Weather variability and conflict forecasts: Dynamic human-environment interactions in Kenya

Abstract

While conflict prediction has gained considerable attention in recent years, the existing literature has relied mainly upon aggregated data for large administrative areas or even entire countries. Such approaches obscure significant geographic variation of conflict dynamics based on household and individual experiences. Conflicts are highly localized, shaped by social and economic contexts that vary across space and change throughout time. We predict two types of conflict reported by respondents in a 2018 Kenyan population survey (N = 1,400) using an identical survey carried out in 2014 in the same enumeration areas (sample locations). We use a conditional random forest (CRF) machine learning method for forecasting. Due to heavy reliance on agriculture in Kenya, we expect that adding weather variability and vegetation health (“environmental”) predictors to a CRF model with 29 demographic and contextual variables will improve the performance of our baseline forecasts. Against our expectations, adding environmental predictors does not enhance our 2018 predictions. Models with only environmental data have the worst fit. A logical extension of many “climate-conflict” studies is that environmental data should improve our ability to predict the location and timing of conflict, yet we find that they generally do not. We interpret this finding through the lens of human-environment interactions research developed in human geography and political ecology. These studies similarly emphasized that circumstantial and historical political, economic, and social relationships have greater credibility for understanding conflict than the weather.