Predicting atrocities. Statistically modeling violence against civilians during civil war

In recent years, conflict research has seen a turn towards micro-level event data which often include casualty counts. This kind of data presents a challenge for statistical analysis, as it most often consists of long periods of no observations interrupted by extreme events with many victims. The present study uses such data and explores whether it is possible to predict the timing and the extent of violent acts perpetrated against civilians. Based on event data of one-sided violence during the war in Bosnia and Herzegovina, it is argued that predicting these events is feasible. Yet, the analysis of such difficult conflict event data requires a more coherent theoretical framework of the data generating process. Based on the assumption that this violence is used strategically, the paper maintains that the violence is unleashed in some situations and halted in others. Furthermore, as planning and execution of orders take place on lower levels of the hierarchy, leadership decisions might determine the occurrence of violent acts, but less so the exact number of victims. Hence, the decision to use violence should be influenced by different factors than the extent of violence. The paper models the violence as a result of two processes, occurrence and extent, and captures the abstract pattern of the data accurately.