partly be attributed to women constituents being disproportionately affected by dysfunctional and corruption ridden public service delivery (Kubbe and Merkle, Reference Kubbe and Merkle2021).

Building on this work, we propose three candidate characteristics that may potentially partly explain why voters prefer women: risk aversion (personality), outsider status (background), and campaigning on social welfare issues (policy priority) – and investigate the effect of activating these stereotypes on voters’ candidate preferences. We explore whether female candidate advantage persists even if male candidates are also attributed these characteristics. In other words, we seek to contrast these explanations with our baseline expectation that voters simply see women candidates as more honest and trustworthy than male candidates (cf. Dollar et al., Reference Dollar2001; Wiesehomeier and Verge, Reference Wiesehomeier and Verge2020). Thus, given the difficulties involved in investigating the honesty/trustworthy stereotype directly, we follow the approach of Wiesehomeier and Verge (Reference Wiesehomeier and Verge2020) and investigate the extent to which activating these alternative stereotypes can ‘explain away’ the general preference for women candidates.Footnote ⁴

Some recent experimental studies suggest that activating stereotypes about women candidate’s being risk averse or having outsider status make voters perceive women as less corrupt (Barnes et al., Reference Barnes2018; Barnes and Beaulieu, Reference Barnes and Beaulieu2019). Similarly, Wiesehomeier and Verge (Reference Wiesehomeier and Verge2020) suggest that women stand for honesty, but women candidates that are described as being as equally embedded in corruption prone networks as men are seen as just as corruptible as their male counterparts. Interestingly, male candidates may also benefit from being depicted as risk averse (Barnes and Beaulieu, Reference Barnes and Beaulieu2019). Studies also provide more indirect support for the risk aversion theory by showing that women voters punish women candidates more harshly, electorally, for engaging in corruption (Eggers et al., Reference Eggers2018). However, this does not necessarily suggest that voters perceive women as more risk averse, but rather that it is particularly risky for women candidates to be implicated in corruption scandals because they are held to higher standards.

Fewer studies investigate if women candidate advantage can be explained by a social welfare policy profile. Studies suggest that citizens are more likely to perceive that women candidates are more competent in dealing with public service provision that directly affects citizens, such as health care and education (e.g., Alexander and Andersen, Reference Alexander and Andersen1993; Huddy and Terkildsen, Reference Huddy and Terkildsen1993; Dolan, Reference Dolan2010), and that women are also more caring and have a stronger communality compared to men (Ellemers, Reference Ellemers2018).Footnote ⁵ We do not know, however, if campaigning on social welfare issues explains women candidate advantage in fighting corruption in the eyes of voters.

Thus, recent studies provide some direct or indirect evidence for the link between these candidate characteristics and a candidate’s expected capacity to contain corruption. Our suggested shift of dependent variable, from investigating whether candidates are seen as likely to be implicated in corruption scandals to whether candidates are seen as effective when it comes to fighting corruption more broadly, may have implications for preferred candidate characteristics. The numerous risks involved in fighting corruption are well documented (Okonjo-Iweala, Reference Okonjo-Iweala2018), not least since fighting corruption often comes at the cost of gaining powerful enemies among corrupt elites. Voters may, therefore, expect that candidates that take on this task may, in part, benefit from networks and connections to the elite if they also aim to stay in office (cf. Bauhr and Charron, Reference Bauhr and Charron2021). It is therefore far from given that voters will see risk averse candidates as more likely to fight corruption, nor that they will be seen to benefit from having outsider status. Thus, despite the prominence of marginalization theory in explaining why women are excluded from participating in corrupt transactions (e.g., Goetz, Reference Goetz2007; Bjarnegård, Reference Bjarnegård2013) and women being perceived as less corrupt (Barnes and Beaulieu, Reference Barnes and Beaulieu2019; Wiesehomeier and Verge, Reference Wiesehomeier and Verge2020), outsider status is not necessarily seen as a desirable trait for a candidate that sets out to fight corruption.

However, we anticipate that voters prefer candidates that are seen as having the right profile to enact change, but that such candidates also possess the capacity to remain in office while doing so. In other words, if voters are sensitive to the numerous insights and accounts of the important personal and political risks involved in tackling corruption (Persson et al., Reference Persson2013; Okonjo-Iweala, Reference Okonjo-Iweala2018; Bauhr and Charron, Reference Bauhr and Charron2021), and therefore expect that only candidates that are careful and cautious will be able to fight corruption effectively while also remaining in office, risk aversion can be seen as a desirable candidate trait. Furthermore, candidates that campaign on social welfare issues may be seen as more likely to combat corruption, partly because this signals a commitment to improving access to public services and caring about the welfare of citizens (c.f. Bauhr et al., Reference Bauhr2019). Expectations on the role of outsider status when it comes to explaining women candidate advantage are more mixed, however. On the one hand, an outsider may be more willing to challenge insider elites. Yet, being too much of an outsider can also impede a candidate’s capacity to challenge the status quo and also remain in office.

Thus, building on the last few decades of work on the link between gender and corruption, we investigate the extent to which the theoretical frameworks proposed by this literature contribute towards explaining voters’ perceptions of women candidate advantage in combating corruption. If this is the case, we expect the pro-woman bias to be reduced (or perhaps even disappear) when such stereotypes are activated. Conversely, if women candidates are preferred over male candidates, irrespective of what attributes a candidate is assigned on average, this would mean that these theories do not fully explain the advantage of women candidates. Given important evidence on voters perceiving women candidates to be generally more honest and trustworthy, we expect that these frameworks may not necessarily fully explain women candidate advantage.

Thus, we hypothesize:

A sizable literature also propounds that voter stereotypes and expectations about women candidates may vary across contexts as well as respondents. The positive effects of elected bodies more closely mirroring the population from which they are drawn has been widely observed (Phillips, Reference Phillips1998; Mansbridge, Reference Mansbridge1999) and studies suggest that women’s descriptive representation can improve perceptions of women’s ability to govern (Alexander, Reference Alexander2012). Furthermore, although findings are somewhat mixed, several studies confirm that women show a stronger aversion towards corruption than men (see Swamy et al., Reference Swamy2001; Torgler and Valev, Reference Torgler and Valev2010; Rivas, Reference Rivas2013). Relatedly, recent studies investigating gender differences in voter responses to corruption find that women are more likely to refrain from voting for a party and candidate involved in corruption scandals (Alexander et al., Reference Alexander2020). Studies also show that women voters, in particular, may have higher standards for female candidates (Eggers et al., Reference Eggers2018).

Drawing from this literature, we anticipate that women citizens may be more likely to perceive and prefer women candidates as ‘cleaners’. Furthermore, we also anticipate this effect to be particularly strong in contexts of greater political gender equality, where citizens would also be more likely to see women as possessing the necessary characteristics to effectively enact change. Although the notion of whether a critical mass of women is needed for women to effectively enact policies has been debated for decades (Childs and Krook, Reference Childs and Krook2006; Funk et al., Reference Funk2022), voters in contexts where there is a higher share of women representatives may be less likely to hold stereotypical perceptions of women being unfit for leadership or unable to enact change, i.e., to be independent agents rather than token women. Studies also advise that higher gender equality may be associated with larger gender differences in preferences and values (Schwartz and Rubel-Lifschitz, Reference Schwartz and Rubel-Lifschitz2009; Falk and Hermle, Reference Falk and Hermle2018). It is also possible that citizens that live in contexts with a higher share of women representatives may have more experiences with women acting with integrity, and therefore expect more of women candidates.Footnote ⁶ Thus, building on this previous work, we hypothesize that women voters and voters in contexts of higher gender equality will be more likely to believe that women candidates will combat corruption.

Data and experimental design

The three hypotheses are tested using recently collected data from the European Quality of Government Index (EQI) Survey (see Charron et al., Reference Charron2022). The survey was launched in October 2020 and concluded in the first week of February 2021 in all EU27 countries. A total of 129,991 respondents from all EU-27 member countries were included, with 50% being randomly selected and administered by computer assisted telephone interviews (CATI), while the other 50% responded via online administration, with our study using only the online sample (see Table A1 in the online Appendix for the full list and sample size by country). The survey’s main purpose is to gauge citizen attitudes about their local public services (such as health care and education) and the degree to which they rate them as high quality and administered with impartiality and low corruption. There are also a number of questions about citizens’ experiences with corruption, as well as some questions on elections and their own personal partisan preferences, along with standard demographic questions.

We embedded a conjoint survey experiment following the main EQI questions. For practical purposes, the experiment was only given to the online respondents, where a total of 48,080 completed the survey and answered the conjoint experiment without a ‘don’t know’ response, and each respondent was given the experiment twice, equaling 96,160 observations in our main analysis. Smaller countries Malta and Cyprus are not included in our analyses, thus our analyses have up to 25 EU member countries. As online respondents self-select into the survey, they tend to over-sample higher educated and younger respondents, inter alia (see Pew Research, 2018). However, treatment effects between online, convenience and randomly drawn, nationally representative samples tend to be similar (Coppock, Reference Coppock2019). In addition, the EQI’s target sample is between 500–600 respondents at the NUTS 1 and NUTS 2 regional (sub-national) levels within countries,Footnote ⁷ and thus respondents from less populated, rural regions have a much higher probability of being included in the survey compared with respondents from urban regions. We thus employ post-stratification and design weights, which adjust differences between the sample and population by gender, age, education and regional population. More on the survey and the respondents can be found in the online Appendix.

The research design employs a dual profile conjoint experiment (Hainmueller et al., Reference Hainmueller2014). Similar to a factorial vignette, the experiment allows researchers to identify causal effects of several variables of interest concurrently and has been used in several recent publications on the effects of corruption on voting preferences (e.g., Breitenstein, Reference Breitenstein2019; Agerberg, Reference Agerberg2020; Mares and Visconti, Reference Mares and Visconti2020). A key difference in our approach is that rather than testing the degree to which corruption matters in candidate choice by randomizing the presence or absence of a scandal in the candidate profiles, we imply to the respondents that the corruption scandal has already occurred, which allows us to focus on which candidates are most stereotypically seen as cleaners. The respondents are presented with two different mayoral candidate profiles, from which they are asked which candidate they perceive would be best able to combat corruption in their municipality. The question reads as the following:

The candidate profiles consisted of six different factors. The first was the candidate’s gender, male or female. We randomize three factors to capture the mechanism elucidated from the literature – risk aversion, issue specialization, and insider/outsider status. Admittedly, these are not straightforward concepts to operationalize, yet we take a pragmatic approach and make simple binary or categorical variables to capture these concepts to provide respondents with simple contrasts, to test effects from these factors. The ‘risk aversion’ mechanism is operationalized on a line labeled ‘personality type’ in the conjoint set up, where we randomize two traits ‘cautious and careful’ and ‘willing to take risks’. The insider/outsider mechanism is captured on a line titled ‘background’, where a candidate is randomly assigned either ‘well-connected politician’ or ‘political outsider’. As per the issue specialization factor, we randomly assign each candidate with one of four main issues on the line ‘main election issue’ – education, health care, infrastructure or economic development.

Additionally, we include two other factors. The first is candidate age, for which we limit variation to just three years – 51, 52 and 53Footnote ⁹ – as we are less interested in this factor. The final factor is the one in which the literature on corruption and voting highlights as most salient – that of partisanship (Anduiza et al., Reference Anduiza2013; Eggers, Reference Eggers2014). We include a randomized list of the largest 2–5 parties in each country and assign them randomly to each candidate. We then take advantage of the respondent’s previous answer on the question of which party they would support ‘if the election were held today’ and create a variable for a ‘partisan match’ – whereby the variable equals ‘1’ if the candidate belongs to the party that the respondent claimed to support, and ‘0’ if otherwise. This allows us to compare the magnitude of the effects of candidate gender to the most salient heuristic on which voters rely. The full visual of the conjoint table is found in the Appendix, section 2.

With this test, whether gender has an independent effect on candidate choice given these three factors, along with partisanship, are simultaneously accounted for and thus we can compare the relative effects across all treatments simultaneously. The outcome question distinguishes between ‘voice’ (selecting one of the available candidates, H1a) and ‘exit’ (refraining from selecting any of the available candidates, H1b, c.f. Hirschman, Reference Hirschman1970; Bauhr and Charron, Reference Bauhr and Charron2018). We argue that allowing respondents to abstain from choosing any of the candidates is a more realistic scenario and provides more valid estimates (see Haaijer et al., Reference Haaijer2001; Agerberg, Reference Agerberg2020 for more on this point). In the sample, 34.2% selected ‘candidate a’, 34.6% selected ‘candidate b’ while 31.2% selected ‘neither’. The survey was run twice for each respondent, randomizing the candidate profiles each time.

Our empirical strategy of the main conjoint experiment is to estimate the average marginal component effect (AMCE) of the quantities of interest (de la Cuesta et al., Reference De la Cuesta, Egami and Imai2022). The AMCE is conceptually similar to the average treatment effect (ATE) and provides the causal effect of a single candidate characteristic, while averaging over the remaining characteristics. Our estimation relies on the assumptions of the AMCE, which are that there are a) no carryover effects across profiles, b) no profile order effects, and c) profiles are completely and independently randomized (Hainmueller et al., Reference Hainmueller2014). To mitigate any biases from violations for the latter two assumptions, we randomized the ordering of the candidate attributes, and an algorithm was used to completely randomize candidate profiles across and within respondents. While some critique the use of the aggregated AMCE in analyzing electoral preferences in conjoint experiments (Abramson et al., Reference Abramson, Koçak and Magazinnik2022), we are not interested in the electoral results per se, but about anti-corruption stereotypes among the general population. Thus, for H1a and H1b, the AMCE is the best available estimand. H2–H3 require additional estimation over the AMCE.

The additional advantages of this design are that we can estimate two types of interaction effects that are of interest to our theoretical propositions. One, we can estimate the average component interaction effect (ACIE), which tests whether the causal effect of one attribute (in our case candidate gender) varies based on other characteristics, for example a candidate’s background or personality type. In this case, we specify interactions between candidate characteristics of interest to test H2. Two, to test H3 we rely on the conditional AMCE (Hainmueller et al., Reference Hainmueller2014), in which we are interested in the interaction between fixed respondent characteristics (their own gender) and randomized candidate characteristics.

Are the effects driven by context of gender equality? Testing H3b

As noted in the hypothesis, we anticipate that the effect of female candidacy will be strongest in areas where the levels of political gender equality are highest, both because more women in office may improve citizens’ beliefs in women’s ability to govern (Alexander, Reference Alexander2012) and because citizens in these contexts may share beliefs and experiences of women’s capacity to enact integrity reforms. To test H3b empirically, we thus ran several cross-level interactions with candidate gender and variables that serve to proxy for contextual political gender equality, including share of women in sub-national parliaments. As political gender equality varies significantly within countries, we report the cross-level interaction with the sub-national measure to account for the within-country variation of the moderating variable and to increase the number of second-level observations in the model from 25 countries to 200Footnote ¹⁵ .

Figure 5 reports the results of the interaction model, whereby we highlight the preference for ‘candidate A’ on the y-axis as a function of the randomized gender of candidate interacted with the proportion of female MPs in sub-national, regional parliaments as a moderating variable (x-axis). In this case, we observe clear moderating effects of gender equality; where female representation is low – roughly less than 20% – there is no observable effect of candidate gender on voter’s preferences to combat corruption. This constitutes approximately 19% of the sample. Yet at 20% of female representation and higher, there is a significant difference in line with H1 whereby respondents prefer female candidates over male ones. At the highest levels of gender equality there is a roughly 5.1%-point difference in the preference for female candidates over male ones.

Figure 5. Women’s representation and preferences for women candidates.

The results in Figure 5 are robust to the alternative measure of female share of MPs in national parliaments (Figure A2) as well as alternative estimation methods such as hierarchical estimation whereby we specify random regional intercepts and random slope for the candidate gender variable. In Figure A3, we show the AMCE test of H1 by country to further elucidate the degree of context-level effects of the main variable. In sum, we find that respondents prefer women candidates across most countries studied, yet that the effect is insignificant in some of the smaller countries in our sample. In no case do we find that respondents in any country significantly prefer male candidates over females to fight corruption.