A multilevel Bayesian approach to climate-fueled migration and conflict

Do climate conditions and extreme events fuel conflict and migration? This question has been widely studied using causal designs that exploit natural variation in climate variables, often analyzed with linear fixed-effects models. Yet in this setting, nonlinear relationships, distributional features of outcomes, and spatial heterogeneity can cause these models to violate core assumptions and yield unreliable inferences. We propose a multilevel Bayesian framework that accommodates such features while retaining identification strategies from natural experiments. We illustrate its potential with a representative analysis from the literature of the effect of temperature anomalies on conflict in Somalia. When outcome distributions suited to event counts are combined with partial pooling across regions, the apparent aggregate climate effect disappears and marked regional heterogeneity emerges, with positive associations in only a few southern regions and negative or uncertain effects elsewhere. Extending pooling across time further improves predictive ability. More broadly, the multilevel Bayesian framework offers a general strategy for strengthening both explanatory and predictive inferences about climate and social outcomes, supporting internal and external validity while efficiently accommodating heterogeneity even with small samples. This methodological bridge between econometric identification strategies and statistical modeling provides a robust foundation for interdisciplinary climate-conflict-migration research.