The Utility of the Ethnic Shortcut

As a shortcut, demographics like ethnicityFootnote ²² have the advantage of being extremely cheap for the voter to collect.Footnote ²³ Voters can use easily observable aspects of a competitor's identity to draw conclusions about at least three things: preferences, capabilities and electoral viability. The first two matters constitute the voter's prediction of the competitor's likely behaviour and performance as an office holder, while the latter is important as a strategic consideration.

A voter might use a competitor's ethnicity to draw conclusions about likely behaviour as an office holder in one of at least two ways. First, the voter might access stereotypes about the preferences of members of some groups – their positions on, say, political and social issues – or capabilities – whether they have produced ‘strong’ or ‘weak’ leaders in the past, whether they are trustworthy when handling finances, etc. In Uganda, for example, certain ethnic groups that live primarily in the country's north-east, such as the Karamajong, are widely snickered at by their fellow citizens as ‘uncivilized’, due to their dress and nomadism.

Secondly, in circumstances where the voter believes that the distribution of goods is dependent upon potential recipients’ ethnic identities, that voter will tend to assume that a co-ethnic's preferences would be more favourable. In Africa, where access to state resources often determines an individual's or community's economic success, there are pressures to limit the size of pork-seeking coalitions, so as to maximize each recipient's share.Footnote ²⁴ Identifying members on the basis of their ethnicity is a particularly useful strategy for limiting coalition size, mainly because it is difficult, if not impossible, to change one's ethnicity; this limits opportunistic bandwagoning after elections.Footnote ²⁵

Finally, voters will also weight their considerations of competitors’ potential capabilities and preferences according to these competitors’ electability. Supporting a highly capable candidate whose preferences align perfectly with a voter might not be an optimal strategy for the voter if the candidate's probability of election is near zero. In contexts in which political competition is already ethnicized, this might mean that candidates from demographically small or historically powerless groups might be deemed to have a low probability of electoral success. As Chandra argues, voters consider not only the ethnic makeup of competitors, but also of the relevant electorate.Footnote ²⁶ After considering the identity of both the competitor and the likely voting population, a rational voter might conclude that supporting a co-ethnic is not the best strategy. In sum, strategic voters will determine their final choice on the basis of competitor preferences, capabilities and electability. Ethnicity can provide cues about all three.

However, while ethnicity is a competitor attribute that the voter can identify extremely cheaply, its utility is limited by its informational quality. Generally speaking, cues vary in the quality of the information they provide to the voter. Here, I define the quality of the information in terms of how much it helps the recipient to reduce errors in predicting competitors’ preferences, capabilities and viability. When compared to alternative sources of information about candidates, ethnic cues are of relatively low quality; in other words, voters relying exclusively on ethnic cues are more likely to make decision errors. For example, in their campaign simulations based on elections in the United States, Lau and Redlawsk find that subjects’ use of candidate appearance heuristics is negatively associated with ‘correct voting’.Footnote ²⁷

As a predictor of office-holder preferences, behaviour and viability, ethnicity is relatively problematic. Certainly, group-based stereotypes are poor predictors of individual members’ attributes, be they intelligence, leadership capacity or scrupulousness. And while voters often assume that rule by co-ethnics will result in favourable distributional policies for their group, formal and empirical work have demonstrated that rulers do not necessarily reward co-ethnics, or even core supporters.Footnote ²⁸ Instead, other information, such as past policy positions and distributional behaviour, can be more useful in predicting possible incumbents’ preferences, while factors such as education and past performance as office holders will likely prove better predictors of capabilities. Viability can be assessed with opinion poll data and other measures of the electorate's dynamic sentiments. All of these types of information are arguably of higher quality than ethnicity.

As the price of such higher-quality information declines – due to factors such as the improvement and diversification of mass media infrastructures, launch of civic education campaigns or increase in government transparency – increasing numbers of citizens will be able to broaden the types of information they access. The decreased cost of higher-quality political information will not be likely to lead voters to eschew ethnic considerations completely, but it will mean that these voters, who now have broader informational repertoires, will not rely on ethnic cues exclusively in their electoral decision making.

Within a given ethnic group, rates of ethnic voting will then decline when higher-quality information is present, provided that such information suggests that a non co-ethnic competitor is significantly superior in one or more key aspects (capabilities, preferences, electability) than a co-ethnic one. When voters have higher-quality information that suggests that a co-ethnic would be a more capable governor, possesses preferences nearly identical to their own or would be more likely to win, their rates of voting for a co-ethnic are likely to be indistinguishable from what they would be in a comparable situation in a low-information environment. Only when that higher-quality information suggests relative deficiencies on the part of the co-ethnic competitor will ethnic voting rates within the group decline.

This leads to the following hypotheses:

An important caveat must be entered here regarding Hypothesis 1, however. As previously discussed, most theories on information and ethnicity do not predict that individuals will automatically vote for co-ethnics in low-information environments. Rather, theories on ethnicity and strategic voting, like Chandra's, hold that voters’ support for co-ethnic competitors is conditional on those competitors’ viabilities.Footnote ²⁹ From an instrumental perspective, a competitor with an extremely low probability of winning is of limited utility to the voter, no matter the ethnic identity. Instead, a voter might conclude that supporting a more viable non co-ethnic is a better strategy, particularly in a context in which institutional rules or demographics make it unlikely that any competitor can win with the support of its own ethnic group alone. Even if such a non co-ethnic is likely to provide his or her ethnic brethren with the largest share of spoils, voters might still conclude that they can maximize utility by supporting a likely winner, since communities and individuals who support losers are often the least likely to receive patronage goods.Footnote ³⁰ Finally, as previously discussed, in contexts in which ethnicity is already politically salient, a competitors’ ethnic identity can provide an important, albeit imperfect, cue about electoral viability, with demographically larger or historically more powerful groups deemed more electable.

From this, we can generate an additional hypothesis:

Findings supportive of Hypothesis 1 cannot necessarily lead us to conclude that the informational approach is superior to a purely social-psychological one, since both theories predict the same behaviour in low-information environments: support for co-ethnics. However, support for Hypothesis 1a or Hypothesis 2 would suggest that ethnic voting is not driven primarily by expressive considerations, as a pure social-psychological theory would suggest. Rather, if voters use competitors’ ethnic identities instrumentally, to predict possible incumbents’ preferences, capabilities and viability, then we should expect that rates of support for co-ethnics will be elastic according to broader informational contexts.

Case Selection

The experiment was conducted in Uganda, which was selected due to the country's history of ethnic salience in politics and (relatively) open environment for discussion of political issues. Uganda is one of the most ethnically diverse countries in the world. The largest group – the Baganda – only comprises 16.9 per cent of the population (2002 Census),Footnote ³² while six groups claim more than 5 per cent of the population, and sixteen between 1 and 5 per cent. Much of the country's post-independence history has been marred by violent conflict organized around religious, ethnic and regional divisions. Early political parties developed around schisms between Catholics and Protestants, the country's first severe political crisis (in 1966–67) centred on disagreements over the role of Buganda's kabaka (king) in national politics,Footnote ³³ and certain groups – particularly the Acholi, Baganda and Langi – were victims of mass killings in the 1970s and 1980s.Footnote ³⁴

Even though the Movement system, which banned electoral participation by parties between 1986 and 2005, was ostensibly organized to spare the country the spectre of communalism, ethnicity remains salient in Uganda.Footnote ³⁵ Under the current regime, there has been resentment, particularly amongst the Baganda, that President Yoweri Museveni's home region in the West and his ethnic group, the Banyankole, have benefited unfairly. Uganda's independent press has begun a more systematic examination of these allegations, particularly in the shadow of inter-ethnic violence in neighbouring Kenya, and has found some evidence of biases against certain regions, in terms of government hiring practices, ministerial positions and budget allocations.Footnote ³⁶ The realities of the country – that ethnicity is important in both everyday interaction and in political competition, and that the country is now stable enough as to allow for open discussion of such issues – make it a propitious setting for such research.Footnote ³⁷

Design of the Ugandan Experiment

The experiment involved within-subject manipulations, in that each subject received every level of the treatment. Subjects were asked to listen to vignettes about hypothetical, two-candidate races for sub-county (LC3) chairmanships;Footnote ³⁸ all subjects were informed at the outset that the information they were being given was about imaginary candidates and races.Footnote ³⁹ Subjects were then asked which of the two candidates they would support. Eight of the vignettes included information about candidates’ ethnicities. Here, one candidate was always a co-ethnic of the subject – the subjects were asked to self-identify before the experiment began – while the other candidate was from a randomly assigned Ugandan group with which the subject did not identify.Footnote ⁴⁰

Under the first condition (i.e., the control), the only distinguishing information provided about the candidates was their ethnic identities. The subjects’ responses here served as a baseline for their general likelihood of supporting a co-ethnic. In each of seven subsequent conditions, one piece of information about each of the candidates in addition to ethnic identity was provided. In other words, apart from the first condition, all conditions contained two pieces of information on which the candidates differed: ethnicity and one other item. These conditions followed a pattern, in which the other information that was provided painted the subjects’ co-ethnic negatively and the non co-ethnic positively.

This information fell into seven categories: party affiliation; popularity, as evidenced by crowd size at a recent rally; past involvement in corrupt practices; level of education; issue position on a locally relevant political controversy (land use); past performance as an office holder; and status as a distributor of goods during the campaign. The co-ethnic was always described as from a party to which the subject was not affiliated, generally unpopular, corrupt, undereducated, holding political opinions incompatible with those of the subject, having performed poorly in previously held positions, or as not having distributed goods to voters during the campaign. Conversely, the non co-ethnic was described as being from the same party as the subject, generally popular, scrupulous, well-educated, holding opinions compatible with those of the subject, having performed well in previously held positions, and as having distributed goods to voters.

The null hypothesis was in each case that there would be no significant difference in subjects’ levels of support for the co-ethnic candidate between the baseline and subsequent treatment; rejecting the null hypotheses would suggest that the presentation of certain other types of information impacted the marginal effect of ethnicity on subjects’ vote choices. Such a finding would conflict with pure social-psychological approaches. They would also, however, suggest that extant literature suggesting that the political salience of ethnicity arises from information scarcity needs to be amended to account for possible effects of changes in informational contexts.

All subjects received these treatments in the same order. Because the survey was conducted using pencil and paper, and because interviewers had to tailor vignettes on the basis of subjects’ responses to earlier questions, randomization of question order would have increased interviewer error. Indeed, this does raise possible concerns over spillover effects, which cannot be mitigated entirely. The specific fear here would be that, once a subject had ‘defected’ from a co-ethnic, he or she would be more likely to defect under a subsequent treatment. However, we can be somewhat more confident that the ordering of treatments did not determine results under later conditions because, out of all the subjects who supported a co-ethnic at the baseline, only 2 per cent defected to a non co-ethnic under some future treatment and then never returned to support a co-ethnic in a subsequent round. In other words, the design does not seem to have precluded ‘de-defection’ once ‘defection’ had occurred. In order to try to minimize possible spillover effects, the research assistants reminded subjects at the outset of each round that they were now discussing an entirely different election.

Subjects were selected via multistage, area probability (‘cluster’) sampling; 186 interviews were carried out in Makindye Division (Kampala District) and 184 in Bbaale County (Kayunga District), for a total N of 370. To be sure, the sample's limited size and geographic location (i.e., drawn only from the Central Region) means that it cannot be considered representative of all Ugandans, let alone of Africans generally. However, the sampling protocol was designed explicitly to minimize the likelihood of drawing urban and rural samples that were significantly different from country-wide averages on key economic and social development indicators. Details on these protocols can be found in Appendix A.

Treatments

The information provided under the different conditions was as follows. English-language wordings of all vignettes can be found in Appendix B, while a summary of treatments can be found in Table 1.

Table 1

Summary of Vignette Constructions

Hypotheses 1 and 1a: Ethnic Voting under the Lowest Information Condition

In the first condition, when no distinguishing information other than ethnicities was given, just over four-fifths (80.3 per cent) of subjects supported the candidate whose identity matched their own. This baseline finding presents two important considerations. First, it establishes that, in low-information settings in which a co-ethnic faces a non co-ethnic, most subjects preferred the co-ethnic candidate. Alone, the finding does not suggest whether ethnic voting is driven primarily by expressive or instrumental logics. Secondly, it suggests that most subjects here had no problem with indicating preference for a co-ethnic in the interview setting. In other words, subjects did not seem overly concerned with projecting a certain level of ‘political incorrectness’. This should provide further confidence that the findings below can be attributable to a genuine weighting of ethnic and co-ethnic candidate options on the part of subjects, rather than to a widespread desire to appear open-minded in the presence of an interviewer (i.e., Hawthorne effects).

Next, Hypothesis 1a drew on theories of strategic voting and suggested that all subjects would not be equally predisposed to ethnic voting, and all competitors would not be equally likely to attract non co-ethnic supporters. Subjects who are ethnic minorities in their relevant local areas should be less likely to support a co-ethnic candidate than members of larger groups will be, and competitors from larger or historically more powerful groups should be better positioned to attract the votes of non co-ethnics.

The results of the experiment tend to support Hypothesis 1a. Under the lowest-information condition, about one in six (16.2 per cent) subjects chose a non co-ethnic. Several patterns are notable: (1) Baganda subjects were less likely than others to defect, (2) members of smaller groups were generally more likely to defect than members of larger groups, and (3) subjects were more likely to defect to candidates from powerful groups than to candidates from more marginalized ones. In Figure 1, selected defection rates under the first condition, according to subject group and identity of the non co-ethnic candidate in the vignette construction, are presented.

Fig. 1

Election 1 defections, by group, according to identity of non co-ethnic candidate *Any group with less than 5.0 per cent of the population in the LC4 of research. Banyankole excluded, even though they only comprise 2.7 per cent of the Bbaale population, because of their coding throughout the analysis as a ‘powerful’ group.

First, the data show significantly higher rates of defection among non-Baganda (24.8 per cent) than among Baganda (12.9 per cent) (z = 2.97, p < 0.001) in the first round. One possible explanation for this finding is that Baganda, as the largest group at both research locations, are generally under less pressure during elections to enter into coalitions behind non co-ethnic candidates. In Makindye, where Baganda comprise a majority (55.0 per cent) of the population, pure ethnic voting, assuming ethnic turnout rates proportional to their population distributions, would always elect a Muganda. In Bbaale, although Baganda comprise far less than a majority (20.7 per cent), their status as the largest group (the second largest is Basoga, with 15.9 per cent) makes it more likely that other groups would have to ally with a Muganda, rather than the other way around. In other words, in elections where ethnicity is the only salient factor, as under the first constructed condition, it would be a wise strategy for a Muganda to support a Muganda.

Next, members of groups that make up a very small (less than 5.0 per cent)Footnote ⁵⁴ portion of the total population of their respective survey area were slightly more likely to defect (21.4 per cent) than members of groups that make up at least 15.0 per cent of the local population (14.0 per cent),Footnote ⁵⁵ although the difference in proportions was not quite statistically significant (z = 1.55, p = 0.12).

Furthermore, not only did the identity of the subject matter, but the identity of the non co-ethnic in the contest seemed to make a difference. For example, non-Baganda whose co-ethnic faced a Muganda had a higher defection rate (27.8 per cent) than non-Baganda whose co-ethnic faced a non-Muganda (17.3 per cent), although this difference in proportions is not statistically significant (z = 1.14, p = 0.13). Additionally, members of those ‘micro’ groups were more likely to defect (25.9 per cent) when their co-ethnic faced a member of either of the two largest and most politically powerful groups in the country – a Muganda or a Munyankole – than when he faced a member of another micro group (19.7 per cent), although the sub-samples are too small for this difference to be statistically significant (z = 0.63, p = 0.52). Finally, subjects overall were significantly less likely to defect to an Acholi candidate (9.6 per cent, z = 2.32, p = 0.02) or a Langi (11.6 per cent, z = 1.79, p = 0.07) than to a Muganda; the Acholi and Langi groups are two of the most marginalized in all of Uganda.

These patterns suggest a possible dynamic in which members of smaller groups might have acted strategically in supporting a member of a more powerful group, such as a Muganda (the largest group nationally, and in both Makindye and Bbaale) or a Munyankole (the second-largest group nationally, and the group of President Museveni). In real-world elections, allying with a larger group, or groups, in order to maximize the probability of being in a winning coalition, is a better survival strategy for micro groups than pure ethnic voting.

As a final test of Hypothesis 1a, I conducted logistic regression analysis in which the outcome variable – a subject's choice in the first election (supporting a co-ethnic vs. a non co-ethnic) – was regressed on a number of predictors, including dummies measuring subject membership in a ‘powerful group’ (i.e., Baganda or Banyankole), or a ‘mid-sized group’ (all groups representing at least 5 per cent of the relevant LC4 population, according to the 2002 Census, excluding Baganda and Banyankole). Here, the reference category is membership in a ‘micro group’. Additionally, dummies were included indicating that the randomly assigned non co-ethnic was a member of a powerful group (i.e., Baganda or Banyankole), or a non-powerful, yet non-especially marginalized group. The reference category here was the non co-ethnic competitor being an Acholi or Langi, which are arguably two of the most-marginalized groups under the current regime. Controls are included for subject age,Footnote ⁵⁶ political sophistication (media consumptionFootnote ⁵⁷ and educationFootnote ⁵⁸), household wealth,Footnote ⁵⁹ sex, and local-level ethnic fractionalization.Footnote ⁶⁰

Results are presented in Table 2. As Hypothesis 1a predicts, membership in a ‘powerful group’ is associated with a statistically significant 16.0 per cent decrease in probability of defection in the first condition, vis-à-vis membership in a ‘micro group’. And vignette construction in which the non co-ethnic was a member of a powerful group was associated with a statistically significant 26.2 per cent increase in probability of defection in the first condition, vis-à-vis a construction in which the non co-ethnic candidate came from a marginalized group. In summary, while most subjects supported a co-ethnic in the low-information baseline election, there is evidence that some subjects might have been ‘voting’ strategically, by supporting a more viable non co-ethnic.

Table 2

Determinants of Support for Non Co-Ethnic under First Condition

Notes: Dichotomized outcome variables are listed (1 = support for non co-ethnic). *Significant at 10%; **significant at 5%; ***significant at 1%. Clustered robust standard errors by enumeration area (the PSU). Valid observations total 351, because subjects that reported ‘don't know’ for any of the predictor variables were dropped. For subject group indicators, reference category is Micro Group. For vignette construction categories, reference group is Marginalized Group (i.e., Acholi or Langi). ^†>5.0%, not Baganda or Banyankole ^‡Any group other than Baganda, Banyankole, Acholi or Langi.

Hypothesis 2: Ethnic Voting with Negative Co-Ethnic Information

Under conditions depicting the co-ethnic negatively vis-à-vis the non co-ethnic, support for the co-ethnic declined significantly, thus supporting Hypothesis 2. Here, I discuss subject support for co-ethnics under the party, popularity, corruption, education, issue and performance treatments, before dealing separately with the treatment involving the distribution of goods, due to special considerations. On average, an individual supported a non co-ethnic under 79.7 per cent of the first seven conditions (i.e., all treatments, excluding the distribution of goods treatment, which is discussed separately). Excluding the lowest-information condition, that rate jumps to 89.9 per cent. In other words, the average treatment effect of exposure to negative information about co-ethnics was to raise the defection rate from 16.2 per cent to 89.9 per cent.

Subjects’ preferences for candidates are presented in Figure 2. Orthogonal contrasts between the lowest-information condition and each of the subsequent treatment conditions are all statistically significant.Footnote ⁶¹ Multiple analysis of variance (MANOVA) yields an F statistic of 39.96 (p ⩽ 0.001), suggesting we can reject the null hypothesis of no significant difference in preferences towards co-ethnic candidates across treatments.

Fig. 2

Subject preferences, with ethnic cues Note: Valid percentages for the party treatment are reported, excluding those subjects who were not asked the question because they self-identified as non-partisans.

Vast majorities of those who were loyal to co-ethnics in the baseline (N = 297) defected under every subsequent condition: 81.7 per cent on party, 77.6 per cent on crowds, 94.2 per cent on corruption, 93.6 per cent on education, 97.0 per cent on land use, and 97.6 per cent on performance. In this sub-group, orthogonal contrasts between the baseline and each of the treatment conditions are all statistically significant at the 0.01 level.Footnote ⁶² In fact, only two people (0.5 per cent) refused to defect from their co-ethnic under any circumstance. At least in this experimental setting, Ugandan citizens showed tremendous willingness to cross ethnic lines to vote for a candidate who shared their party preference or position on land use, or was more popular, scrupulous, educated or competent than their co-ethnic.Footnote ⁶³

Special Case: Ethnic Voting and Goods Distribution

The final treatment involved testing ethnic and distributional cues simultaneously. As discussed previously, voters might draw conclusions about competitors’ capabilities, preferences and viability by observing what competitors do (or do not) distribute, and to whom. Following Hypothesis 2, if this information suggests that a non co-ethnic is more capable of distribution, more in favour of distributional practices that would benefit the voter, or is more electorally viable, vis-à-vis a co-ethnic competitor, we should expect that support for the co-ethnic will suffer. However, due to the apparently ambiguous conclusions subjects draw about goods distribution, the findings here (presented in Figure 3) merit separate discussion.

Fig. 3

Preferences for simultaneous ethnic and distributional cues

Prior to the treatment that exposed subjects to ethnic and distributional cues simultaneously, subjects were asked to respond to a vignette about an election in which there were two candidates: one known for distributing bags of sugar at his rallies, the other known for distributing nothing. This was the only piece of distinguishing information provided about the candidates. I refer to this condition as the baseline distributional treatment. Here, nearly two-thirds (65.4 per cent) of subjects did not support the distributing candidate.

This result was not anticipated. Most of the subjects rejected the distributing candidate, even though such campaign-time distributions are common in Uganda. There are three possible reasons for this. First, while one third of the subjects indicated support for the goods distributor, others might have viewed goods distribution as normatively wrong. Secondly, some subjects might have equated distribution with corruption; distributable resources must come from somewhere, after all.Footnote ⁶⁴ Finally, some subjects who might support distributors in the privacy of the polling station might have been uncomfortable admitting that they would do so in front of an interviewer. While the ubiquity of such practices suggests that it is not an extreme taboo in Uganda, some subjects might not have wanted it to appear that their vote could be purchased. No matter the reason, we can still draw interesting conclusions from the results of this treatment.

Fortuitously, the design here does allow, in some instances, for a measure of something not feasible under the other treatments: support for co-ethnics in situations in which positive information about co-ethnics and negative information about non co-ethnics is provided.Footnote ⁶⁵ Those individuals who had indicated support for a non-distributor in the distribution baseline (i.e., ‘anti-distributors’) received (ostensibly) positive information about co-ethnics in the simultaneous treatment. Among those anti-distributors who had supported a co-ethnic in the baseline (N = 192; Column C, Figure 3), it is not surprising that, with further positive information about co-ethnics, almost everyone (89.5 per cent) continued to support a non-distributing co-ethnic under the simultaneous treatment.

However, for those anti-distributors who had ‘defected’ from their co-ethnic in the ethnic baseline condition (N = 43; Column D, Figure 3), over three-quarters (76.7 per cent) supported the co-ethnic, non-distributor in this treatment. This suggests that ‘positive’ information about a co-ethnic can result in increased support for co-ethnic candidates, at least among individuals who had not shown a strong inclination towards ethnic voting under the lowest-information setting.

For those individuals who had, under the distribution baseline, indicated support for the distributing candidate (i.e., ‘pro-distributors’), the simultaneous treatment mimicked the design of the other cross-pressuring treatments, in that (ostensibly) negative information was provided about co-ethnics, and positive information about non co-ethnics. Amongst those pro-distributors who had also supported a co-ethnic under the ethnic baseline (N = 99; Column A, Figure 3), nearly two-thirds (65.7 per cent) supported the non co-ethnic distributor under the simultaneous treatment. In other words, distributional considerations seemed to trump co-ethnicity here, suggesting that distribution could be a viable strategy for winning non co-ethnic supporters, if distributors can identify populations that look kindly upon such campaign activities. Amongst all pro-distributors, over two-thirds (67.2 per cent) supported the distributing non co-ethnic, over the non-distributing co-ethnic.Footnote ⁶⁶

A Real-World Application

Certainly, external validity is a concern with the growing popularity of survey experiments.Footnote ⁶⁷ With regards to this research, perhaps most obviously, it is problematic to conclude that the subject–research interaction and the stimuli applied fully approximate how subjects would act in real-world settings. ‘Voting’ against a co-ethnic in a survey is different from doing so in a real-world election, where community pressures might be overwhelming, and possible punishments for defection severe. Furthermore, the stark construction of the vignettes here, with the co-ethnic always being portrayed purely negatively and the non co-ethnic always purely positively,Footnote ⁶⁸ particularly raises such concerns, since such clear-cut differences between candidates on certain dimensions are often not present. While the construction of the vignettes in this way biases results in favour of confirming the hypotheses, since this is the first research into these questions in Africa of which I am aware, it was more appropriate to risk Type I than Type II errors and, therefore, possibly close off an area of potentially fruitful future research.

In an attempt to address these concerns, I focus here on a real-world case that closely approximates the experimental design – namely, an election in an ethnically heterogeneous district, between candidates representing different ethnic groups, in which available mass media present information about one candidate that is decidedly negative, in comparison to challengers. If the theory is correct, we should see that, for a group with a negatively portrayed co-ethnic in the race, increased access to those mass media should be associated with decreased support for that co-ethnic, and increased support for a more positively portrayed non co-ethnic. In other words, the negatively portrayed co-ethnic should draw lower support from his or her own ethnic group in areas where access to mass media is high, but higher support in areas where it is low (i.e., where voters rely more exclusively on the cues that ethnicity provides).

Electoral cases that meet all of the previously enumerated criteria – particularly that of an overwhelmingly negative informational context for one major candidate, given increased media pluralism – are somewhat sparse, particularly if we limit ourselves to Uganda. However, the 2006 parliamentary election in the constituency of Soroti, a rural county near the shores of Lake Kyoga, in Uganda's Eastern Region, seems an appropriate test case. First, the electorate is ethnically heterogeneous: according to the 2002 Census, the population was 69.1 per cent Iteso and 29.1 per cent Kumam. Secondly, its parliamentary elections in 2006 reflected this diversity to some degree – two of the candidates were Kumam, while two were Iteso.

Finally, the mass media environment in 2006 was uniformly negative towards at least one candidate: Samuel Anyolo, the incumbent. Although Anyolo had ascended to Parliament by defeating the incumbent, Col. William Omaria Lo Arapai, in 2001, the local media environment was decidedly hostile to his re-election efforts five years later. Radio is overwhelmingly the most important mass medium in Soroti County; the 2002 Census reported that 39.9 per cent of households relied on radio as their primary source of news, while 59.5 per cent relied on word of mouth.Footnote ⁶⁹ And local radio was not particularly kind to Anyolo in the run up to the 2006 election. On top of being the subject of numerous media reports about problems with the law, including his alleged involvement in a 2005 shooting at a local bar, Anyolo was unlikely to have received favourable coverage from any of three then-extant local radio stations. Two – the local outlet of the state-run Uganda Broadcasting Corporation (UBC) and the NRM-supporting Voice of TesoFootnote ⁷⁰ – were not friendly to an Anyolo re-election, since the member, a former pro-Museveni stalwart, left the ruling party to run as an independent. And the opposition-leaning Kyoga Veritas FM has provided a more favourable venue to FDC candidates, such as the ultimate victor, Peter Omolo.Footnote ⁷¹ Ultimately, Anyolo saw his vote share decline precipitously, from the 33 per cent he had claimed in 2001, to just 15 per cent.

If the theory is correct, we should expect that the likelihood that Iteso would vote for the incumbent, a co-ethnic, would decline with increasing media access. In order to estimate the relationship between radio access (i.e., exposure to negative information about Anyolo) and support for him by ethnic group, I first use Gary King's ecological inference (EI) methodFootnote ⁷² to obtain point estimates of the share of co-ethnics and non co-ethnics voting for each candidate at each polling station in Soroti County (N = 102 in 2006).Footnote ⁷³ The estimated means, weighted for the number of votes cast per polling station, for each candidate are reported in Table 3. The EI algorithm estimates the means of Iteso and non-Iteso (i.e., mostly Kumam) support for Anyolo in 2006 as 0.19 (SD = 0.002) and 0.06 (SD = 0.004), respectively. These were down significantly from 2001, when mean Iteso support was estimated at 0.39 (SD = 0.003) and non-Iteso support at 0.18 (SD = 0.007) (N = 96 polling stations).

Table 3

Estimated Weighted Mean Shares of Co-Ethnic and Non Co-Ethnic Support for 2006 Soroti County Parliamentary Candidates

Notes: N = 102 polling stations. Means weighted by number of votes cast per polling station. Standard deviations in parentheses.

Table 4 reports correlations between the prevalence of radio reliance in an area and estimates of co-ethnics’ and non co-ethnics’ support of 2006 candidates by polling station.Footnote ⁷⁴ Support for Anyolo in his re-election campaign seems to have been contingent upon the prevalence of radio reliance in an area. The point estimates for Iteso support for Anyolo and radio reliance measures were negatively and significantly correlated (r = −0.227, p = 0.022), suggesting that Anyolo performed better amongst co-ethnics who had less exposure to negative information about him, as the theory would suggest. In areas where less than 40 per cent of households relied on radio (N = 53), the mean estimate of Anyolo's share of the Iteso vote was 0.26 (SD = 0.189), while in areas where prevalence exceeded 40 per cent (N = 49), it was significantly lower, at 0.17 (SD = 0.204). In contrast, Anyolo's support amongst fellow Iteso was not significantly correlated with radio prevalence in 2001 (r = −0.129, p = 0.212), when, as one of several Museveni supporters in the race,Footnote ⁷⁵ he was less likely to have received negative coverage from the only two then-operating radio stations: UBC and Voice of Teso.Footnote ⁷⁶

Table 4

Correlations between Estimated Support for Candidates, by Ethnic Category, and Radio Prevalence (2006 Soroti County Parliamentary Election)

Notes: * Significant at 10%; ** significant at 5%; *** significant at 1%. N = 102 polling stations. P values in parentheses.

Certainly, many of these Iteso might have supported another Etesot, in Engirot Lawrence Okae of the UPC, but the fact that he, as a UPC member, was unlikely to have received favourable media coverage also seems to have harmed him. Indeed, there is a negative correlation between Iteso support for Okae and radio reliance (r = −0.137, p = 0.169), but one that falls short of statistical significance. In general, among Iteso, there is negative and significant correlation between radio reliance and voting for any Iteso candidate (r = −0.267, p = 0.007).Footnote ⁷⁷

In contrast to Anyolo and Okae, the two Kumam candidates in the race enjoyed, on average, more favourable coverage during the campaign, although neither received universal radio support. NRM candidate Ateker Ejalu is likely to have benefited from the UBC and Voice of Teso, and Omolo from Kyoga Veritas. This comparatively positive coverage seems to be associated with increased Iteso support: as radio reliance increased, Kumam candidates’ support from non co-ethnics increased (r = 0.229, p = 0.020).

In sum, the 2006 parliamentary race in Soroti is supportive of the informational theory of ethnic voting tested in the survey experiment, in that Iteso candidates’ performance among co-ethnics suffered significantly as exposure to negative information about them increased. Iteso voters living in areas with higher reliance on radio were significantly more likely than their counterparts with less mass media reliance to support a Kumam candidate. In this case, then, comparatively negative information about co-ethnics does seem to have resulted in decreased ethnic voting, at least among Iteso. Finally, the Soroti case suggests that it is not media access per se that is associated with a diminution of ethnic voting – radio reliance was positively associated with support from co-ethnic candidates amongst Kumams – but the provision of comparatively negative information about co-ethnics.