Measuring Swing Voters with a Machine Learning Approach
Chris Hare, Assistant Professor of Political Science, University of California, Davis
In this paper, we use a nonparametric machine learning method (boosted decision trees) to derive a more valid and finely-grained measure of swing voter propensity and identify characteristics of swing voters. As past work has demonstrated, boosted trees excel in their discovery of complex interactions and nonlinearities present in the relationship between response and predictor variables. This makes boosted trees a promising candidate for identifying niche subgroups of swing voters, and modeling how swing voter characteristics and behavior varies across subgroups. We use this approach to analyze three-wave panel data from the 2012 Cooperative Campaign Analysis Project. The boosting model generates individual predictions that outperform existing measures of swing voters. It also uncovers a range of meaningful two, three, and even four-way interaction effects that can be used to assess how the influence of ideological moderation and cross-pressures on swing voting vary across groups of voters.