The Greater Effects of Sexual Harassment: A Conjoint Analysis Assessing Variation in Political Scandal Effects on Candidate Evaluations in the UK

The definition of valence issues is in fact divided into two categories: “issue ownership” (policy-based) and “character-based” (non-policy based) (Curini, 2015). The former indicates how a politician’s ability and expertise is in dealing with a certain policy issue (Clarke et al., 2004); the latter focuses on politicians’ character-based images, often referring to perceptions of trustworthiness, integrity, etc. (Clark, 2009). Throughout this article, we align the concept of valence issue with character-based images, including but not limited to honesty, competence, charisma, unity, and so on (Adams, 2012).

To that point, Welch and Hibbing (1997) code scandals into the following categories: campaign violations, bribery, “other crimes,” conflict of interest, abuse of congressional prerogatives, moral charges, and a residual “other” category.

Please refer to online Appendix for the distribution of in-lab and online respondents’ demographic characteristics.

In other words, we cluster our standard errors by respondents given that each individual respondent appears twelve times—twice per round of the experiment (6) given that during each of the six rounds candidate pairings are selecting between two (2) hypothetical candidate pairings (6 rounds × 2 candidates evaluated per round = 12 candidate profile evaluations/candidate-choice selections)—in our analysis of candidate choice. Of course, respondents only select one of the two candidates per each round, thus articulating one candidate preference per each of the six rounds.

There are more women, Labour identifiers, EU supporters, and young persons among sampled participants. Please refer to the online Appendix for details.

In fact, the respondents were asked to make choices over three different sessions, and in each session, every respondent is forced to make decisions between two candidates six times. Given that the number of variables increases with sessions, the number of treatments and the level of complexity of this design inevitably increases, too. The analysis of this paper does not include the results from the latter two sessions; nevertheless, what we have found is that sexual harassment has the largest negative effect on voters’ preferences to candidates across all sessions, and the estimates of the same variables across all sessions are similar, too. The demonstration of consistent estimates across different sessions with different levels of complexity alleviates possibilities that the largest effect of sexual harassment is due to priming effects. We apologize that we are not allowed to provide the results from the latter two sessions as proof to reviewers at this moment given that those results have been published elsewhere. We are pleased to send those results to the editor upon request.

Full results are articulated in manuscript’s Appendix.

Note that according to Leeper et al. (2020), the point estimates for sexual harassment (and for other types too) through AMCE is not its “absolute effect” on how voters evaluate candidates, and therefore, the difference between the estimates of sexual harassment and no scandal is not “absolute difference of effects.” The estimates in Fig. 3 (and Fig. 4 too) only reflect the relative changes among the subgroups in the “percentage points of voting for a candidate when that candidate has been accused of sexual harassment relative to the subgroup’s probability of voting for the candidate in the counterfactual reality where that candidate has no scandal.”.

Indeed, for analytical leverage and consistent with our theoretical framework, we collapse our partisanship measures to code this dichotomous pairing of shared partisanship, coded as “different” if the voter-candidate pairing are opposing-partisans or coded “same” if the voter-candidate pairing are congruent co-partisans. Full model results can be found in manuscript’s Appendix.

As a robustness check, we also assess this hypothesis within the context of shared and diverging genders. This robustness check, found in the Appendix, compliments the forthcoming results.

In other words, we specify a model interacting each scandal-type candidate attribute with candidate attribute gender and run the model on data subsetted by respondent gender. The top panel of Fig. 4b shows these interactive effects for female experimental respondents while the bottom panel shows the same interactive effects for male experimental respondents.