A0515
Title: Bayesian single and multiple index models with additive regression trees
Authors: Ruijin Lu - Washington University in St. Louis School of Medicine (United States) [presenting]
Abstract: In analyzing data of environmental mixtures, single (SIM) and multiple index models (MIM) are powerful tools given their nonparametric links and interpretable index coefficients. A Bayesian additive regression tree (BART) perspective is taken for these models, and variable selections are considered, particularly the selection of exposures in the indices. The challenge of applying BART to SIM/MIM is tackled by using a sigmoid gating function in place of the binary routine at each splitting node and that of variable selection by placing a sparsity-inducing Dirichlet hyperprior. The performance of the proposed approach is examined by conducting extensive simulations and applying them to commonly used benchmark data sets. For real data application, the link between birth weight and exposures to environmental pollutants, dietary intakes, and physical activities during pregnancy is investigated using data from the NICHD Fetal Growth Study.
