Daily Mortality and Air Pollution in Atlanta: Two Years of Data from ARIES

Abstract

Associations between daily mortality and air pollution were investigated in Fulton and DeKalb Counties, Georgia, for the 2-yr period beginning in August 1998, as part of the Aerosol Research and Inhalation Epidemiological Study (ARIES). Mortality data were obtained directly from county offices of vital records. Air quality data were obtained from a dedicated research site in central Atlanta; 15 separate air quality indicators (AQIs) were selected from the 70 particulate and gaseous air quality parameters archived in the ARIES ambient air quality database. Daily meteorological parameters, comprising 24-h average temperatures and dewpoints, were obtained from Atlanta's Hartsfield International Airport. Effects were estimated using Poisson regression with daily deaths as the response variable and time, meteorology, AQI, and days of the week as predictor variables. AQI variables entered the model in a linear fashion, while all other continuous predictor variables were smoothed via natural cubic splines using the generalized linear model (GLM) framework in S-PLUS. Knots were spaced either quarterly, monthly, or biweekly for temporal smoothing. A default model using monthly knots and AQIs averaged for lags 0 and 1 was postulated, with other models considered in sensitivity analyses. Lags up to 5 days were considered, and multipollutant models were evaluated, taking care to avoid overlapping (and thus collinear) AQIs. For this reason, PM_2.5 was partitioned into its three major constituents: SO₄²⁻, carbon (EC + 1.4 OC), and the remainder; sulfate was assumed to be (NH₄)₂SO₄ for this purpose. Initial AQI screening was based on all-cause (ICD-9 codes < 800) mortality for those aged 65 and over. For the (apparently) most important pollutants—PM_2.5 and its 3 major constituents, coarse PM mass [CM], 1-h maximum CO, 8-h maximum O₃—we investigated 15 mortality categories in detail. (The 15 categories result from three age groups [all ages, < 65, 65+] and five cause-of-death groups [all disease causes, cardiovascular, respiratory, cancer, and other “remainder” disease causes]). The GLM model outputs that were considered included mean AQI effects and their standard errors, and two indicators of relative model performance (deviance and deviance adjusted for the number of observations and model parameters). The latter indicator was considered to account for variations in the number of observations created by varying amounts of missing AQI data, which were not imputed. The single-AQI screening regressions on all-cause 65+ mortality show that CO, NO₂, PM_2.5, CM, SO₂, and O₃, followed by EC and OC, consistently have the best model fits, after adjusting for the number of observations. Their relative rankings, however, vary according to the smoothing knots used, and there is no correspondence between mean AQI effect and overall model fit. (Other regression runs often show that the best model fits are obtained with no AQI in the model.) There is no correspondence between mean AQI effect and statistical significance or between mean effect and serial correlation. There is a highly significant (.001 level) relationship between overall model fit and serial correlation; the best fitting models have the most frequent knot spacing and the most negative serial correlation. The regression analyses by cause of death find elderly circulatory deaths to be consistently associated with CO for all models.

The authors thank EPRI for sponsoring ARIES and this work and Brian Lehman and Elisabeth Walden for the creation and reviewing of tables and figures.