Error in estimating area‐level air pollution exposures for epidemiology

In air pollution epidemiology, measurements of relevant exposure concentrations are typically made at point locations, resulting in spatial misalignment between the exposure data and health outcomes aggregated at the area level. To obtain values that match the spatial units of the health data, observations can be averaged directly or prediction models developed. We present a framework for evaluating the error in aggregating exposure concentrations to the area unit. We present estimators of mean squared error that can be used for model selection. We find that exposure prediction models, even when misspecified, outperform monitor averages in settings with realistic numbers of monitors and that important reductions in error of the health effect estimate can be obtained when restricting to areas with a monitor. In an analysis of long‐term particulate matter concentrations across the United States, we estimated the error of the prediction model approach to be less than that of the monitor averaging approach on average across counties. We present health effect estimates about particulate matter exposure and pediatric asthma morbidity in the Medicaid population using each approach. Our findings support the use of a prediction model for estimating area‐wide averages, even when restricting to areas that contain a monitor.