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 with a history of civil war. Yet, only in some cases do we see new eruptions of violence lead to civil war outbreak or recurrence. In the longstanding tradition of conflict research to examine on-the-brink cases to understand what prevents war, endogeneity remains a key challenge. This paper develops a novel inference-by-design framework in the context of conflict forecasting to build theory around why some at-risk cases do not see civil war. The paper makes the theoretical argument that victory and compromise are two observable outcomes of the escalation spiral before reaching the stage of armed conflict. If victory and compromise work as causal mechanisms to halt escalation before armed conflict onset, it holds, an empirical implication should be that such indicators capture signal in early-stage contention. To test this claim, the paper combines UCDP GED data with unpublished data on low-intensity violence that never reached the UCDP threshold for armed conflict, to obtain a case universe of potential conflict onsets. It makes onset predictions using random forests, and applies SHAP values and inclusion/ablation as interpretability tools to identify the contexts in which indicators of victory, compromise or societal resilience contribute meaningfully to improve onset predictions, meaning they capture important signal otherwise lost. Then, drawing on the recent literature using machine learning to estimate the missing potential outcome in inference designs, the paper demonstrates how false and true positives with similar trajectories in onset forecasts can be used to enable inference about what makes the difference between war and non-war.