Mitigating corruption voting

Under what circumstances do voters sanction corrupt politicians? To begin, standard accounts of corruption voting – where corruption is understood as a misuse of public resources for personal and/or political gains – suggest that less corruption is more desirable (e.g. Besley Reference Besley2007). Thus our first, baseline, hypothesis is:

Existing evidence in support of this basic prediction is mixed. One mitigating factor may be the prevalence of corruption in the wider context. When corruption is widespread, voters may choose to ignore it and focus on other aspects of a politician’s performance or character (e.g. Rose and Peiffer Reference Rose and Peiffer2015). They may also believe that there are few or no clean alternatives on offer (Meirowitz and Tucker Reference Meirowitz and Tucker2013) or that corrupt politicians are more effective than clean politicians (Klašnja et al. Reference Klašnja, Little and Tucker2018):

Other studies suggest the opposite. Informing voters that corruption is widespread may increase the salience of corruption in voters’ minds (Klašnja et al. Reference Klašnja, Tucker and Deegan-Krause2016), a process observed more generally for political phenomena (e.g. Iyengar Reference Iyengar1990):

It is difficult to establish empirically the effect of perceived corruption prevalence with observational data. Voters in higher-corruption contexts may conceptualize corruption differently than voters in low-corruption contexts (Pavão Reference Pavão2018). And voter indifference to corruption may be both a cause and a consequence of its prevalence. Our research design helps to address these challenges.

The second mitigating factor we examine is the provision of side benefits to voters. Corrupt politicians may be forgiven if they share some of the rents with voters (Barberá et al. Reference Barberá, Fernandez-Vazquez and Rivero2016). More broadly, punishment may be lower when incumbent performance is otherwise good (e.g. Klašnja and Tucker Reference Klašnja and Tucker2013; Zechmeister and Zizumbo-Colunga Reference Zechmeister and Zizumbo-Colunga2013):

This expectation is also difficult to establish observationally. Corruption-induced benefits may be systematically different from benefits provided by a clean politician, eliciting distinct reactions from voters. Voters may forgive corrupt politicians during good times but not during bad times because of their preference for ability over honesty rather than because of their willingness to trade off corruption for benefits. Our approach allows us to control these aspects. ^{Footnote 4}

Our survey experiment also benchmarks the two potential factors mitigating corruption voting, by comparing them to two other mitigating factors originating at the individual level: (a) co-partisanship, whereby an individual may be less likely to sanction a corrupt politician from their preferred party (Anduiza et al. Reference Anduiza, Gallego and Muñoz2013; Solaz et al. Reference Solaz, De Vries and De Geus2019) and (b) an individual’s general tolerance for corruption (e.g. Barr and Serra Reference Barr and Serra2010).

Conjoint experiments in the southern cone

We fielded conjoint candidate choice experiments embedded in nationally representative surveys in Argentina, Chile, and Uruguay. All three surveys were fielded between March and May 2017 as part of LAPOP’s AmericasBarometer and include just over 1,500 respondents each. ^{Footnote 5}

We focus on these three countries because they offer useful contextual variation. All three have similar political systems and demographics, allowing us to use similar candidate vignettes. Yet they vary in theoretically meaningful ways: Argentina and Chile have lower levels of mass partisanship but higher levels of corruption perceptions than Uruguay, while bribery is considerably more frequent in Argentina than it is in Chile or Uruguay. Given such variation, to the extent that we find similar results across these different settings, we can be more confident that they are not just unique to a particular context.

In our experiment, we presented survey respondents with a short vignette about two hypothetical mayoral candidates, an incumbent and a challenger, running in a local election. Within the text of the vignette, we randomly varied six characteristics of the candidates and the electoral environment: (1) candidate gender, (2) party affiliation (left party, right party, or independent), (3) corruption record (accused of taking bribes or praised for efforts to stamp out bribery in their administration), (4) the information source for the bribery allegation/praise (left or right newspaper, or judicial officials), (5) a potentially mitigating corruption factor (corruption prevalence or the creation of construction jobs; applicable only when a candidate is accused of corruption), and (6) the state of the economy (improved or worsened since the last election; applicable only to the incumbent). We randomized each attribute independently for each candidate, allowing us to simultaneously estimate the causal effect of each characteristic (Hainmueller et al. Reference Hainmueller, Hopkins and Yamamoto2014). After showing respondents the vignette, we asked our key outcome question: “If you had to choose between these two candidates, for whom would you vote?” ^{Footnote 6}

When do voters sanction corrupt politicians?

To estimate treatment effects, we treat each hypothetical candidate as a unique case (i.e. there are two candidates for every respondent), following Hainmueller et al. (Reference Hainmueller, Hopkins and Yamamoto2014) (standard errors are clustered by respondent). We then estimate OLS models relating respondents’ choices to indicator variables for each treatment. ^{Footnote 7} We pool estimates across the three countries and include country dummies in our specifications. ^{Footnote 8} Since the economic performance attribute only applies to the incumbent candidate, we also control for incumbency and an interaction between economic performance and incumbency. ^{Footnote 9}

The top estimate in Figure 1 is strongly consistent with H1: corruption causes a large drop in respondents’ support, from about 53% for clean candidates to 18% for corrupt candidates – a 65% reduction. ^{Footnote 10} The effects of the economy and co-partisanship are also not surprising: a poor economy decreases the probability of support; ^{Footnote 11} belonging to a respondent’s preferred party increases it. ^{Footnote 12}

Figure 1 Conjoint experiment results.

Note: Values represent the difference in respondents’ propensity for supporting a hypothetical candidate based on each vignette characteristic. Lines represent 95% confidence intervals estimated using standard errors clustered by respondent. Estimates are based on OLS regressions reported in Supplementary Table A4.

While the corruption effect is sizable, it is similar in magnitude to other experimental studies from the region (for a review, see Boas et al. Reference Boas, Hidalgo and Melo2019). We suspect that the smaller magnitude of the partisanship effect is a consequence of weak partisanship, particularly in Argentina and Chile, where the effect is close to zero and statistically null (see Supplementary Figure A2). ^{Footnote 13} Corruption voting also seems to be large relative to the effect of the economy, in contrast to some previous findings (Klašnja and Tucker Reference Klašnja and Tucker2013). But it is difficult to know how respondents envisioned the improved and worsened economic conditions signaled in the vignette, or how the magnitudes of these economic changes compared to the magnitude of corruption. Probing these contrasts is an interesting question for future research.

Prevalence and Side Benefits as Mitigating Factors

In Figure 2, we evaluate the evidence for hypotheses H2a (and H2b) and H3 on the mitigating effects of corruption prevalence and side benefits. As benchmarks, we also examine the mitigation due to co-partisanship and tolerance of corruption.

Figure 2 Contextual and individual factors mitigating corruption voting.

Note: Values represent the difference in respondents’ propensity for supporting a hypothetical candidate based on each vignette characteristic. Lines represent 95% confidence intervals estimated using standard errors clustered by respondent. Estimates are based on OLS regressions reported in Supplementary Table A4. Brackets list the difference between effects. *p < 0.1, **p < 0.05, ***p < 0.01

Informing respondents that corruption is widespread in a candidate’s province (the treatment we refer to in the Figures as “bribes common”) did nothing to mitigate the sanctioning of corruption: the difference between the top two values, the basic corruption treatment, and the prevalence treatment, is not statistically significant. This evidence is thus inconsistent with H2a and H2b. ^{Footnote 14}

Information about jobs created through corruption (the treatment we refer to in the Figures as “bribes but jobs”), however, noticeably mitigates the sanctioning of corruption: the difference between the first and third values, the basic corruption treatment and the side benefits treatment, is positive and statistically significant. The mention of corruption reduces the likelihood of voting for a candidate by 36 percentage points, but the side benefit of jobs decreases the corruption penalty from 36 to 27 percentage points (a 25% decrease). This mitigating effect is present in all three countries and to a very similar extent (Supplementary Table A7), bolstering our confidence in uncovering a general pattern.

The second and third sets of values in Figure 2 show the effects of co-partisanship and corruption tolerance. We do not observe co-partisan bias in the propensity to punish corrupt politicians (the difference between the two values under Co-partisan bias); quite the opposite: respondents were on average more likely to punish co-partisan corrupt candidates, by nearly 11 percentage points. ^{Footnote 15} This result is less surprising, however, given that partisan attachments are fairly weak in our three countries.

On the other hand, citizens who are more tolerant of corruption are on average less likely to sanction corrupt politicians. To measure corruption tolerance, we use a binary survey item (asked before the experiment) that measured the extent to which respondents find it justifiable to pay a bribe (see Section A5 of the appendix for wording). The extent of mitigation produced by corrupt side benefits is virtually identical in size to that arising from individuals’ tolerance of bribery. Note that these two mitigating factors appear to be additive: when the corrupt jobs treatment is interacted with individuals’ bribe tolerance, the extent of mitigation is essentially doubled and statistically significant.

Conditions Amplifying the Mitigating Effect of Side Benefits

Figure 3 explores the conditions that may make the mitigating effect of corruption’s side benefits particularly pronounced. In our vignette, the corrupt side benefit takes the form of construction jobs, typically positions held by individuals with lower levels of education and wealth. ^{Footnote 16} We indeed find that such respondents are more willing to trade off corruption for jobs than other respondents, as shown in Figure 3. The mitigating effect of side benefits on sanctioning corruption is six percentage points higher among respondents with lower levels of education, a result that is in line with Truex (Reference Truex2011). And it is five percentage points higher among less wealthy respondents. ^{Footnote 17} Both of these effects persist even when controlling for the other characteristic.

Figure 3 What conditions amplify the mitigating effect of side benefits?.

Note: Values represent the difference in respondents’ propensity for supporting a hypothetical candidate based on each vignette characteristic. Lines represent 95% confidence intervals estimated using standard errors clustered by respondent. Estimates are based on OLS regressions reported in Supplementary Table A4. Brackets list the difference between effects. *p < 0.1, **p < 0.05, ***p < 0.01

Other measures of economic vulnerability generally seem to amplify the mitigating effect of side benefits on corruption sanctioning, though with less statistical precision. Supplementary Figure A6 shows that economic downturns (based on our experimental manipulation) and being unemployed correlate positively with accepting corruption in exchange for construction jobs. Overall, these findings suggest that beneficiaries from the side benefits of corruption are somewhat less likely to sanction it, although the mitigating effect of corrupt side benefits is quite broad, applying to a variety of scenarios and respondents. ^{Footnote 18}