The Divided (But Not More Predictable) Electorate: A Machine Learning Analysis

• In Event: Reconsidering Polarization

Abstract

If a voting bloc is "too reliable," one concern is that the group's interests will not be represented adequately relative to its importance. But how reliable are demographic labels as predictors of presidential vote choice? An emerging body of literature has argued that social groups have polarized in their voting behavior as a byproduct of growing ideological divisions and social sorting. We use public opinion surveys from 1952 to 2020 and apply tree-based machine learning models to calculate out-of-sample predictions of respondents' voting decisions. Contrary to the sorting hypothesis, we show that voters' demographics are as informative about vote choice today as they have been throughout the second part of the 20th century. Partisanship, on the other hand, becomes more predictive of vote choice. Nevertheless, detailed information about voters - including their policy stances, and levels of political interest and knowledge - continues to be necessary for obtaining out-of-sample error rates of 5% or less.

Authors

• Jan Zilinsky, New York University

• Seo-young Silvia Kim, American University