Predicting Conflict Events in Africa at Subnational Level

Abstract

This study reviews the contribution in predictive accuracy of a number of geographic and socio-economic factors that are commonly linked to conflict incidence. A logit model is fitted to sub-national data for Africa at grid-cell level covering the years 2000-2009, generating an out-of-sample forecast for the period 2010-2015. Results show that the strongest predictor of future conflict is current conflict incidence in the grid-cell and neighbouring cells. Additionally, the infant mortality rate, which serves as a proxy for socio-economic well-being, shows some prowess in contributing to accurate predictions. This in contrast with factors such as the share of mountainous terrain. Travel time to the nearest city, to proxy for urban-rural differences, is also a strong predictor, but it must be noted that this could be the result of reporting bias in the outcome variable. In general the results highlight that it is difficult to improve accuracy beyond the contribution of conflict dynamics. Finally, the presented results are based on a relatively simple regression model commonly used in the literature and more sophisticated statistical techniques such as machine learning could improve predictions.