Examining Urban Political Violence and Social Disorder Using Random Forests for Predictive Accuracy

• In Event: Poster Session: New Perspectives in Political Science Research: Scholarship by Vanderbilt REU Students

Abstract

More than half of the world population now lives in cities, and scholars have consistently found higher rates of political violence in urban areas compared to rural areas. A growing body of research has begun exploring the patterns and dynamics of urban violence using increasingly disaggregated conflict-event data, culminating in several versions of the Urban Social Disorder (USD) dataset. While scholars have tended to estimate statistical regression models to describe causal and predictive relationships, I argue for the value of using alternative, machine learning-based approaches to specifically analyze predictors of future urban social unrest. In this project, I present a use-case proof of concept for a supervised machine learning model utilizing random forests to predict lethal urban violence based on a series of political, population, and economic predictors. When compared against parametric models, I find these ensemble learning models have better predictive accuracy and estimate notably different results for the individual predictive power of each indicator. I conclude with reflections on the model and adaptations for future substantive inquiry.

Author

• Kai Keltner, Vanderbilt University