Fig. 1: Area under ROC (left) and PR (right) curves from forecasts of conflict incidence in Africa at different administrative levels: unique predictive power of indicator groups.
From: Extreme weather impacts do not improve conflict predictions in Africa
The figures show model performances—ROC and PR curve results—of conflict predictions made using a Generalized Random Forest (GRF) algorithm with different indicator groups at different administrative levels from the most aggregate (level 0 is the national level) to the most granular (level 2). The figures highlight the effect of removing each of the four indicator groups—extreme weather events (ewi-red), conflict history (conf-green), socioeconomic indicators (socecon-blue), and governance (gov-gold)—from a predictive model with all of the indicator groups. The difference between a model with all indicator groups and a model with all except one captures the unique information found in that indicator group. Battles from the Uppsala Conflict Data Program (UCDP) are used as an outcome variable.
