A0843
Title: Biostatistical extensions of Bayesian additive regression trees for categorical outcomes
Authors: Jaeyong Lee - Chung-Ang University (Austria)
Hyunmin Lee - Chung-Ang University (Korea, South)
Beomseuk Hwang - Chung-Ang University (Korea, South) [presenting]
Abstract: Bayesian additive regression trees (BART) are a versatile nonparametric method that has received substantial attention for its adaptability and robust performance in complex data settings such as clinical trial endpoints, diagnostic classification, and disease-severity staging. Three novel extensions of BART are presented for these settings: Ordered Probit BART (OPBART), which integrates an ordered probit regression structure to handle ordinal measurements; semiparametric OPBART (semi-OPBART), which parametrically models key covariates while nonparametrically adjusting for remaining confounders via tree ensembles; and t-BART, which incorporates a generalized t-distribution to enhance robustness against outliers and model misspecification across binary, multinomial, and ordinal endpoints. Efficient MCMC algorithms enable scalable inference in large patient cohorts with high-dimensional genomic and clinical predictors. Through extensive simulation studies and real data applications, superior predictive performance and clearer interpretability of covariate effects are demonstrated, offering robust, flexible, and interpretable tools for biostatistical analysis of categorical outcomes.
