Does consideration of larger study areas yield more accurate estimates of air pollution health effects? An illustration of the bias-variance trade-off in air pollution epidemiology.

Background: Spatially-resolved air pollution models can be developed in large areas. The resulting increased exposure contrasts and population size offer opportunities to better characterize the effect of atmospheric pollutants on respiratory health. However the heterogeneity of these areas may also enhance the potential for confounding. We aimed to discuss some analytical approaches to handle this trade-off.

Methods: We modeled NO2 and PM10 concentrations at the home addresses of 1082 pregnant mothers from EDEN cohort living in and around urban areas, using ADMS dispersion model. Simulations were performed to identify the best strategy to limit confounding by unmeasured factors varying with area type. We examined the relation between modeled concentrations and respiratory health in infants using regression models with and without adjustment or interaction terms with area type.

Results: Simulations indicated that adjustment for area limited the bias due to unmeasured confounders varying with area at the costs of a slight decrease in statistical power. In our cohort, rural and urban areas differed for air pollution levels and for many factors associated with respiratory health and exposure. Area tended to modify effect measures of air pollution on respiratory health.

Conclusions: Increasing the size of the study area also increases the potential for residual confounding. Our simulations suggest that adjusting for type of area is a good option to limit residual confounding due to area-associated factors without restricting the area size. Other statistical approaches developed in the field of spatial epidemiology are an alternative to control for poorly-measured spatially-varying confounders.