Shaking the trees: Abilities and Capabilities of Regression and Decision Trees for Political Science

Abstract

When committing to quantitative political science, a researcher has a wealth of methods to choose from. In this paper we compare the established method of analyzing roll call data using W-NOMINATE scores to a data-driven supervised machine learning method: Regression and Decision Trees (RDTs). To do this, we defined two scenarios, one pertaining to an analytical goal, the other being aimed at predicting unknown voting behavior. The suitability of both methods is measured in the dimensions of consistency, tolerance towards misspecification, prediction quality and overall variability. We find that RDTs are at least as suitable as the established method, at lower computational expense and are more forgiving with respect to misspecification.