38: Tailoring BART for Environmental Mixture Studies

Music City Center 

Various methods have been developed to investigate complex and collective effects of environmen-tal mixtures on human health. Tree ensemble methods are known for their stability and accuracy in identifying highly correlated and high-dimensional features in the statistical literature, but their use has not been well studied for environmental mixtures analysis.
We tailored the Bayesian Additive Regression Trees (BART) model for environmental mixtures analysis, which allowed a smooth response surface and incorporated confounder adjustment, for both continuous and binary outcomes. We further encompassed component-wise and hierarchical variable selection to accommodate scientific grouping of chemicals. Additionally, we proposed to quantify the marginal contributions of each chemical in the mixture through the Generalized Additive Model (GAM) approximations. A thorough investigation on the proposed approaches was conducted through simulations and a case study with the National Health and Nutrition Examination Survey (NHANES) 2001-2002 data on how persistent organic pollutants influenced leukocyte telomere length, in comparison with the Bayesian Kernel Machine Regression (BKMR) which is one of the most popular mixtures methods.
Our simulation studies demonstrated that the modified BART produced results comparable or superior to BKMR in recovering the true exposure-response surface for both continuous and binary outcomes, especially when chemical groups were considered, with significantly reduced computational time. Both methods effectively identified relevant chemical groups under hierarchical variable selection, but modified BART better distinguished important components within groups. Our case study confirmed these findings, with similar groups identified but different within-group importances estimated. GAM plots accurately summarized individual exposure effects for both modified BART and BKMR fitted results.
We recommend the modified BART as a stable and fast response surface model for environmental mixture analysis, particularly for large sample sizes, binary outcomes and grouped chemicals. GAM approximation is a practical tool for interpreting individual chemical effect in mixtures analysis.

Soft BART

BKMR

Environmental mixture analysis

GAM approximation