Detecting interactions using Bayesian additive regression trees

Bayesian additive regression trees (BART) is an ensemble prediction method that is able to model complex, highly non-linear relationships in continuous and discrete data. Previous methods have utilized BART’s flexible, sum-of-trees model to perform variable selection that is non-parametric with respect to the functional form of the outcome. In this paper we introduce a method to detect interactions in a similar manner. Additionally, we present a BART-based pipeline for variable and interaction selection to build interpretable machine learning models that improve inference over existing methods named BARTselect. Simulation results demonstrate that our method is able to detect a high proportion of true interactions across a range of data settings. We conclude with two example data settings, examining interactions in medical expenditure data and newborn birth weight data.