Replicability of Machine Learning Models in the Social Sciences

Abstract

Abstract. Machine learning tools are increasingly used in social sciences and policy fields due to their increase in predictive accuracy. However, little research has been done on how well the models of machine learning methods replicate across samples. We compare machine learning methods with regression on the replicability of variable selection, along with predictive accuracy, using an empirical dataset as well as simulated data with additive, interaction, and non-linear squared terms added as predictors. Methods analyzed include support vector machines (SVM), random forests (RF), multivariate adaptive regression splines (MARS), and the regularized regression variants, least absolute shrinkage and selection operator (LASSO), and elastic net. In simulations with additive and linear interactions, machine learning methods performed similarly to regression in replicating predictors; they also performed mostly equal or below regression on measures of predictive accuracy. In simulations with square terms, machine learning methods SVM, RF, and MARS improved predictive accuracy and replicated predictors better than regression. Thus, in simulated datasets, the gap between machine learning methods and regression on predictive measures foreshadowed the gap in variable selection. In replications on the empirical dataset, however, improved prediction by machine learning methods was not accompanied by a visible improvement in replicability in variable selection. This disparity is explained by the overall explanatory power of the models. When predictors have small effects and noise predominates, improved global measures of prediction in a sample by machine learning methods may not lead to the robust selection of predictors; thus, in the presence of weak predictors and noise, regression remains a useful tool for model building and replication.