Be Wary of Those Who Ask: A Randomized Experiment on the Size and Determinants of the Enumerator Effect

During survey data collection, respondents' answers may be influenced by the behavior and characteristics of the enumerator, the so-called enumerator effect. Using a large-scale experiment in Uganda randomly pairing enumerators and respondents, the study explores for which types of questions the enumerator effect may exist. It is found that the enumerator effect is minimal in many questions, but is large for political preference questions, for which it can account for over 30 of the variation in responses. The study then explores which enumerator characteristics, and which of their combination with respondent characteristics, could account for this effect. Finally, the conclusion provides some practical suggestions on how to minimize enumerator effects, and potential bias, in various types of data collections.


Policy Research Working Paper 8671
During survey data collection, respondents' answers may be influenced by the behavior and characteristics of the enumerator, the so-called enumerator effect. Using a large-scale experiment in Uganda randomly pairing enumerators and respondents, the study explores for which types of questions the enumerator effect may exist. It is found that the enumerator effect is minimal in many questions, but is large for political preference questions, for which it can account for over 30 percent of the variation in responses. The study then explores which enumerator characteristics, and which of their combination with respondent characteristics, could account for this effect. Finally, the conclusion provides some practical suggestions on how to minimize enumerator effects, and potential bias, in various types of data collections. This paper is a product of the Strategy and Operations Team, Development Economics Vice Presidency. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/ research. The authors may be contacted nathan.fiala@uconn.edu.
3 asked. However, they find large enumerator effects in answers to questions concerning the respondent's opinion about opposition political parties, a very sensitive question in Uganda. The authors then test for the determinants of the enumerator effect on political support. They find that whether the enumerator comes from an urban versus rural area is positively associated with openness to vote for the opposition parties, though not for the ruling party. This result is consistent with the political situation in Uganda: support of the opposition parties comes from mostly urban areas, while rural communities vote overwhelmingly for the ruling party. Moreover, the study's results indicate that respondent's answers are influenced by the gender of the enumerator and by the enumerator's experience as a surveyor. In addition, the study provides some evidence that the enumerator effect is also influenced by the distance between the characteristics of the enumerator and the respondent.
These results have practical implications for conducting surveys on sensitive issues.
In the context of randomized controlled trials, to reduce the likelihood that the enumerator effect will cause bias researchers should balance enumerator characteristics between treated and control groups when assigning enumerators to subjects. When this is not possible, enumerators themselves should be balanced between treated and control groups.
In the cases of non-randomized studies using previously collected survey data, the results of this study indicate the importance that data documentation should provide, at a minimum, a clear description of how enumerators and teams were assigned. This would allow researchers to check for the presence of the enumeration effect and estimate its size before conducting any analysis.
4 This paper presents three contributions to the literature. First, this the first study to document the enumerator effect on a range of outcomes -including stated political opinions -using an experimental design in a developing country. Second, the study collects detailed information on enumerators demographics, work history, and a range of psychological measures. This makes it possible to explore in greater detail than in previous studies the extent to which political opinion responses vary depending on the enumerator's characteristics and on the distance between the characteristics of the enumerator and the respondent. Third, the randomization is based on an unusually large sample for this type of experiment, reinforcing confidence in the robustness of the results.
The remainder of this paper is structured as follows. Section 1 briefly discusses the previous literature on the enumerator effect. Section 2 presents the experimental design conducted on the enumeration team. Section 3 describes the data, while the results are presented in Section 4. Section 5 discusses the implications of our results for survey data collection. Section 6 concludes.

Literature
The literature studying the enumerator effect draws from different disciplines, including economics, political science, statistical methods, anthropology, and psychology (for recent surveys, see Schaeffer et al., 2010;West and Blom, 2017) 2 . The enumerator effect can 2 A related set of studies looks at survey design as a possible determinant of data quality. Researchers investigated different aspects, such as the method of data collection (Blattman et al., 2016), the treatment of hard-to-measure concepts (Laajaj and Macours, 2017), the wording of the questions (Anker et al., 1987;Beaman and Dillon, 2012;Serneels et al., 2016;), questions placing within the survey questionnaire (Karlan and Zinman, 2012), and the length and level of detail of the questionnaire (Kalton and Schuman, 1982;de Mel, et al. 2009;Beegle et al. 2012;. Other studies look at how data quality is affected when household members report for others in the household (Bardasi et al., 2011), when the answers are to be reported publicly (Carlson, 2016), or when the repetition of surveys can be exhausting (Zwane et al., 2011). materialize because of enumerator behavior affecting non-response (Couper and Grove, 1992;West and Olson, 2010;Randall et al., 2013) or due to enumerator characteristics affecting the answers of the respondent (Brunton-Smith et al., 2017).
There is also evidence that the enumerator effect varies with the type of question, being more salient for questions concerning gender-related issues, religion, ethnicity, and politics (Schaeffer, 1980;Davis and Silver, 2003;Himelein, 2016;Laajaj and Macour, 2017).
One of the main challenges for this literature is to avoid confounding enumerator interviewer and respondent characteristics. The only way to exclude this possibility is to randomly assign interviewers to respondents. However, practical factors often make design unfeasible. Thus, much of the literature on the enumerator effect consists of either telephone surveys with small numbers of random interviewers or face-to-face surveys with non-random assignment of interviewers (West and Blom, 2017). Only a few studies have rigorously documented the causal impact of the enumerator on survey responses and all these studies have been conducted in the United States. A first group of studies looked at how the gender of the interviewer affected the response to gender and sex-related questions (Johnson and Delamater, 1976;Catania et al. 1996;Huddy et al., 1997). Another important set of studies has documented the presence of the enumerator effect in surveys on political opinions, showing that it depends on enumerator characteristics such as race, behavior, and 6 political views (see Williams, 1964;Schuman and Converse, 1971;Hatchett and Schuman, 1975;Reese et al. 1986;Davis, 1997;Davis and Silver, 2003).
Studies of the enumerator effect in developing countries are few and all present a quasi-experimental design due to problems with randomization in the field (Bischoping and Schuman, 1992;Flores-Macias and Lawson, 2008) 3 . The number of studies looking at enumerator effects in shaping answers to political questions in developing countries is extremely limited. Bischoping and Schuman (1992) document the impact of perceived enumerator political opinions on voting intensions in Nicaragua in 1990 while Adida et al. (2016) show that co-ethnicity between respondent and enumerator influences answers to voting behavior questions in the Afrobarometer survey. To the best of the authors' knowledge, this is the first paper which provides evidence of the enumerator effect affecting the stated political preferences in a developing country using an experimental design.

Experimental design
The enumerator experiment described in this paper was conducted during the data collection of a separate research project, described in detail in Fiala and Premand (2018).
That project was an experimental evaluation of a large-scale local accountability program conducted in 2016 that included 1064 villages and 8403 respondents. The authors of the present paper utilize the sample of participants in that study for the experiment that is described here.

7
The research on the local accountability project included several economic and political outcomes, and so the questionnaire asked a range of different questions, including ones concerning demographics, asset ownership, consumption of alcohol and tobacco, and political party preferences.
The survey was conducted by 90 enumerators, divided into 4 teams, one for each language group of Uganda. Given the length of the survey and the fact that 8 individuals were to be interviewed each day, two enumerators were needed per village. During the data collection the authors randomized, within the teams, which village an enumerator would visit to conduct interviews. Thus, the pairing of enumerators visiting each village was random. Yet, it was not possible to randomize which specific individual within a village an enumerator interviewed. This could bias the analysis if enumerators could systematically choose individuals to interview of a specific type. However, in the analysis to follow, it is shown that this is unlikely to be the case.
During the morning of the day of data collection, the field manager or team leader assigned a village which was randomly determined using pre-developed randomization lists to each enumerator with the list of households to be interviewed. 4 Randomization was stratified by distance to ensure that if an enumerator went to a distant village the day before, he or she would go to a closer one the current day, and vice versa. That is, the field manager or team leader, before randomizing, split them into two groups: near and distant. An enumerator who went to a distant village the day before was given a randomly selected village from the near group. This process was done for each enumerator, making sure the 8 distance between surveyed villages was well balanced for each person. This ensured that no enumerators felt they were traveling significantly more than their colleagues. 5 As is common in data collections, there were times when some enumerators could not complete all of the surveys, and so a team leader had to conduct the survey. In other cases, the enumerator decided to leave the survey team. As the team leader was not randomly selected and enumerators who left the teams usually completed only a small number of surveys, surveys were dropped done if they were done by enumerators who conducted fewer than 70 interviews 6 . This reduced the sample from 8403 interviews to 6895 and the enumerators from 90 to 47.
Information was collected on the enumerators who participated in the survey operations. This information included demographics (age, gender, whether their home is in an urban or rural area), education level, previous work experience as enumerator, and behavioral and psychological traits. The enumerator survey was administered after enumerators were selected to be part of the survey team, was voluntary, and was covered under the Institutional Review Board (IRB) for the main project.
Since enumerators were randomly assigned to villages within teams, enumerator data was used to validate the randomization by conducting a within-team balance test on the pairing of respondent and enumerator demographics. In practice, the project regressed a set of the respondent's characteristics (age, gender, marital status, if she can read, if she can write, and education level) on enumerators' characteristics (gender, age, education 5 Note that nearness and farness are not about the location of the village, but the location of the survey team. That is, how far a village is from where the surveyors were staying the previous night. The teams did not always stay in one location or in the main town within a region, but moved continuously. 6 Seventy interviews is the 10 th percentile of the distribution of the number of interviews by enumerator. The study also shows later that the results are not sensitive to this selection. 9 level) controlling for team fixed effects. Results reported in Table 1 are reassuring for the randomization as they show that respondent and enumerator characteristics are notindividually or jointly -significantly correlated.
Finally, the possibility was considered that the enumerator effect could be related to the length of tenure in the survey operation. The mean number of work days is 20 and the standard deviation is 11. The distribution of work days is reported in Supplementary Appendix S1, table S1.1. 7 Based on these data, a long-term enumerator is defined as one who worked at least 25 days (which is the 75 th percentile of the total number of work days distribution for the population of enumerators). 8 The results reported in Table 2 indicate that the observable demographics of an enumerator do not predict whether an enumerator is long or short term (column 1) nor -more generally -the length of tenure in the survey operation (column 2).

Data summary
The data collected for this study are presented in Table 3 and include data on the respondents and the enumerators. 9 The enumeration team was 35 percent male 10 , with 92 percent reporting having attended at least some university. There was low variation in age, with the average enumerator being 28 years old. 61percent of the enumerators selfidentified as being from an urban area. Respondents are on average 44 years old, with 49 percent being men. Only 19 percent were single, with 5.8 years of education on average.
Households had on average 8.2 members, with 2.3 heads of cattle. 41 percent of the sample reported having consumed alcohol or tobacco in the last week. They report total weekly spending on alcohol and tobacco, conditional on consuming any, of 3,800 Ugandan Shilling (approximately $1.20).
The focus of the analysis is on questions related to the respondents' support to political parties (the last four rows in Table 3). 11 For each of four political parties, respondents were asked the following question:

4) Very open
The enumerators were instructed to read the entire question and ensure that no one else was listening to the conversation. 12 About 91 percent of the people interviewed responded 11 to this question. This is a very high rate compared to similar questions in the literature. 13 In Uganda, people do not hesitate to state in public their support for the ruling party, the National Resistance Movement (NRM), as this is highly encouraged by the government.
However, people are generally much more reluctant to report support for opposition parties, which include the Democratic Party (DP), Forum for Democratic Change (FDC) and Uganda People's Congress (UPC). This is in part due to perceived and actual government action against supporters of these parties, including harassment, arrests and torture, although torture is much less common. Nevertheless, it is reported in some international press and is believed by many opposition supporters. 14 The distribution for the answers to the political opinion questions are reported in Table   4. Average openness to vote for opposition parties is approximately 1.9 (on a 1-4 scale). This is half of the average value for the ruling party. Voting preferences are highly correlated within communities: stated preference for opposition parties has an intra-class correlation between 0.30 and 0.33.

12
The paper next presents the results on the effect of enumerator characteristics on respondent outcomes. The study first looks at how much variation in responses is accounted for by an enumerator fixed effects model. Next, the focus is on the political questions, which are found to have high enumerator effects, to explore which specific enumerator characteristics may matter for the observed effects.

Measuring the enumerator effect
The main method used to identify the presence of an enumerator effect and its magnitude is to look at how much enumerators themselves contribute to any variation in observed responses. The authors begin by testing for the predictive power of enumerators on respondent answers by examining the R 2 in an enumerator fixed effects regression, similar to Himelein (2016) and Laajaj and Macour (2017). A high R 2 is interpreted as the enumerator effect picking up a large amount of the variation in responses, while a low R 2 indicates that there is no or very low enumerator effect. 15 Figure 1 reports for each question the R 2 (the red dots) with the corresponding (bootstrapped) 95% confidence interval for each of the four teams. These values are obtained from regressing -separately for each team -each outcome of interest (such as respondent's gender, age, household assets, political opinion, etc.) on a constant and the enumerator dummies, with no other controls (detailed results are reported in Table S1.2).

13
Looking at the average R 2 across all the teams, very low values are found for most of the questions, including whether the respondent is single (0.005), respondent gender (0.011), age (0.014), education level (0.012), the number of cattle a household owns (0.024), asset index (0.027), household size (0.029), and even whether the individual consumes and how much they spend on alcohol (0.011 to 0.038). These low values suggest two conclusions. First, individual enumerators do not systematically impact the way respondents report these answers. Second, it confirms that the enumerator randomization strategy worked well. In line with the results of the tests discussed in Section 2, the low values associated with respondent characteristics indicate that enumerators are not systematically choosing respondents within villages based on age, gender, education level, etc. The R 2 's for political party support questions, however, are substantially different from those for the other questions. A low value is found for NRM support (0.34), but the R 2 for opposition parties is high (between 0.234 and 0.267). These values are very high and are not suggestive of random noise or unsystematic reporting of support.

Determinants of the enumerator effect
The next step in the analysis is to look at what specific enumerator characteristic may be affecting individual responses. To this end, the study focuses on the questions in which the enumerator effect seems to matter most, namely those related to the respondent's political opinions. Specifically, for each party, the following OLS regression model is estimated: (1) where is the openness (on a 1 to 4 scale) of individual i living in community c to vote for the party. is a matrix of basic enumerator j characteristics (gender, age, education level and whether the place of residence is urban). is a matrix of enumerator j ability and psychological traits. Finally, c is the set of community fixed effects and is the error term. All standard errors are clustered at the enumerator level. As a robustness check, the authors also include a set of characteristics for respondent i living in community c captured in vector (age, gender, education level, and marital status). Table 5, columns 1-4, presents the regression results on stated support for different political parties for the baseline specification, that is only including basic enumerator characteristics. Some interesting results emerge. The coefficient for the enumerator being a man is always significant and negative, even though the magnitude of the coefficient is significantly smaller for the ruling party. This indicates that respondents are more likely to express a positive opinion on all parties if the enumerator is female. The indicator for the enumerator being from an urban area is remarkably significant for all opposition parties.
Respondents report significantly higher support for opposition candidates to urban enumerators. There is no effect from urban status on stated support for the ruling party.
Columns 5 to 8 report the results when the authors include as additional controls a set of enumerator abilities (including previous work experience as enumerator) and psychological traits. While there is no clear pattern in the way enumerator psychological traits affect respondent's answers, it should be noted that the enumerator previous work experience as a surveyor has a positive effect on the respondent's expressing openness to vote for each of the opposition parties, but it has no effect on expressing openness to vote for the ruling one.

Robustness checks
The results are robust to various checks. First, the results are robust to the inclusion of respondent's characteristics such as age, gender, education level, and marital status (see Table S1.3). Second, all of the results are robust to the sample of enumerators. In particular, the main analysis was re-run considering enumerators who conducted at least 24 interviews (the bottom 1% of the enumerator's number of completed questionnaires), rather than less than 70 (as in the main analysis). Results -reported in table S1.4 -do not change with respect to those obtained using the main sample. Third, the authors explore how the results may be affected by how many days the individual worked as an enumerator. To this end, the main regression was estimated separately for short-term and long-term enumerators.
The results reported in Table S1.5 indicate that the main results do not depend on this enumerator characteristic. Fourth, the results are unchanged when a different econometric model is used, namely an ordinal probit model estimate (see Table S1.6). Fifth, results reported in Table S1.7 and S1.8 show that the main results are robust to including controls for the day and the time of the interview (that is, morning vs afternoon). 16 Sixth, the results are not affected by enumerator fatigue (as proxied by the length of the survey) (see Table   S1.9). This is not surprising considering that the survey was only about 40 minutes long and that all political questions are asked around the same time during the survey.
Finally, a placebo is run test to show that enumerator characteristics do not affect all respondent answers. To this end, model (1) is re-run using as alternative outcomes two non-political questions. These are: (1) the age of the respondent; and (2) if the respondent consumes alcohol or tobacco. Results reported in Table 6 show that enumerator characteristics do not explain answers regarding age, which is clearly a non-sensitive question. On the contrary, it is found that respondents are more likely to report alcohol consumption to enumerators who are older. These results suggest that enumerator characteristics do matter more for sensitive questions but that -a priori -it is not obvious which characteristics matter for each question.

Discussion and additional results
The results suggest that individual enumerator characteristics may have important implications for respondents answering political questions. The first characteristic of interest is whether the enumerator comes from an urban area rather than from a rural one.
The results show that respondents are more open to express their support to opposition parties when in the presence of an enumerator from an urban area. It is possible that this effect captures the fact that most opposition to the ruling party comes from urban centers. 17 This suggests the presence of a social desirability bias at work. Social desirability bias generally refers to the tendency of a respondent to provide responses that she believes will be viewed favorably by others: in this case, the respondent would be anticipating the views of the enumerator and thus answers as to please him or her. This implies that answers to the political questions may be the result of a desirability bias in respondents and do not 17 See http://www.theeastafrican.co.ke/OpEd/comment/Museveni-NRM-party-still-has-huge-support-inrural-Uganda-/-/434750/3036604/-/syo070/-/index.html for a discussion of the results from the 2016 election that occurred 3 months before this data collection and the role of rural voters in the NRM win. reflect their true preferences. 18 These tendencies may be exacerbated on sensitive issues where fear and the desire to avoid embarrassment and criticism are stronger (Blaydes and Gillum, 2013).
Second, the authors find a strong and robust effect of the gender of the enumerator on stated voting preferences of the respondent. This result is in line with several studies that explored the role of gender in the context of survey data collection (see West and Blom, 2017). The results show that respondents are less likely to declare support for any party if the enumerator is male, that is they report a worse opinion for each party. This is in accordance with Axinn (1989), who suggests that female enumerators may be perceived as less frightening. Interestingly, this effect is smaller for the ruling party.
The third enumerator characteristic that turns out to be significant is the number of months previously worked as enumerator, which can be interpreted as a proxy for the enumerator's ability. The positive effect of enumerator's experience on openness to vote for the opposition parties is in line with the evidence presented in Randal et al. (2013) and Jäckle et al. (2013), showing a positive effect of experience on co-operation rates. Yet, this result must interpreted with caution as the design does not allow us to identify whether experience produces less or more biased answers.
Next, the authors explore the possibility that what may explain the enumerator effect are not only enumerator characteristics, but also their interaction with respondent characteristics or the distances between the two. To test for this, a specification is run that includes all the significant enumerator characteristics as they emerge from table 5, (namely gender, urban, and number of months of experience as enumerator) interacted with respondent characteristics such as gender, age, education level, and household wealth. 19 Results reported in Table S1.10 indicate that the main results are unchanged while there is no significant difference in the effect of enumerator characteristics for any of the respondent characteristics. 20 Then, the possibility is explored that what matters instead is the distance in the respondent-enumerator characteristics. To this end, the study estimates the main model (1) including indicator variables for all the possible combinations of the enumerator-respondent characteristics such as gender, age, and education level. Results are reported in Table 7. An interesting finding is that the enumerator effect associated with the enumerator being highly educated is mitigated by the education level of the respondent.
The education level of the enumerator positively correlates with the respondent's openness to vote for an opposition party and negatively for the ruling party. Yet, the interaction term between the enumerator and the respondent education level indicates that this effect decreases with the latter. This suggests that the enumerator effect is smaller when the education distance is smaller, that is when the social desirability bias is likely to be less important for the respondent. Looking at the gender difference, it is found that if the enumerator and the respondent are both male, the negative effect of the enumerator being male on reporting to be open to vote for an opposition party is smaller. This suggests that, while a female enumerator is likely to elicit more positive attitude towards any party, also being of the same gender helps to create trusts between the respondent and the enumerator.
These results add to the mixed results from the literature regarding interaction effects (see for instance Catania et al., 1996;Huddy et al. 1997;Himelein, 2016) and provide suggestive evidence that responses may be biased not only by enumerator characteristics but also on how they interact with enumerator characteristics. But caution should be made when interpreting these results. The autors did not pre-specify this analysis, and so this should be considered to be very exploratory. Further research on the differences between enumerators and respondents would help to clarify this potential effect.

Concluding remarks
Using a large-scale experiment in Uganda, this study has documented that the enumerator effect matters for questions about voting preferences. While it is likely impossible to eliminate enumeration effects entirely, there are ways to minimize them. This section now discusses what the results mean for improving data collection on potentially sensitive topics, such as political opinions, depending on the identifying method used.
In the case of randomized controlled trails, the enumerator effect (that is, enumerator characteristics are correlated with the outcome) is a concern if enumerator characteristics are also correlated with treatment status. In this case, the impact of the treatment on the outcome could be biased. 21 It is important to note that, if enumerator assignment is affected by treatment, simply controlling for enumerator fixed effects would not solve the 21 Note that if enumerator characteristics are correlated with the outcome but not with the treatment status the presence of an enumerator effect does not per se imply that the effect of the treatment will be biased. 20 problem. 22 To minimize the bias, researchers should check the balance in the distribution of enumerator characteristics with respect to the treatment status when assigning enumerators to subjects. 23 Yet, balancing enumerator characteristics can be difficult to implement in the field. When this is not possible, at least researchers should balance enumerators themselves between treated and control groups.
In the case of surveys collecting sensitive information and for which there is not a clear identification strategy (for example Afrobarometer, World Bank Living Standard Measurement Survey (LSMS), etc.), the results suggest that these data need to be interpreted cautiously. Also in this case, simply including enumerator fixed effects in analysis is not enough. Data documentation for these surveys should provide, at a minimum, a clear description of how enumerators and teams were assigned. This would allow researchers to check for the presence and the size of the enumeration effect before conducting any analysis.
Note that, unlike other improvements in measurement, the suggestions from this study do not necessarily involve significant additional survey costs and does not generate tradeoffs between accuracy or bias and cost. 24 On the contrary, for most studies, limiting the possible bias from the enumerator effect can be done quickly, relatively easily, and inexpensively, and so represents a low-cost improvement in the quality of collected data. 22 In this case, controlling for enumerator dummies, would not prevent the enumerator effect from biasing the program impact. This bias could be even more problematic if combined with self-report data (e.g. see Barrera-Osorio et al. (2011) and Baird and Özler, (2012)). 23 For example, the study for this experiment -in which enumerators were randomly assigned -has an R 2 for treatment status on observable and personality enumerator characteristics of 0.020 and on fixed effects of only 0.024. This indicates a minimal -if any -bias due to the enumerator effect. 24 Other strategies that could improve the quality of enumerator data collection include using alternative sources of data (i.e. administrative data) (Baird and Özler, 2012) and employing list experiments (see for instance Blair and Kosuke, 2012). Yet, the former it is not always a viable option while the latter does not allow for individual level analysis. Future work on how to incorporate private reporting of information, such as handing an electronic data collection device to a respondent to complete a question, could be fruitful.

Figure 1: Enumerator effect: R 2 and confidence intervals
Source: Data for the study comes from a survey of farmers in northern Uganda. The survey was conducted by Innovations for Poverty Action.
Notes: The graph reports for each question the R 2 and the corresponding (bootstrapped) 95% confidence intervals for each of the four teams.  Notes: Team (language group) fixed effects and a constant are included in all regressions but not shown. Robust standard errors in parenthesis are clustered at the enumerator level. Notes: Team (language group) fixed effects and a constant are included in all regressions but not shown. Long-term enumerator is a dummy which takes value 1 if the enumerator worked at least 25 days (which corresponds to the 75 th percentile in the distribution of the number of work days during the survey) and 0 otherwise. Number of days of work is the total number of days the enumerator worked during the survey period. Robust standard errors in parentheses are clustered at the enumerator level.     (1). For each column, the outcome variable is (an increasing) categorical variable (1-4) which measures the openness to vote for a political party, namely, the Democratic Party (DP), the Forum for Democratic Change (FDC), the National Resistance Movement (NRM), and the Uganda People's Congress (UPC). Work experience as enumerator is the number of months previously worked as an enumerator. Enumerator high education is a dummy that takes value 1 if the enumerator has completed university, teaching institute, or master's program. Want people to know how good can be at work; Strongly motivated by the wage can earn; Enjoy handling new problems; Enjoy trying to solve complex problems; Curiosity is a driving force for your actions are categorical variables taking value from 1 to 4 (decreasing). Each regression includes community fixed effects and a constant. Robust standard errors in parenthesis are clustered at the enumerator level. *** p<0.01, ** p<0.05, * p<0.10.  (1). For each column, the outcome variable is indicated in the first row. Alcohol and tobacco consumption is a dummy that takes value 1 if the respondent consumes alcohol or tobacco and 0 otherwise. Work experience as enumerator is the number of months previously worked as an enumerator. Each regression includes as additional control variables: Want people to know how good can be at work; Strongly motivated by the wage can earn; Enjoy handling new problems; Enjoy trying to solve complex problems; Curiosity is a driving force for your actions (for the definitions see Table 5). as in Table 3. Each regression includes community fixed effects and a constant. Each regression includes community fixed effects and a constant. Robust standard errors in parenthesis are clustered at the enumerator level. *** p<0.01, ** p<0.05, * p<0.10.  (1). For each column, the outcome variable is (an increasing) categorical variable (1-4) which measures the openness to vote for a political party, namely DP, FDC, NRM, and UPC. Enumerator high education is a dummy that takes value 1 if the enumerator has completed university or teaching institute, or has master. Enumerator young (old) is a dummy that takes value 1 if the enumerator's age<=27 (>27) and 0 otherwise. Respondent young (old) is a dummy that takes value 1 if the respondent's age<=32 (>32) and 0 otherwise. Enumerator high education is a dummy that takes value 1 if the enumerator has completed university, teaching institute, or master. Work experience as enumerator is the number of months previously worked as an enumerator. Each regression includes as additional control variables: Want people to know good can be at work; Strongly motivated by the wage can earn; Enjoy handling new problems; Enjoy trying to solve complex problems; Curiosity is a driving force for your actions (for the definitions see Table 5). Each regression includes community fixed effects and a constant. Robust standard errors in parenthesis are clustered at the enumerator level. *** p<0.01, ** p<0.05, * p<0.10.

Supplementary Appendix
Supplementary Appendix S1: Additional Tables   1,2,3, and 4) the R 2 of a regression of each respondent's answer (column 1) on the full set of enumerator fixed effects and a constant. The political preference questions refer to: the Democratic Party (DP), the Forum for Democratic Change (FDC), the National Resistance Movement (NRM), and the Uganda People's Congress (UPC).
Enumerator male -0.584*** -0.537*** -0.093*** -0.547*** -0.587*** -0.481*** -0.073* -0.501*** -0.562*** -0.537*** -0.030 -0.535*** -0.532*** -0.472*** -0.075** -0.494***  (1). For each column, the outcome variable is (an increasing) categorical variable (1-4) which measures the openness to vote for a political party, namely, the Democratic Party (DP), the Forum for Democratic Change (FDC), the National Resistance Movement (NRM), and the Uganda People's Congress (UPC). Enumerator high education is a dummy that takes value 1 if the enumerator has completed university or teaching institute, or has master. Respondent young is a dummy that takes value 1 if the respondent's age<=32 and 0 otherwise. Enumerator high education is a dummy that takes value 1 if the enumerator has completed university, teaching institute, or master. Respondent high educated is a dummy that takes value 1 if the respondent has completed primary education and 0 otherwise. Enumerator high education is a dummy that takes value 1 if the enumerator has completed university, teaching institute, or master. Work experience as enumerator is the number of months previously worked as an enumerator. Respondent poor is a dummy that takes value 1 if the value of the asset index for the household is lower than the mean value of the asset index for the population and 0 otherwise. Each regression includes as additional control variables: Want people to know how good can be at work; Strongly motivated by the wage can earn; Enjoy handling new problems; Enjoy trying to solve complex problems; Curiosity is a driving force for your actions. Each regression includes community fixed effects and a constant. Robust standard errors in parenthesis are clustered at the enumerator level. *** p<0.01, ** p<0.05, * p<0.10