Ethnicity is an important political cleavage in Africa, yet the degree of its influence on voting is contested. Selection biases from restricted choice sets complicate micro-level analyses, while bias from ecological inferences and unobserved confounders hamper meso and macro-level approaches. We develop the Co-Voting Regression (CVR) model to tackle these challenges. It estimates the effect of co-ethnicity on the probability that pairs of voters co-vote for the same party/candidate while conditioning on other characteristics that connect voters. In doing so, CVR mirrors the micro-foundations of widely-used aggregate indicators, such as the effective number of parties and the Herfindahl-Hirschman index of ethnic homogeneity. Our data consists of dyadic comparisons between respondents from Afrobarometer surveys. Pooling across 28 countries, our results show that co-ethnicity increases co-voting intentions by 16 percentage points. The effect of co-ethnicity is consistent across institutionally diverse countries and at least five times larger than that of other cleavages. Beyond ethnicity, the approach we propose addresses key methodological concerns in studies of the electoral consequences of socio-economic cleavages and bridges gaps between levels of analysis.