Bridging prediction and theory: Introducing the Bayesian Partially-Protected Lasso

Abstract

This work examines the application of regularization methods in political science, both theoretically and practically. The focus of the study is on the use of regularization to specify models for predicting political outcomes. Specifically, what is the predictive price for protecting theoretically important variables from shrinkage? While many machine learning applications prioritize maximizing predictive accuracy, political science relies on reliable prior findings as a theoretical foundation. The literature shows that regularization methods can shrink crucial theoretically-driven variables to insignificance. To address this, we propose a protected Lasso approach in a Bayesian framework that safeguards these variables, balancing theoretical robustness and predictive power. Our analyses, applied to the American National Election Study, demonstrates the effectiveness of this approach.