Bayesian Inference on Chemical Exposure and Sample Feature-Informed Covariance

Abstract

Environmental exposure studies increasingly collect measurements on many correlated chemicals together with subject-specific demographic and socioeconomic information. These data are often high-dimensional, partially observed, and zero-inflated due to non-detection, creating substantial challenges for inference on both exposure levels and dependence structure. In this thesis, we develop a Bayesian factor modeling framework for multivariate chemical exposure data that allows subject-level covariates to influence both the mean and covariance structure of the exposures while incorporating chemical-level meta-covariates to guide structured shrinkage of the loading matrix. To accommodate excess zeros, we model detection indicators through a probit component that shares the same latent factor structure as the continuous exposure model, yielding a unified and computationally tractable framework for jointly analyzing detection and concentration. Posterior inference is carried out using a Gibbs sampler with data augmentation, block updating, and marginalized regression steps. Simulation studies show that the proposed method accurately recovers group-specific correlation patterns, avoids detecting spurious heterogeneity under a null setting, and remains robust under partial observation and zero inflation, with well-calibrated predictive uncertainty. We apply the method to semi-volatile organic compound exposure data from the HOPE 1000 study and observe that both average exposure levels and the correlation structure of chemical co-exposures can vary with subject-level covariates. These findings highlight the potential value of modeling covariate-dependent dependence structures in environmental mixture studies.