Influence of Prediction Approach for Spatially-Dependent Air Pollution on Health Effects Estimation

Abstract

ISEE-338 Objective: Air pollution studies increasingly estimate individual-level exposure from area-based measurements near participant residences. Common exposure prediction approaches either assign the measurement at the closest monitor or predict concentration at individual sites using kriging. Although many studies have adopted either prediction method, little is known about their validity for health effects estimation. Predicted exposure may produce bias and/or increase uncertainty of health effect estimates. This simulation study explores how prediction approaches for fine particulate matter (PM2.5) affect relative risk estimates for cardiovascular events in a single geographic area. Material and Methods: We used 2002 PM2.5 data (mean = 18.59 μg/m3) from the LA area to select parameters to define a correlation structure based on distance. In addition to 2 structures estimated from measured PM2.5, 3 artificial structures with different correlation features were selected to test sensitivity. Based on each of the 5 structures, annual average PM2.5 was generated at 22 existing monitoring sites in greater LA and 2000 preselected individual locations in urban LA. Five subjects were assumed to live at each individual location. Associated survival time until cardiovascular event (RR = 1.24 per 10 μg/m3 increase in PM2.5) was simulated for 10,000 subjects for 10 years. Using only generated PM2.5 at monitoring sites, we predicted PM2.5 concentrations at subject locations by nearest monitor and kriging interpolation. Finally, relative risks (RR) for time to cardiovascular event were estimated in the Cox proportional hazards model using predicted annual average PM2.5 exposures. Results: By comparing RR estimates for cardiovascular events using predicted exposure in each structure, nearest monitor prediction was more biased but less variable than kriging prediction. As variance-dominated bias, nearest monitor predictions gave RR estimates with lower mean square error; this was consistent in all sensitivity analyses. Conclusions: In our simulation study, nearest monitor prediction performed better than kriging prediction for health effects estimation.