Bayesian modeling for spatially misaligned health and air pollution data through the INLA-SPDE approach

In air pollution studies a key issue concerns the change of support: pollutant concentrations are continuous phenomena in space but their measurements are typically available at a finite number of point-referenced monitoring stations or result from numerical models. When linking exposure to health outcomes, the latter are usually available at administrative level, hence on an irregular lattice, providing challenges in terms of data misalignment. In this paper we tackle the change of support problem for air pollution and health studies through a two-stage Bayesian approach; in the first stage our model estimates the air pollution concentration at the area level and then in the second stage it links the exposure to the health outcome, accounting for the uncertainty on the exposure estimates. We show through an extensive and realistic simulation that our model is able to predict the concentration accurately at the administrative level as well as estimate the association between exposure and health outcome. We use the Integrated Nested Laplace Approximation, coupled with the Stochastic Partial Differential Equation method for model implementation. Finally we apply the proposed model to evaluate the effect of NO concentration on hospital admissions for respiratory diseases in the Piemonte region (Italy). We found that the upscaling method and the approach used to propagate uncertainty from the first to the second stage have an impact on the posterior distribution of the relative risk. Moreover, we found a significant increased risk of 1.6% and 1.8% associated with an increase of 10 in NO concentration