Health effects of long-term air pollution: influence of exposure prediction methods

Abstract

Background: Air pollution studies increasingly estimate individual-level exposures from area-based measurements by using exposure prediction methods such as nearest-monitor and kriging predictions. However, little is known about the properties of these methods for health effects estimation. This simulation study explores how 2 common prediction approaches for fine particulate matter (PM2.5) affect relative risk estimates for cardiovascular events in a single geographic area.

Methods: We estimated 2 sets of parameters to define correlation structures from 2002 data on PM2.5 in the Los Angeles area, and selected additional parameters to evaluate various correlation features. For each structure, annual average PM2.5 was generated at 22 monitoring sites and 2000 preselected individual locations in Los Angeles. Associated survival time until cardiovascular event was simulated for 10,000 hypothetical subjects. Using PM2.5 generated at monitoring sites, we predicted PM2.5 at subject locations by nearest-monitor and kriging interpolation. Finally, we estimated relative risks of the effect of PM2.5 on time to cardiovascular event.

Results: Health effect estimates for cardiovascular events had higher or similar coverage probability for kriging compared with nearest-monitor exposures. The lower mean square error of nearest monitor prediction resulted from more precise but biased health effect estimates. The difference between these approaches dramatically moderated when spatial correlation increased and geographic characteristics were included in the mean model.

Conclusions: When the underlying exposure distribution has a large amount of spatial dependence, both kriging and nearest-monitor predictions gave good health effect estimates. For exposure with little spatial dependence, kriging exposure was preferable but gave very uncertain estimates.