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 two common prediction approaches for fine particulate matter (PM2.5) affect relative risk estimates for cardiovascular events in a single geographic area.
Methods: We estimated two sets of parameters to define correlation structures from 2002 PM2.5 data in the Los Angeles (LA) area and selected additional parameters to evaluate different correlation features. For each structure, annual average PM2.5 was generated at 22 existing monitoring sites and 2,000 pre-selected individual locations in LA. 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, relative risks (RRs) of the effect of PM2.5 on time to cardiovascular event were estimated.
Results: Health effect estimates for cardiovascular events had higher or similar coverage probability for kriging compared to 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 geographical 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.
Kim, Sun-Young; Sheppard, Lianne; and Kim, Ho, "Influence of prediction approaches for spatially-dependent air pollution exposure on health effect estimation" (June 2008). UW Biostatistics Working Paper Series. Working Paper 331.