Epidemiologic studies have provided evidence that long-term exposure to airborne fine particulate matter, as indexed by the concentration of particulate matter less than 2.5 μm in aerodynamic diameter (PM2.5), is associated with chronic health effects including cardiovascular and respiratory diseases, and also with increased mortality.1–6 Chronic effects of air pollution potentially encompass both the cumulative effects of long-term exposures and persistent effects of acute exposures.7,8 Chronic effects of air pollution on human health have been estimated primarily with prospective cohort studies, but also with the case-control design (using retrospective estimation of exposure), particularly for lung cancer.9,10
The question of whether long-term exposure to air pollution increases risk for mortality has been addressed in cohort study data by testing whether the hazard is positively associated with time-averaged indicators of pollution exposure. The study data have typically included individual-level covariate information on potential confounding and modifying factors [eg, age, smoking, socioeconomic status (SES), and obesity], and exposure has been classified for geographically defined groups based on common residence location. For example, in the Harvard Six Cities Study (SCS), individual-level data were collected at baseline from samples of residents of 6 US cities and at periodic follow-up; in the initial report on mortality, pollution exposures were assigned to participants based on mean concentrations for the 6 cities over a 17-year follow-up.1 The design of the American Cancer Society's (ACS) Cancer Prevention Study II was similar, but covariate data were collected only on enrollment, and exposure was assigned to the 50 metropolitan areas based on pollution concentrations measured during 1979–1983, over a follow-up interval that extended initially from 1982 to 1989. In the more recent follow-up reports of both cohorts, exposure information has been updated and the chronic effects of pollution over time have been re-estimated to assess whether the risk had changed.3,5 Other cohort studies on long-term exposure to air pollution and mortality have been carried out.11–14 However, the analyses reported to date from these studies have assessed the pollution-mortality relationship cross-sectionally, as did the initial analyses of data from the Six Cities and ACS Studies.6
We have used the Medicare files for 2000–2002 to create a cohort of persons age 65 years and older to assess morbidity and mortality risks associated with air pollution.15 We have previously shown its potential utility in investigating acute exposure to PM2.5 and risk for hospitalization.15 Using this large data set, we have constructed 2 cohorts for the locations corresponding to those included in the ACS and Six Cities Studies, denoted as Med-ACS and Med-SCS, respectively. These 2 cohorts differ from the original analyses with respect to study design, study period, availability of individual and area-level confounders, and modeling approach (Table 1). The original cohorts were “closed” so that new members were not added after their inception, whereas Medicare has new enrollees each month and the Med-ACS and Med-SCS cohort populations are dynamic. In this paper we compare findings from these 2 new cohorts, drawn from Medicare participants, with the results from the 2 earlier cohort studies.
From the Medicare enrollment files, we have constructed a cohort of about 40 million people with individual-level information on age, sex, race, and county of residence for the period 2000–2002. To account for residential history, we have included in the analysis only enrollees who did not change their address during the study period 1999–2002. The cohort is dynamic, including all new enrollees to Medicare during this interval. Using county of residence, we linked the Medicare participants to air pollution monitoring data from the US Environmental Protection Agency Air Quality System on PM2.5. We calculated the mortality rates for the same geographical locations included in the original studies.
Table 2 summarizes the numbers of geographic locations, numbers of deaths and people at risk, mean levels of PM2.5, and exposure periods of the Med-ACS, ACS, Med-SCS, and Harvard Six Cities. For the Med-ACS, we identify the 110 counties corresponding to the 50 metropolitan areas in the ACS.2 These were identified from the list of counties that includes the cities and towns within the metropolitan areas' boundaries as given by Krewski et al.16 For the Med-SCS, we identified the counties that include the 6 cities in the Harvard Six Cities.1
Figure 1 shows the locations of the 6 and 110 counties included in the Med-SCS and Med-ACS, respectively. We estimated yearly county-specific averages of PM2.5 using only data based on measurements for at least 10 months of the year and at least 4 days per month, for the years 2000–2002. Specifically, we computed the 10% trimmed mean of the monitor-specific PM2.5 concentrations for each county (excluding source-based monitors) and then computed the average of these concentrations across the study interval. For the ACS counties, PM2.5 concentrations were not available for Chattanooga, TN and Omaha, NE. We assigned to these counties the PM2.5 concentrations measured in the neighboring counties: McMinn County, TN, and Polk County, IA. For the Harvard Six Cities counties, PM2.5 concentrations were not available for Roane County, TN and Columbia County, WI, which include Harriman, TN and Portage, WI, respectively. For these 2 counties, we assigned the PM2.5 values in the neighboring counties of Knox County, TN and Dane County, WI, respectively. As a sensitivity analysis, we estimated the chronic effects using the SMA as the geographical unit by aggregating the 110 Med-ACS counties to constitute the 50 metropolitan areas used in the ACS.
As indicators of area-level SES, we used: (1) the proportion of women and men with a college degree or higher; (2) the proportion of women and men with a high school degree or higher educational attainment; (3) the proportion of women and men who are unemployed; (4) the proportion of individuals below the poverty level in each age-group; and (5) the median income in the county, from the 2000 US Census. We selected these covariates to maximize comparability with the SES area-level covariates that were used in ACS and Harvard Six Cities.
As surrogates for area-level long-term smoking exposure, we calculated county-specific age-sex-race adjusted standardized mortality rates for chronic obstructive pulmonary disease (COPD) and lung cancer for the period 1992–2002 from the National Center Health Statistics.17
For each year, county of residence, and age-sex stratum, we calculated the number of all-cause deaths and the number of people at risk (Table 2). We stratified the study population in 6 strata defined by sex and 3 age groups (65–74, 75–84, and ≥85 years). Let
be the observed and expected number of all-cause deaths, and let
be the number of people at risk in county c and age-sex stratum s. Let
be the average PM2.5 for the 2000–2002 period. We fit the following log-linear regression model, allowing for age-sex stratum-specific intercepts and including the SES factors
The parameter 10 × θ denotes the log relative risk of mortality associated with a 10 μg/m3 increase in longer-term average PM2.5 adjusted for socioeconomic factors. We fit Model 1 to the Med-ACS and Med-SCS data, separately for each year and also for all 3 years combined. Estimation of Model 1 needs to account for 2 sources of correlation due to the repeated observations by age and sex within a county, and for the clustering of county-specific observations within geographical regions (spatial correlation). Robust standard errors were calculated to account for the correlation due to the repeated measures.18 Poisson regression models with random effects were fitted to account for clustering of county-level observations within larger geographical regions.
As a sensitivity analysis, we explored the impact of exposure period on the results separately by year, regressing county-specific average of PM2.5 for the period 1999–2001 on mortality rates for the period 2000–2002.
Data manipulation and analyses were performed using R, version 2.2.1 (R Development Core Team, 2005) and SAS software, version 9.1 (SAS Institute Inc., Cary, NC 2000).
Figure 2 shows the average PM2.5 concentrations used in the ACS for the 1979–1983 period, plotted against the average PM2.5 concentrations over the 2000–2002 time interval in the Med-ACS. As expected, in most counties average PM2.5 was higher in the earlier ACS data compared with the later Med-ACS data. We found agreement among these measurements except for the noticeable outlier of Fresno, CA, which experienced an increase in PM2.5 concentration.19
Figure 3 shows age- and sex-adjusted mortality rates plotted against average PM2.5 over the 2000–2002 period for the 6 and 110 counties included in the Med-SCS and Med-ACS, respectively. These adjusted mortality rates were obtained by fitting Model 1 without including PM2.5 and the SES covariates in the model. For both studies, the mortality rates tend to be higher in counties with higher average PM2.5 values, although the 2 counties with the highest PM2.5 values among the Med-ACS counties have relatively low mortality rates.
Table 3 compares the results of the analyses of Med-SCS and Med-ACS with the first reports of the Harvard Six Cities and ACS, the reanalyses conducted for the Health Effects Institute, and the follow-up of the Harvard Six Cities and ACS. In the Med-SCS, we found that a 10 μg/m3 increase in average PM2.5 is associated with a 20.8% increase in the overall mortality rate [95% confidence interval (CI) = 14.8–27.1], adjusted for age and sex. In the Med-ACS, we estimated that a 10 μg/m3 increase in average PM2.5 is associated with a 10.8% increase in the mortality rate (8.6–13.0), adjusted for age and sex. Without adjustment for area-level covariates, the Med-SCS and Med-ACS provide results similar to the ACS and Harvard Six Cities. In the Med-ACS, with adjustment for area-level covariates, we estimated a 10.9% increase in the mortality rate (9.0–12.8), slightly larger than the ACS estimates from the reanalysis.16 Estimates from the Med-ACS with the 50 SMAs as geographical units were smaller than from the Med-ACS with the 110 counties and much closer to the ACS estimates.
Table 4 summarizes the results of the Med-SCS and Med-ACS by year and for all years combined, with and without adjustment for area-level SES covariates and with and without adjustment for the SMRs for COPD and lung cancer. The results were similar across years and robust to adjustment for these area-level variables. We also used the average PM2.5 over the period 1999–2001 as the exposure measure, and found that the results were unchanged (results not included). Table 5 gives the sensitivity of the estimates to alternative models for the adjustment of area-level SES covariates. Results were not sensitive to the adjustment.
We have compared risk estimates from 2 well-known cohort studies (ACS and Harvard Six Cities) with estimates from 2 cohorts drawn from Medicare enrollees in the same geographic locations (Med-SCS and Med-ACS). In comparison with the original studies, the Med-SCS and Med-ACS cohorts differed in the exposure data used and in the timing of data collection in relation to outcome events. Risk estimates for the Medicare cohorts were adjusted only for age and sex at the individual level and for smoking and SES at the county level. As surrogate measures of area-level exposure to smoking, we constructed age- and sex-adjusted SMRs for COPD and lung cancer.17 Both diseases reflect the long-term use of tobacco by the population.20
Additionally, the time periods for follow-up were not the same. Nonetheless, the risk estimates were reasonably close and CIs were overlapping (Table 3). Notably, the estimates were comparable even though the estimates for the Medicare cohorts were not adjusted for confounding at the same individual level of detail as in the ACS and Harvard Six Cities analyses.
Prospective cohort studies of air pollution and health with collection of data from individual participants, including the original ACS, have the advantage of collecting extensive information on individual-level covariates. However, they are expensive and time consuming to carry out. In addition, their results continue to be challenged by critics because of imprecise effect estimates and the potential for residual confounding at individual- and area-levels. Because data from Medicare, the EPA, and the census are available nationally, we were able to construct a data set on outcome, air pollution exposure, and confounders for the same geographical locations included in the ACS and Harvard Six Cities. With this data set, we investigated whether Medicare data provide evidence on the chronic effects of air pollution similar to the original ACS and Harvard Six Cities studies, even though confounding cannot be controlled as tightly at the individual level. In general, the estimates from the 2 Medicare cohorts were similar to those from the original cohorts. A useful next step will be a similar replication of other US cohorts.
Strengths of the Medicare data are its size and full representativeness of the target population of elderly persons and, because it is dynamic, successive cohorts can be readily selected so that changes in risk can be tracked. Studies based on the Medicare files, however, are potentially limited by lack of individual-level covariate information on residual confounding and modifying factors.
Several risk factors are of particular concern as potential confounding factors in epidemiologic studies of long-term exposure to air pollution; these include cigarette smoking, and obesity and its correlates. The prevalence rates for both risk factors vary spatially and they are strongly associated with risk for multiple causes of death. However, in the original Harvard Six Cities and ACS analyses, estimates of the effects of air pollution measures on mortality were robust to confounder adjustment, including adjustment for smoking, and effect modification was not found. In the reanalysis of data from these studies, more comprehensive treatment of potential confounding in these studies, including use of time-dependent variables in the Harvard Six Cities, showed insensitivity of the findings to the degree of control of potential confounders.21 The reanalysis found effect modification by educational level, with lesser risk at higher levels of education. Because the Medicare data do not include individual-level information on education, we could not assess whether there was similar effect modification in the Med-SCS and Med-ACS cohorts.
In the Medicare cohort, the potential for individual-level confounding may be low because of the older age of Medicare participants. At older ages, the strength of several risk factors for mortality declines, including smoking and obesity.22,23 These considerations regarding confounding in the elderly and the comparability of estimates from the Medicare and original cohorts imply that cohorts of Medicare participants can provide valid estimates of the effect of air pollution on mortality. We show that the population mortality rates from lung cancer and COPD can be used as indicators of smoking at the population level. We suggest that Medicare enrollees represent a population that can be used to create successive cohorts for tracking the risks of air pollution, on both hospitalization and mortality. We have demonstrated the feasibility of creating such cohorts and comparability of estimates of the effect of particulate air pollution to those from cohorts with better characterization of participants at the individual level.
With adjustment for area-level covariates, in addition to the individual-level risk factors, the Med-ACS estimates were slightly larger than reported in the ACS. Several factors might contribute to the differences. First, chronic effects of PM2.5 on mortality might be underestimated when using exposure indicators aggregated at the larger metropolitan area-level.6 In other words, aggregation bias may have resulted from the differing sizes of geographic units in which estimates of exposure were based. The Med-ACS is based on smaller geographical areas (110 counties) than the ACS (50 metropolitan areas). In the Netherlands cohort study,11 risk estimates almost doubled when local sources of pollution were used in the estimation process versus community-wide concentration. To provide further evidence toward this hypothesis, we have conducted additional analyses using the SMA as the geographical unit instead of the county. Results are presented in Tables 3 and 4. When we used the SMA as the geographical unit, the effect estimates of Med-ACS were similar to the effect estimates from ACS.
Second, the study population in Med-ACS is older than the populations of the ACS and the Harvard Six Cities. The average ages at enrollment in the ACS and Harvard Six Cities were 59 and 50 years, respectively.24 The Medicare cohort comprises only people over 65 years of age, and therefore may represent a more sensitive population to the chronic effects of ambient PM2.5. Previous studies have shown that when pollution levels are high, older people are more likely to be hospitalized for heart and lung problems, and some may die prematurely.25 In contrast, chronic effect estimates in the Harvard Six Cities were actually larger in the younger half of the cohort.1
Third, the estimates from Med-ACS could be biased upward due to the lack of adjustment for individual-level risk factors. Fourth, larger effect estimates observed at lower exposure levels could indicate that the relationship between PM2.5 and mortality is nonlinear.
A further limitation of our analysis is the lack of adjustment for spatially correlated unmeasured confounders. This is an issue of concern in epidemiologic studies that compare adjusted mortality rates with longer-term air pollution exposure across different locations. The Health Effects Institute reanalysis16,24,26–28 developed statistical methods for analyses of spatially correlated data aimed at minimizing the autocorrelation in the residuals.29–32 The original results were confirmed, but the CIs were larger, suggesting that there was substantial residual autocorrelation in the data, a possibility for our analyses as well.
In summary, we have used the Medicare files to assemble cohorts for the same geographic locations included in the Harvard Six Cities and ACS cohorts. Although the analyses of the Med-SCS and Med-ACS cohorts considered only age and sex as individual-level covariates, risk estimates were comparable for the original and Medicare cohorts. We suggest that the periodic assembly of cohorts from Medicare participants can provide tracking of the risks of air pollution in a large, representative, and susceptible population. The Medicare cohorts could provide ongoing evidence to complement data obtained from the small number of fixed-population prospective cohort studies.
We thank Deborah Williams and Charlotte Gerczak for their assistance with the preparation of this manuscript.
1. Dockery DW, Pope CA III, Xu X, et al. An association between air pollution and mortality in six U.S. cities. N Engl J Med
2. Pope CA III, Thun MJ, Namboodiri MM, et al. Particulate air pollution as a predictor of mortality in a prospective study of U.S. adults. Am J Resp Crit Care Med
3. Pope CA III, Burnett RT, Thun MJ, et al. Lung cancer, cardiopulmonary mortality, and long-term exposure to fine particulate air pollution. JAMA
4. Krzyzanowski M, Cohen A, Anderson R. Quantification of health effects of exposure to air pollution. Occup Environ Med
5. Laden F, Schwartz J, Speizer FE, et al. Reduction in fine particulate air pollution and mortality: extended follow-up of the Harvard Six Cities study. Am J Respir Crit Care Med
6. Pope CA III, Dockery DW. Health effects of fine particulate air pollution: lines that connect. J Air Waste Manag Assoc
7. Kunzli N, Medina S, Kaiser R, et al. Assessment of deaths attributable to air pollution: should we use risk estimates based on time series or on cohort studies? Am J Epidemiol
8. Rabl A. Interpretation of air pollution mortality: number of deaths or years of life lost? J Air Waste Manag Assoc
9. Samet JM, Cohen AJ. Air pollution and cancer. In: Schottenfeld D, Fraumeni JF, eds. Cancer Epidemiology and Prevention
. New York: Oxford University Press; 2004.
10. Samet JM, Brauer M, Schlesinger RB. Particulate matter. In: Air Quality Guidelines: Global Update 2005—Particulate matter, ozone, nitrogen dioxide and sulfur dioxide
. Copenhagen: World Health Organization; 2006:217:–305.
11. Hoek G, Brunekreef B, Goldbohm S, et al. Association between mortality and indicators of traffic-related air pollution in the Netherlands: a cohort study. Lancet
12. Lipfert FW, Perry HMJr, Miller JP, et al. The Washington University-EPRI Veterans' Cohort Mortality Study: preliminary results. Inhal Toxicol
. 2000;12(suppl 4):41–73.
13. Abbey DE, Nishino N, McDonnell WF, et al. Long-term inhalable particles and other air pollutants related to mortality in nonsmokers. Am J Respir Crit Care Med
14. Filleul L, Rondeau V, Vandentorren S, et al. Twenty five year mortality and air pollution: results from the French PAARC survey. Occup Environ Med
15. Dominici F, Peng RD, Bell ML, et al. Fine particulate air pollution and hospital admission for cardiovascular and respiratory diseases. JAMA
16. Krewski D, Burnett RT, Goldberg MS, et al. Reanalysis of the Harvard Six Cities Study and the American Cancer Society Study of particulate air pollution and mortality. Investigators' reports parts I and II. 2000. Cambridge, MA, Health Effects Institute.
17. Peto R, Lopez AD, Boreham J, et al. Mortality from tobacco in developed countries: indirect estimation from national vital statistics. Lancet
18. Liang KY, Zeger SL. Longitudinal data analysis using generalized linear models. Biometrika
19. American Lung Association. State of the air: 2004. Technical report. 2004. New York, American Lung Association.
20. US Department of Health and Human Services (USDHHS). The health effects of active smoking: A report of the Surgeon General. 2004. Washington, D.C., U.S. Government Printing Office.
21. Health Effects Institute. Reanalysis of the Harvard Six Cities Study and the American Cancer Society Study of particulate air pollution and mortality. 2000. Cambridge, MA, Health Effects Institute.
22. Burns DM, Garfinkel L, Samet JM. US Department of Health and Human Services (USDHHS), Public Health Service, and National Cancer Institute (NCI). Changes in cigarette-related disease risks and their implication for prevention and control. Bethesda, Maryland, U.S. Government Printing Office (NIH Publication No. 97–4213). Smoking and Tobacco Control Monograph 1997;8.
23. Jee SH, Sull JW, Park J, et al. Body-mass index and mortality in Korean men and women. N Engl J Med
24. Krewski D, Burnett RT, Goldberg MS, et al. Overview of the reanalysis of the Harvard Six Cities Study and American Cancer Society Study of Particulate Air Pollution and Mortality. J Toxicol Environ Health A
25. Pope CA III. Epidemiology of fine particulate air pollution and human health: biologic mechanisms and who's at risk? Environ Health Perspect
. 2000;108(suppl 4):713–723.
26. Krewski D, Burnett RT, Goldberg MS, et al. Validation of the Harvard Six Cities Study of particulate air pollution and mortality. N Engl J Med
27. Krewski D, Burnett RT, Goldberg M, et al. Reanalysis of the Harvard Six Cities Study, part II: sensitivity analysis. Inhal Toxicol
28. Krewski D, Burnett RT, Goldberg M, et al. Reanalysis of the Harvard Six Cities Study, part I: validation and replication. Inhal Toxicol
29. Jerrett M, Burnett RT, Goldberg MS, et al. Spatial analysis for environmental health research: concepts, methods, and examples. J Toxicol Environ Health A
30. Jerrett M, Burnett RT, Willis A, et al. Spatial analysis of the air pollution-mortality relationship in the context of ecologic confounders. J Toxicol Environ Health A
31. Burnett R, Ma R, Jerrett M, et al. The spatial association between community air pollution and mortality: a new method of analyzing correlated geographic cohort data. Environ Health Perspect
. 2001;109(suppl 3):375–380.
32. Ma R, Krewski D, Burnett RT. Random effects Cox models: a Poisson modelling approach. Biometrika