In our large study we analyzed over 90,000 emergency department visits for pediatric asthma in relation to ambient air pollutant concentrations. We controlled tightly for meteorology and seasonal asthma trends, and we observed several positive, statistically significant associations between ambient air pollutant concentrations and the rate of pediatric asthma emergency department visits in Atlanta. Ozone was associated with emergency department visits for asthma during the warm season and during the temperate cold season months (November, March, and April). We also observed associations with several traffic-related primary pollutants during the warm season. These pollutants have been found to cause asthma exacerbations and airway inflammation in observational and experimental studies (32
); as supported by both the quintile analysis and the smooth estimates of dose–response, we observed evidence that associations were present at relatively low ambient concentrations. Further, results from two-pollutant models support the conclusion that ambient concentrations of both ozone and traffic-related primary pollutants independently contribute to the burden of asthma exacerbations. Among the three markers of primary traffic pollution that we investigated in two-pollutant models, the rate ratio of the highest magnitude was for carbon monoxide. Because levels of carbon monoxide present in ambient air do not pose appreciable health risks, carbon monoxide concentrations are likely a surrogate for other pollutants emitted from combustion sources more plausibly linked to asthma. Estimates from distributed lag models suggested there were both immediate and lagged effects for these pollutants, with the association of highest magnitude tending to occur on the day of the emergency department visit.
We also observed associations with 3-day moving average concentrations of warm season sulfur dioxide, warm season PM2.5
sulfate, warm season PM2.5
organic carbon, and cold season coarse particles (PM10–2.5
). None of these warm season results were significant in two-pollutant models that also contained 3-day moving average ozone concentrations; however, many of the lag-specific point estimates (from the distributed lag models) were positive at lag 3 and longer, thereby raising the possibility that some pollutants might have shown effects had we created multipollutant models that spanned longer lag periods. We are particularly suspicious of the sulfur dioxide result, because local plume touchdowns strongly impact measured sulfur dioxide concentrations, and consequently it is challenging to develop a daily sulfur dioxide metric that could be considered representative of the urban airshed based on measurements from only five monitoring stations. Further, in previous epidemiologic and experimental studies, ambient concentrations of sulfur dioxide and PM2.5
sulfate have not been consistently associated with impaired pediatric respiratory function (6
). Respiratory function decline and increased risks of asthma exacerbation associated with ambient PM10–2.5
organic carbon concentrations have been reported in previous studies (39
), although there have been relatively few investigations of these pollutants compared with the body of work on PM10
. Both PM10–2.5
organic carbon are comprised of several different compounds, with PM10–2.5
concentrations in Atlanta being largely comprised of metal oxides and crustal material (43
), and PM2.5
organic carbon consisting of mixture of compounds of both primary and secondary origin (26
Our tendency to find stronger associations during the warm season is consistent with previous findings (10
), and although we are unclear about the underlying mechanism for these apparent seasonal differences, it may simply be that during the warm season a greater proportion of asthma exacerbations are caused by air pollution. Rates of emergency department visits for pediatric asthma increase by 60% during the cold season; this increase is largely attributable to exacerbations triggered by viral infections. If the additive effect of air pollution is similar year round, then the attributable fraction (and, correspondingly, the rate ratio) will appear higher during the warm season because there are fewer competing causes of asthma exacerbations during the warm season. Alternatively, it may be that children actually respond more severely to air pollutants during warmer temperatures, perhaps because of some unidentified synergism between the pollutant and a meteorologic or physical factor. Additional contributions to the observed seasonal differences may include nonlinear dose–response functions (e.g., air pollutant concentrations typical during warmer months may be on a steeper part of the dose–response curve) and behavior differences that impact personal pollutant exposures. For example, during the summer children are more likely to play outside, which may lead to a higher correlation between measured ambient concentrations and personal exposures, and consequently result in higher estimated effects of ambient pollutants.
We relied on codes from hospital administrative databases to identify emergency department visits for pediatric asthma. Our definition was relatively broad and included codes for both asthma and wheeze among children aged 5 to 17 years. We excluded children younger than 5 years from our analysis because young children frequently experience transient wheeze, and asthma diagnoses may be suspect (46
); however, even among children age 5 years and older, we observed significant hospital-to-hospital variability in the proportion of emergency visits coded as “asthma” as opposed to “wheeze.” Further, we observed variability in the coding of primary versus secondary diagnoses; for asthma, this typically occurred when a patient presented with both asthma symptoms and a respiratory infection. We conducted subanalyses limited to emergency department visits where asthma or wheeze was reported as the primary diagnosis and observed results similar to those from our primary analytic approach. Because comorbidities were not coded completely and consistently across hospitals, we deemed these data to be of inadequate quality to support analyses where individual visits were stratified according to the presence or absence of a respiratory infection as a comorbidity, even though at the aggregate level the daily count of emergency department visits for respiratory infections was likely an adequate surrogate for the actual burden of respiratory infections in Atlanta. We controlled for the daily count of upper respiratory visits in our statistical models and found it to be an extremely strong predictor of the rate of emergency department visits for pediatric asthma; further, we observed evidence of confounding by respiratory infections, because control for this covariate tended to attenuate the rate ratio estimates, particularly during the warm season.
Although we chose our primary statistical model carefully, all statistical models are misspecified to some degree. Therefore, we reported results from sensitivity analyses using alternative model specifications. Our primary model is based on the case-crossover design, with implementation by Poisson time-series models that account for overdispersion, given that under certain formulations these approaches are nearly identical (27
). Traditionally, investigators have implemented the case-crossover design by matching either on day-of-week (48
) or temperature (50
) within a given month; matching on both day-of-week and temperature is typically not feasible, because data become sparse with too many matching factors. To implement a case-crossover approach in a time-series framework requires terms for the main effects of year, month, and the matching-factor (e.g., a term for each day-of-week); terms for the two-way interactions between year and month, year and day-of-week, and month and day-of-week; and terms for the three-way interactions between year, month, and day-of-week. In developing our primary analytic approach, we explored case-crossover models with matching on year, month, and either day-of-week or lag 0 maximum temperature; however, regardless of the approach, we observed evidence of confounding by within-month trends (e.g., the increasing trend in asthma exacerbations during late August and September because of the “back-to-school” effect) (51
). To control smoothly for these within-month trends we included a cubic polynomial for day-of-season in the regression models. Given this cubic polynomial, inclusion of the three-way interaction terms no longer meaningfully changed the point estimates for the air pollutant effect. Therefore, we abandoned the three-way interactions (and thereby removed hundreds of parameters from the model) and instead implemented a case-crossover analysis by matching only on month and year. In addition to matching on these factors, we controlled tightly for both day-of-week and lag 0 maximum temperature; our base model included indicator variables for year, month, day-of-week, and lag 0 maximum temperature (for each degree of Celsius), and selected two-way interactions (between month and year; month and day-of-week; and month and lag 0 maximum temperature) that we found to be highly predictive of the pediatric asthma emergency department visit rates.
Although we controlled tightly for meteorology and temporal trends and used a case-only analytic approach, confounding by an unmeasured or inadequately modeled risk factor that varied in a systematic way with short-term fluctuations in ambient air pollutant concentrations could have biased our results. Whereas we cannot dismiss the possibility of confounding by an unmeasured factor, we did conduct extensive analyses to understand the relationships with meteorology, temporal trends, and ambient pollen concentrations in our data. Further, we investigated associations with the lag-1 pollutant concentration (the concentration on the day after the emergency department visit), while controlling for the average concentration on lags 0 to 2, as an approach to evaluate model misspecification (14
) because we know that tomorrow's pollutant concentrations are not causally related to today's count of emergency department visits, and any association not caused by chance must be biased.
Measurement error is inherent in all large epidemiologic studies of urban air pollution health effects. Although studies of personal exposures to air pollutants help to advance understanding of biologic responses, from a regulatory standpoint the ambient concentrations are of greatest relevance. One prominent component of error in our study is, therefore, how well the population-weighted spatial average of measurements from urban monitoring stations approximates the ambient concentration across the entire metropolitan Atlanta area. The extent of this measurement error likely varies by pollutant, with primary pollutants (e.g., those from traffic sources) tending to have more measurement error than secondary pollutants (e.g., ozone and PM2.5
). Indeed, in previous work, we observed associations between emergency department visits for cardiovascular disease and spatially heterogeneous pollutants (carbon monoxide and nitrogen dioxide) using measurements from several different air pollution monitors located within 20 miles of the Atlanta population center; however, we did not observe associations when measurements were used from a rural monitor located 38 miles away. Conversely, we observed associations for the spatially homogeneous pollutants (ozone and PM2.5
) regardless of whether the measurements were from the rural or urban monitors (18
). Also contributing to the measurement error issue is the number of air pollutant monitoring stations, which ranged from only one central monitor (for PM10–2.5
water-soluble metals) to 11 monitors (for PM2.5
). Interpretation of two-pollutant models is complicated by these measurement error issues; the pollutant with the stronger estimated effect (e.g., ozone in our analyses) may not be the more harmful pollutant but may instead be the pollutant that has less measurement error (20
). This measurement error also impacts the statistical power to detect effects. In our study we did not find evidence of synergism between ozone and any of the other air pollutants, perhaps because of issues involving measurement error and statistical power.
The findings from our large, population-based time-series study in Atlanta complement previous findings from multicity studies (13
). Whereas multicity designs offer a statistically powerful approach for investigating the health effects of ambient air pollutants, large single-city studies provide the opportunity for investigators to better understand and account for the nuances of local data. Further, the SOPHIA study, which has amassed data on over 10,000,000 emergency department visits in metropolitan Atlanta since 1993, affords ample statistical power to detect subtle health effects of ambient air pollutants, including the health effects of PM2.5
components. In our study we observed evidence that ambient concentrations of ozone and primary pollutants from traffic sources independently contributed to the burden of emergency department visits for pediatric asthma. Further, these associations were present at relatively low ambient concentrations, reinforcing the need for continued evaluation of the Environmental Protection Agency's National Ambient Air Quality Standards to ensure that the standards are sufficient to protect susceptible individuals.