Forecasting Civil Conflict with Zero-Inflated Count Models

Advances in the study of civil war have led to the proliferation of event count data, and to a corresponding increase in the use of (zero-inflated) count models for the quantitative analysis of civil conflict events. Our ability to effectively use these techniques is met with two current limitations. First, researchers do not yet have a definitive answer as to whether zero-inflated count models are a verifiably better approach to civil conflict modeling than are ‘less assuming’ approaches such as negative binomial count models. Second, the accurate analysis of conflict-event counts with count models – zero-inflated or otherwise – is severely limited by the absence of an effective framework for the evaluation of predictive accuracy, which is an empirical approach that is of increasing importance to conflict modelers. This article rectifies both of these deficiencies. Specifically, this study presents count forecasting techniques for the evaluation and comparison of count models' predictive accuracies. Using these techniques alongside out-of-sample forecasts, it then definitively verifies – for the first time – that zero-inflated count models are superior to comparable non-inflated models for the study of intrastate conflict events.