The War That Never Was: How Model Interpretability in Forecasting Helps Explain What Prevents Civil War

When civil strife turns violent, there is an imminent risk of escalation into armed conflict, especially in societies that have a history of civil war. Yet, only in some cases do we see new eruptions of violence lead to civil war outbreak or recurrence. While previous research has developed solutions to tackle the selection biases associated with studying cases at risk, endogeneity remains a key challenge in existing explanations for why some cases remain peaceful even after violence erupts. Building on the longstanding tradition in peace and conflict research to examine on-the-brink cases to understand what prevents war, the present study uses model interpretability in conflict forecasting to build theory around why some at-risk cases do not see civil war. Leveraging data on low-intensity violence in the UCDP Candidate Events Dataset, I analyse cases where conflict models over-predict a transition into civil war. I use these false positives of war to assess alternative explanations for why civil war is contained. First, I use tools for interpretable machine learning to understand the risk settings in which civil war is contained and examine the conditional effects of potential triggers in these high-risk cases. Second, I systematically test how preventive factors at varying levels of analysis – elite bargaining and restraint at the conflict-party level, third-party involvement, and factors of societal resilience – adjust the prediction of civil war in these contexts. The findings demonstrate the potential of using conflict forecasting to identify the predicted civil wars that never materialise, and build theory around their patterns of successful prevention. The implications have direct real-world application in conflict early-warning systems and prevention.