Multipollutant modeling issues in a study of ambient air quality and emergency department visits in Atlanta

Multipollutant models are frequently used to differentiate roles of multiple pollutants in epidemiologic studies of ambient air pollution. In the presence of differing levels of measurement error across pollutants under consideration, however, they can be biased and as misleading as single-pollutant models. Their appropriate interpretation depends on the relationships among the pollutant measurements and the outcomes in question. In situations where two or more pollutant variables may be acting as surrogates for the etiologic agent(s), multipollutant models can help identify the best surrogate, but the risk estimates may be influenced by inclusion of a second variable that is not itself an independent risk factor for the outcome in question. In this paper, these issues will be illustrated in the context of an ongoing study of emergency visits in Atlanta. Emergency department visits from 41 of 42 hospitals serving the 20-county Atlanta metropolitan area for the period 1993-2004 (n=10,206,389 visits) were studied in relation to ambient pollutant levels, including speciated particle measurements from an intensive monitoring campaign at a downtown station starting in 1998. Relative to our earlier publications, reporting results through 2000, the period for which the speciated data are available is now tripled (6 years in length). Poisson generalized linear models were used to examine outcome counts in relation to 3-day moving average concentrations of pollutants of a priori interest (ozone, nitrogen dioxide, carbon monoxide, sulfur dioxide, oxygenated hydrocarbons, PM10, coarse PM, PM2.5, and the following components of PM2.5: elemental carbon, organic carbon, sulfate, and water-soluble transition metals). In the present analysis, we report results for two outcome groups: a respiratory outcomes group and a cardiovascular outcomes group. For cardiovascular visits, associations were observed with CO, NO2, and PM2.5 elemental carbon and organic carbon. In multipollutant models, CO was the strongest predictor. For respiratory visits, associations were observed with ozone, PM10, CO, and NO2 in single-pollutant models. In multipollutant models, PM10 and ozone persisted as predictors, with ozone the stronger predictor. Caveats and considerations in interpreting the multipollutant model results are discussed.