Semiparametric approaches for mitigating spatial confounding in large environmental epidemiology cohort studies

Epidemiological analyses of environmental risk factors often include spatially-varying exposures and outcomes. Unmeasured, spatially-varying factors can lead to confounding bias in estimates of associations with adverse health outcomes. Several approaches for mitigating this bias have been developed using semiparametric splines. These methods use thin plate regression splines to account for the spatial variation present in the analysis but differ in how to select the amount of spatial smoothing and in whether the exposure, the outcome, or both are smoothed. We directly compare current approaches based on information criteria and cross-validation metrics, and additionally introduce a hybrid method to selection that combines features from multiple existing approaches. We compare these methods in a simulation study to make a recommendation for the best approach for different settings and demonstrate their use in a study of environmental exposures on birth weight in a Colorado cohort.