Robustness of association of annual PM2.5 concentration with mortality when incorporating exposure uncertainty

Even at historically low air pollution concentrations in the United States, pollutants are consistently associated with adverse health. To improve exposure assessment, spatio-temporal predictions are used, which possess inherent uncertainty. This uncertainty, however, is not commonly accounted for in health effects models. Uncertainty propagation is critical for comprehensively assessing air pollution impacts on health. We characterized the association between estimated annual PM2.5 concentrations and Medicare death rates across the entire United States while fully propagating PM2.5 uncertainty. We leveraged nationwide annual all-cause death records for people ≥65 years (N=10,697,944) at the ZIP Code level from Medicare claims (2010–2015). We used the Bayesian Non-parametric Ensemble (BNE) to generate spatio-temporal PM2.5 concentrations and uncertainties, then used 500 random samples of annual ZIP Code-level PM2.5 concentrations from the BNE posterior distribution estimates. For each of the 100 nationwide sample estimates, we fit a health model with a Bayesian formulation of the conditional Poisson model, adjusting for potential confounders. We then collected 2000 samples from each posterior health estimate (for 200,000 total samples) into a single pool for the final overall association incorporating uncertainty. We estimated an overall increase in death rates of 3.3% [95%CrI 1.1%, 5.8%; 99.8% posterior probability of positive association] per 1µg/m3 increase in PM2.5 concentrations. 98 of 100 (98%) estimates had a 95% posterior probability of a positive association. During 2010–2015, when propagating the full spatio-temporal uncertainty in PM2.5 predicted concentrations, estimates of the positive association between PM2.5 concentrations and death rates were still positive indicating a consistently harmful association even at PM2.5 levels well below the newly proposed national standards.