Predicting Large-Scale Violence in Criminal Wars (VIRTUAL)

Presenter:
Natan Skigin
Kellogg PhD Fellow, Political Science

Discussant:
Jacob Turner
Kellogg Doctoral Affiliate, Political Science and research affiliate of the Kellogg Institute's Notre Dame Violence and Transitional Justice Lab (V-TJLab).

Abstract:
Because criminal wars – the multiple armed conflicts between drug cartels and betweenthem and the state – can be even more deadly than civil wars, recent research has madeimportant progress in understanding their underlying causes. Yet even prominent studiesoften present poor predictive capacity. Four challenges make traditional statistical methods inappropriate to accurately classify out-of-sample instances of large-scale violence:stringent functional form assumptions, class-imbalanced data, multicollinearity and overfitting, and unspecified interactions between covariates. I address these drawbacks by applying machine learning techniques to fine-grained subnational data from Mexico—the country with the most lethal drug war worldwide. I show that ensemble methods offer superior accuracy in predicting local violence than any of the logistic regression models; that they compare favorably to those employed by political conflict scholars; and that risk factors differ from those stressed by studies aimed at causal inference. In particular, democratization and the decapitation of criminal groups have little predictive power,whereas population size, poverty levels, and state presence become reliable predictors of intercartel wars.