Methods for effect modification with multivariate environmental exposures

Abstract

Humans are exposed to a multitude of environmental insults every day. Exposures such as air pollution, heat and extreme weather, heavy metals, and environmental chemicals are known to be linked to adverse health outcomes. There is interest in understanding how multivariate exposures, including repeated measures of the same exposure over time for the same observation and measures of multiple exposures at a single time point, impact health. Several statistical approaches have been proposed for the analysis of multivariate exposure data. Two commonly used methods are distributed lag models (DLMs) for repeated measures of exposure and Bayesian kernel machine regression (BKMR) for multiple exposures observed at a single time point. These methods and their variants are widely used in environmental health studies. However, they lack flexibility to estimate effect modification in most settings. In this dissertation, I develop methods to include effect modification in both DLMs and BKMR. The first method is the distributed lag interaction model (DLIM) that extends the standard distributed lag framework to allow for modification of the exposure-time-response function by a single continuous variable. I use a cross-basis, or bi-dimensional function space, inspired by the distributed lag non-linear framework to simultaneously describe the temporal and modification structure. Next, I developed a distributed lag interaction model with index modification (DLIM-IM) that allows for modification of the exposure-time-response function by multiple modifiers via a data-derived modification index. I use a Bayesian hierarchical framework to simultaneously estimate the exposure-time-response function and a weighted modification index, and I allow for selection on the candidate modifiers. Finally, I propose and evaluate extensions of the BKMR framework to include effect modification by a categorical modifier. I propose a new version of BKMR with a separable covariance function that allows for increased flexibility to estimate effect modification as well a comparing alternative ways to apply BKMR for assessing modification For each of these methods, I validated these methods through simulation and applied them to multiple data sets to demonstrate their application. I have made open-source software for the methods publicly available on CRAN and GitHub.