It used to be a truism among political scientists that African voters support coethnic candidates and African parties target coethnics in election campaigns (e.g., Bayart Reference Bayart2009; Horowitz Reference Horowitz1985; Rabushka and Shepsle Reference Rabushka and Shepsle1972).Footnote 1 Increasingly, however, researchers question this hypothesis. Some micro-level studies propose a broader sociological understanding of voting by identifying alternative cleavages like religion or urban–rural differences (e.g., McCauley Reference McCauley2014; Nathan Reference Nathan2016); others adopt rationalist perspectives that emphasize individual economic interests and the quality of information available to voters (e.g., Bratton, Bhavnani, and Chen Reference Bratton, Bhavnani and Chen2012; Casey Reference Casey2015; Ferree, Gibson, and Long Reference Ferree, Gibson and Long2021). Meso- and macro-level analyses identify electoral systems and ethnic inequality as conditioning factors of ethnic voting (Huber Reference Huber2012; Huber and Suryanarayan Reference Huber and Suryanarayan2016). Some studies even suggest that the effects of ethnicity on vote choice and party systems in Africa are in part spurious, simply reflecting underlying geographic clustering or shared economic interests (Boone et al. Reference Boone, Wahman, Kyburz and Linke2022; Ferree and Horowitz Reference Ferree and Horowitz2010).
Yet, micro-, meso-, and macro-level analyses of the electoral effects of ethnic cleavages suffer from potentially severe methodological problems. At the micro-level, the interdependence between observed voting patterns and the fixed menu of parties and candidates in any one election complicates inference. If researchers conceptualize ethnic voting as the support of voters for coethnic candidates, they risk selection bias if some voters exclusively face (non-)coethnic candidates (see also Ferree Reference Ferree2022). Moreover, the idiosyncratic and ever-changing menu of parties or candidates in individual elections prevents comparisons across countries and elections. Meso- and macro-level comparative approaches address this challenge by analyzing the degree to which ethnic groups differ in their voting patterns (Houle, Park, and Kenny Reference Houle, Park and Kenny2019; Huber Reference Huber2012; Huber and Suryanarayan Reference Huber and Suryanarayan2016). Yet, inferring individual-level voting motivations from group- or country-level analyses constitutes a clear case of ecological inference. Since ethnic cleavages frequently correlate with other social divisions such as geographic or economic differences, this approach might be biased by omitting non-ethnic determinants of vote choice (Boone et al. Reference Boone, Wahman, Kyburz and Linke2022; Ferree and Horowitz Reference Ferree and Horowitz2010).
We introduce a covoting regression (CVR) model as a new analytical approach. The CVR combines the complementary strengths of micro-level voting studies and comparative work at the meso- and macro-levels to address several of the problems affecting either. Similar to aggregate indices employed by comparativists, the model assesses the likelihood of covoting among pairs of individuals, which we operationalize as the shared intention to vote for the same party or candidate.Footnote 2 Yet, instead of building aggregate country or group-level measures, we model covoting at the micro-level of individual pairs of voters. Footnote 3 This allows us to estimate the effect of coethnicity (or any other cleavage) between individuals conditional on other dyadic socioeconomic characteristics with a standard linear probability model. The main methods we present are implemented in the R-package CVReg.Footnote 4
Conceptually, the estimated conditional effect of coethnicity parallels Huber’s (Reference Huber2012) group- and party-based understanding of ethnic voting. It increases with greater differences in voting patterns between ethnic groups and greater differences in the ethnic composition of parties’ voters. Aggregating across these two dimensions, the CVR thus explicitly considers ethnic bloc voting for candidates from other ethnic groups as an integral part of ethnic voting. Conceptually, this is a major departure from prevalent micro-level approaches, which typically rely on coethnicity between voters and candidates. Beyond the broad micro-level interpretation, we use the CVR to derive covariate-adjusted estimates of group-based and party-based conceptualizations of ethnic voting at the meso-level (Huber Reference Huber2012). At the macro-level, the model’s estimates reflect the classic Herfindahl–Hirschman index (HHI) and can be interpreted as the elasticity of party system concentration to marginal changes in ethnic homogeneity. The CVR thus bridges the gap between the differing levels of analysis in studies of electoral behavior with possible applications to cleavages beyond ethnicity.
The CVR solves several of the aforementioned methodological problems of micro-, meso-, and macro-level approaches. By modeling pairwise covoting, parties and candidates disappear from our formulation. This abstraction reduces selection biases from restricted choice sets and allows for comparative analyses across countries and over time. It does, however, come at the cost of the model’s inability to yield group- or party-specific inferences. In addition, other forms of selection bias, resulting from missing data and endogenous party formation, still affect the CVR.Footnote 5 At the same time, estimating the conditional association between coethnicity and covoting addresses ecological inference problems and bias from confounders, which can affect meso- and macro-level studies, while increasing statistical power.
Empirically, we use the CVR to provide comprehensive evidence on the effect of coethnicity on covoting across sub-Saharan Africa. Drawing on multiple rounds of the Afrobarometer surveys from 28 states across sub-Saharan Africa, we recast each country round into pairwise comparisons between respondents. We proxy unobserved actual covoting by measuring our outcome variables as covoting intentions in presidential elections and shared preferences for a political party.Footnote 6 Coethnicity, the main explanatory variable, is measured as the shared mother tongues of respondents. Shared demographic, economic, and geographic characteristics as well as survey-round-fixed effects constitute our controls. Building on our conceptual derivation of the CVR, we estimate the probability of covoting intentions in linear probability models and adjust standard errors to address the repeated inclusion of individuals from a limited number of ethnic groups across dyads.
Our results show strong support for the dominance of coethnicity in shaping covoting intentions in most African states in our sample. We estimate that coethnicity by mother tongue increases the probability that two respondents have shared voting intentions by 17 percentage points. Zooming in on specific countries and elections, we discuss variation in the estimated effect of coethnicity on covoting intentions over time and across cases. Additional analyses account for cross-cleavage interactions, alternative measures of coethnicity and linguistic distance, different sampling procedures, and modeling choices. We even find substantial, if smaller, effects of coethnicity on covoting intentions within comparatively homogeneous villages or neighborhoods, suggesting that coethnicity captures more than shared, spatially clustered political preferences.
In a set of descriptive analyses, we demonstrate how the CVR can be applied to test explanations of ethnic voting or changes therein at the three levels of analysis. At the micro-level, we find that higher ethnocentric trust and shared perceptions do not robustly explain coethnic covoting. At the meso-level, the effect of coethnicity is augmented in politically relevant ethnic groups and strongest for covoting for ethnic parties. At the macro-level, we find no substantive moderating effects of electoral, democratic, or traditional institutions.
Our results also show positive effects of religious, educational, occupational, and geographic similarities on covoting, thus supporting conjectures on socioeconomic characteristics beyond ethnicity (e.g., Boone et al. Reference Boone, Wahman, Kyburz and Linke2022; Bratton, Bhavnani, and Chen Reference Bratton, Bhavnani and Chen2012; Koter Reference Koter2016; McCauley Reference McCauley2014). Yet, across our broad sample of African electorates, the effects of these alternative cleavages are at least four times smaller than those of coethnicity.
We conclude by discussing the implications of our new approach for the study of electoral cleavages. In doing so, we highlight the importance of bridging levels of analysis to achieve inferences of high internal and external validity and discuss the utility of the CVR in studying the electoral effect of socioeconomic cleavages. Potential extensions can apply the CVR to local election results and individual attitudes and choices beyond voting.
ETHNICITY AND VOTING IN SUB-SAHARAN AFRICA
Political scientists have come a long way from the once paradigmatic view that elections in Africa constituted an ethnic census (Horowitz Reference Horowitz1985, 196). Classic works on vote choice in sub-Saharan Africa either stress psychological or instrumental motivations for ethnic voting and the corresponding existence of ethnic parties (Mozaffar, Scarritt, and Galaich Reference Mozaffar, Scarritt and Galaich2003). In short, the psychological approach entails that voters reaffirm their identity through voting for coethnic candidates and attempt to avoid discrimination by ethnically distinct rulers. Political leaders cannot escape the logic of ethnic outbidding, in which more extreme political demands on behalf of coethnics increase electoral support (Horowitz Reference Horowitz1985; Rabushka and Shepsle Reference Rabushka and Shepsle1972).Footnote 7 In turn, instrumentalists suggest that African voters support coethnic candidates to receive economic benefits through clientelist exchanges during the election period and patronage distribution afterward, if “their” candidate joins the ruling coalition. Political elites themselves prefer to build ethnically based support coalitions to limit access to state funds to ethnic insiders (Bates Reference Bates1974; Chandra Reference Chandra2007).
Political scientists investigate individual-level demand and party-level supply at the micro-, meso-, and macro-levels. Micro-level studies probe both the mechanisms that link individuals’ ethnic identity to vote choice and alternative explanations for party preferences. Experimental research supports an instrumentalist interpretation of ethnic voting that portrays the ethnicity of candidates as an informational shortcut for voters to learn about the likelihood of future economic benefits (Carlson Reference Carlson2015; Casey Reference Casey2015; Conroy-Krutz Reference Conroy-Krutz2013; Ferree Reference Ferree2006; Ferree, Gibson, and Long Reference Ferree, Gibson and Long2021).Footnote 8 In contrast, survey-based research frequently explores alternative voter motivations, such as economic performance evaluations and education (Bratton, Bhavnani, and Chen Reference Bratton, Bhavnani and Chen2012; Bratton and Kimenyi Reference Bratton and Kimenyi2008), partisanship (Ferree and Horowitz Reference Ferree and Horowitz2010; Hoffman and Long Reference Hoffman and Long2013), populism (Resnick Reference Resnick2012), and urban–rural differences (Nathan Reference Nathan2016; Reference Nathan2019). Others stress that ethnic voting depends on local factors characteristic of many African countries, including traditional authorities (Baldwin Reference Baldwin2013; Koter Reference Koter2016) and local ethnic geography (Ichino and Nathan Reference Ichino and Nathan2013).
Comparativists working at the macro-level complement micro-level studies by highlighting supply factors. Guided by Duverger’s law, cross-country comparisons suggest that the effective number of ethnic groups in a country correlates with the effective number of parties (ENPs) under permissive electoral rules (Clark and Golder Reference Clark and Golder2006; Lublin Reference Lublin2017; Mozaffar, Scarritt, and Galaich Reference Mozaffar, Scarritt and Galaich2003). In contrast, case studies show that party competition goes beyond ethnicity (Elischer Reference Elischer2013) and that presidential candidates pursue non-coethnic supporters (Horowitz Reference Horowitz2022). Larger groups are indeed more likely to have coethnic candidates on the ballot (cf. Posner Reference Posner2004b) and to see intra-group competition Ferree (Reference Ferree2010). Echoing micro-level insights on the importance of alternative cleavages, Boone (Reference Boone2024) argues that the correlation between ethnicity and vote choice hides the effects of regionally concentrated economic interests (see also Ishiyama Reference Ishiyama2012). Analyses of ethnic voting at the group- or meso-level fall in between these two endpoints by highlighting group-level economic inequality and electoral rules as moderators (Houle, Park, and Kenny Reference Houle, Park and Kenny2019; Huber Reference Huber2012; Huber and Suryanarayan Reference Huber and Suryanarayan2016).
Studies at each level of analysis complement each other. Substantively, research at all levels raises questions about the influence of other cleavages including urban–rural differences (Harding Reference Harding2020; Nathan Reference Nathan2016; Wahman and Boone Reference Wahman and Boone2018), regional coalitions (Boone Reference Boone2024; Ferree and Horowitz Reference Ferree and Horowitz2010), and economic class interests (Bratton, Bhavnani, and Chen Reference Bratton, Bhavnani and Chen2012; Houle, Park, and Kenny Reference Houle, Park and Kenny2019; Huber and Suryanarayan Reference Huber and Suryanarayan2016; Resnick Reference Resnick2012). Micro-level research trades in generalization against detailed, case-specific insights with the potential of identifying causal underpinnings of vote choice. Meso- and macro-level comparisons potentially offer greater external validity but rely on ecological inference and risk omitted variable bias.Footnote 9 Taking groups as units of analysis or building blocks of aggregate indices, meso- and macro-level analyses assume the primacy of ethnicity rather than testing it.Footnote 10
Importantly, micro-, meso-, and macro-level approaches differ in their conceptualization of ethnic voting itself. Both macro- and micro-level studies commonly analyze the degree of ethnic voting as the probability that an individual chooses a coethnic candidate or party. Meso-level research, in turn, analyzes the extent of ethnic divergence in voting patterns based on the electoral distance between groups and ethnic distance between parties (Huber Reference Huber2012). Correspondingly, we label these alternative strategies the voter-candidate correspondence and the divergence approach. In the correspondence approach, ethnic voting increases as members of a group increasingly vote for parties that represent their group and their proportion among all voters of these parties increases. In contrast, ethno-political divergence increases under two alternative conditions: as a political party increasingly draws support from one ethnic group in what Huber (Reference Huber2012) coins the “party-based approach” or as voters from one ethnic group vote together (“group-based approach”).
This contrast brings into relief three important conceptual and methodological issues. First, empirical analyses that employ the correspondence approach prioritize a party-based rather than group-based understanding of ethnic voting. This can lead to an underestimation of ethnic voting when members of a group vote en bloc for a non-coethnic candidate. While the correspondence approach would not regard those votes as “ethnic,” the divergence approach would detect this ethnicization of political preferences.Footnote 11 Second, the correspondence approach classifies votes for coethnics as ethnic even if they were cast for a “non-ethnic” reason or even due to pure randomness. This produces an upward bias in aggregate estimates of ethnic voting.Footnote 12 Third, constrained choice sets can leave voters with either exclusively coethnic or non-coethnic candidates (Ferree Reference Ferree2022). This produces selection bias in the correspondence approach.Footnote 13
This selection bias arises in three analytical situations. First, potential candidates do not run in expectation of low support from their group, for example, where groups are too small to lift candidates to victory (Posner Reference Posner2004b). When excluding such groups from an analysis, analysts will potentially overestimate ethnic voting in a society. Second, the utility of ethnic voting might be stable across groups, but some groups might still have no coethnic candidate. Here, a “0” classification of ethnic voter–candidate correspondence mixes the absence and the impossibility of ethnic voting. If members of such “unrepresented” groups voted en bloc for a “second-best” non-coethnic who represents their interests, correspondence approaches would underestimate the overall extent of ethnic voting. Huber’s divergence approach, however, can capture group-based voting for non-coethnic candidates. In a third situation, a broad choice set is available to voters, but the analyst artificially limits vote choice, for example, by only evaluating support for the incumbent (e.g., Bratton, Bhavnani, and Chen Reference Bratton, Bhavnani and Chen2012; Bratton and Kimenyi Reference Bratton and Kimenyi2008). In this case, the degree of ethnic voting may be over- or underestimated depending on the utility that individuals derive from ethnic voting and the distribution of votes for excluded candidates.Footnote 14
Because we only observe real but not hypothetical alternative candidate choice sets, those analyzing observational data currently have two options.Footnote 15 First, voting studies that rely on the correspondence approach would need to adjust for factors that influence the existence of “ethnic” candidates (Ferree Reference Ferree2022), and candidates’ expected vote shares under non-ethnic voting. However, this strategy is difficult to implement in broadly comparative work. Second, the divergence approach developed by Huber (Reference Huber2012) offers the possibility to model group-based voting across contexts even where choice sets are constrained. It frees the researcher from specifying correspondences between voters and parties and from taking arbitrary decisions on which groups to include in the analysis. However, the aggregation of individual choices to the group level makes this approach vulnerable to omitted alternative cleavages.
In sum, a new approach to observational research on ethnic voting should: (1) account for alternative motivations and identity categories that might serve as a basis for voting to avoid omitted variable bias; (2) avoid the pitfall of artificially limited choice sets to overcome selection bias; and (3) facilitate coverage of a broad number of countries over time to ensure external validity.
COVOTING REGRESSION: A NEW WAY TO ESTIMATE THE EFFECTS OF (ETHNIC) CLEAVAGES
Our new CVR bridges prevailing approaches to ethnic voting by combining their strengths, which together address their respective weaknesses. In short, we follow macro- and meso-level measurement strategies to conceptualize the effect of ethnic cleavages as the effect of coethnicity on covoting intentions. Overcoming reliance on aggregate indicators, which leads to problems of ecological inference and omitted variable bias, we study covoting intentions among pairs of individuals contained in survey data with standard regression models. In the following, we introduce our method, explain the transformation of the Afrobarometer data into voter pairs, and discuss the empirical CVR specification.
Bridging Micro-, Meso-, and Macro-Approaches
Macro-approaches often study the structure of party systems as the result of an interplay between institutional determinants and societal cleavages. Empirically, they employ aggregate measures to operationalize the main variables of interest, such as the HHI and the ENPs to measure party system concentration and ethnic homogeneity. Yet, while the empirical focus is on the macro-level, these measures have explicit meso- and micro-foundations in theory and measurement. In particular, the HHI of party concentration and its inverse, Laakso and Taagepera’s (Reference Laakso and Taagepera1979) ENP, are often computed with meso-level data on parties’ vote shares as
$$ HHI(party)={ENP}^{-1}=\sum \limits_{p=1}^K{\left(\frac{N_p}{N}\right)}^2, $$
where
$ \frac{N_p}{N} $
is the vote share of party
$ p\in K $
among voters
$ N $
. Party concentration thus decreases with more and more equally sized parties, which, in turn, increases the ENPs. As scholars of ethnic fragmentation note, at the micro-level the HHI has clear micro-foundations: it reflects the chance that two randomly drawn individuals belong to the same group or category (e.g., Alesina, Baqir, and Easterly Reference Alesina, Baqir and Easterly1999). In terms of voting, the relative vote shares of one party simply reflect individuals’ voting probability for that party. The square of that probability then yields the chance that two random voters voted for the same party. We can thus reformulate
$ HHI(party) $
as the chance that two randomly chosen voters covote:
$$ HHI(party)=\frac{1}{N^2}\sum \limits_{i,j}^N\;{\mathrm{covoting}}_{i,j}, $$
where
$ i $
and
$ j $
are individual voters drawn from all voters
$ N $
and
$ {\mathrm{covoting}}_{i,j} $
is an indicator that returns 1 if
$ i $
and
$ j $
covote for the same party
$ p $
and 0 otherwise.
Rarely discussed, the covoting formulation in Equation 2 includes comparisons within the same individual
$ i=j $
, inducing downward bias when computed on a finite sample of individuals
$ N $
. This is because comparisons within the same voter
$ i=j $
must yield
$ {\mathrm{covoting}}_{i,j}=1 $
.Footnote
16 Drawing on Simpson (Reference Simpson1949), this bias can be corrected by avoiding “within-individual” comparisons when computing the HHI:
$$ HHI(party)=\sum \limits_{p=1}^N\left(\frac{N_p}{N}\frac{N_p-1}{N-1}\right) $$
$$ =\frac{1}{N^2-N}\sum \limits_{i,j,i\ne j}^N{\mathrm{covoting}}_{i,j}. $$
Equation 4 shows how the unbiased expectation of
$ HHI(party) $
is equivalent to the average covoting rate between all voters.Footnote
17 In turn, ethnic homogeneity among voters as measured through the HHI is equivalent to the average rate of coethnicity between all voter pairs. Moving beyond measures of diversity, measures of dispersion such as the Gini coefficient can be similarly reformulated as comparisons between pairs of individuals.Footnote
18
We hone in on these micro-foundations of macro-level approaches and propose to model the effect of ethnic cleavages on voting by estimating the effect of coethnicity on covoting in pairs of individuals
$ i $
and
$ j $
. We start deriving our CVR model as
$$ {\mathrm{Covoting}}_{i,j}={\beta}_0+{\epsilon}_{i,j} $$
with
$ i,j\in N,i>j $
.Footnote
19 In this formulation, estimate
$ {\beta}_0 $
captures the average rate of covoting among all pairs of individuals and is thus equivalent to
$ HHI( party) $
in Equation 4.Footnote
20
The pairwise regression model in Equation 5 can be easily extended by adding a covariate matrix
$ {\mathbf{x}}_{i,j} $
which measure individuals’
$ i $
and
$ j $
’s similarities or differences on important socioeconomic cleavage dimensions, as well as other control variables such as cross-cleavage interactions.Footnote
21 We thus propose to estimate the effect of coethnicity on covoting as
$$ \begin{array}{c}{\mathrm{Covoting}}_{i,j}={\beta}_0+{\beta}_1\hskip0.1em {\mathrm{coethnicity}}_{i,j}\\ {}\hskip0.6em +\gamma \hskip0.1em {\mathbf{x}}_{i,j}+{\epsilon}_{i,j}.\end{array} $$
The focus of the CVR on covoting between
$ i $
and
$ j $
abstracts from the identities of candidates and parties. Their disappearance from the formulation circumvents the selection bias that arises in the correspondence approach when linking candidates to ethnic groups. Yet, the CVR remains vulnerable to biases from endogenous candidate supply when analyzing observational data. However, we find that its accounting for ethnic bloc voting decreases the size of this bias in comparison with the correspondence approach.Footnote
22 We also note that the CVR can be used to evaluate data from experimental studies with exogenous candidate supply.
Due to the equivalence of the
$ {\mathrm{coethnicity}}_{i,j} $
indicator in Equation 6 with the micro-level interpretation of the HHI index, the estimate for
$ {\beta}_1 $
has interpretations at all levels of analysis. At the macro-level and conditional on covariates
$ {\mathbf{x}}_{i,j} $
, it is the elasticity of the party concentration in response to marginal changes in ethnic homogeneityFootnote
23 such that
$$ \frac{\delta \hskip0.1em HHI(party)}{\delta \hskip0.1em HHI(ethnic)}=\frac{\delta \hskip0.1em {\mathrm{covoting}}_{i,j}}{\delta \hskip0.1em {\mathrm{coethnicity}}_{i,j}}={\beta}_1. $$
This feature of the CVR extends to other dyadic comparisons between voters that reflect macro-level measures, such as pairwise wealth differences, which together constitute the Gini coefficient.
At the meso-level of pairs of ethnic groups,
$ {\beta}_1 $
captures the difference in average covoting within groups and covoting between groups, with group pairs weighted by the product of their size. As a consequence, Equation 6 can also be used to derive the expected rate of covoting among coethnics as well as the expected rate of coethnicity among covoters at average covariate values
$ \overline{\mathbf{x}} $
:Footnote
24
$$ \hskip-3pc \mathit{\Pr}\left(\mathrm{covoting}|\mathrm{coethnicity}=1\right)={\beta}_0+{\beta}_1+\gamma \overline{\mathbf{x}}, $$
$$ {\displaystyle \begin{array}{l}\hskip-0.5pc \mathit{\Pr}\left(\mathrm{coethnicity}|\mathrm{covoting}=1\right)=\frac{\overline{\mathrm{coethnicity}}}{\overline{\mathrm{covoting}}}\left({\beta}_0+{\beta}_1+\gamma \overline{\mathbf{x}}\right),\end{array}} $$
where the rate of coethnicity among covoters in Equation 9 is the rate of covoting among coethnics scaled by the relative share of coethnic pairs compared to covoters in the population. These measures closely resemble the inverse of Huber’s (Reference Huber2012) measures of ethnic group vote fractionalization and party vote fractionalization (see Appendix D.2 of the Supplementary Material). Importantly, the conditional rates of coethnicity and covoting are driven not only by the effect of coethnicity on covoting
$ {\beta}_1 $
but also by the intercept as well as average covariate values and effects.
Finally, at the micro-level,
$ {\beta}_1 $
can be interpreted as the marginal effect of coethnicity on the probability of covoting between individuals, again conditional on covariates. In principle, this covoting probability can also be recovered from voting probabilities derived from a multinomial logistic or linear regression model (Appendix A.4 of the Supplementary Material). Doing so yields largely equivalent results but does not yield standard uncertainty estimates and comes with estimation problems where data are sparse. Reformulating covoting as the product of two individual probabilities to vote for the same party further reveals how to address biases that stem from the endogeneity of one cleavage
$ {x}_1 $
to another
$ {x}_2 $
. Where such endogeneity exists, it becomes necessary to not only model voter characteristics of
$ i $
and
$ j $
but also their interactions
$ {x}_{1,i}\hskip0.2em {x}_{2,j} $
and
$ {x}_{2,i}\hskip0.2em {x}_{1,j} $
, thereby capturing cross-cleavage covoting.
By providing interpretations at all levels of analysis based on dyadic, micro-level modeling, the CVR combines the advantages of previous approaches. In particular, it avoids making ecological inferences from macro-level modeling but at the same time allows for inter-temporal and cross-country comparisons of the conditional association of coethnicity with covoting. While the conceptualization of ethnic voting as coethnic covoting entails some loss of detail, it prevents biases from narrower but potentially more informative definitions of ethnic voting, which require ex ante coding or standardization of parties or candidates. We show in our analysis of heterogeneous effects and mechanisms how the CVR can be used to test theoretical arguments about the drivers of ethnic voting at all three levels of analysis. To facilitate understanding of our main empirical strategy, we first introduce our data structure and then present our regression model.
Dyadic Data on Covoting, Coethnicity, and Other Cleavages
To operationalize our analysis of covoting intentions among individuals, we transform survey data into pairs of individual respondents. For each pair, we encode whether respondents share voting intentions and measure ethnic and other cleavages through respondents’ pairwise shared ethnicity and similarity in other socioeconomic characteristics.
Our main data source consists of the nationally and, in expectation, locally representative Afrobarometer survey series, which contains data on political preferences across an increasingly large set of states in sub-Saharan Africa since 1999. For the most part, our main analyses rely on the survey’s seventh round fielded between 2015 and 2018 in 28 states with 1,200 respondents per country.Footnote 25 To gauge variation in the effect of ethnic cleavages over time, we draw on rounds 3–7.Footnote 26 In addition to surveying preferences for presidential candidates and political parties, the surveys cover a large range of demographic and economic items and provide geographic information on respondents’ place of residence. The resulting information allows us to capture covoting intentions along with non-ethnic cleavage dimensions discussed by existing work. For replication data, see Müller-Crepon and Bormann (Reference Müller-Crepon and Bormann2026).
Unit of analysis: Following the logic introduced in Equation 4, we transform survey rounds for each country into undirected dyadic comparisons between all respondents
$ i,j\in {N}_{c,t} $
with
$ i>j $
. This gives rise to a total of
$ \left({N}_{c,t}\left({N}_{c,t}-1\right)\right)/2 $
observations per country round. After dropping observations with missing data, our main analysis of preferences for presidential candidates (parties) draws on a median number of 717 (533) respondents and 256,686 (141,778) dyadic comparisons between them per country surveyed.Footnote
27
Measuring covoting intentions: Because actual voting behavior remains unobserved in the Afrobarometer, we encode our two main measures of covoting intentions by drawing on answers to questions on respondents’ preferences over presidential candidates and parties:
Voting intention: If a presidential election were held tomorrow, which party’s candidate would you vote for?
Party preference: Do you feel close to any particular political party? Which party is that?Footnote 28
Drawing on these items, we record two dummy variables that take the value of 1 if respondents
$ i $
and
$ j $
share a preference for the same candidate or party and 0 otherwise.Footnote
29 The result is visualized for a sample of 10 respondents from Ghana in Figure 1. Each dot represents one respondent, with its color reflecting their preferred candidate/party. Lines between respondents are drawn in black (1) where they share a preference and in gray (0) where they do not. On average, we observe that respondents share their voting intention and party preference in 46.0 and 45.9% of all dyads, which corresponds directly to the cross-country average HHI of party concentration or the inverse ENP (see Equation 4).
Figure 1. Covoting Dyads from 10 Respondents in Ghana, Round 7
Note: In Panel b, NPP, light blue; NDC, red. Covoting as black edges.
Coethnicity: We capture our main explanatory variable of interest—respondents’ pairwise coethnicity—in a binary variable that records whether they share the same mother tongue (1) or not (0).Footnote 30 As visualized for the randomly drawn 10 Ghanaians in Figure 2a, this leads to many coethnic dyads among respondents from large language groups (e.g., the Akan in red). On average, respondent pairs share their mother tongue at a rate of approximately 25 percent, which corresponds to the average HHI of ethnic homogeneity across the countries in the sample.
Figure 2. Encoding of Main Explanatory Variables on Example Graph of 10 Respondents from Ghana
Respondents’ reported mother tongue is among the least malleable ethnic identity indicators and therefore least likely affected by reverse causality or omitted variable bias. In contrast to the language spoken at respondents’ home or their self-proclaimed ethnic identity, reported mother tongues should be least affected by respondents’ political concerns (e.g., Green Reference Green2021) and assimilation (e.g., Müller-Crepon Reference Müller-Crepon2025).Footnote 31 We employ three different strategies to address the remaining potential for omitted variable bias through, for example, economic factors affecting political preferences as well as ethnic identities, and reverse causation, such as multi-generational assimilation that aligns ethnic to political identities. First, we condition our estimates on several other individual-level covariates which might affect individuals’ stated ethnic origin and their political preferences. Second, we analyze variation in effects at short and large linguistic distances, which are harder to overcome through assimilation or misrepresentation. Third, a set of robustness checks zooms in on covoting intentions among respondents from the same enumeration area, thus holding geographic factors constant.
Control variables: With regard to our first strategy of conditioning on observables, we encode a set of pairwise comparisons between respondents that capture prominent political cleavages and might affect individuals’ reported language. All are visualized for our exemplary 10 Ghanaian respondents in Figures 2b to 2i. For reasons of consistency, we construct our measures such that larger positive values denote greater similarity between respondents, which should, in expectation, come with higher probabilities of covoting intentions.
First, we complement our measure of shared mother tongues by accounting for whether respondents share the same religion. Second, we capture demographic similarities between respondents by recording age and gender similarities. Third, we approximate economic cleavages by adding dummy variables for shared educational and occupational background as well as wealth similarity, measured as one minus absolute wealth differences.Footnote 32 Lastly, we capture purely geographic cleavages by including as-the-crow-flies proximity between respondents (in 1,000 km)Footnote 33 and a dummy variable capturing whether respondents share their urban or rural status.
Combining data across countries and rounds: Because our measures of covoting intentions, coethnicity, and additional control are measured as binary or continuous indicators of similarity, the data can be stacked and analyzed across countries and rounds without any additional processing. This is a substantive advantage over standard approaches of modeling the effect of ethnic (or other) cleavages on party or candidate preferences, which require harmonization across context with the selection biases this gives rise to.
Modeling the Effect of Ethnic Cleavages on Covoting
With the undirected dyad of respondents
$ i $
and
$ j $
as our main unit of analysis, we employ a linear probability model to estimate the probability of covoting intentions between respondents along the lines of the CVR proposed in Equation 6.Footnote
34 We move beyond the confines of one survey sample of individuals and generalize the model across countries as
$$ {\mathrm{Covoting}}_{i,j,c}={\alpha}_c+{\beta}_1\hskip0.1em {\mathrm{coethnic}}_{i,j}+\gamma \hskip0.1em {\mathbf{x}}_{i,j}+{\epsilon}_{i,j}, $$
where dyads are constructed only among individuals observed in the same country
$ c $
.
$ {\alpha}_c $
is a fixed effect for each country
$ c $
, capturing the average rate of covoting in a country and the effects of unobserved country-level factors, such as states’ population size, history, or electoral institutions.
$ {\beta}_1 $
then captures the average effect of coethnicity by mother tongue on covoting intentions among individuals across all countries in the sample, conditional on the covariates
$ \mathbf{x} $
introduced and visualized in Figure 2 above.
Since all socioeconomic factors underlying
$ {\mathbf{x}}_{i,j} $
can plausibly be causes and consequences of respondents’ ethnic identity, adding these controls could trigger post-treatment bias. At the risk of omitted variable bias, we therefore first estimate a baseline model without any controls. We then compare the estimates across the baseline and the model with controls. Although it is theoretically possible that the potential biases from omitted variables and post-treatment controls lead us to overestimate the effect of coethnicity in both cases, small differences between the two specifications and subsequent robustness checks would suggest that we approximate the true effect of coethnicity.
By construction, the CVR model is estimated on interdependent data. We thus consider various strategies to adjust standard errors. In our main analyses, we rely on the conservative two-way clustering on the ethnicity of individuals
$ i $
and
$ j $
that constitute each dyad. These clusters correspond to the level of “treatment assignment” if we consider ethnic groups and their members to be jointly treated as groups. We explore alternative clustering strategies in Appendix F.7 of the Supplementary Material: clustering at the country level yields just as conservative standard errors, while clustering on the level of individuals or their enumeration area yields smaller error bands. Lastly, we show that Aronow, Samii, and Assenova’s (Reference Aronow, Samii and Assenova2015) cluster-robust variance estimator for dyadic data yields comparatively small uncertainty estimates when employed at the level of individuals, their ethnicity, or their locations of residence.Footnote
35
RESULTS
Our analysis yields strong support of the hypothesis that coethnicity increases the rate of shared voting intentions and party preferences among Afrobarometer respondents. Table 1 presents our main estimates, showing the unconditional and conditional effect of coethnicity on respondents’ covoting intentions, measured as shared voting intentions and preferences for parties. We find that pairs of respondents who share their mother tongue are between 16.2 and 17.7 percentage points more likely to have the same voting intention and party preference than pairs that do not. The effect is very stable across specifications, does not vary much between our two outcomes, and is associated with little uncertainty (
$ p<0.001 $
).
Table 1. Covoting and Shared Mother Tongue
Note: Standard errors clustered at the level of each respondent’s mother tongue in parentheses. Significance codes: ***p < 0.001, **p < 0.01, *p < 0.05, +p < 0.1.
Substantively, these effects are large. We observe shared covoting intentions in 46 percent of all survey respondent dyads. The conditional increase in covoting intentions resulting from shared mother tongues of around 17 percentage points thus amounts to 37 percent of the average rate of covoting. The estimated effect of coethnicity also swamps that of any other pairwise similarity measure between respondents, the substantively largest being that of shared occupation with an estimated effect of 2.6 percentage points. We compare the effect of coethnicity with other cleavage dimensions more thoroughly below.
At the meso-level, the coefficient of coethnicity together with the base rates of covoting intentions and coethnicity allows for computing the group- and party-based measures of ethnic voting (Equations 8 and 9). Based on Model 4, 57.7 percent of coethnics covote compared to 41.5 percent of non-coethnics. In turn, adjusted for covariates, we estimate that there are 34.3 percent of coethnics among covoters, compared to 21.4 percent of coethnics among pairs of non-covoters.
At the macro-level, our results indicate that, on the margin, increasing ethnic homogeneity translates into a higher concentration of parties with an elasticity of 0.17 (see also Equation 7).Footnote 36 This positive elasticity stands in drastic contrast to the negative bivariate relationship observed when using country-level data (see Figure A.8 in the Supplementary Material), highlighting the caveats of ecological inferences drawn from aggregate data.Footnote 37
We observe little systematic change in the aggregate effect of coethnicity on covoting intentions over time. When repeating our analysis for Afrobarometer rounds 3 to 7 in Figure 3, we find a slight upward trend in the full sample, which includes increasingly many countries.Footnote 38 Yet, there is no significant increase in the effect of coethnicity once we subset the sample to countries that have always been surveyed. In other words, the upward slope observed in the upper panels in Figure 3 results mostly from the increasing sampling of countries with more extensive ethnic voting.
Figure 3. Effect Over Time, by Afrobarometer Survey Round
Note: Coefficients result from the fully specified model in Equation 10 estimated separately for each Afrobarometer survey round, using respondents’ self-identified ethnicity to construct the coethnicity indicator. “Full sample” refers to all countries included in any one survey round, while “consistent sample” refers to countries included since Afrobarometer round 3. Gray lines plot country-by-country estimates over time; see Figures A21–A23 in the Supplementary Material for full results.
In contrast, the estimated effects of coethnicity on covoting intentions vary within countries over time. We discuss three cases that feature prominently in previous studies, sometimes as examples that demonstrate the weakness or even absence of ethnic voting. Figure 4 displays results across Afrobarometer rounds 3–7 for Kenya, Malawi, and Mali. Political scientists typically describe Kenyan elections as classic cases of ethnic voting (Bratton and Kimenyi Reference Bratton and Kimenyi2008). Our analysis supports this interpretation. Over time, the estimated effect of coethnicity increased on average, with a small decline in 2016 when shared ethnicity is associated with a 30 percentage point increase in covoting intentions, almost two times our main estimate.
Figure 4. Ethnic Voting Over Time in Kenya, Malawi, and Mali
Note: In Panel a, coefficients result from the fully specified model in Equation 10 estimated separately for each Afrobarometer survey round since round 3. We use respondents’ self-identified ethnicity as an ethnicity indicator and cluster standard errors at the level of respondents
$ i $
and
$ j $
. See Figures A21–A23 in the Supplementary Material for all countries in the sample. In Panel b, ratios are computed by dividing the covariate-adjusted rate of coethnicity among covoters by the rate of coethnicity of non-covoters (red) and by dividing the rate of covoting among coethnics by the rate of covoting among non-coethnics (blue).
The Kenyan illustration serves well to illuminate both the comparative advantages and remaining limitations of the CVR. The coethnicity coefficient
$ {\beta}_1 $
can be compared across survey rounds to reveal aggregate changes in the patterns of coethnic covoting conditional on covariates. It does not, however, reveal whether these aggregate changes result from individual changes in voting behavior, candidate entry and exit, or both. The right panel of Figure 4 shows the derived group- and party-based components of coethnic covoting. While their values can be affected by changes in other quantities besides coethnicity (
$ {\beta}_1 $
), including covoting (
$ {\beta}_0 $
) and covariate effects (
$ \gamma $
), both measures run largely in parallel to
$ {\beta}_1 $
. The results suggest that the Kenyan increase in coethnic covoting came with a greater increase in the ethnic homogeneity of candidates’ voters as compared to a smaller increase in the homogeneity of groups’ voting intentions.Footnote
39
Next, we turn to Malawi, a country for which prominent studies diagnosed weak ethnic voting patterns. In line with Ferree and Horowitz’s analysis of Malawi, we indeed observe relatively small coefficients for shared ethnicity in the run-up to the 2009 election, in which the ethno-regional voting “pattern broke down in dramatic fashion” (Reference Ferree and Horowitz2010, 535). However, since then, our results indicate a strengthening of the effect of coethnicity that reaches the estimated average effect for all elections in sub-Saharan Africa in Afrobarometer round 7. Increasing voting differences between ethnic groups and rising ethnic differences between parties’ voters run parallel to this trend. This development challenges recent work that identifies Malawi’s persistent regional voting blocs and underlying shared economic interests as the better-fitting explanation of covoting intentions (Boone et al. Reference Boone, Wahman, Kyburz and Linke2022, but see Robinson Reference Robinson2024).Footnote 40
Finally, in a widely cited study, Dunning and Harrison (Reference Dunning and Harrison2010, 21) “help explain why ethnicity has a relatively minor role in Mali … [a] country in which ethnic identity is a poor predictor of vote choice.” Our analysis confirms Dunning and Harrisons’ verdict when they wrote their study in the late 2000s before the renewed outbreak of civil war in 2012, which also produced increasing communal violence. Since then, the effect of shared ethnicity on Malians covoting intentions increased to a comparatively small effect of 6 percentage points in 2017. This development comes with more covoting within ethnic groups and heightened ethnic differences between parties’ voter pools. We estimate similarly small effects of shared ethnicity in Burkina Faso and Senegal (see Figures A21 and A23 in the Supplementary Material), two other countries that Africanists have singled out as examples of weak ethnic voting (e.g., Koter Reference Koter2013a). As in Mali, however, both cases see increasing covoting intentions among coethnics.
Together, these trends underline the importance of broad comparative work like ours. It is important to note that changes in the effects of coethnicity on covoting over time and differences between countries can be driven by many factors not discernible from CVR estimates such as general changes in political preferences, party systems, or electoral institutions. Explicit analyses of effect heterogeneity can, however, shed light on such drivers, as we show in exploratory analyses below.
Robustness Checks
We systematically test the robustness of our results to the omission of cross-cleavage interactions, the measure of ethnicity used, accounting for the potentially biasing effect of geography, and changes in the setup of the data and estimation. Our discussion below summarizes the results presented in Appendix F of the Supplementary Material.
Saturated interaction and multinomial regression models: Our main explanatory variable “shared mother tongue” is based on an interaction between linguistic characteristics of respondents’
$ i $
and
$ j $
. The main analysis might therefore be biased by the omission of substantively important cross-cleavage interactions, interactions between all characteristics (ethnicity, wealth, education, etc.) of voter
$ i $
with that of voter
$ j $
(Appendix A.4 of the Supplementary Material). As a remedy, we estimate fully interacted models.Footnote
41 We also compute the effects of coethnicity on covoting from multinomial choice models for each country in our sample. Appendices F.1 and F.2 of the Supplementary Material, respectively, show that the results closely align with the estimates from our specification with simple control variables (Equation 6).
Accounting for potentially endogenous ethnicity: Endogenous ethnicity (Müller-Crepon Reference Müller-Crepon2025; Pengl, Roessler, and Rueda Reference Pengl, Roessler and Rueda2022) or identity misreporting (Green Reference Green2021) might bias our main estimates. In Appendix F.3 of the Supplementary Material, we therefore assess whether our results hold for respondents’ current home language and ethnic self-identification, both of which are more malleable to change but potentially more precise. Supporting our main results, we find that a shared home language and ethnic self-identification increase covoting by 12 and 19 percentage points, respectively. We additionally estimate the effect of the pairwise linguistic proximity between respondents, finding that covoting increases with linguistic proximity. Varying the coarseness of language classification on the basis of the phylogenetic language tree yields a positive effect of coethnicity even if it is solely based on speaking a language belonging to one of six language families. In combination, these results suggest that strategic misreporting or endogenous ethnic change is unlikely to substantively affect our results.
Accounting for geographic variation: Because many ethnic groups’ feature regionally distinct “homelands,” coethnic voting might simply emerge from an alignment of economic interests and political preferences of individuals from the same administrative region or even location (Boone Reference Boone2024; Boone et al. Reference Boone, Wahman, Kyburz and Linke2022), geographies which might have partially shaped ethnic geography itself (Müller-Crepon Reference Müller-Crepon2025; Posner Reference Posner2005).Footnote 42 Appendix F.4 of the Supplementary Material shows that our results are robust to adding more fine-grained controls for geographic distance and joint residence in enumeration areas, districts, and regions. These do not substantively interact with coethnicity. In addition, results from estimating our CVR only for the small, non-representative subset of respondent pairs residing in the same region or enumeration area show that coethnicity increases covoting in these selected pairs by 10–14 percentage points. This estimate decreases by another 2–4 percentage points once we include region and enumeration fixed effects, which soak up significant variation in ethnic and political homogeneity across locations. Overall, these results suggest that ethnic cleavages—while certainly influenced by geographic factors—produce strong electoral effects that are largely independent of geography.Footnote 43
Data construction: We vary a number of choices made in the construction of our dyadic comparisons between survey respondents (see Appendix F.6 of the Supplementary Material). We first sequentially reduce the number of comparisons to the point of leaving only one comparison per respondent. This yields stable coefficients and uncertainty estimates. Second, we account for variation in the number of dyadic comparisons per country by weighting each dyad by the inverse number of dyads from its country such that every country receives the same weight.Footnote 44 This increases coefficient estimates slightly. Lastly, we recode preferences for “other” candidates and parties such that each such response is coded as its own candidate or party instead of being dropped. Doing so does not materially change the results.
Model specification: We furthermore test the robustness of our results regarding the most important modeling decisions (see Appendix F.7 of the Supplementary Material). We first reestimate the main specifications in Table 1 using logistic regression models, which yields equivalent results (Table A21 in the Supplementary Material).Footnote 45 Second, we test various ways of clustering our standard errors to account for the interdependence between dyadic comparisons, which yields less conservative estimates with the exception of clustering at the country level, which yields marginally more conservative uncertainty estimates. Third, we implement different fixed effect specifications to account for potential sources of bias at the level of language groups, enumeration areas, and individual respondents on each side of a comparison. Doing so drastically improves model fit but does not substantively change the estimated effect of a shared mother tongue on covoting.
MECHANISMS AND HETEROGENEOUS EFFECTS
We can use the CVR to further test for the mechanism and heterogeneity underlying the effect of coethnicity on covoting. Building on previous theoretical and empirical approaches at the micro-, meso-, and macro-level, the following exploratory, correlational analyses show the strongest support for meso-level arguments about the importance of political mobilization of ethnic groups by “ethnic” parties. Evidence for micro-level arguments about ethnic biases and shared interest and macro-level theories on the importance of political institutions is mixed. While we focus on explaining cross-sectional heterogeneity, we note that similar analyses can be used to study drivers of temporal changes in the effect of coethnicity on covoting.
Micro-Level: Ethnocentrism and Shared Interests
Arguments in the micro-level literature can be divided into psychological explanations of ethnic voting that highlight preferential bias for coethnics (Horowitz Reference Horowitz1985), instrumental approaches that point to expectations of ethnic favoritism (Bates Reference Bates1974), ethnicity as an information shortcut in low-information environments (Ferree Reference Ferree2006), and shared preferences among coethnics (Lieberman and McClendon Reference Lieberman and McClendon2013). We find no substantive moderating effect of ingroup bias proxied as respondents’ ethnic rather than national identification (Appendix E.1 of the Supplementary Material). Expectations of favoritism are positively but noisily (
$ p<0.1 $
) related to coethnic voting. Greater political knowledge and news consumption do not decrease the effect of coethnicity on covoting.Footnote
46
Our test of the importance of ethnically structured interests proceeds in two steps. First, we extend the CVR to test whether coethnicity increases shared perceptions of pressing problems in a country and of governments’ performance across 19 issue areas ranging from the economy to youth policy.Footnote 47 As Models 1 and 2 in Table 2 show, coethnicity comes with small, statistically significant increases in shared perceptions, yet these effects amount to less than 5 percent of the respective outcome means. Second, we show that shared perceptions increase the likelihood of covoting between respondents. However, these effects do not diminish the estimated effect of coethnicity in Model 3. These results suggest that individual-level explanations do not explain the bulk of coethnic covoting, at least not with the available proxies.Footnote 48
Table 2. Shared Mother Tongue, Perceptions
Note: Standard errors clustered at the level of each respondent’s mother tongue in parentheses. Significance codes: ***p < 0.001, **p < 0.01, *p < 0.05, +p < 0.1.
Meso-Level: Politicized Ethnic Groups and Ethnic Parties
At the meso-level, we examine heterogeneity in the effect of coethnicity on covoting in two regards. First, we ask whether the political mobilization of (clusters of) ethnic groups as coded by Vogt et al. (Reference Vogt, Bormann, Rüegger, Cederman, Hunziker and Girardin2015) and Posner (Reference Posner2004a) increases covoting above and beyond the mere effect of speaking the same mother tongue. Suggesting substantive effects of ethnic groups’ political relevance, Figure 5a shows that joint membership in a politically relevant group increases covoting by 10 percentage points, even if two respondents do not share the same mother tongue. Sharing the same mother tongue increases covoting by an additional 10 percentage points.
Figure 5. Variation in the Effect of Coethnicity on Covoting by Groups’ Political Relevance and Parties’ Reliance on Ethnic Constituencies
Notes: Panel a: covoting within politically relevant “super-groups” from EPR and PREG. Politically relevant “super-groups” from EPR (Vogt et al. Reference Vogt, Bormann, Rüegger, Cederman, Hunziker and Girardin2015) and PREG (Posner Reference Posner2004a) linked to mother tongues via Müller-Crepon, Pengl, and Bormann (Reference Müller-Crepon, Pengl and Bormann2022). Estimates from Appendix Tables A6 and A7. Panel b: covoting for parties with differing reliance on ethnic support groups. Data on ethnic support groups from the VDEM V-Party data (Lührmann et al. Reference Lührmann, Düpont, Higashijima, Kavasogly, Marquardt, Bernhard and Döring2020) linked to Afrobarometer Round 6 via Döring and Regel (Reference Döring and Regel2019). Estimates from Appendix Table A8.
Second, we disaggregate the coding of covoting to reflect whether respondents share a preference for a party that the V-Party data (Lührmann et al. Reference Lührmann, Düpont, Higashijima, Kavasogly, Marquardt, Bernhard and Döring2020) codes as having a high or medium level of ethnic support or none at all.Footnote 49 The results in Figure 5b suggest that two-thirds of the overall effect of coethnicity on covoting is driven by covoting for parties with a high level of ethnic support, with the remainder being due to parties with a medium level of ethnic support. We find no substantive effect of coethnicity on covoting for parties with no ethnic support base. Their voter bases are spread uniformly across ethnic groups, which shows that our results align closely with experts’ assessments of parties’ core constituencies.
Macro-Level: Political Institutions
Prior research highlights theoretical reasons to expect substantive variation between countries in the extent to which ethnic cleavages structure electoral politics (e.g., Huber Reference Huber2012; Mozaffar, Scarritt, and Galaich Reference Mozaffar, Scarritt and Galaich2003). We analyze such heterogeneity along electoral systems, countries’ level of democracy, and the strength of traditional institutions. We do not find substantively or statistically significant variation in the effect of coethnicity on covoting intentions across these three arguably relevant institutional dimensions (Figure 6 and Appendix E.3 of the Supplementary Material). While these findings shed doubt on the importance of institutional variation, our findings can only be understood descriptively as we do not account for potential endogeneity of the moderating factors. It is also possible that existing analyses find empirical support for theoretical mechanisms that operate at the meso- and macro-levels. It is particularly important to note that institutions can shape strategic candidate entry. This would introduce differential selection bias into our analysis and bias estimated differences in the effect of coethnicity on covoting.
Figure 6. Effect of Coethnicity on Shared Voting Intentions by Countries’ Institutional Characteristics
Note: Coefficients result from the fully specified model in Equation 10 estimated separately for countries with different electoral rules, levels of the V-Dem’s polyarchy index, and constitutionalization of traditional institutions. See Tables A9–A11 in the Supplementary Material for underlying models.
Comparing Cleavages
Finally, we compare the effect of coethnicity and that of other socioeconomic similarities between respondents with respect to their rate of covoting intentions. To facilitate a fair comparison that takes account of differences between conditional and unconditional effects, Figure 7 plots the results of baseline models of the effect of each variable without any additional controls, as well as coefficient estimates from the fully specified models (see Table 1, Models 2 and 4).
Figure 7. Results by Cleavage Indicator
Note: Coefficient estimates from (1) baseline model that only include the respective variable and country-fixed effects equivalent to the setup of Models 1 and 3 in Table 1, and (2) fully specified models with controls (Equation 10), equivalent to the setup in Models 2 and 4 in Table 1. Error bars denote 95% CIs.
Among identity cleavages, shared mother tongues seem to be by far the strongest and most stable predictor of covoting. The first column in Figure 7 depicts our main results from Table 1. Next, effects associated with shared religion are positive but decrease once we condition on covariates. We presume that the unconditional effect of shared religion captures some of the effect of (correlated) sharing of mother tongues. Across stated support for presidential candidates and legislative parties, we find no substantive effects of age and gender similarities. For age, we find a small negative effect of being close in age on covoting intentions, which suggests that party preferences within age groups are marginally more diverse than across them.
Economic similarities show substantively smaller effects on convergent voting intentions than those associated with shared mother tongues. Shared levels of education and occupation between respondents translate into an increase in the chance of supporting the same party by about 1.5 and 2.2 percentage points, respectively. These effects are robustly estimated. Interestingly, proximity in wealth levels between respondents does not significantly relate to covoting intentions between them.Footnote 50 This finding speaks to previous findings on economic voting in sub-Saharan Africa (Bratton, Bhavnani, and Chen Reference Bratton, Bhavnani and Chen2012; Bratton and Kimenyi Reference Bratton and Kimenyi2008).
Lastly, we find the geographic proximity correlates with shared support for presidential candidates and parties. In the unconditional baseline models, increasing geographic proximity by 1,000 km, or moving from the 97th to the 1st percentile in our data, comes with an increase in covoting intentions by 10 percentage points, consistent with the existence of regional voting blocs (Boone et al. Reference Boone, Wahman, Kyburz and Linke2022).Footnote 51 Yet, once we condition on all other cleavage measures, the effect of geographic proximity drops by half. This decrease supports the interpretation that geography correlates with voting preferences because of its reflection of economic incentives and ethnic identities. Shared urban or rural status has a consistent and statistically significant effect on covoting of approximately 2 percentage points when including controls. This is consistent with work on rural–urban cleavages on the continent (e.g., Harding Reference Harding2010; Koter Reference Koter2013b).
CONCLUSION
In this article, we introduced the CVR model as a novel analytical approach to study the electoral effects of social cleavages in general and ethnic voting in Africa in particular. Shifting from individual support for coethnic candidates toward shared voting intentions between two individuals allows us to address two key methodological issues in existing work. First, we reduce selection bias that plagues micro-level studies when the supply of candidates does not allow survey respondents to express support for coethnic candidates or forces them to do so in the absence of non-ethnic rival candidates (e.g., Ferree Reference Ferree2022). The CVR does not overcome the selection bias that results from strategic choices of candidates (not) to run, a problem that will be more severe under plurality and majoritarian electoral rules. By considering ethnic bloc voting, however, the CVR reduces such bias in many contexts in comparison with previous approaches. Second, we avoid ecological inference inherent in meso- and macro-level research that examines ethnic covoting but fixes ethnic groups as the main unit of analysis while disregarding other cleavages. Coincidentally, we retain the advantages of micro- and macro-studies. The CVR model we introduce captures both individual-level effects and recovers country-wide concentration indices such as the ENP and the HHI of ethnic concentration. Finally, CVR operates at scale and enables broad comparisons over time and across countries without preventing country-specific insights.
Our empirical analysis of 28 countries and five survey rounds from the Afrobarometer indicates that language-based ethnicity continues to be the dominant electoral cleavage across sub-Saharan Africa. The effect of coethnicity on voting intentions is at least four times larger than alternative cleavages including religion, shared urban or rural residence, geographic regions, and educational and occupational background. Although we find that coethnicity does not influence covoting equally across all survey rounds and countries, our results indicate that prominent case studies that question the effect of ethnicity in vote choices describe exceptions rather than broader trends across the African continent (Boone et al. Reference Boone, Wahman, Kyburz and Linke2022; Dunning and Harrison Reference Dunning and Harrison2010; Ferree and Horowitz Reference Ferree and Horowitz2010). Finally, our analysis reveals the importance of group- and party-level characteristics in explaining the incidence of coethnic covoting but yields no robust evidence for micro-level explanations around ethnocentric bias or policy perceptions, or for macro-level arguments on the impact of electoral rules, democratic institutions, or traditional authorities (Baldwin Reference Baldwin2013; Huber Reference Huber2012; Rabushka and Shepsle Reference Rabushka and Shepsle1972).
Our study opens up new avenues for studying ethnic and more generally cleavage-based political behavior in sub-Saharan Africa and beyond. More precisely measured data on economic income, production patterns, and partisanship would enable us to gain deeper insight into class, economic, and psychological explanations of voting—three core concerns of voting research outside Africa. For example, incorporating local economic production data could be used to approximate residents’ economic interests and model the extent to which they structure the formation of regional voting blocs that can span across ethnic divides (Boone Reference Boone2024). While economic-instrumentalist and psychological factors have already received much attention in the study of ethnic voting, one major theory of voting, its sociological basis (Lazarsfeld, Berelson, and Gaudet Reference Lazarsfeld, Berelson and Gaudet[1944] 1968), has been widely overlooked by students of sub-Saharan Africa (though see work on traditional institutions and norms, Baldwin Reference Baldwin2013; Holzinger et al. Reference Holzinger, Haer, Bayer, Behr and Neupert-Wentz2019). Given appropriate data, the CVR model could, for example, test the effect of different social networks on voting by capturing the overlap in (the homogeneity of) social contacts.Footnote 52
Beyond sub-Saharan Africa, our analytical focus on covoting intentions lends itself to the study of the relative strength of different cleavages, such as the re-emergence of urban–rural divides (Cramer Reference Cramer2016), and the increasingly dominant nationalist-cosmopolitan division across Western democracies (Kriesi et al. Reference Kriesi, Grande, Dolezal, Helbling, Höglinger, Hutter and Wüest2012). Our method might also benefit existing analyses of vote shares in small-scale spatial units, such as municipalities (e.g., Cagé and Piketty Reference Cagé and Piketty2023). These analyses face similar challenges as the ones we discussed in the context of research on voting intentions in sub-Saharan Africa. Rather than estimating the likelihood of covoting intentions at the individual level, we would require compositional similarity scores between spatial units in terms of voting results as a function of similarities in their social structure. Lastly, our work shows that the CVR might benefit studies of the effects of cleavages on political attitudes and behavior more generally, as it can accommodate more outcomes than just covoting, for example, shared (sets of) political attitudes and policy preferences between individuals. After all, social and political cleavages are an inherently relational concept, and their effects on political preferences should be operationalized and studied as such.
SUPPLEMENTARY MATERIAL
To view supplementary material for this article, please visit https://doi.org/10.1017/S0003055426101579.
DATA AVAILABILITY STATEMENT
The research documentation and data that support the findings of this study are openly available at the American Political Science Review Dataverse: https://doi.org/10.7910/DVN/1F74D0.
ACKNOWLEDGMENTS
We are grateful for helpful comments by Paola Galano Toro, Sara B. Hobolt, Michael Laver, Emmy Lindstam, Martha Wilfahrt, two anonymous reviewers, the APSR editors, and panel participants at the EPSA General Conference in 2023, the APSA Annual Meeting in 2025, and the Political Behavior Work in Progress Seminar at the London School of Economics.
FUNDING STATEMENT
N.-C.B.’s work was supported by the European Union’s Horizon 2020 research and innovation program under Grant Agreement No. 950359.
CONFLICT OF INTEREST
The authors declare no ethical issues or conflicts of interest in this research.
ETHICAL STANDARDS
The authors affirm this research did not involve human participants.
Open access
Comments
No Comments have been published for this article.