Error in air pollution exposure model determinants and bias in health estimates

Abstract

BACKGROUND: Land use regression (LUR) models are commonly used in environmental epidemiology to assign spatially resolved estimates of air pollution to study participants. In this setting, estimated LUR model parameters are assumed to be transportable to a main study (the ''transportability assumption''). We provide an empirical illustration of how violation of this assumption can affect exposure predictions and bias health-effect estimates. METHODS: We based our simulation on two existing LUR models, one for nitrogen dioxide, the other for particulate matter with aerodynamic diameter <2.5 μm. We assessed the impact of error in exposure determinants used in the LUR models on resultant air pollution predictions and on bias in an exposure-health-effect estimate assessed in a hypothetical cohort. We assigned error to predictors at monitoring sites (sites used to develop the LUR model) and at prediction sites (sites for which exposure predictions were needed), allowing for different error levels between site types. RESULTS: Realistic error in the exposure determinants of the selected LUR models did not induce large additional error in exposure predictions and resulted in only minor (<1%) bias in health-effect estimates. Bias in the health-effect estimates strongly increased (up to 13.6%) when exposure determinant errors were different for monitoring sites than for prediction sites. CONCLUSIONS: These results suggest that only modest reductions in bias in estimated exposure health-effects are to be expected from reducing error in exposure determinants. It is important to avoid heterogeneous errors in exposure determinants between monitoring sites and prediction sites to satisfy the transportability assumption and avoid bias in estimated exposure health-effects.