Predicting uncertainty in stages: Using a semiparametric hierarchical hurdle model for predicting distributions of conflict fatalities

Abstract:

Forecasting armed conflict incidence and intensity is of significant interest to researchers, practitioners, and policymakers. However, these user groups are not only interested in the most likely outcomes (point estimates) but also in the range of possible scenarios (uncertainty). Providing a large set of possible outcomes is especially useful given the skewed distribution of conflict fatalities, as more than 99% of observations at the geographically and temporally disaggregated PRIO-grid-month (pgm) level exhibit no fatalities. Meanwhile, the processes determining the intensity of violence in a conflict location are different from the processes that determine whether a location experiences any violence in the first place. We address this extreme zero-inflation within a very skewed distribution via a hierarchical hurdle count model. This framework extends the approach of Fritz et al. (2022) to probabilistic forecasts. Stages one and two predict whether any fatalities will occur at the country-month and pgm level, while stage three uses a truncated count regression model to predict the number of deaths conditional on the previous stages. By decomposing the prediction of fatalities into three respective problems that can be comprehended as two Bernoulli and one truncated Poisson random variables, we obtain an estimate of the aleatoric uncertainty from aggregating the stage-wise uncertainty. Our model has three core benefits. First, it offers full predictive distributions of future violence intensity at the pgm level. Second, it is easily interpretable, with substantive effect estimates for all violence predictors. And third, it is computationally lightweight compared to most other machine learning algorithms.

Authors:

Cornelius Fritz, Christoph Dworschak, and Marius Mehrl