This paper proposes a theoretical model of electoral violence which is tested and validated empirically. The paper studies if electoral violence is driven by (i) political competition (ii) the distribution of violent supporters and (iii) the ethnic distribution. The paper develops a model which predicts that electoral violence is more likely to emerge if two conditions are jointly satisfied. First, political competition should be tight, such that violent parties expect to win the election if their violent move is successful. Second, parties only engage in violence if the likelihood to be successful is sufficiently high. A key implication is that bilateral violence is more likely to emerge if parties have similar violent capacities, such that both parties have a reasonable chance to win the violent contest. By contrast, electoral violence is unlikely in the presence of parties with very different violent capacities.
The predictions of the theoretical model are tested using a unique dataset on electoral violence from the 2010 Burundian electoral process. The empirical analysis examines whether electoral violence at the municipal level is correlated with political competition, polarization between violent supporters, and various control variables including ethnic fractionalization. In line with the model, empirical results show, first, that political competition between parties is a key driver of electoral violence. A one-standard-deviation increase in political competition induces a 35 to 66% increase in the predicted number of violent episodes. This effect is stronger in the presence of numerous violent supporters. The second major finding is that violence is more likely to emerge in places where polarization between ex-rebel groups is high, that is, where parties have a similar potential for violence. A one-standard-deviation increase in ex-rebels’ polarization induces a 40 to 50% increase in the predicted number of violent episodes. The latter effect is especially strong for bilateral violence. These findings are consistent with the model and robust to numerous specifications.