Development and Testing of Response Surface Methods for Investigating the Epidemiology of Exposures to Mixtures

Abstract

We are extending Bayesian kernel machine regression (bkmr) to examine health outcomes as a function of a multi-dimensional space of exposures, adding capabilities needed by environmental epidemiologists such as mediation, multiple imputation and the ability to analyze large data sets. Borrowing ideas from toxicology, we are developing an approach related to bkmr that constructs exposure summary scores and tests when they are appropriate, an important goal for mixtures epidemiology. We are testing bkmr and other other mixtures methods and comparing results using both synthetic data and real world data sets, including the Environment And Reproductive Health (EARTH) Study.A second focus of our project is the application of causal models to mixtures epidemiology, including the construction of synthetic data sets and interpretation of results. We have extended our earlier analysis of co-exposure amplification bias to methods in addition to regular linear regression. Special care in causal methods must be used when constructing synthetic data sets with increasing correlation between exposures.