The autocratic bias: self-censorship of regime support

ABSTRACT Because of a perceived (and real) risk of repressive action, some survey questions are sensitive in more autocratic countries while less so in more democratic countries. Yet, survey data on potentially sensitive topics are frequently used in comparative research despite concerns about comparability. To examine the comparability of politically sensitive questions, I employ a multilevel analysis with more than 228,000 respondents in 37 African countries to test for systematic bias when the survey respondents believe (fear) that the government, rather than an independent research institute, has commissioned the survey. The findings indicate that fear of the government induces a substantial and significant bias on questions regarding trust, approval and corruption perceptions in more autocratic countries, but not in more democratic countries. In contrast, innocuous, apolitical questions are not systematically influenced by regime type.


Introduction
When Zimbabweans were asked in 2018 how much they trusted their President Emmerson Mnangagwa, on average 68% said "a lot" or "somewhat." This is considered strong approval by most accounts, but is it true? Dividing respondents into two groups we get a different picture. In the one that believes that the interviewer was sent by the government 77% indicated trust in Mnangagwa, and in the one that does not, some 57% shared this sentiment. In contrast, in democratic Ghana, the difference between these two groups of respondents was only 4 percentage points, compared to 20 points in considerably less democratic Zimbabwe. 1 Does autocracy bias certain survey questions?
Given that much of our current knowledge about politics and everyday life in autocratic countries are informed by public opinion surveys, 2 it begs to questionare we misinformed? What we know about the effects and causes of, for example, trust in government, democratic attitudes, corruption perceptions, regime support, and political legitimacy, rely to a large extent on survey research comparing countries with varying regime types, 3 where data is derived through direct questions on these (in some countries, but not in others) sensitive topics. Such studies all make the assumption that survey respondents across the sampled countries aresomewhat equallywilling to express their true opinions with respect to political topics. In this article, I show that this assumption is not safe to make as bias from self-censorship do not operate uniformly across regime type.
In autocratic countries, survey questions can be sensitive for reasons beyond privacy and social adaptionin particular, questions regarding the citizens' attitudes towards and evaluations of the authorities. Respondents subjected to autocratic rule may practice "preference falsification" to align their answers with the perceived wishes of the regime. 4 Given that authoritarian regimes often pay close attention to what their citizens do and say in order to sanction those who challenge the official discourse, 5 there is a real risk that respondents will associate public opinion surveys with government intelligence gathering. Respondents can therefore be expected to appease the regime with their responses out of fear that failure to do so may result in repression, physical or otherwise.
Existing empirical evidence of self-censorship in authoritarian regimes are primarily drawn from single-country studies and their findings are mixed. A handful of studies show self-censorship to be problematic for measuring public opinion, 6 while others find self-censorship to be less of a concern 7 or even non-existing. 8 In a recent study Shen and Truex 9 estimate self-censorship across a large number of countries by utilizing differences in non-response rates between sensitive and non-sensitive items. They find self-censorship to be highly variable across autocratic regimes but conclude that it is oftentimes not an obstacle to capturing public opinion in autocratic settings. With an approach utilizing actual responses (as opposed to non-responses) this article finds self-censorship to be a severe issue in most autocracies.
I test the variation in levels of self-censorship in countries where respondents experience different perceived (and real) risks of repressive action, by employing a simple research design that utilizes data on whether respondents think the government sent the survey enumerator to interview them, or that the enumerator works for an independent research organization. I then analyse if these two groups of respondents answer systematically differently to potentially sensitive questions and whether this difference is a function of the climate of political repression. Drawing upon data from more than 228,000 respondents across 37 African countries over 6 survey rounds the results show that there is indeed an autocratic bias. Responses to questions related to the citizen-state relationship, such as whether respondents trust the president or prime minister and corruption perceptions of key state institutions, are systematically biased with the level of democracy in the country, while apolitical questions, such as "How much do you trust your neighbours?" are not. Thus, caution is warranted in employing the former types of survey items, but not the latter, in comparative studies across different regime types.
Lastly, I engage with the literature on response bias induced by the ethnic (miss)match of the enumerator and the respondent. 10 I show that the bias introduced by enumerators who are coethnic with the country's key leader but non-coethnic with the respondent operates along the logic of the autocratic bias: that the ethnic (miss)match prompts regime friendly responses in autocratic countries but not in more democratic countries.

Self-censorship and how to estimate it
Survey respondents can feel the need to censor themselves if a question elicits responses that can be socially undesirable or politically incorrect, or if the respondent fear that their responses can have consequences if disclosed. 11 Questions related to, for example, income, voter turnout, prejudice against ethnic or religious groups, and drug abuse can cause the respondent to hide the truth because of concerns about their prestige, fear of social sanctioning from peers, or fear of punishment. This can lead to high rates of systematic non-responses and/or biased answers, resulting in poor data. Concerns about prestige and social sanctioning can induce social desirability bias in surveys conducted in both democratic and autocratic regimes, while fear of punishment is of greater concern in autocratic and semi-autocratic contexts, where the perceived risk of repressive action is likely to be higher. Kuran 12 argues that citizens subjected to authoritarian rule have strong incentives to practice "preference falsification" and Schedler 13 raises concerns about the possibility of obtaining reliable measures of regime legitimacy through representative public-opinion surveys or qualitative interviews in autocracies because of the opaque and repressive features of those regimes. Fear of repercussions for failing to give the officially desired answer is expected to have an effect on responses, especially when respondents are uncertain about their anonymity.
What does this mean for cross-country comparative studies if we are interested in, for example, approval ratings, regime support, or the legitimacy of the regime attributed by citizens? If the levels of self-censorship are more or less equal across countries on proxies for or components of said question or indexes, the issue is less severe. We would simply have to deal with either inflated or deflated numbers across the board. However, if the propensity to self-censor depends on some traits that are heterogeneous across countries, such as the level of democracy or political repression, the size of the bias differs systematically between countries, and thus restricts the possibility of comparative analysis.
Recent findings warrant caution regarding the reliability of survey responses in repressive and non-democratic settings. In Zimbabwewhere government repression were commonplace -Garcia-Ponce and Pasquale 14 find reported level of trust in the president and the ruling party to be affected by recent experiences of state-led repression. Kalinin 15 employs a series of list experiments and finds that Russians' electoral support for Vladimir Putin is inflated by about 20 percentage points. In contrast, also using list experiments, Frye et al. 16 estimate Putin's approval ratings at about 10 percentage points below those received through direct questioning but conclude that direct survey questions largely reflect the attitudes of Russian citizens. Shockley et al. 17 show score inflation of elite governance evaluations among executive survey informants in autocratic regimes, and in particular that executives of firms with headquarters in autocratic Qatar inflated scores vis-a-vis those with out of country headquarters. In the Chinese context, Jiang and Yang 18 show an increase in preference falsification in the aftermath of a major political purge in Shanghai, and using list experiments Robinson and Tannenberg 19 find respondents to falsify regime support by up to 25 percentage points. These studies support concerns that individual respondents inflate their approval in autocratic settings.
Does this mean that we cannot trust surveys to measure citizens' trust in government, or their political preferences in general? To answer this question, we need to complement single-country studies and test for systematic bias across a larger sample of countries. To test for response bias due to perceived fear of the government, I turn to Afrobarometer data, and specifically the last item of the Afrobarometer battery, which asks, "just one more question, who do you think sent us to do this interview?" Even though the enumerators conducting the survey introduces themselves as affiliated with "an independent research organization" that does not "represent the government or any political party," over the second to seventh rounds of the Afrobarometer survey, 49% of respondents believed that the survey was sponsored by the government, while 35% considered it to be independent and 16% stated that they did not know. With the help of this survey question, I divide respondents into three groups: non-suspecting (those who believe the survey to be independent), suspecting (those who believe the government was sponsoring the survey), and a "Don't know" group.
This item has been featured as a proxy for the "costliness of dissent" or "fear of the government" in models predicting vote choice in 16 African countries and voting intentions in Zimbabwe. 20 In contrast to these authors, I argue that the propensity to suspect the government as sponsor of the survey is not informative of political fear in that country. This is illustrated by the fact that in the most democratic and the most autocratic countries in the most recent round, Cape Verde and Cameroon, virtually the same percentage (41) of respondents believed the government to be sponsoring the survey. In the full sample of 157 country years, suspecting the government as survey sponsor is only correlated with the country's level of democracy at a level of 0.1.

Government sponsorship and self-censorship
Existing findings on the effect of perceived and real survey sponsorship on politically sensitive items are mixed. In a study of 20 African countries, Zimbalist 21 uses this perceived survey sponsor to show that a fear-of-the-state, on average, bias survey responses, and illustrates with case studies why this bias is more pronounced in autocratic Mozambique than in democratic Cape Verde. In the context of Communist Poland, Sulek 22 compares responses to political items from surveys run by a government opinion pollster to responses in independent academic surveys in mid-1980s. He finds more critique of the government in the independently run surveys, and show that this difference disappears late 1980s with the fall of the Communist regime. In 1970s communist Bulgaria, Welsh 23 finds substantial effects on items of regime support of being interview by a Party-cadre instead of a university-affiliated interviewer. In contrast, Calvo et al. 24 do not find beliefs about survey sponsor to induce bias when comparing government-sponsored household surveys to independently run Afrobarometer surveys in eight African countries. In the opposite direction Lei and Lu 25 find that Chinese respondents in fact are slightly more critical of the regime when the survey enumerator convey cues of Chinese Communist Party membership.
To move this debate forward I expand the scope of previous studies by analysing responses to more than 40 survey items across 37 countries over 6 survey rounds. Taking stock of Zimbalist's 26 case studies I argue that beliefs about survey sponsorship should have an impact on sensitive questions only to the extent that the respondents also fear punishment from the authorities. In sum, you should be more likely to falsify your preferences when you believe the regime will learn what you say and care about it. Suspicion of the survey sponsor should lead to preference falsification on potentially sensitive topics, but not on questions that are apolitical in nature. From this I derive the following hypotheses: . H1: Respondents who believe the government has commissioned the survey will give responses that are more favourable to the regime compared to citizens who believe the survey to be independent if they live in more autocratic countries, but not if they live in more democratic countries. . H2: Respondents who believe the government has commissioned the survey will not answer differently on non-sensitive questions, compared to citizens who believe the survey to be independent, irrespective of regime type.

Research design
To test if political survey items suffer from greater sensitivity bias in authoritarian contexts, I employ a simple research design that compares respondents who think the government sent the survey enumerator to interview them to those who believe the enumerator works for an independent research organization. I then analyse if these two groups of respondents answer systematically differently a set of 41 sensitive (H1) and non-sensitive (H2) survey questions, and test if this difference is conditioned by the climate of political repression. For example, if respondents who believe the government has sent the enumerator are significantly and substantively more likely to indicate that they trust the President to those who believe the enumerator works for an independent research organization, if they live in a country more autocratic country compared, but not if they live in a more democratic country, this would be indicative of H1. If, in contrast, there is no difference between the suspecting and non-suspecting respondents for trust in their relatives regardless of regime type, this would be indicative of H2. The level of democracy functions as a proxy for the "fear-of-the-government" mechanisms theorized to induce self-censorship.
In assessing the potential sensitivity of question items I rely on Blair, Coppock and Moor's 27 Social Referent Theory of Sensitivity Bias which state that we should expect sensitivity bias when all four of the following are present: the respondent (1) has a social referent in mind when responding; (2) believes the referent can infer his/her response; (3) has a perception of the referent's preferred response; and (4) has a perception that failure to provide the preferred response would entail a cost. In this application, the state is the social referent (1). Perceived survey sponsor indicates if the respondent believes that the referent can know the respondent's answer (2). Respondents are assumed to know the referent's preferred response (3). Lastly, the level of democracy works as a proxy for whether or not failing to provide the preferred response the social would entail costs to the respondent (4). This setup predicts all items probing for an evaluation of the state or state institutions to be sensitive whereas evaluations of private or non-state actors and institutions are not (see Table  1). Most items are clear cut, but a brief discussion is warranted on the potential sensitivity of questions regarding preference for democracy, army rule, and traditional leaders, where it may be more difficult for respondents to know the referent's preference. First, it is possible that respondents do not perceive it sensitive to report that they prefer democracy if they live in autocratic countries that frequently (ab)use the term democracy to describe their rule or even incorporate it in their name, such as in the People's Democratic Republic of Algeria. Second, given that the vast majority of autocratic regimes in the sample are not ruled by military juntas, disagreement with the proposition of Army rule should generally be non-sensitive. This should not be the case in Egypt in 2015 where I expect the question to be particularly sensitive in the aftermath of the 2013 military coup. This is also what the data show: in Algeria perceived survey sponsorship has no bearing on reported preference for democracy, and in Egypt perceived survey sponsorship is a particularly strong predictor of acceptance of military rule. Lastly, the degree to which traditional chiefs are included in the governing structure varies substantially between and within states, 28 which makes a general prediction of the referent's preferred answer difficult.
This design can inform us about variation in the level of self-censorship but it does not allow estimates of the absolute level of self-censorship at hand. Even among respondents who believe that the survey truly is independent, self-censorship may be taking place. They may still be wary that the authorities can use the survey to trace unsanctioned opinions to an individual, a neighbourhood, or a village. To the extent that respondents adopt a better-safe-than-sorry approach, the overall response bias will be larger, as this would reduce between-group differences. The results reported in this article include this potential built-in downward bias of the estimates. Bearing in mind that the absolute levels of bias cannot be established, the findings do show that between-group differences are clear and meaningful.
One assumption of the research design is that the suspecting respondents in autocracies and democracies do not differ on any dimensions other than those I can account for in the analysis. This would be violated if, for example, regime supporters in autocracies (but not in democracies) are more likely to believe the regime is powerful and therefore also more likely to sponsor the survey. I cannot test this assumption, but the fact that state capacity does not exhibit a relationship with sponsorship belief offers an indication that the assumption is not violated.
Another assumption is that respondents' belief about survey sponsorship is stable throughout the survey, and not formed towards the end of the survey after having answered the potentially sensitive items. Given that the question is always asked last this assumption is not testable.
Lastly, the research design does not allow me to determine whether the effects stem from bias caused by believing that the government has sent the enumerator or by believing that the enumerator is from an independent organization, or from a combination of the two. There are, however, clear theoretical reasons to suspect that the bias stems from the former.

Data and modelling strategy
Individual-level data are taken from the second, third, fourth, fifth, sixth and seventh rounds of the Afrobarometer. 29 I match the various survey rounds with country-level data for the corresponding year from the Varieties of Democracy data set. 30 This pooled dataset provides more than 228,000 respondents nested in 157 country years, nested in 37 countries.

Dependent variables
To test the autocratic bias hypothesis, I employ all variables probing for trust, corruption perception and approval rating available in the data sets, as well as a handful of questions on democratic values. In total I have 41 dependent variables (DVs), most of which are theoretically sensitive and a smaller share that are not. The guiding premise is that questions evaluating the state or state institutions are potentially sensitive whereas evaluations of private or non-state actors and institutions are not. The DVs fall under four categories: Trust; Corruption; Approval; and Values and are listed in Table 1.
For the Trust DVs respondents are asked "How much do you trust each of the following, or haven't you heard enough about them to say." And are provided the following answer options resulting in a "Not at all"; "Just a little"; "I trust them somewhat"; "I trust them a lot"; and "Don't know/Haven't heard enough." In the main analysis I drop respondents who chose "Don't know/Haven't heard enough" (DK), resulting in a 4-point scale. The only exception is trust in Most people, with the binary choices "Most people can be trusted" and "Must be very careful." The Corruption questions ask respondents "How many of the following people do you think are involved in corruption, or haven't you heard enough about them to say," with the answer options: "None; Some of them; Most of them; All of them; and Don't know." Dropping DKs in the main analysis gives a 4-point scale. The one exception is Increase in past year which is on a 5-point scale ranging from "Increased a lot" to "Decreased a lot." For the Approval DVs respondents are asked "Do you approve or disapprove of the way the following people have performed their jobs over the past twelve months, or haven't you heard enough about them to say," with 4 answer options ranging from "Strongly disapprove" to "Strongly approve," plus a Don't know-option.
For the Values DV Prefer democracy, respondents are asked which statement is closest to your own opinion. "For someone like me, it doesn't matter what kind of government we have; In some circumstances, a non-democratic government can be preferable"; and "Democracy is preferable to any other kind of government," resulting in a 3-point scale with higher values indicating a preference for democracy. For the three other DVs, One-party; Army; and Strong-man rule respondents are asked "There are many ways to govern a country. Would you disapprove or approve of the following alternatives," with the options ranging from "Strongly disapprove" to "Strongly approve," including a neutral middle option resulting in a 5-point scale.
Note that all DVs are not available for all countries and all rounds. The number of respondents and country years available for each of the 41 DVs are included in the regression tables in Appendix A.

Independent variables
The main independent variable, Survey sponsor, is generated from the question item "Just one more question: Who do you think sent us to do this interview?" Respondents who indicate that they believe the enumerator was sent by the local, regional, or national government, or any of its agencies, are coded as 1, while those who believed the survey to be commissioned by a non-governmental organization (NGO), a university, a research company, etc. are coded as 0. Of the complete sample, 49% of respondents reported that they believed that the government was behind the survey, while 36% believed it to be independent and 15% said that they did not know. The variance of suspecting the government as sponsor at a country year level is between 6.5% in Liberia in 2015 and 82% in Madagascar in 2005. Figure 6 Appendix F, displays the share of suspecting respondents in each country over time. Regressing the share of suspecting respondents on a number of plausible predictors shows only corruption to be associated with a larger share suspicion. State capacity, human rights abuses and the level of democracy are not associated with suspicion at the country level. The latter is important for the research design and it is reassuring that the between-country variation of suspicion only correlates with the level of democracy at a 11-level. The "Don't know" group of respondents is excluded in the main analysis but are coded as 1 together with the "suspecting" group in robustness checks of all model specifications, the rationale being that not knowing who the sponsor is likely to also induce preference falsification in repressive settings, albeit to a lesser degree than suspecting the government.
A note is warranted on the issue of item non-response: are people who believe that the enumerator is asking a question on behalf of the government more likely to refuse to answer or simply state that they "Don't know" when asked political questions in autocracies? I find no evidence for that. This is consistent with theories of preference falsification, which propose that fearful respondents will resort to the safest possible answer. 31 Saying that you do not know or refuse to answer whether you, for example, trust the president can raise suspicion and risk signalling dissent (see Appendix E, Table 21, 22 and 23 for details).
While there is no correlation between the country level averages of suspecting that the government sponsored the survey and countries' levels of democracy, the next issue is to examine whether there is something particular about those individuals who suspect the survey sponsor that also makes them more likely to state high trust in the key leader, etc. than the population at large. Note that it would need to be a peculiarity that alters their behaviour only in more autocratic countries and not in more democratic countries. So, who suspects the government? The balance of individual-level covariates for the different groups of respondents shows minor differences. If we split the sample into autocratic and democratic countries using V-Dem's categorical Regimes of the World index, 32 and look at the covariate balance between the two groups of respondents, the pattern of differences is the same. For example, in the democratic sample, the mean level of poverty for those who suspect the government as survey sponsor is 0.14 higher than for those who believe in the survey's independence; in the autocratic sample, this difference is 0.11 (see Appendix D, Table 18). The differences in means between the two groups are similar in both samples, and in the same direction for all demographic variables. There are no observable individual characteristics that can explain why only those who live in authoritarian countries and who suspect the government to be behind the survey are more likely to provide regimefriendly responses to politically sensitive questions. As a robustness test, I follow Calvo et al. 33 and use propensity score matching to account for selection into treatment (see Appendix G, Tables 25 and 26 for details). Matching on pre-treatment covariates do not change the results.
In all model specifications, I control for the following set of individual level control variables: age, gender, education level, urban/rural residency, and for an index of lived poverty. 34 To avoid post-treatment bias I do not include variables that are themselves sensitive items and hence might be affected by sponsorship belief. For example, whether you believe the country is "moving in the right or wrong direction," may predict whether you trust the key leadership figure, but it may also be affected by sponsorship belief.

Country level
As a proxy for the perceived risk of repressive actions at the country level, I employ the Varieties of Democracy's Electoral Democracy Index 35 (see Coppedge et al.,and Marquardt and Pemstein 36 for detail on aggregation rules and methodology). The rationale for using a highly aggregated index of the level of democracy is to be able to test whether existing studies that draw conclusions from comparative survey data from countries at vastly different levels of democracy are suffering from biases. To get closer to the mechanism of fear that the mediating variable is theorized to induce I substitute the Electoral Democracy Index with a sub-index of Freedom of Expression for all models (see Appendix B). The two indexes are highly correlated and it is no surprise that the results hold to this substitution. I note that the theorized patterns are in fact even more pronounced using freedom of expression as the mediating variable. This makes sense as states that have a have a low score on its Democracy index partly due to low state capacity may not invoke the same fear as states that have capacity and chooses to use it to suppress political opponents.

Estimation
Because respondents are not randomly distributed but clustered within countries, I employ multilevel models that take these data hierarchies into account and allow testing for the effect of a two-level interaction between perceived survey sponsor (individual level) and level of democracy (country year level). The model is a linear random slope model. The specification of the baseline multilevel model is as follows: Adding the two-level interaction term between individual-level suspicion of survey sponsor and country-level democracy, we get the full multilevel model specification (see Aguinis et al. 37 ): where y ic is the dependent variable for individual i in country c, γ 00 is the average individual level intercept, dem c is the country-level democracy, spons ic is an individual's perception of survey sponsor, X´i c and Z´c are vectors of individual-and country-level controls, U 0c is the intercept variance, R 1c is the slope variance (for spons ic ), and η ic is the individual-level error term. I do not discuss nor present model building, the slope, and intercept variance in the main text and tables. In short, the intercept variance is reduced in each step of building up the models and the slope variance is reduced when introducing the two-level interaction, i.e. it explains some of the between-country variance (for details, see replication code). To facilitate easy interpretation, I proceed by graphing the interaction effects and visualizing the effects of perceived sponsor on each dependent variable, conditioned on the level of democracy in order to contrast more autocratic and more democratic countries. I de-mean and standardize all outcome variables by each country and survey round so that effect sizes are in country-round-specific standard deviation units. Thus, a coefficient of 0.25 indicates that believing you are interviewed by a government agent, rather than an independent researcher, is associated with a quarter of a standard deviation change in the dependent variable.
Empirical findings I first look at the Trust variables. Figure 1 shows the marginal effects plots for all 20 trust variables. The bar charts at the bottom of each graph display the distribution of country-years (level-2 units). For simplicity, detailed regression output is not displayed here, but can be found in the regression tables in the Appendix A: Tables 2-8. All of the following DVs have substantial and significant interaction effects in the expected direction: President; Members of Parliament; Local government; Ruling party; Opposition party (opposite direction from the other); Electoral commission; Police; Courts; Army; Traditional leaders; Government News; and Government Broadcasts. In contrast the following trust variables are not biased along regime type: Independent News; Independent Broadcasts; Tax officials; Religious institutions; Neighbours; Relatives; Vendors; Most people.
The results are in line with H1: that sponsorship belief bias responses to sensitive questions in favour of the regime in more autocratic countries but not in more democratic countries. To take an example, respondents who believe it's the government who is asking is estimated to say that they trust their President 0.2 standard deviations more than those who believe an independent institute is asking, when they live in autocratic countries. An effect of that magnitude is expected in countries with an electoral democracy score in the range of 0.2-0.4, such as Burundi, Eswatini, Egypt and Zimbabwe. The estimated difference is low or indistinguishable from zero in more democratic countries on top of the scale, such as Ghana, Cape Verde and South Africa. Only one sensitive variable does not exhibit the hypothesized pattern: reported trust in Tax official is only marginally affected by sponsorship belief and the effect does not depend on the level of democracy.
The results are largely in support of H2: the respondents who believe the government has commissioned the survey will not answer differently to non-sensitive questions, compared to citizens who believe the survey to be independent, irrespective of regime type. Save for trust in Traditional leaders, responses to non-sensitive questions, such as trust in Neighbours, Relatives, Vendors, Most people etc. are unaffected by sponsorship beliefs and regime type.
Moving on to the Corruption variables, all the following have substantial and significant interaction effects in the expected direction: President; Members of Parliament; Local government; Police; Courts; Army; Increase in past year; and Businesses. In contrast, the following variables are not biased along regime type: Bureaucracy; Tax officials; Traditional leaders; Religious institutions; and NGOs (see Appendix A Tables 5 and 6 for details). Figure 2 shows how respondents' reported corruption perception is systematically biased across the level of democracy. All else equal, respondents who believe it is the government that is asking are less likely to answer that corruption is widespread among various state institutions, compared to those who believe the question is coming from a research organization. With the exception of Bureaucracy and Tax authority, the patterns for all sensitive variable support H1. Similarly, the absence of an effect with regards to perceptions of Traditional leaders, Religious institutions and NGOs is in line with H2. Yet, the apparent interaction-effect for Businesses goes against the expectation of H2.
The significant effects of the four sensitive Approval variables offer additional support for H1 (see Figure 3). Believing the government to sponsor the survey substantially increases reported approval of the President, Member of parliament, and Local government in more autocratic countries, while having an insignificant effect as the level of democracy surpasses values around 0.7. Thus, countries at a level of democracy on par with Benin and Tunisia are not expected to see an upward bias. While the interaction effect of the supposedly non-sensitive item Traditional leader is insignificant, it does exhibit a pattern counter to H2. The relationship is consistent with how respondents react to Traditional leaders with regards to trust and corruption and may indicate that it may in fact be a sensitive topic.
Moving on to democratic values, Figure 4 shows how respondents' preference for democracy is systematically biased across the level of democracy. All else equal, respondents who believe it is the government that is asking are around one-tenth of a standard deviation less likely to answer that democracy is a preferable system, compared to those who believe the question is coming from a research organization. Studies concerned with public demand for democracy 38 likely underestimate the true demand due to this downward bias in more autocratic countries. We see a similar interaction effect with regards to acceptance of One-party rule, with over reporting in more autocratic countries when the government is the perceived sponsor. The effect of survey sponsor is uniform across regime types for acceptance of Army rule and Strong-man rule. The first is unsurprising given that most autocratic regimes in the sample are not military regimes, but the latter is counter to my expectation.

(How)does ethnicity matters?
Given that ethnicity is politically salient in many of the countries in the sample, and that this article is concerned with political questions, I test influence of respondent' and enumerators' ethnic identities. This is especially relevant since Adida et al. 39  show how a large number of the Afrobarometer survey items suffer from response bias stemming from the ethnic (miss)match of the enumerator and the respondent. In particular, the authors show that being interviewed by a non-coethnic generate more socially desirable responses, andpertinent to this studythat respondents interviewed by an enumerator who is a coethnic of the country's key leader are more likely to give regime friendly responses, such as approve of the government's handling of the economy and express lower trust in opposition parties. An alternative explanation is thus that the autocratic bias put forward in this article is driven by an ethnically induced social desirability bias. This is not the case: the models are robust to controlling for the ethnic match between the enumerator and respondent, as well as for the enumerator being a coethnic with the key leader. Moreover, by replicating a number of Adida et al. models and adding an interaction between the variable on leader-interviewer coethnicity and level of democracy I show in Figure 5 that the induced biased appears to operate through the logic of the autocratic bias (for full regression tables see Appendix C, Tables 16 and 17): i.e. that the bias is primarily a concern in more autocratic countries and generally not in democratic countries. These interaction graphs display the marginal effect of being interviewed by a non-coethnic enumerator who is coethnic with the country's key leader (potential ethnic threat) for: trust in the president; the ruling party; opposition parties; vendors; corruption perceptions of the office of the president; approval of the president; evaluation of the government handling of the economy; and approval for strongman rule, conditioned on level democracy. The models include all controls used throughout the article and serve as a replication of four models included in Adida et al.'s study (see figure 7B in their paper).

Discussion
The empirical findings suggest that more repressive regimes indeed enjoy the autocratic bias. Autocratic countries receive inflated evaluations of trust, corruption, approval and regime-friendly values. This becomes particularly clear looking at the stark contrast between the otherwise identical questions on trust in government newspapers and independent newspapers, as well as government broadcasting services and independent broadcasting services (see Figure 1, row 4). Sponsorship belief impacts reported trust in the government's information channels in autocratic countries but not in more democratic countries. Trust in the independent information channels remains unaffected by sponsorship belief irrespective of regime type. These results have implications for the comparative use of survey data. As the bias is consistent across a large number of survey items commonly used in comparative politics, previous studies that rely on survey research comparing countries with varying regime types may need re-evaluation. 40 Future work would do well to consider the obstacle self-censorship pose to their ability to compare public opinion data across regimes, and whenever possible test for its presence. Indirect questioning techniques, such as the list experiment or the randomized response technique are powerful tools to estimate sensitive opinions but are not without cost. The indirect approaches require a larger sample, 41 are cognitively taxing, and sacrifice individual-level data for aggregated estimates. Because of these drawbacks, researchers will often have to rely on direct questions for sensitive questions. When direct questioning is necessitated and when information on perceived sponsorship is unavailable, researchers need to be explicit about the assumptions they are making with regards self-censorship and how a violation of these may alter their inferences. This is especially prudent when the potential bias works "in favor" of the hypothesis being tested.
In addition to across-regime comparative work, the autocratic bias can distort results of longitudinal studies of approval ratings of leaders and ruling parties, 42 and demand for democracy 43 within the same autocratic regime. Even in the absence of significant political changes, the share of respondents that perceive the state to sponsor the survey can differ between survey rounds leading to seemingly dramatic public opinion shifts. Mozambique provides an illustrative example of this. Between 2012 and 2015, the share of respondents that reported trust in the ruling party dropped 18 percentage points (from 74 to 56). Disregarding the autocratic bias, one conclusion might be that support is rapidly wittering for the dominant one-party regime. However, the share of respondents that perceived the interviewer to be sent by the government was a full 28 percentage points lower in 2015 than 2012. Given that in 2012, all else equal, that group of respondents were 2.4 times more likely to indicate trust in the ruling party, the high levels of trust in FRELIMO in 2012 is likely a product of the autocratic bias. This illustrates that it is important and also that it is possible to increase the perceived independence of a survey. In the Mozambique case, two different firms were responsible for the field work in 2012 and 2015. It is possible that the reputation of firms, names, logos, attire, etc. influence perceived independence. Given the large variation in sponsorship perception between and within countries (see Appendix F, Figure 6), it is clear that sponsorship belief is not static and future research into how to minimize respondents' suspicion will help to advance survey research in autocracies.

Conclusion
This article shows that respondents' belief about who has commissioned the survey influences answers on politically sensitive questions in more autocratic countries while having no impact on responses in very democratic countries. In more politically repressive environments, respondents who believe (fear) that the government has sent the interviewer are more likely to state that they trust the country's leader or ruling party, and less likely to state that they believe rulers and state institutions to be corrupt, as compared to respondents who think that the interviewer works on behalf of an independent research organization.
This study provides a significant contribution to comparative public opinion research in general and to the study of political behaviour and public opinion in authoritarian regimes in particular. I provide evidence that a large set of commonly used survey itemsranging from regime legitimacy and popular support for incumbents to corruption perceptions and preferences for democracy suffer from systematic bias. Sensitive questions evaluating rulers suffer from a larger bias than do questions evaluating those exercising public power. This is evident by the larger effect on trust in or corruption perception of in the key leadership figure or ruling party compared to the effect vis-a-vis the bureaucracy and tax authorities. Innocuous apolitical questions, such as trust in your neighbours or relatives, provide perfect tests for the autocratic bias hypothesis, as one would expect no difference between the two groups of respondents' answers, no matter the level of political repression. Indeed, there is none.
The autocratic bias is not only of methodological concern. Insofar as good-governance or democracy-promotion initiatives are informed by survey items measuring corruption perceptions or demand for democracy, this bias is also of direct political relevance. The usefulness of a survey item probing respondents' sense of anonymity should be evident. Given the low implementation cost, such an item should be added to surveys with the ambition to be comparable across countries where the doubt of anonymity is likely to produce different response behaviour. Save for the Afrobarometer, large-scale public opinion surveys do not include a similar item, 44 and thus fail to provide data users with an easy and effective method of estimating the varying sensitivity of survey items across countries. Studies using data where such a variable is available should ideally include an interaction term in the analysis to determine the presence of systematic bias. In the presence of bias, running the analysis using only the sub-sample of respondents who believe the survey to be independent provides a first robustness check for the problem of preference falsification. Having estimated the bias for a set of sensitive items, another avenue forward would be to construct reliability weights to enable the researcher to account for biases in the analysis while retaining the full sample.