Performance of Machine Learning Algorithms and Accuracy of Regression Estimators

• In Event: Poster Session: Political Methodology & Political Science Education & Civic Engagement
  In Poster Session: Poster Session: Political Methodology

Abstract

Recent applications of machine learning techniques in political science research assess the performance of machine learning algorithms applying either the Receiver Operating Characteristic approach or the Precision-Recall approach. While there is an agreement about what is a perfect and excellent machine learning performance as suggested by practitioners of machine learning, there is no scholarly agreement on how imperfect machine learning algorithms affect our regression estimators. This is important because political science research uses machine learning techniques to examine causal relationships between variables rather than simply predicting a quantity of interest. Given that most applications of machine learning techniques in political science rely on algorithms that produce good enough or not bad performances, it is particularly hard to assess the robustness of regression estimators when machine learning algorithm does not perform well as intended. Using Monte-Carlo simulation and real-world data, we provide a practical guide on how to evaluate our machine learning algorithms for political science research.

Authors

• Jongwoo Jeong, Texas A&M University

• Yohan Park, Texas A&M University