A self-controlled approach to survival analysis, with application to air pollution and mortality

Background: Many studies have reported that long-term air pollution exposure is associated with increased mortality rates. These investigations have been criticized for failure to control for omitted, generally personal, confounders. Study designs that are 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 PM2.5, O3 and NO2 exposures predicted in which follow-up period the death occurred. We used conditional logistic regression stratified on person and controlled for nonlinear terms in calendar year and age. By design slowly varying covariates such as smoking history, BMI, diabetes and other pre-existing conditions, usual alcohol consumption, sex, race, socioeconomic status, and green space were controlled by matching each person to themselves. Results: There were 6,452,618 deaths in the study population in the study period. We observed a 5.37% increase in the mortality rate (95% CI 4.67%, 6.08%) for every 5 μg/m3 increase in PM2.5, a 1.98% (95% CI 1.61%, 2.36%) increase for 5 ppb increment in O3, and a 2.10% decrease (95% CI 1.88%, 2.33%) for a 5 ppb increase in NO2. When restricted to persons whose PM2.5 exposure never exceeded 12 μg/m3 in any year between 2000 and 2015, the effect size increased for PM2.5 (12.71% (11.30, 14.15)), and the signs of O3 and NO2 reversed (-0.26% (-0.88, 0.35) for O3 and 1.77% increase (1.40, 2.13) for NO2). Effect sizes were larger for Blacks (e.g. 7.71% (5.46, 10.02) for PM2.5). Conclusion: There is strong evidence that the association between annual exposure to PM2.5 and mortality is not confounded by individual or neighborhood covariates, and continues below the standard. The effects of O3 and NO2 are difficult to disentangle