2615

Background: Many studies reported that long-term exposure to air pollution is associated with increased mortality rates. These studies have been critiqued for failure to control for omitted, generally personal, confounders. Studies robust to such confounders can address this issue. Methods: We used a self-controlled design for survival analysis. We stratified on each person in the Medicare cohort between 2000 and 2015 who died and examined whether exposure predicted in which follow-up period the death occurred. We controlled for nonlinear terms in calendar year and age. Slowly varying covariates such as smoking history, BMI, diabetes, usual diet and alcohol consumption, sex, race, socioeconomic status, and green space were controlled by matching. Analyses were restricted to persons entering the study at age 65. We used machine learning models of PM_2.5, NO₂, and O₃ that predict exposure on a 1km grid for the entire U.S. Results: There were 6,452,618 deaths in the study population. In multipollutant models we observed a 5.37% increase in the mortality rate (95% CI 4.67%, 6.08%) for every 5 μg/m³ increase in PM_2.5 , a 2.10% decrease in mortality rate (95% CI -1.88%, -2.33%) for each 5ppb increase in NO₂, and a 1.98% increase in mortality rate (95% CI 1.61%, 2.36%) for each 5ppb increase in O₃. When restricted to people whose PM_2.5 exposure never exceeded 12 μg/m³ the effect size increased for PM_2.5, became positive and significant for NO₂, and disappeared for O₃. Conclusion: There is strong evidence that the association between PM_2.5 and mortality is not confounded by individual or neighborhood covariates, which were controlled by matching. The effects of NO₂ and O₃ were more ambiguous and may reflect differential confounding by each other in different atmospheric conditions.