Arena, Vincent C. PhD; Mazumdar, Sati PhD; Zborowski, Jeanne V. PhD; Talbott, Evelyn O. DrPH; He, Shui PhD; Chuang, Ya-Hsiu MPS; Schwerha, Joseph J. MD, MPH
An important public health concern is the effect of air pollution on both acute and chronic health problems. This issue has been a topic of much debate and investigation over the last 3 decades. A particular focus of air quality research has been geographic areas with a known history of heavy industrialization. Allegheny County, located in southwestern Pennsylvania, is one such community. In a previous study conducted by Mazumdar and Sussman,1 a relationship was demonstrated among air pollution, total mortality, and heart disease in Allegheny County. Later, a 10-city study from 1986 to 1993 indicated a positive relationship between deaths (excluding external causes) among person ≥65 years of age and PM10, overall and in Pittsburgh.2 The 20-city study of Samet et al3 that included data from 1987 to 1994 showed a positive association with the death rates from all causes, cardiovascular, and respiratory categories for most of the cities, including Pittsburgh.
In the last 2 decades, federal regulators have adopted more stringent air pollution standards, additional air pollution controls have been installed, vehicles are cleaner, heavy industry has declined, and air quality in Allegheny County has generally improved. The question to be addressed now is whether current air pollution levels remain associated with health effects. Although the air quality has improved, the county still houses several chemical plants as well as coking facilities, a carbon filter regenerator, a cement kiln, and steel and other heavy manufacturing sites. In April 1990, the Environmental Protection Agency (EPA) declared portions of Neville Island, an area located within Allegheny County on the Ohio River, as a Superfund Waste Site. Current releases of hazardous chemicals from plants on Neville Island as reported in the EPA’s Toxic Release Inventory (TRI) account for 25% of the annual toxic chemical releases in Allegheny County.
Recent epidemiologic studies of adverse health effects associated with environmental exposure to air pollution have assessed a variety of pollutants and their relationship with various pulmonary disease outcomes using hospital admissions information.4–14 In most of these studies, temperature and humidity were controlled.4–8,12–14
Air pollutants such as particulate mass (PM10 and PM2.5), sulfur dioxide, coefficient of haze (COH), ozone (O3), and aerosol strong acidity (H+) have been investigated. Fine particulate mass, specifically PM10 (and more recently PM2.5), has been identified as able to significantly increase respiratory illness hospital admissions independently of other pollutants. Elevated levels of PM10 pollution have also been associated with increases in reported symptoms of respiratory disease and use of asthma medication. Associations between compromised respiratory health and elevated PM10 pollution have been observed even when PM10 levels were well below the 24-hour national ambient air quality standard of 150 μg/m3.9 In addition, PM10 has been associated with exacerbations of asthma and emergency room visits (P < 0.05).8,10 Other studies have suggested that PM10 and ozone appear to be risk factors for several pulmonary diseases such as pneumonia, chronic obstructive pulmonary disease, and asthma.11,12 The vast majority of studies have shown a positive correlation between air pollution levels and hospital admissions for cardiopulmonary diseases (P < 0.05).6,12,14,15
The objectives of our present study were to evaluate more recent trends in air pollutant levels and cardiopulmonary hospital admissions in Allegheny County from 1995 through 2000 and to investigate the potential association between pollutants at these lower levels and cardiopulmonary events. In this article, we present our findings from the Allegheny County Air Pollution Study (ACAPS).16 We focus on three questions in this study: 1) Are temporal trends evident in the release of air pollutants in the Allegheny County area from 1995 through 2000? 2) Are temporal trends also evident in hospital admissions for cardiopulmonary diagnoses during this same time period? 3) Are hospital admissions for pulmonary and/or cardiopulmonary diseases among residents of Allegheny County correlated with air pollution in the Allegheny County area?
Materials and Methods
Our target population for the present article used adults 65 years of age and older in Allegheny County that are most susceptible to environmental insults. As of December 2000, a total of 1.28 million individuals were residents of Allegheny County. Although small changes in population have been noted (approximately 4% decrease during 1990–2000), the assumption was made that the Allegheny County population was relatively stable during the sampling frame of interest (1995–2000). Approximately 228,000 older adults ≥65 years resided in the county in 2000 and were considered to be the at-risk population.
Cardiopulmonary Hospital Admissions.
Cardiopulmonary hospital admission data for Allegheny County were obtained from the Pennsylvania Hospital Cost Containment Council, Special Request Unit (Harrisburg, PA). To protect the privacy of the patient, the available standard data set did not include certain information required for the ACAPS such as date of admission or discharge. After approval from the University of Pittsburgh Institutional Review Board, a special request was submitted to the Executive Board of the Pennsylvania Cost Containment Council for the inclusion of these variables in the dataset. A database was created that included all inpatient cardiopulmonary hospital admissions of Allegheny County residents within the targeted age group of ≥65 years. These admissions included all records containing a discharge diagnosis of the circulatory system (International Classification of Diseases, 9th Revision [ICD-9] codes 390–459) or respiratory system (ICD-9 codes 460–519).17
Daily Meteorologic Data.
The meteorologic data were derived from the U.S. National Climatic Data Center database for the monitoring site located at the Pittsburgh International Airport, Allegheny County (coopid: 366993, wbandid: 94,823, latitude 40°:30′, longitude −80°:14′, elevation 350.5 ft). This database contains extensive weather information collected on a daily basis. In our study, we abstracted information for daily mean temperature, daily mean dew point temperature, daily mean barometric pressure, and daily mean relative humidity for the time period 1995 through 2000.
Air Pollutant Concentrations.
Ambient air levels of the criteria pollutants were obtained in electronic format for all Allegheny County monitoring sites for December 1994 through December 2000 from the Allegheny County Health Department. The location of the individual sites was established before the conduct of our study and are used by the county for regulatory and compliance purposes. As such, the choice of locations was not determined by the needs of an epidemiology study assessing human health. Data were available from 10 different sites located throughout the county. However, two of the sites did not continuously monitor PM10, with data only available for every sixth day. These data were not suitable for the present investigation. The remaining eight sites did provide continuous monitoring and daily measurements of PM10 and were available for the entire study period. Based on these eight monitoring locations, site-specific daily minimum, maximum, and average values were computed for PM10. Two composite measures were derived to derive a summary daily measure representing exposure for Allegheny County as a whole. Our analyses used the mean of the site-specific daily average PM10 values across all monitoring sites (henceforth referred to as the mean PM10). This is consistent with one of the approaches taken.18
Statistical Modeling of Cardiopulmonary Hospital Admissions
Time-Series and Generalized Additive Models.
Time-series methodologies are needed to examine short-term associations between urban air pollution exposure and indicators of health such as daily mortality and morbidity. The importance of removing the effects of long-term trends and seasonality, meteorologic variability, and autocorrelation has been well recognized in air pollution research. Using specific time-series techniques such as exponential smoothing, autoregressive integrated moving average (ARIMA), and seasonal decomposition, air pollution investigators have previously observed short-term associations between air pollution and health indicators with pollution levels measured on the same day and within 5 days of the health event (ie, mortality, hospital admissions, emergency room visits).9 In recent years, the use of generalized additive models (GAMs)19 and distributed lag models2 have increased; these models have been recognized as the state-of-the-art approach in air pollution research. GAMs allow examination of the possibility of association between specific factors when nonlinear relationships cannot be ruled out and also allow control for potential nonlinear covariates.20 Distributed lag models allow for the effect of an increase in pollution concentration on a single day to be distributed over the same day and several subsequent days.
During the conduct of the ACAPS, concerns were raised in the air pollution research community related to the use of GAMs in the assessment of pollution–health care outcome associations. The results from several large air pollution studies, including the Health Effects Institute-funded National Morbidity, Mortality and Air Pollution Study (NMMAPS),21 have been questioned as a result of specific methodological issues. The GAM is often fitted by the S-Plus software package (Insightful Corp., Seattle, WA). It has recently been demonstrated that the default convergence criteria in S-Plus (version 3.4) do not assure convergence of its iterative estimation procedure and can provide biased estimates of regression coefficients and standard errors.22 The assurance of the convergence of the iterative procedure can be achieved by using a more stringent convergence criterion, but the underestimation of the standard error and the presence of bias in the estimate of the regression coefficient remained.
It has been shown that this underestimation can occur if concurvity, the nonparametric analog of multicollinearity, is present in the data and might lead to significance tests with inflated type 1 error (ie, rejection of the null hypothesis when it is in fact true).23 This may result in erroneously declaring a statistically significant effect when none exists. These researchers argue that some degree of concurvity between the transformed smooth functions and temperature and air pollutant levels is likely to be present in all epidemiologic time-series data, especially when time is used as an independent confounding variable. Dominici, Ramsay, and others suggested that other parametric approaches such as natural splines might be used in place of nonparametric GAM.22–24 A recently developed S-Plus package, gam.exact, allows a more robust assessment of uncertainty of air pollution effects.25
In the present investigation, we used GAM and distributed lag models to evaluate the association between the air pollutants and daily cardiopulmonary hospital admissions because this was considered to be the state-of-the-art methodology when we initiated the current study. We used the stringent convergence criterion for the iterative procedure in the S-Plus software. In our exploratory modeling, we evaluated different meteorologic variables as well as same-day and distributed lag models. The Akaike information criterion (AIC) was used for the comparison and selection of the best-fit models. Mean daily temperature and mean daily relative humidity were associated with the outcome and ambient air pollutant levels and were retained in the selected models as potential confounders. Loess (locally weighted regression smoother) was used to form nonparametric smooth functions required in GAM.26 In our analyses, we did not use gam.exact to fit our models because it requires the smoother to be symmetric and loess is not a symmetric smoother.25
Modeling of Cardiopulmonary Admissions.
Examination of all the data series revealed seasonal variations. Although some of the common seasonal patterns of the hospital admissions, pollutant level, and weather variables (temperature, humidity) may be casual, valid estimate of short-term effects can be obtained by eliminating the longer cycles in the data. We used the GAM to fit the logarithm of the number of daily hospital admissions as a sum of smooth functions of the predictor variables (eg, weather and season). Loess was used to obtain the smooth function of the covariates. The model is given by
where Y is the daily cardiopulmonary admissions, and Xi’s are the covariates. In addition, six dummy variables representing the days of week are also included as linear terms.
Both an unconstrained lag model and a polynomial distributed (constrained) lag model were considered. We decided on a second-degree polynomial distributed lag model, but explored the possibilities of different lags.
The unconstrained lag model is given by
Equation (Uncited)Image Tools
where q is the number of days before the event
Z0 is the PM10 concentration on the concurrent day
Z1 is the PM10 concentration on the previous day
Zq is the PM10 concentration on the qth day before the event.
In the second degree distributed lag model, the coefficient for the effect previous to j days of the event is given by
Equation (Uncited)Image Tools
With the weather variables of daily mean temperature and relative humidity, and days of the week covariates, the final form of the distributed lag model is given by
Equation (Uncited)Image Tools
where the covariates W0, W1, and W2 are calculated as
Equation (Uncited)Image Tools
Descriptives and Trend Analyses for Hospital Admissions
Among Allegheny County residents 65 years and older, a total of 253,151 hospital admissions were observed for the 6-year study period. The mean age of elderly admitted to the hospital was 77.5 years. The mean length of hospital stay was 6.7 days (range, 0–401 days). The vast majority of these hospitalizations were for cardiovascular diseases (72.6%). Females comprised slightly more than one half (56.4%) of the total admissions. In addition, whites comprised most of the admissions (88.7%), followed by blacks (9.0%), a proportion consistent with overall county estimates of race.
Pulmonary admissions followed a seasonal pattern for the elder residents. The highest numbers of admissions for pulmonary diseases occurred in the winter and fall months. The lowest numbers occurred in the summer months. This pattern was observed during all 6 of the study years (1995–2000) and appeared to be constant.
Although cardiovascular hospital admissions did follow a seasonal pattern, it was not as consistent as that for pulmonary-related diseases. Higher numbers of cardiovascular admissions occurred during the winter and fall months for all of the years examined. The fewest admissions occurred during the summer months (eg, June, July, and August). However, this pattern was not consistent over all years. Overall, there appeared to be a slight downward trend in the number of admissions as time progressed during the study period.
An analysis of hospitalization by resident zip code was conducted to determine if the patterns of hospitalization for cardiopulmonary diseases changed during the study period (1995–2000). Because of the time-consuming nature of this analysis, we restricted it to the hospitals with the top three numbers of admissions for adults. In addition, only the most common zip codes were examined because each hospital treated residents from at least 50 different zip codes. Admissions were sorted by zip code and year for 1995 through 2000 to determine where those admitted to the facility resided. A summary table was generated displaying the individual zip codes feeding the three hospital systems by year. Although certain hospitals experienced some increases or decreases in admissions during the study period, overall, the percentage of yearly cardiopulmonary admissions for individuals ≥65 years appeared to be relatively consistent across these zip codes from 1995 through 2000, suggesting that, at least for cardiopulmonary events, a shift to emergency room treatment without subsequent admission did not occur in response to the changing health care environment (ie, HMOs). If a significant shift had occurred, it would be difficult to investigate hospital admission trends over time in relation to pollutants. We are reasonably confident that the Pennsylvania Hospital Cost Containment Council dataset suggests similar admission/treatment patterns for cardiopulmonary diseases across the entire 1995 through 2000 timeframe.
In an effort to determine how many of the hospital admissions were readmissions for a preexisting condition, data from one of the largest area hospitals was analyzed to estimate the percent readmissions. Frequencies of the patient identifier number from 1995 through 2000 were generated. The number of unique identifiers was counted and divided by the total number of admissions to determine the proportion of readmissions that occurred during the study period. Based on the data from this hospital, it is estimated that approximately 20% of the admissions are readmissions. The percentage approaches 40% to 50% if the elderly (65 years and older) are assessed separately.
Cardiopulmonary admission rates (admissions per 1000 residents in the area of interest) were calculated for the years 1995 through 2000 among the elderly (65–74 years and 75+ years) residents of Allegheny County. The Health Care Financing Administration (HCFA) beneficiary denominator was used to calculate admission rates for Allegheny County. Cardiopulmonary admission rates among the elderly 65 to 74 years decreased during the study period, whereas admission rates among the very elderly (75 years and older) increased. It is unknown whether this divergence is related to differences in treatment patterns, increased life expectancy with increased morbidity among the 75-year and older group, aging of the population, or other unidentified factors. Age adjustment of the rates may provide further information concerning the related factors. Race-specific rates of admissions were computed and no distinct differences in admission patterns from 1995 through 2000 were evident. Admission rates were lower among both whites and blacks in 2000 compared with 1995.
Figure 1 shows the distribution of daily hospital cardiopulmonary admissions for our target population. The average was approximately 115 and the median was 119, the lower and upper 25th percentiles were 93 and 135, and a maximum of 212 admissions per day was noted (Table 1).
Descriptives and Trend Analyses for PM10
Table 1 presents the descriptors of PM10, daily mean temperature, and daily mean humidity during the study period (1995–2000). The PM10 concentration was moderate, with an average of 27.9 μg/m3 and a range from 4.8 to 102.4. The mean temperature and relative humidity during the study were 51.9°F and 69.4%, respectively. The Pearson correlation coefficients among the air pollutant and weather variables are shown in Table 2. PM10 demonstrated a statistically significant positive correlation with daily mean temperature and negative correlation with daily mean relative humidity. Monthly mean daily average values for PM10 were plotted for the entire 1995 through 2000 period to graphically display trends in air pollutant levels over time (Fig. 2). As expected, a distinctive cyclic pattern is readily seen with higher values occurring in the summer/fall months and lower values in the winter/spring. During the 6-year period, there was an overall attenuation of the mean PM10 values with an approximate 17% decrease during the 1995 through 2000 period. This is also apparent in the decrease in the peaks over time.
Model Fitting for Cardiopulmonary Hospital Admissions
Although a variety of models were fit with various combinations of weather variables, we present the results for those including mean daily temperature and mean daily relative humidity. Other weather variables when included in the model were found to be not significantly associated with admissions or did not fit the data as well as the combination of temperature and humidity.
The selection of the smoothing parameter (span) for loess was based on the AIC criteria, with smaller AIC values indicating better model fit and the residual plots indicating after removal of the seasonal variation. We tried to keep the degrees of freedom somewhere between 3 and 8 per year. The residual plots are shown in Figure 3 with the smoothed function depicted in red. A strong seasonal pattern is obvious with span = 0.15 (upper left panel) and span = 0.10 (upper right panel). The seasonal patterns have been clearly removed with span = 0.06 and span = 0.02. With 6 years of data, we chose the span as 0.06 with relative smaller AIC value and the degree of freedom per year as 4.8. The default span of 0.5 was used for loess (temperature) and loess (relative humidity).
A weather model was first fit including terms of loess (mean daily temperature), loess (mean daily humidity), time, and day of the week using the previously described specifications. We then assessed the impact of PM10 by adding a term that reflected the pollutant level for the same day. We then fit five second-degree distributed lag models with lag lengths of one to five. Each of the lag models was compared with the weather model to assess the improvement in their fit using the change in deviance and the associated P values. We should note here that because the distributed lag models were not “nested,” the fit of the models cannot be tested between themselves.
For the same-day model, a borderline improvement in model fit (P = 0.06) was noted over the weather model with a significant (t = 4.21) PM10 effect (Table 3). The distributed lag models with lags 1, 2, and 3 did not show better fits over the weather model. For these three models, significant coefficients were seen only when the same-day pollutant level was included together with other lag values (ie, the coefficients for w01, w02, and w03 were significant). The distributed lag models with lags 4 and 5 showed significantly better model fit over the weather model. As mentioned before, the coefficients of pollutant terms in which the same-day level was included were found to be significant (ie, the coefficients for w04, and w05 were significant). These results indicated that a simple sum (or average) of PM10 values for the same day and several previous days is related to the cardiopulmonary admissions.
Because the results of fitting the distributed lag model suggested the effect may be the result of a simple sum of the pollutant levels of the current day and previous days, we wanted to further delineate the relationship with respect to specific days. We fitted unconstrained models using equation 1 successively for lags one through five. These models included all the terms of the weather models and separate terms for PM10. Table 4 summarized the results. Compared with the weather model, the 4- and 5-day lag models showed improved fits as seen in the significant reduction in deviance.
A common finding in all of these six unconstrained models was the positive statistically significant association of the current day’s PM10 level (Z0) with hospital admissions irrespective of the additional information provided by previous days’ PM10 levels. Furthermore, the beta coefficients and standard errors corresponding to the current day PM10 were similar for all of the models. The coefficients ranged from 0.000652 to 0.000551, same-day to 5-day lag models.
Lag models of 3 days and longer also showed a significant negative association with the last day of PM10 level. The apparent inverse effect of PM10 levels from 4 and 5 days previous is attributed to multicollinearity in the data. A detailed look at these results shows that when the same-day and previous-day PM10 levels (Z0 and Z1) were included, both the coefficients were positive, although the coefficient of Z1 is not significant. For a 2-day lag model, with Z2 in the model, the same association with Z0 and Z1 was seen with a negative, nonsignificant coefficient for Z2. For a 3-day lag model, the coefficient of Z1 changes sign but was not significant. Also, the coefficient of Z3, which was negative and significant, becomes not significant and positive in the 4-day lag model and not significant and negative in the 5-day lag model. For the 4-day lag model, the coefficient of Z4 was negatively significant but loses the significance when Z5 was considered with it in the 5-day lag model. This pattern of changes in the significance and the direction of the variables when considered with other variables is a common feature of the inferential problems associated with the multicollinearity in data. Hence, in the absence of any other strong reason for this phenomenon, we attribute it to multicollinearity.
Until recently, Allegheny County did not meet the National Ambient Air Quality Standards (NAAQS) established pursuant to the Clean Air Act of l970. In l998, however, for the first time, all federal standards were met except for the new 8-hour ozone and PM2.5 standards for which the data are currently limited. In the 2 decades, federal regulators have adopted more stringent air pollution standards, additional air pollution controls have been installed, vehicles are cleaner, heavy industry has declined, and air quality in Allegheny County has generally improved from l988 to the present. This current assessment of air pollution trends for PM10 from 1995 through 2000 attempts to validate these improvements over time. In general, air pollutant levels have decreased in Allegheny County during the 6-year study period. Specifically, PM10 concentrations decreased approximately 17% over the study period.
A characteristic pattern of cardiovascular admissions was observed for Allegheny County residents 65 years old and older. Higher numbers of admissions were observed during the spring months. The lowest number of admissions occurred during the late summer months of July and August. Pulmonary admissions among Allegheny County residents 65 years of age and older followed a seasonal patterns with the highest numbers of admissions for pulmonary diseases occurring in the winter and fall months. The lowest numbers occurred in the summer months. These trends were observed during all 6 of the study years (1995–2000).
Overall, cardiopulmonary admissions among the elderly remained relatively constant in Allegheny County from 1995 through 2000; no large decrease occurred to suggest a shift to emergency room or outpatient visits versus hospital admissions for cardiopulmonary treatment. If a significant shift had occurred, it would be difficult to investigate hospital admission trends over time in relation to pollutants. We are reasonably confident that the Pennsylvania Hospital Cost Containment Council dataset suggests similar admission/treatment patterns for cardiopulmonary diseases across the entire 1995 through 2000 timeframe.
Cardiopulmonary admission rates among those elderly individuals 65 to 74 years decreased during the study period, whereas admission rates among the very elderly (75 years and older) increased. It is unknown whether this divergence is related to treatment patterns, increased life expectancy with increased morbidity among the 75 years and older age group, or other unidentified factors.
As suggested by the hospitalization patterns noted here, time-series methodologies were required to examine short-term associations between urban air pollution exposure and indicators of health such as daily mortality and morbidity. The importance of removing the effects of long-term trends and seasonality, meteorologic effects, and serial autocorrelation has been well recognized in air pollution–health outcome research.
Like with many ecologic time-series studies, most monitoring sites are strategically placed to assess compliance of point source emitters and not to determine overall air quality on a countywide basis. The regression models developed assume that exposure to ambient air pollution is uniform across the population at risk and is a composite index from multiple monitoring locations. Health outcomes research related to air pollutant concentrations is difficult under these circumstances. Future health research may require the placement of temporary air quality monitors with the concurrent use of geographic information systems (GIS) to spatially model potential exposure to ambient air pollutants adjusting for wind direction and other possible mitigating factors.
Generalized additive models allow for the examination of the association between variables when nonlinear associations cannot be ruled out and also allows for the control of potential nonlinear covariates. Distributed lag models allow for the effect of an increase in pollutant concentration on a single day to be distributed over the same day and several subsequent days in a nonlinear fashion. Our exploratory modeling of the data suggested that there was an absence of a nonlinear effect of PM10. Compared with the weather model, which contained mean daily temperature, mean daily humidity, time, and day of the week, only the model that included up to 4 and 5 previous days of PM10 levels showed an improved fit of the data. Thus, after controlling for seasonality, weather (average daily temperature and average daily relative humidity), and the day of the week through GAM, our findings suggest that the positive association of PM10 and hospital admissions was related to the sum of the current pollutant level and up to 5 previous days of pollutant levels.
Further examination of the data by fitting lag models containing separate PM10 levels for the current day and previous days showed a consistent pattern that the current day’s pollutant level was positively associated with daily hospital admissions (Table 4). Consistent with the results from the distributed lag models, unconstrained models of 4 and 5 days showed significant improvement in model fit compared with the weather model. On closer inspection of these models, only the same-day PM10 term (Z0) and the most distal-day PM10 term (Z4 from the 4-day and Z5 from the 5-day lag model) were statistically significant. This suggests that the finding of the significant improvement in model fit of the 4- and 5-day distributed lag models over the weather model, in which only the w04 and w05 terms were significant, is driven by the current PM10 and the most distal PM10 levels. That is to say, current hospital admissions are affected by current-day PM10 levels. As discussed in the “Results” section, we attribute the negative association with the pollutant levels beyond the same-day or previous-day levels to multicollinearity.
The coefficients for the current-day PM10 level term were relatively stable among the unconstrained models. The average value and standard error of the coefficient across the six models was 0.000609 and 0.000176. This translates into an increase of 0.0609% (95% confidence interval, 0.0263–0.0955) in cardiopulmonary hospital admissions above that which is predicted by seasonal trend, daily temperature and relative humidity, and day of the week for 1 μg/m3 of PM10. Considering 20 μg/m3 of PM10, which is the interquartile range observed during the study time period, this would translate into a percent increase in hospital admission of 1.2256 (95% confidence interval, 0.8692–1.5833).
In the conduct of our study, we confined our analyses to the evaluation of PM10 and did not consider the possible effects of gaseous copollutants. It is possible that the observed relationship, in part, could be attributed to another pollutant that is correlated with PM10. However, with the consistent strong significant association with the same-day PM10 level, we believe that the other gaseous copollutants, if measured, would not have changed the results. Hence, within this limitation, we conclude that even with the lower concentrations of fine particulate matter (PM10) in Allegheny County during the recent time period of 1995 through 2000, there are still associated adverse health outcomes. Daily cardiopulmonary hospital admissions among elderly residents were associated with countywide average PM10 measures. In particular, our findings suggest that there is a positive association of current-day PM10 levels with cardiopulmonary hospital admissions in this population independent of season, weather (average daily temperature and average daily relative humidity), and the day of the week.
1.Mazumdar S, Sussman N. Relationships of air pollution to health: results from the Pittsburgh study. Arch Environ Health. 1983;38:17–24.
2.Schwartz J. The distributed lag between air pollution and daily deaths. Epidemiology. 2000;11:320–326.
3.Samet JM, Dominici F, Curriero FC, Coursac I, Zeger SL. Fine particulate air pollution and mortality in 20 US cities, 1987–1994. N Engl J Med. 2000;343:1742–1749.
4.Burnett RT, Dales RE, Raizenne ME, et al. Effects of low ambient levels of ozone and sulfates on the frequency of respiratory admissions to Ontario hospitals. Environ Res. 1994;65:172–194.
5.Burnett RT, Smith-Doiron M, Stieb D, Cakmak S, Brook JR. Effects of particulate and gaseous air pollution on cardiorespiratory hospitalizations. Arch Environ Health. 1999;54:130–139.
6.Delfino RJ, Becklake MR, Hanley JA. The relationship of urgent hospital admissions for respiratory illnesses to photochemical air pollution levels in Montreal. Environ Res. 1994;67:1–19.
7.Delfino RJ, Murphy-Moulton AM, Burnett RT, Brook JR, Becklake MR. Effects of air pollution on emergency room visits for respiratory illnesses in Montreal, Quebec. Am J Respir Crit Care Med. 1997;155:568–576.
8.Lipsett M, Hurley S, Ostro B. Air pollution and emergency room visits for asthma in Santa Clara County, California. Environ Health Perspect. 1997;105:216–222.
9.Pope CA III. Respiratory hospital admissions associated with PM10 pollution in Utah, Salt Lake, and Cache valleys. Arch Environ Health. 1991;46:90–97.
10.Romieu I, Meneses F, Sienra-Monge JJ, et al. Effects of urban air pollutants on emergency visits for childhood asthma in Mexico City. Am J Epidemiol. 1995;141:546–553.
11.Schwartz J. Air pollution and hospital admissions for the elderly in Detroit, Michigan. Am J Respir Crit Care Med. 1994;150:648–655.
12.Schwartz J. PM10, ozone, and hospital admissions for the elderly in Minneapolis–St. Paul, Minnesota. Arch Environ Health. 1994;49:366–374.
13.Sunyer J, Saez M, Murillo C, Castellsague J, Martinez F, Anto JM. Air pollution and emergency room admissions for chronic obstructive pulmonary disease: a 5-year study. Am J Epidemiol. 1993;137:701–705.
14.Thurston GD, Ito K, Hayes CG, Bates DV, Lippmann M. Respiratory hospital admissions and summertime haze air pollution in Toronto, Ontario: consideration of the role of acid aerosols. Environ Res. 1994;65:271–290.
15.Bates DV, Sizto R. Air pollution and hospital admissions in Southern Ontario: the acid summer haze effect. Environ Res. 1987;43:317–331.
16.Schwerha JJ, Talbott EO, Zborowski JV, et al. Allegheny County Air Pollution Study (ACAPS): A Retrospective Investigation of Cardiopulmonary Outcomes and Exposure to Air Pollutants (1995–2000). Pittsburgh: University of Pittsburgh; 2003.
17.World Health Organization. International Classification of Diseases, 9th Revision, Clinical Modification. Washington, DC: US DHHS, Public Health Service-Health Care Financing Administration; 1980.
18.Smith RL, Davis JM, Sacks J, Speckman P, Styler P. Regression models for air pollution and daily mortality: analysis of data from Birmingham, Alabama. Environmetrics. 2000;11:719–743.
19.Schwartz J. The use of generalized additive models in epidemiology. Proceedings of the 17th International Biometric Conference. 1994:55–80.
20.Hastie T, Tibshirani R. Generalized Additive Models. London: Chapman and Hall; 1990.
21.Samet JM, ed. National Morbidity, Mortality, and Air Pollution Study. Part I: Methods and Methodologic Issues. Part II: National Morbidity, Mortality, and Air Pollution in the United States. Report 94 Parts 1 and 2. Cambridge, MA: Health Effects Institute; 2000.
22.Dominici F, McDermott A, Zeger S, Samet J. On the use of generalized additive models in time-series studies of air pollution and health. Am J Epidemiol. 2002;156:1–11.
23.Ramsay TO, Burnett RT, Krewski D. The effect of concurvity in generalized additive models linking mortality to ambient particulate matter. Epidemiology. 2003;14:18–23.
24.He S, Mazumdar S, Arena VC. A comparative study of the use of GAM and GLM in air pollution research. Environmetrics. 2005.
25.Dominici F, McDermott A, Hastie TJ. Improved semiparametric time series models of air pollution and mortality. J Am Stat Assoc. 2004;99:938–948.
26.Cleveland WS. Robust locally weighted regression and smoothing scatterplots. J Am Stat Assoc. 1979;74:829–836.