Much of what we know about voting behavior in Africa is based on data from public opinion surveys. However, there has been little investigation into whether reported voter preferences are reliable, or whether they are affected by bias, particularly that which may arise from the social undesirability of “tribalistic” voting.
This study utilises a voting simulation experiment in Uganda and analysis of existing surveys from a number of African countries to show that voters who are observed by others are less likely to report a preference for a coethnic candidate. It also shows that coethnics of the incumbent are less likely than noncoethnics to fail to report a preference at all.
Together, these findings suggest that African respondents intentionally hide their ethnic vote preferences. A counterfactual variable was generated to measure the magnitude of the bias, but the variable estimates coethnic voters’ true, but unstated, preferences. The analysis suggests that approximately 15% of survey respondents must hide an ethnic preference in order to produce significantly attenuated estimates of the effect of coethnicity on the vote.
Key findings:
- The results of the Ugandan experiment presented here raise serious concerns about the reliability of selfreported support for coethnic presidential candidates in Sub-Saharan Africa. A search for bias in existing survey results provides some evidence that this bias exists outside the experimental context, as well as outside Uganda. The results of this analysis suggest that voting behavior measued by Afrobarometer surveys, particularly responses that indicate ethnic preferences, are subject to under-reporting due to social desirabilty bias.
- One solution is to avoid the use of surveys for studying vote choice altogether, and instead to rely on actual voting returns. However, as noted, this places an untenable constraint on researchers. In many African countries, electoral returns are only available at high levels of aggregation, which poses a serious ecological inference problem. In addition, electoral returns may reflect fraud or malfeasance and therefore may be less reflective of true preferences than are survey data. These concerns suggest that the only real solution is to retain the use of surveys, but alter standard survey protocols to allow responses on ethnic questions to be as private as possible, without interaction with either community members or an enumerator.
- These changes have potentially high costs to researchers. In the American context, subjects can have privacy by self-administering questionnaires. However, this usually involves respondents who can read and write, and a substantial minority of African respondents are illiterate. Possible alternatives include having respondents respond to enumerator questions with paper ballots like those used in the experiment presented here, or entirely self-administering parts or all of the the survey using yes/no responses to prerecorded taped questions heard privately through earphones. Explaining these non-standard protocols to respondents adds time to what are already intensive surveys.
- However, these aformentioned protocols may be the only way to ensure reliable results on what is one of the most critical questions on the Afrobarometer survey. Before instituting these changes it may be worthwhile to gather data on voting behavior using smaller purpose-designed surveys that attempt to eliminate survey bias and then compare the results of these studies directly to existing Afrobarometer data to ensure bias is actually present in the particular country being analyzed.
- For the large amount of research that has already been done using survey-based voting indicators, the evidence presented here does not call for scrapping earlier results, nor does it demonstrate that earlier conclusions are wrong. However, it does suggest that conclusions drawn about voting behavior should be taken as tentative until either the data can be shown to be unbiased or the conclusions are confirmed using alternate methods.
