Corrupted Estimates? Response Bias in Citizen Surveys on Corruption

Researchers have long acknowledged that survey responses are sometimes unstable and inaccurate, and that there are clear psychological incentives to shape responses in certain ways (Berinsky 1999; Bertrand and Mullainathan 2001; Zaller 1992). In particular, research on citizens’ evaluations of the economy has uncovered strong response biases stemming from the respondent’s political leanings (Evans and Andersen 2006; Lau et al. 1990; Palmer and Duch 2001; Sears and Lau 1983; Wilcox and Wlezien 1996).

An early strand of research argued that this type of bias can arise because party identification works as a ‘perceptual screen’ through which economic performance is assessed, and that an individual tends to see what is favorable to his or her partisan orientation (Campbell 1960). Later research shows that even beliefs about basic objective political facts can be strongly affected by political identification, partly due to selective information processing (Bartels 2002; Gerber and Huber 2010). Respondents might also reason their way to the conclusion that the economy is doing better when their preferred party or politician is in power. The theory of motivated reasoning holds that all reasoning is motivated in the sense that it is driven by specific motives and goals. Taber and Lodge (2006) argue that these goals often are directional goals (as opposed to accuracy goals) where individuals apply their reasoning powers in defense of a prior specific conclusion. Directional goals are thus often defensive of particular identities, attitudes, or beliefs that are strongly held (Flynn et al. 2017; Leeper and Slothuus 2014).

In political science a large body of research has focused on political bias in economic perceptions. Much less is known about how political affiliations might interact with corruption perceptions. Recent decades have seen a rapid increase in efforts to measure corruption, both via expert surveys and surveys of the public (Fisman and Golden 2017; Holmes 2015; Wysmulek 2019). Many of the survey instruments used in the latter category clearly resemble the instruments used to tap into people’s economic perceptions; Klasnja et al. (2016) even adopt the terms ‘sociotropic’ and ‘egotropic corruption voting’ directly from the economic voting literature. Researchers use such measures to study, for instance, how individual corruption perceptions and experiences are related to incumbent support and vote choice (Agerberg 2017; Gingerich 2009; Klasnja et al. 2016; Xezonakis et al. 2016; Zechmeister and Zizumbo-Colunga 2013), and how corruption perceptions relate to political legitimacy and support for the democratic system (Anderson and Tverdova 2003; Dahlberg and Holmberg 2014; Morris and Klesner 2010; Seligson 2002). To be able to estimate the effect of corruption in such settings it is important that corruption perceptions are (exclusively) determined exogenously, to avoid biased estimates. In this sense, I define ‘political bias’ as a situation where respondents’ reported corruption perceptions instead are partly endogenous to their political identity.

Researchers interested in the effects of corruption would hence hope that these perceptions are only determined by external changes in an individual’s environment. However, recent studies give us reasons to believe that this might not be the case. Anduiza et al. (2013) show that tolerance for corruption can have a clear political dimension. In a survey experiment fielded in Spain the authors show that respondents’ judgment of the seriousness of a political corruption scandal partly is determined by whether the accused politician belongs to the respondent’s preferred party. The authors argue, in line with the literature on economic perceptions, that this is a way for respondents to reduce cognitive dissonance: by downplaying the importance of corruption when it affects the own party respondents make the political world more consistent with their political predispositions.

Jerit and Barabas (2012) show that individual-level motivated bias is present on a wide range of topics, as long as a question has importance or strong political implications (see also Bolsen et al. 2014). The authors show that such bias is particularly likely for issues receiving high media coverage. I argue that corruption perceptions are likely to be affected in this way: people—especially in developing democracies—view corruption in society as a question of great importance (Fisman and Golden 2017; Holmes 2015). For instance, when the World Economic Forum in 2017 surveyed individuals in 186 countries about the most pressing political issue ‘government accountability and transparency/corruption’ ranked 1st (World Economic Forum 2017). About 25% of Europeans say that they are ‘personally affected by corruption in their daily lives’; the number for countries like Romania, Croatia, and Spain is as high as 60–70% (Eurobarometer 2017). In such countries, corruption trends in society and both local and high-level scandals are often intensely covered by the media (see Hajdu 2018). Voters in countries where corruption is widespread also tend to associate current levels of perceived corruption with the incumbent government (Klasnja 2015; Klasnja et al. 2016; Xezonakis et al. 2016), and view ‘the fight against corruption’ as one of the priorities that should be most important for political leaders (Holmes 2015). Government supporters therefore have a ‘preferred world-state’ where corruption levels are decreasing (this supports their political leanings), while opposition supporters have incentives to view the situation as worse (this would be a reason to oust the current incumbent) (Jerit and Barabas 2012). As suggested by the literature on motivated reasoning, voters who sympathize with the government (for whatever reason) might therefore convince themselves that the situation with regard to corruption is more positive than what is warranted by evidence.

In general, a connection between reported corruption perceptions and government support can exist for two reasons: (1) the respondent might experience changing corruption levels in society and adjust his or her support for the incumbent accordingly, (2) the respondent reports perceived corruption levels that are consistent with his or her political affiliations. If the latter is true, making political affiliations more salient should affect reported corruption perceptions, whereas if corruption perceptions are only determined exogenously this should not be the case. In line with the literature reviewed above, I argue that we have reasons to believe that some degree of political bias is present in the reporting of corruption perceptions. In this sense, a respondent’s reported corruption levels can be a way to defend and justify his or her beliefs about the current government. I focus specifically on government supporters since this group is relatively easy to define, also in a multiparty system (I discuss this in more detail below). Drawing on this discussion, my first hypothesis has two parts and can be stated as follows:\(H_{1a}\) predicts an empirical pattern that is more likely to be true if the political bias (PB) hypothesis is true. Finding this pattern does not, however, provide definite evidence for the PB hypothesis; the pattern could, in theory, arise exclusively because people with low corruption perceptions become government supporters. Here, \(H_{1b}\) instead provides a direct test of one of the main implications of the hypothesis.

According to Tourangeau and Yan (2007) a survey question is likely to be ‘sensitive’ if it touches on ‘taboo’ topics, if it induces concerns that the information given will become known to a third party, or if the question elicits answers that are socially unacceptable or undesirable. If this is the case the respondent can be expected to give a ‘socially desirable’ answer. That is, an answer that the respondent thinks will be viewed favorably by others, resulting in under-reporting of ‘undesirable’ attitudes and behavior. Such sensitivity bias (SB) has been shown to be present on a wide range of topics based on self-reports, from questions about drug use (Fendrich and Vaughn 1994) to questions on voter turnout (Holbrook and Krosnick 2010).¹³

Surveys based on self-reported experiences are also common in corruption research. These so called ‘experiential surveys’ are one of the most direct methods for gauging the amount of corruption in society, by simply asking citizens about their experiences of corruption (e.g. “Have you in the past 12 months had to pay a bribe in any form?”). The method is now widely deployed by several large organizations in multi-country surveys (Holmes 2015), including, for instance, the Global Corruption Barometer. Should we expect citizens to truthfully report their first-hand experiences with corruption and bribery?

Kraay and Murrell (2016) argue that some respondents might be ‘reticent’ to tell interviewers about their experiences with corruption. The authors develop a statistical model of potential under-reporting that shows that the true rate of bribe payment in ten developing countries might be substantially higher than the conventional estimates, based on direct questions. One reason that direct questions might be sensitive is that bribery—in the sense of an official demanding a bribe for a service—is illegal in essentially every country in the world (Fisman and Golden 2017, pp. 26–29). In an overview of the research on sensitive questions Krumpal (2013) identifies several studies reporting substantial SB on topics related to criminal behavior and crime-victimization. Admitting to having been part of a corrupt exchange (for example, paying a bribe) is to admit part in an illegal transaction, and in the light of this something that could be considered sensitive. Even a question about whether an individual has been asked to pay a bribe should be sensitive, given that an affirmative answer signals that the individual potentially also completed the corrupt transaction.

Moreover, corruption is something that people find morally reprehensible. This is true even in countries where corruption is ubiquitous (Karklins 2005; Persson et al. 2013; Rothstein and Varraich 2017). World value survey (WVS) has been asking about people’s attitudes towards bribery in several waves and respondents all over the world typically show a strong distaste for corruption.¹⁴ Figure 1 shows data from Romania—one of the most corrupt countries in Europe—based on the most recent WVS wave for this question. It is clear that there exists a very strong norm against bribery, even in this context where corruption is widespread: over 80% of respondents say that accepting a bribe can never be justified.

Given this norm, it is hence reasonable that people would view admitting to being part of a corrupt transaction as something that is ‘socially undesirable’ (Tourangeau and Yan 2007). The topic of vote-buying has been extensively studied and has obvious similarities with the case of bribe taking: both acts are illegal transactions that are generally stigmatized. Traditionally, studies on vote-buying asked direct questions about its occurrence and found limited evidence. In recent years researchers have acknowledged that such direct questions might be sensitive and instead considered more sophisticated survey methods. As a result, several more recent studies have discovered substantial under-reporting of vote-buying due to SB (Carkoglu and Erdem Aytac 2015; Corstange 2012; Gonzalez-Ocantos et al. 2012). Important to note is that severe under-reporting has been found even for questions asking respondents whether someone ‘offered’ them to sell their vote—and not only for question asking if they actually sold their vote. Given these similarities to vote-buying, the illegality of bribery, and people’s almost unanimous distaste for its occurrence, I argue that we should expect a similar pattern with regard to reported corruption experiences:

To test these hypotheses about political bias (\(H_{1}\)) and sensitivity bias (\(H_{2}\)) in corruption reporting I conducted a large survey in Romania, with two embedded survey experiments. Romania is one of the most corrupt countries in Europe, where the problem is very much a current issue. After the legislative elections of 2016 the Social Democratic Party (PSD) and the Alliance of Liberals and Democrats (ALDE) formed the governing coalition. Following massive protests in 2017 against a bill that (among other things) would have pardoned officials imprisoned for bribery offenses (see The New York Times May 4th, 2017), the government resigned and was replaced by a second iteration of the coalition. After an internal power struggle in PSD a third iteration of the PSD–ALDE coalition took office in January 2018. Due to the political turbulence, and partly related to accusations of corruption within the government, the public support for the coalition decreased over the course of 2018. In the end of 2018 many opinion polls showed a support of around 35% for the governing coalition.¹⁵

Similar situations, with strong politicization of corruption, are not uncommon—not least in other eastern European countries (Bågenholm 2013). Researchers even talk about an incumbency disadvantage in many developing democracies, where holding office seems to decrease chances of reelection. Such an effect has been demonstrated in, for instance, Brazil (Klasnja and Titunik 2017), India (Uppal 2009), and post-Communist Eastern Europe (Roberts 2008). Klasnja (2015) shows, in a study of Romanian mayors, that one plausible explanation for this pattern is corruption, where office holders exploit their position to reap private gains—at the cost of subsequent electoral success.

The turbulent situation during the past years in Romanian politics is thus rather typical for many young democracies (e.g. Eastern Europe, Latin America), a group of countries that also tend to struggle with widespread corruption in politics and in society overall (Keefer 2007). This makes Romania a relevant case to study. Moreover, partisanship has generally been relatively weak in Romania in the post-Communist period (Tatar 2013), a pattern that is common in developing democracies. This makes the case a relatively tough test for the political bias hypothesis. Any effects found in a context like this, with weak partisanship, are likely to be more pronounced in contexts where partisanship is stronger. The results from the study at hand should hence be of high relevance for developing democracies in Europe and elsewhere, where questions about governance are on the political agenda and politicization of corruption is common.

The aim of the study is to assess each hypothesis in turn with two different research designs. Codings and analysis-strategies were pre-specified in the pre-analysis plan. To test the political bias hypothesis the first design exploits question order effects (Bertrand and Mullainathan 2001). With regard to economic voting, researchers have shown that question order effects can be substantial (e.g. Wilcox and Wlezien 1996). Sears and Lau (1983) argue that two such effects are common: political preferences might be personalized when assessed immediately after the respondent’s own economic situation has been made salient, or perceptions of the economic situation might be politicized when assessed immediately after important political preferences have been made salient. Given my hypotheses in this paper I focus on the latter. Questions about corruption perceptions subsequent to the political questions are assumed to exhibit stronger politically biased response patterns since asking the political questions make the respondent’s political identity more salient (Klar 2013). In this sense, the question ordering activates a particular political ‘frame’ around the corruption questions (Zaller 1992). If my hypothesis about PB is correct, ‘politicizing’ corruption perceptions in this way will significantly affect the response to these questions.

Following Evans and Andersen (2006) I asked the following political questions: (1) What political party would you vote for if the national parliamentary election were today? (2) Please choose one of the following phrases to say how you feel about the current government of Romania: ‘Strongly against’, ‘Against’, ‘Neither in favor nor against’, ‘In favor’, ‘Strongly in favor’. Based on these questions I coded respondents as government supporters if they said that they would vote for one of the parties in the current ruling coalition in Romania (PSD and ALDE) and answer that they are ‘Neither in favor nor against’, ‘In favor’, or ‘Strongly in favor’ of the current government in Romania. This way I identify government supporters in terms of vote intention, but exclude respondents that state that they are ‘against the government’ from the definition. I consider different coding decisions with regard to this variable below.

To measure corruption perceptions I asked three different questions that are commonly used in the literature and that are of theoretical interest. The questions are regularly asked in multi-country surveys like the GCB and the ISSP. Together, the three questions give a reasonably comprehensive picture of how the respondent perceives current corruption in Romania, both in terms of absolute levels and in terms of recent change. The questions ask about whether the respondent perceive that corruption has increased during the past year (corruption increase), how many politicians that are involved in corruption (corruption in politics), and how worried the respondent is about the consequences of corruption for society (corruption worry). I also included a standard question about economic perceptions as a point of comparison (economy worse). All outcome questions were coded so that high values indicate ‘bad’ outcomes; increased corruption, worsened economy, high political corruption, and high worry about corruption. See the Supplementary Materials for more details about these questions.

The effects of interest with regard to the corruption/economy questions can then be estimated with a simple regression model:where \(y_{i}\) represents the outcome variable of interest, \(x_{i}\) is an indicator variable equal to 1 if a respondent is a government supporter and 0 otherwise, \(T_{i}\) indicates if a respondent is in the treatment group (\(T_{i}=1\)) or the control group (\(T_{i}=0\)), \((x_{i} \times T_{i})\) is an interaction term including \(x_{i}\) and \(T_{i},\) and \(\epsilon _{i}\) represents the error term. The treatment, again, consists of the intervention of ‘priming’ respondents with their political preference before answering one of the three corruption questions. In the interest of space I simply refer to ‘corruption perceptions’ as a catch-all term referring to all three questions (change, level, and worry—coded in the way described above). As per \(H_{1a},\) I expect \(\beta _{1}\) to be \(<0\) (on average, government supporters perceive corruption to be lower) and, as per \(H_{1b},\) I expect \(\delta \) to be \(<0\) (the effect of the prime is negative for government supporters—that is, government supporters report even lower perceived corruption when their political preference has been made salient).¹⁶ I consider a confirmation of these expectations for all three corruption outcomes to be strong evidence in favor of \(H_{1a}\) and \(H_{1b}.\) I consider a partial confirmation of the expectations (finding significant results in the expected direction for one or two of the outcomes) to be somewhat weaker evidence in favor of \(H_{1a}\) and \(H_{1b}.\)¹⁷

To facilitate interpretability of the results I first estimate Eq. (1) using OLS with robust standard errors as the baseline model. The Supplementary Materials include several additional robustness checks.

Kraay and Murrell (2016) develop a method to estimate under-reporting of corruption that is based on modeling respondents’ answers to a combination of a direct corruption question and several ‘random response’ questions (see Blair et al. 2015). While the approach definitely has strengths, it also relies on fairly strong assumptions about which respondents that actually are ‘reticent’ (and hence would provide an untruthful answer to the sensitive question). Moreover, the approach also involves asking ten different ‘random response’ questions, each involving the respondent flipping a coin (see Azfar and Murrell 2009 for an earlier version of this approach).

To test the sensitivity bias hypothesis I instead deploy a list experiment, which was implemented in the middle of the survey.¹⁸ This is a conveniently simple survey method, previously used to estimate the prevalence of sensitive behavior like drug abuse, cheating, and vote buying, where the respondent does not have to directly disclose any information about the sensitive item (see Glynn 2013).¹⁹ The list experiment works by aggregating the sensitive item with a list of non-sensitive items so that the respondent only has to indicate the number of items that apply and not which specific items that are true. To implement this design, I asked the respondents to do the following:

Here is a list with different things that you might have done or experienced during the past 12 months. Please read the list carefully and enter how many of these things that you have done or experienced. Do not indicate which things, only HOW MANY.

• Attending a work-related meeting;
• Investing money in stocks;
• Being unemployed for more than 9 months;
• Discussing politics with friends or family.

The treatment group was shown the same list but with a fifth item added (the item-order was randomized for all lists):

• Being asked to pay a bribe to a public official.

The design protects the respondents’ privacy since as long as respondents in the treatment group answer with anything less than “five”, no one directly admits to answering affirmative to the sensitive question (having been asked to pay bribe). Following the advice in Glynn (2013) the control items were chosen to be negatively correlated to avoid floor and ceiling effects (where respondents would select either 0 or all items). As shown by Blair and Imai (2012), if we assume that the addition of the sensitive item does not alter responses to the control items and that the response for each sensitive item is truthful, then randomizing respondents into the treatment and control groups allows the analyst to estimate the proportion affirmative answers for the sensitive item by taking the difference between the average response among the treatment group and the average response among the control group (i.e. a difference-in-means estimator).²⁰

By asking the sensitive question directly to the control group (who did not receive the sensitive item on their list) I can also model the amount of sensitivity bias by comparing the direct question with the estimated proportion of affirmative answers to the sensitive item in the list experiment. For the direct question I asked: In the past 12 months were you at any point asked to pay a bribe to a public official? The answer alternatives given were ‘yes’, ‘no’, and ‘prefer to not answer’. I coded affirmative answers as 1 and other answers as 0.²¹

For the basic analysis of the list experiment I rely on the linear estimator in Imai (2011), corresponding to a standard difference-in-means estimator. To estimate the overall level of SB I use the procedure described in Blair and Imai (2012) and compare the predicted response to the direct question, modeled with a logistic regression model, to the predicted response to the sensitive item in the list experiment.²² The predicted response to the sensitive item in the list experiment can then be compared to the response to the direct question to get an estimate of the amount of SB (including 95% confidence intervals for the estimates, obtained via Monte Carlo simulations). I consider a SB estimate that is positive and statistically different from 0 to be evidence in favor of \(H_{2}.\)

To check for heterogeneity in SB I also perform exploratory analyses with regard to the following variables (the variables were identified based on previous research; see discussion below): Government supporter, Gender, University degree, Big city inhabitant, Age, High-income household (top 20% of the distribution in the data set). For the exploratory analyses I rely on the non-linear least squares (NLS) estimator developed in Imai (2011) (see Supplementary Materials).

After a pilot study was conducted to test the questions in the survey as well as one of the assumptions underlying the PB experiment (see Supplementary Materials), the final survey was fielded between 19th of December 2018 and 24th of January 2019 in collaboration with the public opinion research company Lucid.²³ This is a survey respondent aggregator that uses quota sampling to target a nationally representative sample. Research from the US find that experimental treatment effects on Lucid match well with effects found in US national probability samples (Coppock and McClellan 2019). Based on two series of power analyses (see Supplementary Materials) the target number of respondents was set to at least 2900. The sample was collected based on nationally representative quotas on gender, age, and region.²⁴ 3027 Respondents in total completed the survey. Descriptive statistics for the sample are available in the Supplementary Materials.

I start by evaluating \(H_{1}\) (a and b). The unpopularity of the current government is reflected in the survey: about 24% of the sample said they would vote for a party in the governing coalition if the national parliamentary election were today. PSD is still the most popular party in the sample, but its share of the total vote decreases as many respondents indicated that they would ‘not vote’. The share true ‘government supporters’ according to the definition above is smaller, at about 14%. While this group is relatively small, it still contains a large number of respondents (about 410) given the large overall sample. Below I also consider alternative ways of coding the ‘government support variable’ that utilizes the sample in a different way.

The PB hypothesis predicts that government supporters, on average, should have a more positive view of corruption in Romania, and that this group should report an even more positive view when primed with their political affiliation. To test this, I estimated Eq. (1) for each of the four outcome variables (the three corruption variables + the economy variable), using OLS.²⁵ The results are reported in Table 1.

The coefficient for Government support shows the baseline difference between government supporters and others for the control group. That is, in the group that answered the outcome questions before the political questions. The first two models show the results for the outcomes corruption change and economy change—the two outcome questions that are the most similar in terms of structure. The results for these outcomes are also very similar: government supporters in general place themselves about one category lower (in the direction of less corruption/better economy) than the rest of the respondents. The corresponding coefficient for the last other two outcomes are somewhat smaller, but still highly significant (\(p<0.001\)). Overall, this is in line with \(H_{1a}\): government supporters report a much less negative view of corruption in Romania in general and say that they are less worried about the problem.

The interaction effect (Gov. support \(\times \) Prime) estimates the effect of the ‘political prime’—e.g. being asked about political affiliation before the corruption questions, rather than the other way around. As shown in the table, the effect is large. For the corruption increase outcome the difference between government supporters and others increases from 0.9 in the control group to about 1.6 (\((-0.89)+(-0.66)\)) in the treatment group (\(p<0.001\)). The pattern is, again, similar to that for the economy worse outcome where the difference increases from 1.2 to 1.7 (\((-1.19)+(-0.47),\) \(p<0.01\)). In both cases are government supporters substantially more positive (or less negative) to begin with, and become even more positive when randomly assigned to the political prime.

The last two outcomes show the same pattern: government supporters think corruption in politics is lower and worry less about corruption, a difference that becomes significantly more pronounced with the political prime (both interaction effects are statistically significant at the 0.05-level). In this experimental condition government supporters answer on average about 0.8 to 1 categories lower. To graphically display the results, predicted responses for all four outcomes are shown in Fig. 2.²⁶

The Supplementary Materials include several robustness checks, including alternative codings of the supporter variable. While the robustness checks in general corroborates the main results, they also confirm that the ‘prime effect’ for the outcome corruption in politics is slightly weaker overall.²⁷

I now turn to the SB hypothesis. As argued above, it is reasonable to assume that the often used direct question about bribe experience is sensitive and hence under-reported. To test this, I started by estimating the proportion of affirmative responses to the sensitive item in the list experiment using the basic difference-in-means estimator (Glynn 2013). I then estimated a logistic regression intercept-only model with the responses to the direct bribe question as the dependent variable. The two estimates show what proportion of respondents giving an affirmative answer when their privacy is protected (in the list experiment) vs when their privacy is not protected (the direct question). I also used the procedure in Blair and Imai (2012) to compute 95% confidence intervals for the estimates and the difference between them. The results are presented in Table 2 and displayed graphically in Fig. 3.

The direct estimate of 19% ‘yes’ is very close to the reported statistic in the 2017 Eurobarometer for Romania at about 18% (Eurobarometer 2017). This stand in stark contrast to the list estimate at over 35%. The difference of more than 16 percentage points is highly statistically significant. This is clear evidence that respondents under-report the sensitive item when asked directly and suggests that the true estimate might be 90% higher than the estimate based on the commonly used bribe question.²⁸ As noted above, these estimates are based on the ‘less sensitive’ version of the bribe question (the other version asking if the respondent actually paid the bribe), and are also based on a survey mode that should be less likely to elicit SB (online survey).

Voters under-reporting their experiences with corruption is obviously a serious problem for researchers or organizations trying to estimate the occurrence of bribery based on direct questions. However, if this sort of measurement error is randomly distributed across the population it would still be possible to use direct questions to explore the dynamics of bribery and assess which individuals or groups that are most likely to be asked to pay bribes. This is for instance done in Mocan (2004). To explore if the extent of under-reporting differs between groups I used the NLS estimator to model the relationship between respondent characteristics and responses to the sensitive item (Blair and Imai 2012), based on the six variables described above. I also estimated a logistic regression model regressing the direct bribe question on the same variables. Comparisons between the direct estimate and the list estimate based on these models are shown in Fig. 4. The figure displays the results based on the variables government supporter, gender, and income. In the interest of space, the results for the variables age, city inhabitant, and education are presented and discussed in the Supplementary Materials.

Figure 4 reveals interesting differences in under-reporting among different subgroups. The left-hand graph indicates that government supporters tend to severely under-report the sensitive item. When asked directly, under 9% of government supporters say that someone asked them to pay a bribe, compared to the list estimate at 58%. Given the relatively small size of this group the point estimate from the list experiment needs to be taken with a grain of salt, given the substantial uncertainty around the estimate.²⁹ The results do suggest, however, that under-reporting is huge among government supporters. This is completely in line with both the SB and PB hypothesis: government supporters might under-report the sensitive item to make their group look better and to depict overall corruption levels as lower (Carkoglu and Erdem Aytac 2015 find a similar pattern with regard to vote buying in Turkey).

It has long been noted that women seem to be less involved in corruption than men. In an overview of the literature, Heath et al. (2016, pp. 67–72) find that women have lower risk of being asked for a bribe in all nine studies reviewed. Some have argued that one reason for this might be that women simple have fewer opportunities to engage in corrupt activities and that women get asked to pay bribes less often than men (e.g. Goetz 2007; Mocan 2004). This is also the pattern shown in the direct estimate of about 13% for women and 21% for men. The list estimates, however, suggest the opposite pattern; when using the indirect questioning method women seem to be asked for bribes more often than men. The list estimate for women is over three times as high as the direct estimate—43% vs 13%. This result is interesting, given that it goes against what much of previous research has argued. At this point I can only speculate about the reasons behind this pattern. One possibility is that women as a group are more affected by sensitivity bias.³⁰ The higher list estimate could reflect the fact that women utilize the health care sector more than men, and that this sector, according to many estimates, is the sector most permeated by corruption (see Eurobarometer 2014, 2017).

Finally, Mocan (2004) argues that we should expect income to be positively related to bribe victimization, given that it should be possible for a rent-seeking official to extract higher bribes from a wealthier individual (see also Heath et al. 2016, pp. 67–72). This is also the pattern found in the study at hand. Interestingly, both the list estimate and the direct estimate are substantially higher for individuals in the top 20% of the income distribution, possibly suggesting a ‘normalization’ of bribe-paying in this group.

Overall, these results provide evidence in favor of the SB hypothesis and suggest not only that bribe victimization is under-reported in general, but also that under-reporting differs substantially between different subgroups. As a consequence, researchers and practitioners should be very cautious in using direct, obtrusive, questions about corruption experiences to gauge overall levels of corruption and to model the dynamics of bribery based on these questions. As in the case of male and female respondents, using different questioning techniques might lead to opposite conclusions. Some researchers have suggested that direct questions (or ‘victim surveys’) might be ‘the least problematic of the currently available [corruption] measures’ (Heath et al. 2016, p. 58). The results in this section suggest that this is far from obvious.