Sensitivity of Regional Health Effect Estimates to Different Approaches to Estimating Long-Term PM2.5 Exposure

Abstract

ISEE-0838 Background and Objective: Estimation of long-term individual-level PM2.5 exposure in health effects studies has made use of prediction approaches such as nearest monitor and kriging. A recent simulation study indicated that health effect estimates using these two approaches differed more in regions where pollutant concentrations were less spatially correlated. This simulation study extends this finding by examining whether nearest-monitor and kriged PM2.5 exposure estimates produce different health effect estimates in 6 U.S. regions with different spatial correlation structures of PM2.5. Methods: Six US study regions (Mid-Atlantic, Northeast, Southeast, Midwest, Northern and Southern California) were defined by overlapping 200 kilometer buffers from the centroids of 24 Women’s Health Initiative Observational Study (WHI-OS) cities. Using year 2000 annual average PM2.5 at EPA monitoring sites, we estimated geostatistical parameters and characterized the spatial correlation structure as the range divided by square root partial sill for each region. Then, we simulated PM2.5 using estimated parameters and survival time to cardiovascular events based on previous findings from the WHI-OS (an incidence rate of 0.032 and relative risk (RR) of 1.24 per 10 μg/m3 increase in PM2.5) in each region. Given the simulated PM2.5 concentration at monitoring locations, we predicted exposure to PM2.5 at hypothesized subject locations using nearest-monitor and kriging approaches. RRs of cardiovascular events were estimated separately given the two prediction approaches in each region. Differences in RR from the two methods were then compared across regions. Result: Long-term PM2.5 spatial correlation was higher in the Midwest and lower in California. Differences in RRs for cardiovascular events from two predicted PM2.5 were higher in the regions with lower PM2.5 spatial correlation. Conclusion: Health effect estimates of long-term PM2.5 exposure are more sensitive to the approach to estimating individual exposure in regions where PM2.5 is less spatially correlated.