This analysis was based on daily counts of emergency hospital admissions for 1999–2005 derived from billing claims of Medicare enrollees from the National Claims History Files. Because the Medicare data analyzed for this study did not include individual identifiers, consent was not specifically obtained. This study was reviewed and exempted by the institutional review board at the Johns Hopkins Bloomberg School of Public Health.
Each billing claim includes age, sex, and race, the date of service, disease classification in accordance with the International Classification of Diseases, Ninth Revision
), and county of residence. In 2006, there were 36.3 million Medicare enrollees aged 65 years or older, representing more than90%of the US population older than 65 years.17
Two broad classes of outcomes were considered based on the ICD-9
codes. Cardiovascular admissions included heart failure (428), heart rhythm disturbances (426–427), cerebrovascular events (430–438), ischemic heart disease (410–414, 429), and peripheral vascular disease (440–448). Respiratory admissions included chronic obstructive pulmonary disease (490–492) and respiratory tract infections (464–466, 480–487). For each outcome, only the primary diagnosis for the hospital admission was considered as the basis for inclusion in the data set. Daily time series of hospitalization rates were constructed by cause for each county by summing the number of emergency hospital admissions for each day in a county for a given outcome.
Our study population consists of approximately 12 million Medicare enrollees living on average 9 miles (14.4 km) from a collocated pair of PM2.5
monitors with data in the EPA’s Air Quality System. The analysis was restricted to 108 counties with a general population larger than 200 000 in 2000 and with at least 210 daily measurements of collocated PM10
data between 1999 and 2005. A map of the 108 counties is shown in . The schedule for measuring PM2.5
was generally 1 every 3 days, while the schedule for measuring PM10
was more commonly 1 every 6 days. A 10% trimmed mean was used when averaging values across monitors within a county, after adjusting for yearly averages within each monitor.18,19
County-specific information is available at http://www.biostat.jhsph.edu/rr/coarse/countyinfo.html
. Temperature and dew-point temperature data were obtained from the National Climatic Data Center on the Earth-Info CD database.
US Counties With a General Population Larger Than 200 000 and With at Least 210 Daily Measurements of Collocated PM10 and PM2.5 Data Between 1999 and 2005
Because PM10-2.5 is not measured directly, its concentration was estimated using the measurements of PM10 and PM2.5 at each location. An indicator of PM10-2.5 was constructed by subtracting the daily measurements of PM10 and PM2.5 at collocated monitors. These differences were averaged across a county using a trimmed mean if the county had multiple collocated monitoring pairs.
Two-stage Bayesian hierarchical models were applied to estimate national and regional average associations between day-to-day variation in PM10-2.5 (at lags 0, 1, and 2 days) and day-to-day variation in county-level hospital admission rates, adjusting for PM2.5, weather, and seasonal and long-term trends in both PM10-2.5 and admission rates.
A power of 80% was estimated to detect a national average relative risk (RR) as small as 0.45% per 10 µg/m3 for cardiovascular disease and 0.81% per 10 µg/m3 for respiratory disease.
In the first stage, overdispersed Poisson models were fit to the county-specific data to obtain estimates of the RR of hospital admissions associated with PM10-2.5. Two parallel time series of admissions were created for those aged 65 to 74 years and for those aged 75 years or older. These county-specific models included (1) the logarithm of the number of people at risk on a given day as an offset; (2) an indicator of the day of the week; (3) age-specific intercept; (4) smooth functions of the current day’s temperature and the mean of the previous 3 days’ temperatures (each using 6 degrees of freedom); (5) smooth functions of the current day’s dew-point temperature and the mean of the previous 3 days’ dew-point temperatures (3 degrees of freedom); (6) a smooth function of calendar time (8 degrees of freedom per calendar year); (7) an indicator for age of 75 years or older; (8) a smooth function of time and age indicator interaction (1 degree of freedom per year); and (9) the daily concentration of PM10-2.5 at a given lag. Each of the smooth functions in the model was represented using natural cubic splines.
For the smooth functions of calendar time, 8 degrees of freedom per year was chosen for the smoother so that little information at time scales longer than 2 months would be retained in estimating the risks. For temperature, 6 degrees of freedom was chosen to give the model sufficient flexibility to account for potential nonlinearity in the relationship between temperature and health outcomes.20
At the second stage, a national average estimate of the short-term association between PM10-2.5
and hospital admissions was obtained by using Bayesian hierarchical models.21–24
These models combine RRs across counties accounting for within-county statistical error and for between-county variability of the true RRs (also called heterogeneity). The posterior probability that the national average effect is positive, as a measure of the strength of the evidence of an association, also was calculated. Significance is
assessed by the posterior probability that the RR is greater than 0 (values greater than 0.95 are considered significant). To produce regional estimates for the eastern and western United States, the county-specific RR estimates across 77 counties in the eastern region and 31 counties in the western region were combined. Counties were defined to be in the eastern region if they had a longitude greater than −100 (), following previous regional comparisons of the health effects of PM2.5
To gauge the potential public health impact of the risk estimates, the annual reduction in admissions (H) attributable to a 10-µg/m3 reduction in the daily PM10-2.5 level for the 108 counties was calculated. H is defined as H= (exp(βΔx)−1) × N where β is the national relative rate estimate for a 1-µg/m3 increase in PM10-2.5, Δx is 10 µg/m3, and N is the number of hospital admissions across the 108 counties for 2005.
Within a county, levels of PM10-2.5
are less homogeneous than for PM2.5
. To assess the potential effect of exposure measurement error, regression calibration25
was performed for a subset of 60 counties with more than 1 pair of collocated PM2.5
Chemical composition data for PM10-2.5
are not available at the national level. The chemical composition of PM2.5
differs between the eastern and western United States9,26
and it is likely this also is true for PM10-2.5
. Therefore, the effects of PM10-2.5
for the eastern and western United States were estimated separately. In addition, the composition of PM10-2.5
is known to vary by degree of urbanicity, 9
but evidence indicating to what extent these compositional differences lead to different health risks is sparse. Therefore, the modification of PM10-2.5
log RRs was explored by a county’s degree of urbanicity by including the percentage of the population living in an urban area or urban cluster within a given county as a second stage covariate in the hierarchical model. An urban area is defined in the US census as a densely settled area consisting of core census block groups that have both a population density of at least 1000 people per square mile and are surrounded by census blocks that have an overall density of at least 500 people per square mile.
The sensitivity of the key findings was assessed with respect to the degrees of freedom in the smooth function of time used to adjust for seasonal and long-term trends, the lag of exposure to coarse particulate matter, and the degrees of freedom in the smooth functions of temperature and dew-point temperature.
The data were analyzed using the statistical software R version 2.6.2 (R Core Development Group). The specific code used for analyzing these data can be viewed at http://www.biostat.jhsph.edu/rr/coarse/