B1750
Title: Hierarchical modeling using Bayesian additive regression trees
Authors: Vittorio Orlandi - Duke University (United States) [presenting]
Alexander Volfovsky - Duke University (United States)
Abstract: Hierarchically structured data, in which units belong to various known or unknown groups, pervades many fields ranging from macroeconomics, to education, to medicine. Hierarchical models attempt to learn group-specific effects, relating units in the same group, while allowing for the sharing of information across groups. We propose using Bayesian Additive Regression Trees (BART), a flexible nonparametric method, to model such data. While BART has previously been used to model data with a group structure, past approaches treat group membership indicators no differently from covariates, which can lead to inefficient tree structures and high variance estimates. We show the advantages of our approach over this simpler alternative on simulated data and also present an extension of our method, based on a latent variable formulation, that addresses the case of unknown groups. We demonstrate this behavior in the United Network for Organ Sharing (UNOS) data.
