WPS5236 Policy Research Working Paper 5236 The Ethnicity Distraction? Political Credibility and Partisan Preferences in Africa Philip Keefer The World Bank Development Research Group Macroeconomics and Growth Team March 2010 Policy Research Working Paper 5236 Abstract Much of the research on ethnicity, development and with the credibility of party commitments to pursue conflict implicitly assumes that ethnic groups act collective ethnic interests: ethnic clustering of political collectively in pursuit of their interests. Collective support is less widespread than expected; members of political action is typically facilitated by political parties clustered ethnic groups exhibit high rates of partisan able to make credible commitments to pursue group disinterest and are only slightly more likely to express a interests. Other work, however, emphasizes the lack of partisan preference; and partisan preferences are more political credibility as a source of adverse development affected by factors, such as gift-giving, often associated outcomes. Evidence presented here uses partisan with low political credibility. These findings emphasize preferences across 16 Sub-Saharan African countries to the importance of looking beyond ethnicity in analyses of distinguish these positions. The evidence is inconsistent economic development. This paper--a product of the Macroeconomics and Growth Team, Development Research Group--is part of a larger effort in the department to understand the institutional roots of growth and governance. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at pkeefer@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team The ethnicity distraction? Political credibility and partisan preferences in Africa Philip Keefer pkeefer@worldbank.org Acknowledgements and Disclaimer: The empirical analysis in this paper was made possible by the Stata wizardry of Marek Hanusch, whose assistance and insightful comments are gratefully acknowledged. I benefited from helpful comments on an earlier version of the paper by Jeffry Frieden, David Laitin, Gary Jacobson and Susanne Mueller and seminar and conference participants at the Universities of Michigan, Essex and Pompeu Fabra, and George Mason, Harvard, and Indiana Universities. The opinions and findings here are those of the author and do not represent the views of the World Bank or its directors. 1 The ethnicity distraction? Political credibility and partisan preferences in Africa Researchers have linked competition between ethnic groups to adverse development phenomena ranging from slow growth and the under-provision of public goods to conflict. Other research, looking at the determinants of ethnic voting, has concluded that uninformed voters use ethnicity as a cue about candidate policy preferences. The power of these arguments hinges on whether political competitors can credibly commit to pursuing the collective interests of co-ethnics. If they cannot, their ability to mobilize electoral support with broad ethnic appeals and their political incentives to pursue development-friendly policies are both attenuated. Similarly, ethnic cues are more important when co-ethnic politicians can make credible commitments to pursue the collective interests of their ethnic groups. Whether political competitors can, in fact, make these credible commitments has not been the subject of detailed research. In a first attempt to examine this issue, this paper asks whether political parties are vehicles for the collective organization of ethnic groups. If they are, and if ethnicity is the most salient dimension of political competition, three phenomena should be observed. Ethnic groups should disproportionately favor one party over others; members of such clustered ethnic groups should exhibit low rates of partisan indifference and be much more likely to express a partisan preference than members of other groups; and other strategies for mobilizing political support, particularly those associated with non-credible political commitments, such as gift-giving, should have a small effect on partisan preferences. 2 Though Africa is a central focus of research on ethnicity, voting and development, data from Afrobarometer surveys in 16 countries are inconsistent with these predictions.1 First, ethnic clustering of partisan support appears to be less pronounced than is implied by the literature on ethnicity in Africa. Five different definitions of clustering are explored here. Under the one which comes closest to capturing theoretical descriptions of the role of ethnicity in politics, only 25 percent of respondents belong to clustered ethnic groups. No ethnic groups at all meet this standard in approximately half of the 16 countries. Second, respondents from clustered ethnic groups are not substantially more likely to express a partisan preference than respondents from non-clustered parties. In country-by- country estimates, half or more of the countries exhibit no statistically significant difference between clustered and non-clustered respondents in the likelihood of expressing a partisan preference. In only two countries where they consistently exhibit a significant difference (South Africa and Mozambique) does the evidence support the claim that parties have made credible commitments. The third indication of the lack of credibility of partisan appeals to ethnic groups is that even respondents from clustered ethnic groups express high rates of partisan indifference. A large fraction of these respondents indicate that they are "not close to any party" (around 35 percent, almost three times the rate at which American voters declare themselves to be independents). Finally, fourth, the magnitude of other determinants of 1 The 16 countries are typical. They exhibit the high levels of ethnic fractionalization (.69) associated in the literature with poor development outcomes, similar to levels in the rest of Sub-Saharan Africa and far higher than in the rest of the world (.37) (Alesina, et al. (2002); fractionalization is the probability that two randomly selected individuals are not from the same ethnic group). 3 partisan preference, such as vote-buying, is similar to the effects of ethnic clustering, again suggesting that appeals to the collective interests of ethnic groups are not the main determinant of electoral mobilization in these countries. The analysis cannot reject the possibility that groups still act collectively through other, non-party means. Individual politicians or political families, by virtue of their personal relationships with co-ethnics (e.g., as the patrons sitting atop large clientelist networks), or non-partisan organizations of other kinds, might be able to make credible commitments to most co-ethnics. If these individuals or organizations have only loose partisan connections, strong ethnic support for co-ethnic candidates could exist together with high rates of partisan indifference. Nevertheless, though plausible in principle, these possibilities imply a breadth of clientelist networks or a level of penetration of non-party organizations that are inconsistent with most depictions of African politics. A second explanation for high clustering and high partisan indifference is precisely that political competitors cannot make broadly credible promises and the clientelist networks to which they can make credible promises are narrow. Clients are more likely to be co- ethnics, giving rise to clustering, but most co-ethnics are excluded from clientelist networks and therefore exhibit high rates of partisan indifference. This interpretation is consistent with most analyses of African politics; several variables associated with clientelist appeals, such as gift-giving, are significant determinants of partisan preference in the analysis here. Fear is a third potential explanation of clustering and high rates of indifference. Politicians cannot make credible commitments to defend the interests of an ethnic group, giving rise to partisan indifference, but the belief that co-ethnic politicians are less likely to predate on voters than others leads to ethnic clustering of political support. To partially account for this, the analysis examines the effect of ethnic grievance on partisan preference. 4 Different explanations for ethnic clustering call for distinct policy responses to the appearance of ethnic conflict. If ethnic groups are collectively organized and policy failures represent conflicts among genuinely divergent ethnic preferences, then appropriate policy responses are institutional arrangements that facilitate policy compromises. In contrast, if ethnicity is a side effect of the lack of political credibility, efforts to build parties or other organizations (ethnic or not) is more appropriate. The next section of the paper briefly reviews the broad literature on ethnicity in development and in the politics of developing countries. The argument is then developed that credible ethnic appeals should influence ethnic clustering of partisan support and rates of partisan preference. The empirical strategy, particularly the construction of variables that distinguish ethnic groups that cluster on parties and ethnic groups that do not, is then described. As the final section explains, the results suggest greater focus on underlying factors in African politics that give rise both to the salience of ethnicity in politics and to policy choices that hinder development. Ethnicity, credibility and development in the literature This paper contributes to two strands of research on ethnicity. One investigates the impact of ethnic fragmentation or polarization on development outcomes. The findings below call into question the implicit assumption in this work, that ethnic leaders credibly represent the preferences of their ethnic groups. The other strand asks why ethnic voting emerges. The analysis supports one, under-examined explanation, the inability of politicians to make credible commitments to broad groups of citizens. Easterly and Levine (1997) show that ethnic fragmentation is associated with slow growth, particularly in Africa, and refer to two arguments in the literature to explain the finding. Ethnicity may lead to a common pool problem in which each group predates on 5 public resources without taking into account consequences for other groups. It may also increase the tendency of each group to oppose reforms in the broader public interest. Miguel and Gugerty (2005) show that levels of locally-provided public goods are lower in more ethnically fragmented communities in Kenya. Alesina, Baqir and Easterly (1999) and Alesina and La Ferrara (2005) emphasize that because of preference differences among ethnic groups, ethnic diversity can reduce public good provision. The assumption underlying these arguments is that politicians represent the collective interests of ethnic groups and those interests conflict. In contrast, the analysis here indicates that African political parties, at least, do not appear to mobilize support with credible commitments to serve the collective interests of any groups, including ethnic groups. Research has also linked ethnicity to conflict, again a phenomenon to which Sub- Saharan Africa is exceptionally vulnerable. In a careful analysis of ethnicity and conflict, Montalvo and Reynal-Querol (2005) argue ethnic polarization, rather than fragmentation, drives conflict. To test these arguments, they devise empirical measures of ethnic polarization that closely reflect the analysis in Esteban and Rey (1994), who conclude that polarization is greatest when groups are approximately equally sized and large. The collective organization of polarized groups has two effects on conflict. One is ambiguous: collectively organized ethnic groups are better able both to attack other ethnic groups and to defend against aggression: defensive capability deters conflict, but the capacity to act offensively encourages it. Another is not: collectively organized ethnic groups are better able to make credible commitments with each other that forestall conflict. For example, Bardhan (1997) concludes that polarization leads to conflict when institutions that allow for credible inter-ethnic agreement are absent. Fearon and Laitin (1996), though, argue that inter-group agreements are more credible when groups 6 themselves can police and punish deviations by own-group members against the other group. Lake and Rothchild (1996) also trace conflict to a breakdown in the collective organization of an ethnic group. They observe that ethnic conflict is most likely when political entrepreneurs, operating within ethnic groups, exploit fears of insecurity. Other research examines how leaders exploit ethnicity to trigger conflict (Fearon and Laitin 2000, Blimes 2006). Such behavior is less likely to emerge when ethnic groups can collectively constrain counterproductive behavior by group leaders. The analysis here can therefore explain why ethnicity and conflict are particularly linked in the African context: intra-ethnic organization, as manifested in the ability of political parties to make credible ethnic appeals, is weak, making inter-ethnic agreements difficult to enforce.2 A large literature examines ethnic voting, particularly in Africa. In a recent contribution, Eifert, Miguel and Posner (forthcoming) show that respondents to Afrobarometer surveys were 1.8 percent more likely to identify themselves ethnically rather than in terms of another social categorization, such as occupation, for every month closer a competitive election was to the timing of the survey in a country. Lindberg and Morrison (2008) question the importance of ethnicity, arguing that in their Ghanaian sample, most respondents explain their political preferences in terms of incumbent performance; few acknowledge ethnic motivations or vote-buying. Ferree, Gibson, Hoffman and Long (2009), also looking at Ghana, also find that performance evaluations are important, but in their analysis ethnicity is still a significant determinant of partisan preferences. 2 This evidence is also consistent with evidence in Keefer (2008), showing that insurgency is more likely when political parties are vehicles of individual leaders and party members are less able to act collectively to sanction leaders for not defending their interests. 7 Ferree (2006) summarizes three common explanations for ethnic voting. Ethnic voting could be a simple expression of ethnic identity, unprompted by politicians' commitments to serve ethnic interests in any tangible way. It could also be coincidental: policy cleavages simply overlap with ethnic differences. Finally, as Chandra (2004), Posner (2005), and Ferree argue, ethnic voting could emerge because uninformed voters use ethnicity as a cue about politician preferences. The second and third differ only with respect to voter information. The effects of both, however, depend on whether voters also believe that politicians have incentives to avoid shirking on commitments they make to co-ethnics. The results below point to a complementary explanation for ethnic voting. In an environment in which political competitors cannot make credible commitments to broad groups of voters, they resort to clientelist appeals to members of their clientelist networks ­ networks defined precisely by the ability of members to make mutually credible commitments (Keefer and Vlaicu 2008). This argument is consistent with a large literature arguing that politics in Africa is largely clientelist and that political appeals are rooted in patron-client ties. Bratton and van de Walle (1997) emphasize clientelism and the dominant role of the "big man" in African politics. Van de Walle (2001) concludes that in many African countries "democratization has had little impact on economic decision making, because the new democratic regimes remain governed by neo-patrimonial logic (p. 18)." Broadly credible appeals to favor an entire ethnic group are generally inconsistent with such neo-patrimonial logic, though the clients of "big men" may primarily be co-ethnics. Specific country studies support the argument that the clients of "big men" are more likely to be members ­ though only a subset ­ of the same ethnic group. Stroh's (2009) field research in Burkina Faso emphasizes that parties look for candidates who are fils du terroir, sons of the earth, who are deeply rooted in their local communities ­ that is, who 8 control a substantial clientelist network, regardless of ethnicity. Erdmann (2007) argues that clientelist networks are likely to be ethnically homogeneous. Political competition rooted in patron-client relationships is less likely to yield political parties capable of making broad appeals to citizens: the large fraction of citizens who are outside of clientelist networks have no reason to believe the promises of political parties that represent the association of clientelist networks, even if those are comprised of co-ethnics. The lack of credibility of political promises can also explain the association of ethnicity and poor development outcomes found, for example, in Easterly and Levine (1997). Keefer and Vlaicu (2008) formally demonstrate that political competitors unable to make broadly credible commitments are more likely to rely on patron-client appeals, restricting their credible commitments to narrow groups of voters. This has policy consequences ­ weak public good provision, high rent-seeking, and significant emphasis on the provision of narrowly targeted policy benefits ­ that the literature has also associated with ethnic politics.3 It also leads to ethnic clustering of partisan support if parties are constructed around ethnically homogeneous clientelist networks. Any hypothesis of ethnic voting rooted in differences in policy preferences between ethnic groups, such as the use of ethnicity as a cue to candidate policy preferences, also depends on the credibility of politicians claims to pursue the collective interests of co- ethnics. If they cannot, they are more likely to pursue their private interests at the expense 3 The policy choices of younger democracies ­ those in which politicians are likely to struggle with credible commitment ­ exhibit the pattern of policy performance associated with non-credible politicians. However, controls for ethnolinguistic fractionalization do not account for this difference. On the contrary, they are usually insignificant in the presence of controls for the years of continuous competitive elections (Keefer 2007). 9 of collective interests, rendering the ethnic cue less valuable. Political parties play a key role in establishing the credibility of these ethnic appeals. Theory and Hypotheses The empirical tests below are based on the argument that groups' collective interests are more likely to be represented in political decision making when a political party can credibly commit to representing them. If political parties can make credible appeals to pursue the collective interests of an ethnic group, members of that ethnic group should exhibit low rates of partisan indifference and be more likely to express a partisan preference than members of unrepresented ethnic groups. Groups, including ethnic groups, confront significant obstacles to ensuring that their collective interests are represented in political decision making. One challenge is to verify that the politicians that they support share the group's policy preferences. Another is to monitor whether they actually pursue those interests once in office. Individually, group members have limited incentive to collect this information. Lacking it, they have no reason to believe that politicians will pursue the group's interests. A further challenge is to punish politicians who renege on commitments to serve group interests. Again, individually, group members have limited capacity to punish politicians who renege on their commitments to the group. Instead, groups that are not organized to act collectively are in the position of voters in Ferejohn's (1986) model: to the extent that they can spontaneously coordinate on a performance threshold, group members can reject politicians who do not deliver a threshold level of welfare. The absence of a credible challenger to the incumbent means, however, that this threshold is low. More importantly, even this level of accountability requires spontaneous coordination on a common performance threshold, which is not assured. 10 Political parties can mitigate these collective action problems.4 Snyder and Ting (2002), for example, describe conditions under which parties' strategies of candidate recruitment provide information to voters about candidate policy preferences. Party affiliation then provides benefits to candidates that they cannot obtain outside the party: individuals vote for them because of their party label and not because of direct information about the candidates' individual characteristics. Parties can also limit excessive rent-seeking by politicians in ways that individual group members cannot, by coordinating group responses to non-performing politicians, ensuring at least the level of performance suggested by models of ex post voting behavior, as in Ferejohn (1986). They can also facilitate non- electoral collective action against politicians who shirk (e.g., demonstrations). Finally, following Aldrich's (1995) argument, parties that credibly represent the interests of a group allow members of the group to punish the party's politicians collectively for policy failures, even when they are uncertain about the contribution of any individual politician to the failure. Knowing this, a party's politicians have an incentive to block rent- seeking by any one of their group unless they all receive a share of the rent that at least offsets the electoral losses that all will experience as a result of rent-seeking. The larger the share of rents that a president or head of a legislative assembly has to surrender to members 4 Non-party organizations can also facilitate collective political action by groups; the analysis below cannot exclude the possibility that they do in Africa. However, if this were true, political parties should still play a large role. Though unions in Great Britain, for example, represented the group interests of workers, political candidates favoring union interests ran as candidates of the Labour Party, to which unions provided the bulk of the financing (Eggers and Hainmueller 2009). In this case, levels of partisan indifference among the working class continues to be a valid signal of the degree to which their collective interests are represented. 11 of the legislature, the less attractive rent-seeking is to them. Even if individual politicians can each make credible commitments to some co-ethnics, the absence of a unifying ethnic party makes it more likely that they will shirk on commitments to the entire ethnic group. Though political parties organized to pursue the collective interests of an ethnic group allow the group to limit shirking by co-ethnic politicians, parties are often not organized for this purpose. Citizens do not believe that parties will discipline members who shirk; party candidates do not believe that their support base hinges on their party affiliation. If such a party does exist, members of the targeted ethnic group should be less likely to indicate, in answer to an Afrobarometer question, that they do not feel "close to any party." They should also be more likely to express a partisan preference than members of ethnic groups that are not the target of appeals, depending on the salience of ethnic preferences. There are three cases. In the first, non-ethnic parties appeal equally to all ethnic groups: the fraction of each ethnic group that "does not feel close" to these non-ethnic parties is similar across all ethnic groups. In this case, members of an ethnic group that is targeted by a credible ethnic party should exhibit lower levels of partisan indifference or distance than members of other groups. The more important is ethnicity, the larger should be the difference. In the second, non-ethnic parties actually appeal more to targeted ethnic groups than to those unrepresented by a party. Here again, as long as the ethnic party is credible, the targeted ethnic group should exhibit significantly lower rates of partisan indifference or distance compared to other ethnic groups. In the third case, non-ethnic parties appeal more to non-targeted ethnic groups. Members of the targeted ethnic group could express higher levels of partisan indifference, depending on whether ethnicity is more salient to the targeted ethnic group than non-ethnic issues are to the non-targeted groups. For example, ethnic groups may not be targeted by an 12 ethnic party because, for these groups, ethnicity is simply less salient than other issues (contrary to the presumption of most observers of African politics). However, it should still be the case that, after controlling for respondent characteristics that make them susceptible to the appeals of non-ethnic parties, ethnic groups targeted with credible ethnic appeals should exhibit higher rates of partisan preference. In sum, if ethnic groups are the target of appeals by credible ethnic parties, members of the ethnic groups must disproportionately vote for those parties (exhibit ethnic clustering). If they do not cluster, either parties cannot mobilize the support of the ethnic group (e.g., with credible appeals, by mobilizing ethnically homogeneous clientelist networks, or with fear) or ethnicity is simply not salient. The foregoing arguments then make two predictions. First, those who cluster should be substantially more likely to express a partisan preference than members of other ethnic groups; and, second, alternative party strategies for mobilizing support should have a weaker impact on expressions of partisan interest than membership in a clustered ethnic group. Specification and data Equation (1) is estimated to assess the degree to which members of clustered ethnic groups members are more likely to express a partisan preference than members of other ethnic groups. It assumes a logistic distribution and relies on survey data on individuals i from 16 countries j. (1) Partisan Preference ij (0,1) 0 1 Ethnic Clustering ij ij 2 ij j The ethnic clustering dummy variable is equal to one for ethnic groups who disproportionately support particular parties; every respondent from an ethnic group shares the same clustering value, regardless of whether the respondent supports the cluster party. If 13 a substantially higher fraction of co-ethnic respondents express a partisan preference, the coefficient 1 will be large and significant. Fixed country effects, j are likely to be (and are) large. The base specification controls for these. In general, the addition of group indicators to conditional likelihood estimators like the logit yields inconsistent estimates. However, this is only a concern when the number of groups is large and the observations in any group are small. In the sample here, with only 16 groups (countries) and more than a thousand observations per group, bias is not a concern (Chamberlain 1980). The effects on rates of partisan preference of ethnic clustering and other control variables may also vary across countries: j j . For a large number of countries, each with a relatively small number of observations, the appropriate econometric response is to take parameter heterogeneity explicitly into account with a generalized least squares approach. However, since the country-specific idiosyncratic effects j are likely to be uncorrelated across countries, it is inefficient to use GLS, particularly because the number of countries is small relative to the number of observations per country. Instead, to examine the possible effects of parameter heterogeneity, the final section of the paper simply reports country-by-country estimates. The data on partisan preferences, ethnic clustering and the controls are taken from the 2005-06, third round of the Afrobarometer surveys. Afrobarometer uses a nationally- representative sample (generally, 1,200 respondents, but in larger countries 2,400). Only this round includes questions key to the analysis here. For example, earlier rounds did not ask for the ethnicity of respondents, but instead asked respondents to name the social group with which they most closely identified (used in Eifert, et al. (forthcoming)), a question that, in turn, was not asked in 2005-06. 14 Measuring expressions of partisan preference The dependent variable, whether respondents express a partisan preference, is based on question 85. All respondents who answered question 85 with "do not feel close to any party", refused to answer or did not know are coded as having no partisan preference. Among these respondents, the vast majority answered "not close to any party".5 Rates of partisan indifference are high, by itself strong evidence that political parties across the 16 countries struggle to mobilize support with credible pre-electoral commitments to voters. Partisan indifference is near or above 50 percent in four of them and above 30 percent in eleven. While Eifert, et al. (forthcoming) find that elections matter for the salience of ethnicity, rates of partisan indifference appear to be high regardless of election proximity. In the six countries with elections in 2004, just before the surveys were conducted (Botswana, Ghana, Malawi, Mozambique, Namibia and South Africa), 31 percent of respondents did not express a partisan preference (Mozambique was lowest, with 18.1 percent). The competitiveness of these elections seems also not to have mattered. Botswana, Ghana and Malawi had the most competitive elections (the government party received 52 percent, 56 percent and 47 percent of the vote respectively), and exhibited rates of partisan indifference of 22, 34 and 40 percent, respectively. Rates of partisan preference are low even in countries where ethnic voting is notable. Although Kenya exhibits high rates of ethnic voting, between 35 and 40 percent of Kikuyu, Kalenjin and Luo express support for no party. These numbers are high when compared to surveys undertaken by the American National Election Studies from 1952-2004, asking "Generally speaking, do you usually think of yourself as a Republican, a Democrat, an Independent, or what?" It is plausible that 5 For example, this is true for all 403 of the Ghanaian respondents coded as having no preference. 15 those who answer Republican or Democrat would also answer "yes" to the Afrobarometer question, "Do you feel close to any particular political party?" The fraction of respondents who do not answer Republican or Democrat ranges from eight to 18 percent (in 1974, after the Nixon administration and the Watergate scandal), and is most typically eleven or twelve percent, well below even the lowest rates of partisan indifference in these 16 countries.6 Measuring ethnic clustering Ethnic clustering is an important indicator in nearly all discussions of ethnicity in politics. Easterly and Levine (1997) motivate their cross-country findings that ethnic fragmentation reduces growth with evidence from Kenya. For example, the Luo candidate for president won 75 percent of the vote in the Luo region and the Kalenjin candidate 71 percent of the vote in the Kalenjin region. The literature is silent on how precisely to measure ethnic clustering, however, generally suggesting only that ethnic groups disproportionately support a party or parties. This "lax" clustering standard allows a single ethnic group to cluster on multiple parties or multiple ethnic groups to cluster on the same party. However, the lax standard is likely to show high rates of clustering in countries in all multi-ethnic countries, regardless of whether parties use ethnicity to mobilize support. The polarization literature, in contrast, emphasizing homogeneous groups with heterogeneous 6 American National Election Studies, http://www.electionstudies.org//nesguide/toptable/tab2a_2.htm. The ANES asks follow-up questions to probe the strength of partisan affiliation more deeply, identifying "leaning independents", "weak partisans" and "partisans". The sum of true independents and leaning independents ranges from 25 to more than 40 percent of the electorate. A comparable procedure in the Afrobarometer survey would likely have a similar effect on measures of partisan indifference, by identifying, among those who say they are "close to a party", a large group that is "less close". 16 preferences, implies a stricter "monogamous" standard, in which clustering occurs if one and only one group disproportionately supports one party and no other.7 The analysis here therefore examines five clustering rules. At one end is the lax definition. An ethnic group clusters when the fraction of respondents from group i who ij i support party j exceeds the ethnic group's share of the population, or , where ij i ij N is the number from ethnic group i who support party j, and N is the total population of respondents who express a partisan preference.8 This admits all manifestations of clustering, including those matching many groups to a party and many parties to a group. In constrast, the strict, monogamous definition allows one and only one ethnic group k to cluster on one kl kj k and only one party l: k , but for all other parties j, and for all other i il N i ij N il groups i, i . i il N Three intermediate rules are also examined. One applies the lax rule, but requires ij i that .25 : the group's share of support for the party must exceed its population i ij N share by at least 25 percent. It is possible that multiple parties represent the interests of an ethnic group. The second intermediate rule, "one group-many parties", captures this by coding ethnic groups as clustered if they disproportionately support one or more parties, but 7 By this "monogamous" standard, if even a very small group clusters on a party also preferred by a large ethnic group, neither group is classified as clustered, but both cluster by the "lax" standard. 8 This is consistent with most discussions of ethnic clustering, which look at the votes a group casts as a fraction of total votes cast rather than of all registered voters. 17 no other groups cluster on those same parties. Finally, some ethnic groups may in fact have common preferences.9 The third intermediate cluster variable, "one party-many groups", therefore classifies ethnic groups as clustered if they disproportionately support one and only one party, whether or not other groups also cluster on that party. Small ethnic groups and small parties can introduce distortions. If a tiny ethnic group clusters on a large party that is also the favored party of a large ethnic group, neither group is classified as strictly clustered. In contrast, if a small fraction of a large ethnic group provides disproportionate support for a very small party, it could be classified as strictly clustered. To minimize such distortions, respondents from ethnic groups that comprise less than five percent of the sample are simply excluded from the analysis. In addition, groups are not recorded as clustering if they disproportionately support parties with less than 5 percent support in the sample. The clustering variables are based on answers to two questions in the Afrobarometer survey. Question 86 of the Afrobarometer survey asks, conditional on respondents expressing a partisan preference, which party they prefer. Respondent ethnicity is based on question 79, asking respondents to self-identify their ethnic group with the open-ended question "What is your tribe?" and the prompt "You know, your ethnic or cultural group." Ethnic classifications are therefore generated by the self-perceptions of respondents and are 9 For example, the salient ethnic groups that emerge from the self-identification question in Afrobarometer are sometimes more disaggregated than those identified by outside observers (e.g., Alesina, et al. 2002). The "many groups-one party" clustering rule then captures whether a party makes a credible commitment to pursue the common interests of these ethnic groups. 18 not driven by external judgments; nor are they influenced by pre-set classifications into which respondents are expected to fit themselves. Self-identification could introduce bias. As Chandra (2004) and Posner (2005) argue, ethnic identification can change depending on the salience of ethnicity in politics: if politicians make ethnicity a political issue, individuals are more likely to self-identify as members of an ethnic group. If this introduces bias, however, it is in favor of the hypothesis that ethnic clustering is associated with a higher rate of partisan preference. The ethnicity that individuals choose for themselves is at least partly determined by the appeals made by parties. If parties make credible appeals to an ethnic group, individuals are more likely to self-select into that ethnic group, increasing the fraction of the respondents from that ethnic group that express support for a political party. In any case, however, the results below do not depend on the proximity of elections in countries. Moreover, only 1,410 out of 23,093 respondents, about 6 percent, did not assign themselves an ethnic identity. Almost half of these are from Tanzania, 619 (out of 1304 respondents). Tanzania, consistent with the arguments here, exhibits little clustering and clustered groups are not more likely to express a partisan preference. Control variables Numerous control variables capture non-ethnic factors that might influence respondents' partisan preferences in the event that ethnicity is less salient. The coefficient on the clustering variable is insensitive to the presence of these controls for these factors, however, suggesting that the relative salience of ethnic and non-ethnic issues injects little or no bias into the estimates here. Non-ethnic appeals could concern citizen wealth, religion, education, gender, occupation, or age, all of which are controlled for. The specifications include a variable, 19 based on Question 93, that is the fraction of five assets that households own: a radio, television, bicycle, motorcycle and car. Results are insensitive to alternative formulations, including the use of five dummy variables. Controls are also included for whether the respondent has a religious affiliation, what religion that is, and how often the respondent participates in religious services (questions 91 and 92). The survey asks for respondent's highest level of educational attainment (question 90), gender, age and occupation; these are also taken into account. Harding (2008) uses citizen access to services as a proxy for whether respondents are urban residents or not, a status that is not captured directly in the Afrobarometer survey. Here, the services variable is the average of dichotomous variables created from answers to question 116, indicating whether respondents have easy access to a post office, school, police station, electricity, water, sewage and/or health clinic. Consistent with the notion that clientelist appeals and networks are weaker in urban than in rural areas, access to services is strongly, negatively associated with expressions of partisan preference. The efficacy of partisan appeals to an ethnic group might also vary with the share of the group in the overall population; each group's population share is therefore taken into account. Though the literature on African politics does not anticipate this, parties could mobilize support through their policy stance on the role of government in the economy, as in established Western democracies. Parties that succeed in doing so would attract the support of citizens who share their views on these issues, leading these respondents to express a partisan preference. Respondent attitudes to government intervention in the economy are reflected in question 19 of the Afrobarometer survey, which asks whether the respondent believes that people should look after themselves and be responsible for their own success or whether the government should bear the main responsibility for their well- 20 being. Approximately 47 percent of all respondents believe that individuals are responsible for their own welfare, while 49 percent believe that government has an important role to play. This variable is equal to one for respondents who believe that individuals are responsible for their own success and zero otherwise. Opinions about democracy are relevant in two ways. First, some parties are associated with the transition to democracy; the appeal of such parties can be captured by controlling for respondents' attitudes towards democracy. Second, a sense of civic responsibility may lead respondents to support some party, independent of the credibility of the party's commitments, in order to fully participate in the democratic process. In question 37, respondents indicate whether they believe that democracy is preferable to any other kind of government, whether non-democratic governments might be preferable, or whether the type of government does not matter. More than 60 percent of respondents agree that it is the best form of government. The democracy variable is coded one for respondents who believe democracy is always better and zero otherwise. In all democracies, partisan preference is strongly associated with ex post evaluations of incumbent performance. On the one hand, good performance should inspire greater partisan attachment. On the other hand, respondents' evaluation of incumbent performance could be influenced by whether the politician is from the party they favor. However, this second possibility is most likely when political parties can credibly represent the interests of broad groups of citizens. One indication that, unlike the United States, partisan preference does not drive incumbent approval, is the high percentage of African respondents ­ 34 percent ­ who approve of incumbent performance, but do not express a partisan preference. The corresponding figure in the United States is only 8.6 percent (American National Election Studies, http://www.electionstudies.org//nesguide/toptable/tab2a_2.htm). 21 In a non-credible environment, the association between incumbent approval and partisan preference should be asymmetric. In line with Ferejohn (1986), citizen reliance on ex post evaluations should lead them to prefer the party of the incumbent who meets a performance threshold. However, when challengers are not credible, disapproval of incumbent performance should not be associated with higher rates of partisan preference. Questions 65a and 65b in the Afrobarometer survey gauge respondent opinions of the incumbent with the questions, "How well or badly is government managing the economy?" and, "How well or badly is the government creating jobs?" Nearly all respondents express a negative or positive opinion (94 and 96 percent, respectively). Two variables are created from these questions, equaling one if respondents answered "fairly well" or "very well" and zero otherwise. If both incumbent and challenger parties are credible, this variable should be insignificant. Ethnic grievances can also influence partisan preferences. They might emerge because the group is unrepresented, or grievances might prompt the emergence of a party that represents the group. The grievance variable is based on two questions. Question 80a asks whether the respondent believes the economic condition of the respondent's ethnic group is worse or better than that of other groups. Question 81 asks whether the respondent believes the government treats the respondent's ethnic group unfairly. Ethnic grievance is then a dichotomous variable that equals one if, in response to question 81, respondents indicated that their group was often or always treated unfairly by the government and if, in response to question 80a, respondents indicated that their group was economically worse off or much worse off than other groups.10 11.5 percent of 10 The majority of those with grievances belong to large ethnic groups. In Ghana, 67 of 140 22 respondents expressed a grievance, which is slightly negative correlated (-.03) with the lax definition of clustering and slightly positively with the strict definition (.03 percent). Gift-giving is an alternative strategy of electoral mobilization that scholars associate with weakly credible politicians. Pre-electoral gifts can be rationalized as efforts by politicians to build credibility (Lindberg 2003 makes this point for Africa), or as emerging when politicians can sanction voter shirking in the gift-vote exchange more easily than voters can sanction politicians for shirking on their commitments to pursue particular policies in exchange for votes.11 Scholars often use pre-electoral gifts as an indicator of clientelist strategies of electoral mobilization (e.g., Harding 2008). Question 57f asks whether, during the last election, a candidate or someone from a party offered something a gift in return for the respondent's vote. 19 percent of respondents report having been offered a gift. In the African context, gift-giving is generally controlled by candidates and usually aimed at building their personal constituency. To the extent that candidates, rather than parties, take credit for gifts, the estimated effect of gift- giving on partisan preferences is likely to be a lower bound on its electoral importance. aggrieved respondents were Akan or Ewe, each of which dominates a political party in Ghana. Among the aggrieved Akan and Ewe, about two-thirds express a partisan preference, about the same as the non-aggrieved. In Kenya, 60 of 154 Luo claimed an ethnic grievance, of whom 27 percent expressed partisan indifference. This is not substantially lower than the 35 percent indifference rate of Luo respondents who did not claim an ethnic grievance. 11 For example, in the Philippines, notorious for its non-programmatic parties, vote-buying in municipal elections, is verifiable: ballots are printed on carbon paper; voters can retain their carbon copies and present them for payment to the candidate whom they supported. Stuti Khemani, personal communication. 23 Table 1: Summary of main variables Obs. Mean Std. Dev. Min Max Any partisan preference (0=disaffection)? 23075 .59 .49 0 1 Prefers democracy? 23093 .62 .45 0 1 Any opinion about democracy? 23085 .72 .45 0 1 Believes individual is more responsible for own 23093 .47 .50 0 1 welfare than government ("self-reliance")? Any opinion about "self-reliance"? 23083 .96 .19 0 1 Approves of incumbent performance on job 23093 .27 .45 0 1 creation? Any opinion about incumbent jobs performance? 23093 .96 .19 0 1 Approves of incumbent performance on the 23093 .49 .50 0 2 economy? Any opinion about incumbent economy 23093 .94 .23 0 1 performance? Respondent offered gifts by candidates? 22822 .19 .39 0 1 Frequency of candidate gift giving (Never, rarely 0, 21951 .72 .45 0 1 Usually, always, 1) Do politicians keep their promises? (Never/rarely, 0­ 22488 .17 .37 0 1 Usually/always,1) Ethnic grievance? 23093 .11 .32 0 1 Fraction of seven public services they receive 23058 .43 .29 0 1 The surveys also ask respondents whether they believe politicians generally offer gifts. Question 78f asks respondents how often politicians offer gifts to voters during campaigns. More than 70 percent of respondents across all surveys report that politicians do this often or always. The analysis also controls for a variable created from question 78c, "In your opinion, how often do politicians keep their campaign promises after elections?" The dichotomous variable created from this question is one if respondents answered always or often and zero if they answered rarely or never. 17 percent of respondents indicate that politicians always or usually keep their promises. Table 1 summarizes all of the variables used in the analysis. 24 Ethnic clustering and partisan preferences Rates of clustering depend heavily on which of the five cluster rules is used, but the rates at which members of clustered and unclustered groups express a partisan preference differ by much less. Table 2 summarizes these differences. Under the lax rule, clustering is nearly universal, suggesting that such a rule might record any country with a diverse population as exhibiting clustering, regardless of the salience of ethnic appeals. The rates of clustering drop significantly under the other four rules. Table 2: Summary of clustered respondents according to different cluster rules Cluster rule Percent of respondents from clustered ethnic groups (# countries with clustered ethnic groups) Lax (many parties-many groups) 89.7% (16) Lax with 25% threshold 44.8 (15) One group-many parties 28.8 (11) One party-many groups 60.5 (16) Monogamous (one party-one group) 23.4 (8) Note: See text for explanation of cluster rules. There are 17,353 observations in which respondents answered questions about ethnicity and partisan preference and belonged to ethnic groups with more than 5 percent of the respondent population. Applying the 25 percent threshold to the lax rule, clustering drops to 49 percent of respondents. By requiring clustered ethnic groups to be represented by only one party ("one party-many groups"), the rate of clustering falls to 60.5 percent. The restriction that only one group can cluster on a party reduces the observed rate of clustering by an additional 30 percentage points to 28.8 percent. The combination of both criteria in the monogamous cluster rule yields a further drop in observed rates of clustering to 23.4 percent. The number of countries that exhibit any clustering follows the same pattern: all 16 countries under the lax rule, but only eight under the monogamous rule. 25 Kenya provides a useful illustration of how the rules work in practice. Discussions of ethnic conflict there always invoke the Kikuyu and Luo. Each of these groups easily meets the lax standard for clustering: Kikuyus constitute 28 percent of respondents who expressed support for the National Rainbow Coalition (NARC), significantly more than the 19 percent of respondents who gave their ethnicity as Kikuyu. Similarly, Luo respondents, who comprise 13 percent of respondents, constituted half of all respondents who indicated that they preferred the Liberal Democratic Party (LDP). However, only the Luo meet the monogamous clustering standard. The Luo were the only ethnic group to give the LDP disproportionate support, but in addition to the Kikuyu, the Kamba and Meru (together, 17 percent of Kenyan respondents) disproportionately supported the NARC. This illustrates the appropriateness of the stricter clustering criteria: it is implausible that the central appeal of the NARC would have been its commitment to defend the interests of Kikuyu, who comprised only 27 percent of its support base and were not the only ethnic group that disproportionately embraced it. Though rates of clustering vary significantly across clustering rules, rates at which respondents express a partisan preference do not. This provides some reassurance that the choice of rule does not drive the results, and additional evidence that ethnic appeals seem to have only a modest impact. Table 3 illustrates this, using the lax and monogamous rules from column 2 in Table 2. The first three rows of Table 3 display the fraction of respondents who express a partisan preference in three groups of countries: those in which all respondents belong to clustered groups, according to the given rule; where some respondents are clustered and others are not; and where no respondents belong to clustered ethnic groups. Rates of partisan preference are similar among the three groups and across the three clustering rules, ranging from .57 to .66. 26 Table 3: Fraction expressing partisan preferences across clustered and non- clustered groups % of population expressing Lax (many One party-many Monogamous a partisan preference: parties-many groups (one party-one groups) group) In countries where: Fraction (number) of respondents all respondents belong to .57 .66 NA clustered groups (10,499) (1,175) some respondents belong .64 .59 .59 to clustered groups (6,841) (16,165) (9,654) no respondents belong to NA .61 clustered groups (7,686) In countries where some ethnic groups cluster and others do not Respondents from .67 .62 .60 clustered groups (5,054) (9,308) (5,594) Respondents from non- .57 .55 .57 clustered groups (1,787) (6,857) (4,060) Respondents excluded .60 .58 .56 because of 5 percent rule (2,845) (5,713) (2,382) Note: Respondents from groups with less than 5% of population are excluded. The last three rows of the table compare respondents who belong and do not belong to clustered ethnic groups, or who are excluded from the analysis by the 5 percent rule, in countries where both types exist. The key point to notice is that the stricter the clustering rule, the smaller is the difference in rates of partisan preference between clustered and unclustered respondents; the difference is 10 percentage points between the 89 percent of respondents who are clustered under the lax rule and the 11 percent who are not; it is seven percentage points between the 60.5 percent of respondents who are clustered under the one party-many groups rule versus the 39.5 who are not; and it is three percentage points for the strictest, monogamous rule. The regression analysis in the next section confirms that these differences are small even when many other factors are taken into account. 27 Clustering and partisan preferences across countries Table 4 presents the results of the regression analysis of the aggregate data. The estimates pool all countries, controlling for invariant, unobserved country effects with country dummies. The differences in partisan preferences between clustered and unclustered groups are identified only by within-country variation in those countries that have both clustered and unclustered groups. Under the lax rule, all 16 countries exhibit clustering, but in only six are some clustered and some not. Applying the 25 percent threshold to the lax rule, 16 countries are still clustered, but in 13 there is within-country variation. Under the one group-many parties rule, 11 countries exhibit clustering; in all of these, clustering is partial. Under the one party-many groups rule, all 16 countries again exhibit clustering, and in 15 it is partial. Finally, under the monogamous rule, nine countries exhibit clustering; in all cases, it is partial. The first (bivariate) and second (multivariate) estimates of the clustering coefficient in Table 4 indicate that clustering is statistically significant in all five specifications, but the magnitude of the effect is generally small. The similarity of the estimates between the bivariate and multivariate specifications indicates that the appeal of non-parties is not a source of bias. Under the lax clustering rule, are the odds that a respondent from a clustered group expresses a partisan preference large (56 percent greater) relative to the odds for a respondent from other groups.12 However, when one restricts attention only to those 12 Even a high odds ratio can obscure a smaller real difference between the cases. For example, although the odds ratio in the first column is high, the difference in the probability that clustered and unclustered respondents will express a partisan preference is much lower: 77 percent of clustered respondents express a partisan preference, compared to 68 percent of non-clustered respondents; the 9 percentage point difference is sufficient to generate an odds ratio of 1.56. 28 instances where the ethnic group's support for a party exceeds its population share by 25 percent, the odds are only 24.8 percent greater. They drop further with the stricter definitions, to 18 percent, 9.6 percent and 13 percent. Statistical significance is unsurprising: any condition from fear to reliance on clientelism could generate such a result. However, if credible appeals to co-ethnics were the dominant form of political mobilization, the magnitude of the clustering coefficient would be large. This is not the case. Another way to see that the effect is not large is to compare it to estimates of the effects of other strategies of electoral mobilization, particularly those associated with political competition under conditions of limited political credibility. The odds that respondents who received a gift also express a partisan preference are 35 percent greater than the odds of respondents who received no gift and constant across clustering rules. The odds that those who approve of both incumbent jobs and economic performance express a partisan preference are 40 percent greater than those who disapprove of both, also larger than most of the cluster coefficients. This is also an indication of low credibility, indicating that those who disapprove have no credible party that they can support.13 Urban respondents (respondents with more access to public services), are less easily reached with clientelist promises. The odds that those who receive all seven services would express a partisan preference were 26 percent less (from odds ratios of approximately .74) than the odds of those who received no services. 13 It could be that citizens from the same ethnicity as the incumbent are more likely to express support for the incumbent, so that the incumbent controls bias downwards the effects of the ethnic clustering variable. However, the estimated coefficients on the ethnic clustering variables are nearly the same, with or without all of the controls, including incumbent approval. 29 Table 4: How much more likely are respondents from clustered ethnic groups to express a partisan preference? (Logit estimation; coefficients are odds ratios; robust z-statistics in parentheses) Dependent variable: Cluster Lax: many Lax, 25% One One Monogamous: Does respondent rule: parties/ threshold group, party, one party, one express a partisan groups many many group preference? (0-1) parties groups Bivariate regression: Respondent 1.62 1.30 1.18 1.10 1.15 ethnic group clustered? (0,1) (7.89) (6.78) (3.54) (2.62) (2.90) Respondent ethnic group clustered? 1.564 1.248 1.181 1.096 1.130 (0,1) (6.61) (5.25) (3.00) (2.28) (2.22) Respondent offered a gift before 1.365 1.371 1.354 1.357 1.351 last election? (0,1) (6.74) (6.83) (6.56) (6.62) (6.52) Do politicians usually make gifts? 0.844 0.843 0.846 0.847 0.846 (0,1) (-3.90) (-3.92) (-3.84) (-3.82) (-3.83) Do politicians usually keep their 1.403 1.403 1.400 1.402 1.401 promises? (0,1) (6.70) (6.72) (6.67) (6.70) (6.68) Favors democracy (0-1) 1.363 1.357 1.363 1.361 1.363 (8.21) (8.10) (8.21) (8.18) (8.21) Approves incumbent performance, 1.236 1.241 1.239 1.239 1.239 jobs (0-1) (4.68) (4.77) (4.75) (4.75) (4.75) Approves incumbent performance, 1.147 1.157 1.155 1.154 1.153 economy (0-1) (3.37) (3.59) (3.55) (3.52) (3.51) Fraction of seven services they 0.739 0.730 0.737 0.734 0.736 receive(0-1) (-4.40) (-4.58) (-4.45) (-4.50) (-4.47) Fraction of five assets they 1.666 1.667 1.672 1.662 1.670 possess(0-1) (5.63) (5.64) (5.67) (5.61) (5.66) Highest level of education? (0, none 1.009 1.006 1.007 1.007 1.007 - 9, post-graduate) (0.78) (0.53) (0.59) (0.64) (0.62) Observations 16017 16017 16017 16017 16017 Pseudo R-squared 0.10 0.0986 0.0977 0.0975 0.0975 Note: Country fixed effects and other controls not reported. Students, those working in the home, those who never had a job and those with no religion are significantly less likely to express a partisan preference. Younger respondents, men, Catholics, evangelical Protestants, Sunni Muslims and other Muslims (not Shiite) are significantly more likely. These controls are insignificant: population share of ethnic group; whether respondent expresses an ethnic grievance; whether respondent believes individual is more responsible for own welfare than government; respondent's level of education; whether respondent belongs to a mainline Protestant church or is Christian with no specific sectarian affiliation; frequency of attendance at church services; and whether respondent is a subsistence or peasant farmer or vendor/hawker. Respondents who believe politicians usually keep their promises are significantly more likely to exhibit a partisan preference. However, only 17 percent of respondents 30 express this belief, suggesting that the ability of parties to make broadly credible commitments is not the source of their optimistic view. At the same time, the question is not specific about the type of promises that are kept. Given the narrow group of respondents who regard politician promises as credible, these are more likely to be clientelist promises associated with political competitors who cannot make broadly credible policy promises. The odds that those who believe democracy is the best form of government express a partisan preference are 36 percent greater. The odds that respondents with all five assets express a partisan preference are 66 percent greater than for respondents with none. Women are far less likely to express a partisan preference than men. The country fixed effects (not reported) are large, but consistent with other information. For example, political parties in Benin and Madagascar are known to be weak. The odds that respondents from either of these countries express a partisan preference are about 58 percent and 69 percent lower than the odds that a Zambian respondent expresses a preference (Zambia is the omitted country to which all others are compared). Mozambique has a well-organized ruling party (Frelimo). Correspondingly, the odds that respondents from Mozambique express a preference are more than 400 percent greater than those for Zambian respondents. The importance of the country fixed effects suggests that the aggregate results may obscure large differences in the effects of ethnic clustering across countries. The next section demonstrates that the modest effect of ethnic clustering on partisan preferences is reflected at the country level, as well. On the one hand, clustering has a statistically insignificant impact on rates of partisan preference in about half of the countries where clustering is partial. On the other hand, even in countries where clustering has a significant 31 effect on partisan preference, it is implausible that credible appeals by political parties to serve the collective interests of ethnic groups can explain it. Individual country results The number of countries that exhibit within country variation with respect to clustering varies by clustering rule. For each of these countries, using the specifications in Table 4, the effects of clustering on partisan preference are separately estimated. However, Table 5 displays, for each country and each specification, the results for just two variables: ethnic clustering and whether the respondent received a gift. Six countries exhibit within-country variation with respect to the lax cluster criteria. In three of these, clustered ethnic groups are more likely to exhibit a partisan preference. Gift-giving is also significantly associated with partisan preferences in three countries. Under the lax rule with the 25 percent threshold, 13 countries exhibit within-country variation, but again, in only three of these is ethnic clustering associated with partisan preferences, compared to seven where gift-giving is significantly associated with having a partisan preference. Under the rule that allows clustering only if one group clusters on one or several parties, 10 countries exhibit within-country variation. In three of these, respondents from clustered ethnic groups are more likely to exhibit a partisan preference; in seven they are not. In four of the ten, gift-giving is significant. There is within-country variation in 15 countries under the fourth rule, which allows clustering if a group clusters on only one party, even if multiple groups cluster on the same party. In only two of these countries is ethnic clustering significantly and positively associated with expressing a partisan preference; in one (Uganda) it is negatively associated; and in the remainder it is insignificant. Gift-giving is significant in six of the countries. Finally, nine countries vary with respect to 32 the monogamous clustering rule (one party/one group). In two of these, clustering is significantly associated with partisanship; in four, gift-giving is significant. For most countries and most clustering rules, therefore, political parties do not make credible commitments to serve the interests of ethnic groups. Gift-giving is a more prominent correlate of partisan preferences, despite the fact that gift-giving in Africa is generally candidate-driven and not organized by political parties. Qualitative evidence from several of these countries further supports the claim that the basis for ethnic clustering of partisan preferences is unlikely to be the credible commitments of political competitors to improve the welfare of ethnic groups. The three countries in which clustering has the most consistent effect on partisan preferences are Kenya, Mozambique and South Africa. In Mozambique and South Africa, ethnic groups cluster on ruling parties that are well-organized and long-lived: FRELIMO in Mozambique was 43 years old in 2005, and the ANC in South Africa 93 years old (Beck, et al. 2001, Database of Political Institutions). These parties have undergone leadership changes and are associated with the interests of particular ethnic groups (e.g., the Xhosa in South Africa, only 22 percent of whom express partisan indifference, compared to 55 percent of Afrikaans respondents and 42 percent of Zulu respondents). They are exceptional in Sub-Saharan Africa: the average age of political parties in the rest of the continent was 17.2 years. A closer look at Kenya, in contrast, indicates the improbability of parties making credible appeals to ethnic groups, suggesting that ethnic clustering must have other explanations. The Luo and Kalenjin are the only two groups of respondents who meet all clustering criteria in Kenya. However, the two parties on which they cluster (the LDP and the Kalenjin's Kenya African National Union, or KANU) do not appear to make broadly credible appeals. Like the unclustered Kikuyu, of whom 35 percent express partisan 33 indifference, 32 percent of Luo and 41 percent of Kalenjin do, as well. The weakness of Kenyan political parties makes these responses entirely understandable. The LDP was only a few years old in 2005 and did not contest the 2007 elections. KANU was the party of the autocrat Daniel Moi, a Kalenjin, but structured to maintain him in power rather than to further the interests of all Kalenjin. Even if Kalenjin were disproportionate targets of clientelist payoffs by the regime, most Kalenjin did not benefit and, correspondingly, a large fraction of Kalenjin are indifferent to KANU's appeal. Instead, it is more likely that grievance or fear, rather than the credibility of party promises to ethnic groups, drive clustering. Mueller (2008) examines endemic inter-ethnic violence or threats of violence associated with electoral competition in Kenya. Consistent with the possibility that fear or grievance, rather than the credibility of the LDP, drove Luo support for the party, 39 percent of the Luo in the sample expressed ethnic grievance, but only two of the (unclustered) Kikuyu respondents. In contrast, although Ghana exhibits ethnic clustering, and has, by the standards of African democracies, institutionalized political parties, having seen changes of leadership within each party, clustered groups are not more likely to express a partisan preference.14 Gift-giving is key. Lindberg (2003) and others emphasize the burden of gift-giving that candidates face, a burden that is reflected in the significance of Ghana's gift-giving coefficient in Table 5. In contrast, overt ethnic appeals seem to play little role in Ghanaian 14 52 percent of survey respondents were Akan, as were 64 percent of the respondents who supported the largest party, the National People's Party. The Ewe comprised 27 percent of the support base of the National Democratic Party (the party of the former dictator, Jerry Rawlings) and 14 percent of all respondents. 34 political campaigns. The Center for the Development of Democracy-Ghana held debates for parliamentary candidates in 24 constituencies across the country in 2004 and recorded the proceedings of 15 of them in detail (CDD, no date). Ethnic appeals were not apparent. Of 14 NPP candidates who appeared in these debates, eight made no mention at all of the national party, despite the generally recognized success of the NPP's first four years in government. Only five of all the NPP candidates mentioned policy issues that extended beyond the constituency; the remainder focused on their individual performance, generally their contribution to local public works projects and to targeted transfers, such as student scholarships, many funded out of their own pockets or out of the MPs' Common Fund. South Africa and Mozambique are two of only four countries in the sample that use proportional representation, which mitigates the incentives of candidates to build a personal constituency and might explain the efficacy of ethnic appeals. Benin and Namibia are the other two. However, Benin, in contrast to Mozambique and South Africa, exhibits the highest rate of partisan indifference in the sample and no evidence that clustered ethnic groups are more likely to express a partisan preference. Again unlike South Africa and Mozambique, political parties in Benin are fragmented; none has experienced a change of leadership and the average party age of the three largest government parties and the main opposition party in 2005 was less than 10 years. 35 Table 5: Partisan preferences, ethnic clustering and vote-buying across 16 countries Lax Lax, 25% One group, One party, Mono- threshold many parties many gamous groups Benin Clustered? .94 .86 .86 (-.27) (-.66) (-.66) NA Gift 1.13 1.15 1.15 offered? (.79) (.87) (.87) Botswana Clustered? 1.14 .83 1.17 (.69) (-.72) (.81) NA NA Gift 3.03 3.02 2.97 offered? (.25) (1.16) (1.14) Ghana Clustered? 1.43 3.25 3.25 (1.40) (1.55) (1.55) NA NA Gift 2.28 2.31 2.31 offered? (3.41) (3.48) (3.48) Kenya Clustered? 2.21 2.01 1.04 2.21 1.04 (4.20) (4.07) (.26) (4.20) (.26) Gift .95 .94 .96 .95 .96 offered? (-.32) (.72) (-.24) (-.32) (-.24) Lesotho Clustered? 1.18 1.10 2.28 1.18 2.28 (.87) (.52) (2.51) (.87) (2.51) Gift .68 .68 .66 .68 .66 offered? (-.67) (-.64) (-.72) (-.67) (-.72) Madagascar Clustered? .84 .98 (-.58) (-.09) NA NA NA Gift 1.15 1.15 offered? (.80) (.81) Malawi Clustered? 1.41 6.29 1.32 6.29 (1.62) (2.21) (-1.35) (2.21) NA Gift .80 .81 .80 .81 offered? (-1.06) (-.99) (-1.03) (-.99) Mali Clustered? 1.23 1.71 .80 .81 (1.15) (2.26) (-1.05) (-.37) NA Gift 1.97 2.01 1.98 1.96 offered? (3.81) (3.93) (3.83) (3.80) 36 Table 5 (con't): Partisan preferences, ethnic clustering and vote-buying Mozambique Clustered? 2.43 1.75 1.11 (2.77) (2.03) (.47) NA NA Gift 2.94 3.24 3.30 offered? (2.32) (2.54) (2.56) Namibia Clustered? 5.49 1.39 5.49 (.55) (1.07) (.55) NA Gift .69 .68 .69 offered? (-1.17) (-1.24) (-1.17) Nigeria Clustered? 1.14 1.14 1.14 1.14 1.14 (.73) (.73) (.73) (.73) (.73) Gift 2.21 2.21 2.21 2.21 2.21 offered? (6.28) (6.28) (6.28) (6.28) (6.28) Senegal Clustered? .87 1.10 1.10 1.10 (-.48) (.48) (.48) (.48) NA Gift 2.05 2.05 2.04 2.04 offered? (2.41) (2.41) (2.41) (2.41) South Africa Clustered? 2.75 1.33 2.74 (5.50) (2.27) (5.50) NA NA Gift 1.56 1.57 1.56 offered? (1.76) (1.74) (1.76) Tanzania Clustered? 1.36 2.57 1.36 (.70) (1.36) (.70) NA NA Gift 1.16 1.15 1.16 offered? (.33) (.31) (.33) Uganda Clustered? .84 .73 (-1.28) (-3.00) NA NA NA Gift .85 .87 offered? (-1.53) (-1.29) Zambia Clustered? 1.02 (.07) NA NA Gift 2.21 offered? (3.89) Note: Country-by-country regressions of the specifications in Table 4. The table reports the estimated odds ratio and z-statistic for the respective clustering variable and whether respondent was offered a gift. No estimates indicate either 100% clustering or no clustering in the country. 37 As in Kenya, ethnic conflict in Nigeria is endemic and parties are fragile. The largest governing party was only seven years old in 2005. Rates of partisan indifference are correspondingly greater than 50 percent among clustered and unclustered groups. However, unlike Kenya, clustered groups are not more likely to express a partisan preference. Instead, gift-giving plays a large role. In fact, 37 percent of Hausas (one of the clustered groups) and more than half the 59 percent of Hausas who express a partisan preference indicated that they had been offered a gift. Vote-buying has a correspondingly strong effect across all definitions of clustering: the odds that a respondent who had been offered a gift by a politician expresses a partisan preference are more than twice those of respondents who had not, substantially higher than the aggregate effects. Religious polarization is another characteristic of Nigeria and could bias the results on ethnic polarization. Specifically, if parties make strong appeals to religious groups, members of those religious groups could exhibit very low rates of partisan indifference, such that rates of partisan indifference among clustered ethnic groups look high by comparison. However, the specifications in Table 4, replicated for Nigeria in Table 5, control for religious affiliation, eliminating this source of bias. Moreover, though Sunni and other non-Shiite Muslims are more likely to express a partisan preference, they do not appear to cluster their support. The two groups disproportionately expressed support for two parties, the PDP and the ANPP, and one of those parties, the PDP, is also disproportionately supported by Catholics and mainstream Protestants. Zambia is the focus of many studies of ethnicity in African politics, with pathbreaking work by Posner (2004) demonstrating the shifting efforts of politicians to target particular ethnic groups depending on the political exigencies and institutional arrangements that they confronted. The very fact that their focus shifted from one type of 38 ethnic cleavage to another suggests that they would have had difficulty making credible appeals to pursue the collective interests of any particular ethnic group, however. The Zambian responses to the Afrobarometer survey are consistent with this. All the Zambian ethnic groups comprising more than 5 percent of respondents, the Bemba, Tonga, Lozi, Chewa and Nsenga, meet the lax definitions of clustering. None meet any stricter definition of clustering that requires some degree of "monogamous" sorting among parties and ethnic groups. All express high rates of partisan indifference, 47 percent on average. Conclusion Analyses of ethnicity argue that development outcomes deteriorate when political competition is a contest between the collective interests of different ethnic groups. Such a contest cannot emerge, however, if ethnic groups are not collectively organized. The evidence here indicates that political parties in 16 African countries do not, with a few exceptions, facilitate that collective organization. Numerous patterns in the data support this conclusion. First, ethnic clustering of partisan support is far from a universal phenomenon; by the strictest standards of clustering, only 25 percent of respondents to the Afrobarometer surveys belong to ethnic groups that cluster their partisan support. Second, members of ethnic groups that cluster their partisan support are not substantially more likely to express a partisan preference. In most countries with both types of groups, there is no significant difference in rates of partisan indifference between clustered and unclustered groups. Absolute rates of partisan indifference are high even where clustered groups are more likely to express a partisan preference than non-clustered groups. Finally, clustering has a smaller 39 effect on partisan preferences than other determinants of partisan preference, some of which are associated with the lack of credibility of political appeals, such as vote-buying. This analysis does not say that ethnicity does not matter in African politics. Rather, it shows that ethnic clustering of political support cannot easily be explained as a manifestation of collective action by co-ethnics to improve their collective welfare; such collective action is difficult when political parties cannot make credible appeals to pursue the collective interests of ethnic groups. Instead, the salience of ethnicity is likely to be a byproduct of other features of the political environment. If politicians exploit voters' imperfect information to exacerbate fear, or rely on coercive authority over their co-ethnics, then the development effects associated with ethnicity are properly attributed to the absence of information and use of coercion. If non-credible politicians exploit ethnically homogeneous patron-client networks to secure the support of narrow groups of voters, then it is the lack of credibility that matters for development rather than ethnicity itself.15 These results are important because many analysts argue for institutional reforms specifically targeted at ethnicity to mitigate the risks of conflict. Reynal-Querol (2002), for example, points to evidence that particular types of electoral and political institutions are likely to mitigate the risks of conflict, precisely because they facilitate inter-ethnic agreement, arguments reflected as well in Bardhan (1997) and institution-building strategies pursued by donors in post-conflict and other fragile countries. The effects of institutional reforms are 15 Franck and Rainer (2009) find that the ethnic group of the leaders of 18 African countries spent more time in primary school and suffered lower rates of infant mortality, on average. This is consistent with the reliance by leaders on ethnically homogeneous patron-client networks, which would, on average ­ but not for all ­ increase incomes of the leader's ethnic group. 40 unlikely to be significant, however, in environments where politicians are unable to make broadly credible commitments to citizens. Keefer (2007, 2008) finds, in fact, that several institutional strategies, such as proportional representation, have a limited impact on the policy choices of democracies or the probability of conflict among poor countries when one controls for measures of political credibility. Instead, the analysis points to the need for research on institutional reforms aimed at reinforcing the ability of citizens to act collectively, and particularly those related to the institutionalization of political parties. 41 References Aldrich, John H. (1995). Why Parties? The Origin and Transformation of Political Parties in America. Edited by B. I. Page, American Politics and Political Economy Series. Chicago: University of Chicago Press. Alesina, Alberto, Reza Baqir and William Easterly (1999). "Public Goods and Ethnic Divisions." Quarterly Journal of Economics 119(2): 565-611. Alesina, Alberto, Arnaud Devleeschauwer, William Easterly, Sergio Kurlat, and Romain Wacziarg. (2002). "Fractionalization." Journal of Economic Growth 8(2):155-94. Alesina, Alberto and Eliana La Ferrara (2005). "Ethnic Diversity and Economic Performance." Journal of Economic Literature XLIII: 762-800 (September). Bardhan, Pranab (1997). "Method in the Madness? A Political-Economy Analysis of the Ethnic Conflicts in Less Developed Countries." World Development 25:9, 1381-98 (September). Beck, Thorsten, George Clarke, Alberto Groff, Philip Keefer, and Patrick Walsh. 2001. "New Tools in Comparative Political Economy: The Database of Political Institutions." World Bank Economic Review 15(1): 165--176. Blimes, Randall J. (2006). "The Indirect Effect of Ethnic Heterogeneity on the Likelihood of Civil War Onset." Journal of Conflict Resolution 50; 4, 536-547 (August). Bratton, Michael and Nicolas van de Walle (1997). Democratic Experiments in Africa: Regime Transitions in Comparative Perspective. Cambridge, England: Cambridge University Press. Center for Democratic Development, Ghana (no date). "Report on CDD-Ghana's Candidates' Debate Forum for Aspiring Parliamentary Candidates ­ 2004 General Elections." Mimeo. Chamberlain, Gary (1980). "Analysis of Covariance with Qualitative Data." The Review of Economic Studies 47:1, 225-238 (January). 42 Chandra, Kanchan (2004). Why Ethnic Parties Succeed. New York: Cambridge University Press. _____ (2006). "What is ethnic identity and does it matter?" Annual Review of Political Science 9, 397- 424. Easterly, William and Ross Levine (1997). "Africa's growth tragedy: policies and ethnic divisions." Quarterly Journal of Economics 112:4, 1203-50 (November). Eifert, Benn, Edward Miguel and Daniel N. Posner (forthcoming). "Political Competition and Ethnic Identification in Africa." American Journal of Political Science. Erdmann, Gero (2007). "Ethnicity, Voter Alignment and Political Party Affiliation ­ an African Case: Zambia." German Institute of Global and Area Studies (GIGA) Working Paper 45 (March). Eggers, Andrew C. and Jens Hainmueller (2009). "MPs for Sale? Returns to Office in Postwar British Politics." American Political Science Review 103:4, 513-533 (November). Esteban, Joan and Debraj Ray (1994). "On the Measurement of Polarization." Econometrica 62:4, 819-51. Fearon, Jams D. and David D. Laitin (1996). "Explaining Interethnic Cooperation." American Political Science Review 90:4, 715-735 (December). _____ (2000). "Violence and the Social Construction of Ethnic Identity." International Organization 54:4, 845-77 (Autumn). Ferejohn, John (1986). "Incumbent performance and electoral control." Public Choice 50: 5-26. Ferree, Karen E. (2006). "Explaining South Africa's Racial Census." Journal of Politics 68:4, 803-815. _____, Clark Gibson, Barak D. Hoffman and James D. Long (2009). "Explaining the African Vote." Mimeo, University of California, San Diego (February). Franck, Raphael and Ilia Rainer (2009). "Does the Leader's Ethnicity Matter? Ethnic Favoritism, 43 Education and Health in Sub-Saharan Africa." Mimeo, George Mason University. Harding, Robin (2008). "Vote Buying Across Africa." Mimeo, New York University (May 12). Keefer, Philip (2007). "Clientelism, credibility and the policy choices of young democracies." American Journal of Political Science 51:4, 804-821 (October). _____ (2008). "Insurgency and credible commitment in autocracies and democracies." World Bank Economic Review. 22:1, 33-61 _____ and Razvan Vlaicu (2008). "Democracy, Credibility and Clientelism." Journal of Law, Economics and Organization 24:2, 371-406 (October). Lake, David A. and Donald Rothchild (1996). "Containing Fear: The Origins and Management of Ethnic Conflict." International Security 21:2, 41-75 (Autumn). Lindberg, Staffan I. (2003). "'It's Our Time to `Chop'': Do Elections in Africa Feed Neo- Patrimonialism rather than Counteract it?" Democratization 10:2 (Summer), 121-140. _____ and Minion K.C. Morrison (2008). "Are African Voters Really Ethnic or Clientelistic? Survey Evidence from Ghana." Political Science Quarterly 123:1, 95-122. Miguel, Edward and Mary Kay Gugerty (2005). "Ethnic diversity, social sanctions, and public goods in Kenya." Journal of Public Economics 89: 11-12, 2325-68 (December). Montalvo, José G. and Marta Reynal-Querol (2005). "Ethnic Polarization, Potential Conflict, and Civil Wars." American Economic Review 95:3, 796-816. Mueller, Susanne D. (2008). "The Political Economy of Kenya's Crisis." Journal of Eastern African Studies 2:2, 185-210 (July). Posner, Daniel N. (2004). "The Political Salience of Cultural Difference: Why Chewas and Tumbukas are Allies in Zambia and Adversaries in Malawi." American Political Science Review 98(4): 529-45. 44 _____ (2005). Institutions and Ethnic Politics in Africa. New York: Cambridge University Press. Reynal-Querol, Marta (2002). "Ethnicity, Political Systems, and Civil Wars." Journal of Conflict Resolution 46:1, 29-54. Snyder Jr., James M., and Michael M. Ting (2002). "An Informational Rationale for Political Parties." American Journal of Political Science 46 (1):90-110. Stroh, Alexander (2009). "The Power of Proximity: Strategic Decisions in African Party Politics." German Institute of Global and Area Studies (GIGA) Working Paper 96 (February). Van de Walle, Nicolas (2001). African Economies and the Politics of Permanent Crisis, 1979-1999. New York: Cambridge University Press.