Bayesian Methods for Characterizing Complex Multivariate ExposuresEPA Grant Number: R831843
Title: Bayesian Methods for Characterizing Complex Multivariate Exposures
Investigators: Herring, Amy H. , Savitz, David A.
Institution: University of North Carolina at Chapel Hill
EPA Project Officer: Klieforth, Barbara I
Project Period: November 1, 2004 through October 31, 2007
Project Amount: $389,248
RFA: Environmental Statistics Research: Novel Analyses of Human Exposure Related Data (2004) RFA Text | Recipients Lists
Research Category: Environmental Statistics , Human Health
Crucial components of environmental data analysis are exposure assessment and the relationship between exposure and a health outcome. Investigators often go to great lengths to obtain careful, detailed measures of exposure, which may have multiple constituents, change over time, and be modified by personal factors. Investigators often replace this multivariate exposure data with simple summary statistics (e.g., cumulative or average exposures in some time window of interest) for analysis, which may or may not be the aspects of exposure most closely related to health outcomes of interest. Indeed, to assess the association between an individual’s exposure history and a health event, it is preferable (and often necessary) to reduce the dimensionality of the exposure history data to calculate aggregate and cumulative exposure effects. Some examples include exposure to drinking water disinfection by-products (DBPs), occupational exposures to agricultural chemicals, and exposures to air pollutants. The primary objectives of this research are to develop useful methods for summarizing complex exposures and for examining the relationship between the exposures and health outcomes, to account properly for resultant uncertainty in the exposure assessment, and to use the models for estimation, prediction, and other inferences about the effects of multi-chemical, multi-pathway exposures.
We propose to analyze data from Right from the Start, a prospective cohort study of DBPs and spontaneous abortion. We use a flexible Bayesian latent variable model to borrow information across the different constituents in the mixture of DBPs in tap water. We relate time-specific constituent levels (estimated from water quality data at the level of the supplier, behavioral modifiers of exposure, and scientific information on uptake of different DBPs by various routes and exposure sources) to latent variables for certain classes of chemicals (e.g. chlorinated or brominated compounds, and volatile or nonvolatile compounds). To link the exposure to a health outcome, we allow the probability of spontaneous abortion at each week of gestation to depend on latent variables representing chemical class. Using Markov chain Monte Carlo (MCMC) computational methods, we will fit the exposure assessment and outcome models jointly and will assess effects of DBPs on spontaneous abortion. Our approach also provides a framework for assessing effects of specific compounds or groups of compounds in the mixture on a health outcome, isolating the most influential contributors to risk.
The main result of this research will be a method for investigators trying to understand the relationship between multi-chemical, multi-pathway exposures and health outcomes to more precisely and comprehensively assess the underlying relationships, compared to conducting separate analyses of exposure mixture components. Risk assessments that address complex multivariate exposures will therefore be more accurate.