Abstract
The Toxicological Evaluation of Realistic Emissions of Source Aerosols (TERESA) study involved withdrawal, aging, and atmospheric transformation of emissions of three coal-fired power plants. Toxicological evaluations were carried out in rats exposed to different emission scenarios with extensive exposure characterization. Data generated had multiple levels of resolution: exposure, scenario, and constituent chemical composition. Here, we outline a multilayered approach to analyze the associations between exposure and health effects beginning with standard ANOVA models that treat exposure as a categorical variable. The model assessed differences in exposure effects across scenarios (by plant). To assess unadjusted associations between pollutant concentrations and health, univariate analyses were conducted using the difference between the response means under exposed and control conditions and a single constituent concentration as the predictor. Then, a novel multivariate analysis of exposure composition and health was used based on Random Forests(™), a recent extension of classification and regression trees that were applied to the outcome differences. For each exposure constituent, this approach yielded a nonparametric measure of the importance of that constituent in predicting differences in response on a given day, controlling for the other measured constituent concentrations in the model. Finally, an R(2) analysis compared the relative importance of exposure scenario, plant, and constituent concentrations on each outcome. Peak expiratory flow (PEF) is used to demonstrate how the multiple levels of the analysis complement each other to assess constituents most strongly associated with health effects.