The data on hospital admissions were extracted from the Health Care Financing Administration (Medicare; Baltimore, MD) billing records, which we obtained for the years 1985–1999. The Medicare system provides hospital coverage for all U.S. citizens ≥ 65 years of age.
We analyzed data on persons who were admitted to the hospital with a primary diagnosis of MI (ICD-9 code 410) between 1986 and 1999. Medicare data provided personal characteristics such as age, sex, and race and the type of admission. Using this information, we selected only emergency admissions to ensure that these were new events and to better ascertain the timing of the event relative to air pollution exposure.
Using a unique identifier for each subject, we traced them through Medicare records to assess whether they had any primary or secondary diagnosis of atrial fibrillation (ICD-9 code 427.3), chronic obstructive pulmonary disease (COPD; ICD-9 code 490–496, except 493), diabetes (ICD-9 code 250), congestive heart failure (CHF; ICD-9 code 428) on previous admissions, and pneumonia (ICD-9 code 480–487) as secondary diagnoses on the index admission. These characteristics were examined as effect modifiers. These diagnoses have previously been suggested as modifiers of the cardiovascular effects of particles (Sunyer et al. 2000
; Zanobetti et al. 2000b
). Previous admissions were traced back to 1985, ensuring at least 1 year of data before the start of the particle data.
Daily monitoring of PM10 is not done in all U.S. cities. We selected the following 21 cities with daily monitoring of PM and representing a geographic distribution across the country: Birmingham, Alabama; Boulder, Colorado; Canton, Ohio; Chicago, Illinois; Cincinnati, Ohio; Cleveland, Ohio; Colorado Springs, Colorado; Columbus, Ohio; Denver, Colorado; Detroit, Michigan; Honolulu, Hawaii; Houston, Texas; Minneapolis–St. Paul, Minnesota; Nashville, Tennessee; New Haven, Connecticut; Pittsburgh, Pennsylvania; Provo-Orem, Utah; Salt Lake City, Utah; Seattle, Washington; Steubenville, Ohio; and Youngstown, Ohio.
For most cities, the metropolitan county encompassed the city and much of its suburbs, but we used multiple counties for Minneapolis-St. Paul (Ramsey and Hennepin, MN), Birmingham (Blount, Jefferson, St. Clair, Shelby, and Walker, AL), Steubenville (Jefferson, OH, and Brooke and Hancock, WV), and Youngstown (Columbiana and Mahoning, OH).
We investigated the association between daily PM10
concentrations and hospital admissions for MI using a case-crossover design. The case-crossover design was developed as a variant of the case–control design to study the effects of transient exposures on acute events (Maclure 1991
). This design samples only cases and compares each subject’s exposure experience in a time period just before a case-defining event with that subject’s exposure at other times. Because there is perfect matching on all measured or unmeasured subject characteristics that do not vary over time, there can be no confounding by those characteristics. If, in addition, the control days are chosen to be close to the event day, slowly varying subject characteristics are also controlled by matching.
Bateson and Schwartz (1999
demonstrated that by choosing control days close to event days, even very strong confounding of exposure by seasonal patterns could be controlled by design in the case–control approach. This makes the approach an attractive alternative to the Poisson models. Levy et al. (2001)
showed that a time-stratified approach to choosing controls, such as sampling control days from the same month of the same year, avoided some subtle selection bias issues and resulted in a proper conditional logistic likelihood. Schwartz et al. (2003)
recently demonstrated with simulation studies that this approach gave unbiased effect sizes and coverage probabilities even with strong seasonal confounding. We used this same stratified approach in our analysis. Matching on day of the week as well as season also controls for the possibility that the day of the week effect varies seasonally.
We defined the hazard period—when a person is at risk for the triggering of an acute MI—as the day of the patient’s hospitalization. Air pollution has short-term serial correlation; to ensure that all of our control days were independent, we chose control days matched on day of the week, in the same month and year as the event day. The data were analyzed using a conditional logistic regression (PROC PHREG, release 8.2; SAS Institute, Cary, NC).
The analysis was conducted for each city separately, and we controlled for day of the week and weather. To control for potential impacts of weather, we used apparent temperature (AT) for the same and previous day, defined as an individual’s perceived air temperature given the humidity. AT was calculated with the following formula (Kalkstein and Valimont 1986
; Steadman 1979
where Ta is air temperature and Td is dew point temperature.
Because risk may vary nonlinearly with AT, we used a regression spline (with 3 df) for both the same day and the previous day. PM10
was modeled linearly. To confirm the report of Braga et al. (2001)
that the association was predominant with PM10
on the day of the event, we examined effects at exposure from lag day 0 to lag day 2. If we could confirm a primary association with lag day 0, we used this for the subsequent analysis described below.
As a sensitivity analysis, we tested an alternate referent selection scheme that matched on AT (rounded to the same degrees Celsius) and used indicator variables to control for day of the week. Because matching on two covariates controls for interactions between the covariates, this controls for the possibility that the temperature effects vary by month. It also renders moot any question of whether the nonlinear dependence of MIs with temperature was modeled correctly. Previous day’s temperature was controlled using a cubic spline in this analysis, as well.
Case-crossover analyses lend themselves to the analysis of effect modification. Factors such as sex are controlled by matching in the design of the study, but we can still test for effect modification with interaction terms or a stratified analysis. We chose stratified analyses, because if a characteristic modifies the effect of PM10, it might also modify the effect of weather or other covariates. A stratified analysis controls for this. Specifically, we conducted stratified analyses by sex, age (< 75 vs. ≥ 75), and previous admission for chronic disease such as atrial fibrillation, COPD, CHF, and diabetes, and secondary diagnosis for pneumonia as an acute modifier.
In a second stage of the analysis, the city-specific results were combined using the multivariate meta-regression technique of Berkey et al. (1998)
. To be conservative, we report the results incorporating a random effect, whether or not there was a significant heterogeneity.
Finally, we assessed the shape of the dose–response relationship by fitting a piecewise linear spline, with slope changes at 20 μg/m3 and 50 μg/m3. We combined these estimates using a random effect meta-analysis as well.