What Drives State-Sponsored Violence?: Evidence from Extreme Bounds Analysis and Ensemble Learning Models

The literature on state-sponsored violence has grown significantly over the last decades. Although scholars have suggested a number of potential correlates of mass killings, it remains unclear whether the estimates are robust to different model specifications, or which variables accurately predict the onset of large-scale violence. We employ extreme bounds analysis and distributed random forests to test the sensitivity of 40 variables on a sample of 177 countries from 1945 to 2013. The results show that GDP per capita, the post-Cold War period, and stable political regimes are negatively associated with mass killings. In contrast, ethnic diversity, civil wars, and previous political turmoils increase the risk of state-led violence. Years since the last episode of mass violence, GDP per capita, urban population, ethnic polarisation, the number of military personnel, and democracy make the greatest contribution to the models' out-of-sample predictive power.