Integrating Spatio-temporal Diffusion into Statistical Forecasting Models of Armed Conflict via Non-parametric Smoothing

Political armed conflict is responsible for thousands of fatalities every year. Facilitated by advancements in conflict event databases, research studies have moved towards predicting conflict and understanding its determinants subnationally. However, existing statistical and predictive models do not (fully) account for the diffusion and thus dependence of armed conflict across both time and space. As a result, predictive performance deteriorates, and predictors of interest are potentially biased. To address these shortcomings, this paper introduces a statistical regression model that captures both the spatial as well as temporal dimension of conflict diffusion, while its effects remain fully interpretable. Using conflict data from Africa, we demonstrate the importance of accounting for conflict diffusion and quantify its effects. We observe that conflict exhibits relevant dependence up to a distance of 522.5 km. Studying more complex diffusion patterns, we find that conflict tends to originate in high population areas and from there diffuses to lower population areas.