Utilizing Regression Trees to Identify Complex Patterns of Multipollutant Joint Effects

Abstract

Background: Of interest in multipollutant research are the joint effects of two or more pollutants on a given outcome. Classification and regression trees are nonparametric approaches that use supervised recursive partitioning to split data into groups that contain similar responses for the outcome. Aims: Estimate the joint effects of ambient CO, NO2, O3, and PM2.5 concentrations on respiratory emergency department visits in Atlanta, Georgia during 1998-2009. Methods: Pollutant concentrations on each day (N=4,139) were categorized into quartiles. Days when all four pollutants were in their lowest quartile (N=155) served as the referent; the remaining 3,984 days were analyzed using regression tree methods. To do this, we specified a Poisson time-series model containing terms for suspected confounders (e.g., time, meteorology) to evaluate the effect of each pollutant parameterized as a dichotomous variable according to quartile cut points (e.g., comparing quartile 1 with quartiles 2-4). The process was repeated for each pollutant across the three possible quartile cut points. The pollutant-specific cut point that resulted in the smallest p-value below a pre-specified alpha was selected as the first splitting variable, and the process was repeated on the resulting subsets of data. This continued until no cut points had a p-value below alpha (resulting in a “terminal node” in the tree). We used the time-series model to estimate the risk ratio (RR) of respiratory ED visits for each terminal node, using the subset of days when all pollutants were in their lowest quartile as the referent. Results: The regression tree is depicted for various levels of alpha. The largest RR was observed for the terminal node corresponding to days when both CO and PM2.5 were in the highest quartile (RR: 1.02, 95% CI: 1.00–1.05). Conclusions: Regression trees can be used to estimate joint effects from the data and can be used to generate hypotheses about the health effects of multiple pollutants.