Long-Term Exposure To Air Pollution And Non-Accidental Mortality: A Reanalysis Of A Large Cohort Study Using Escape Exposure Assessment And Propensity Scores

Abstract

Introduction Large cohorts based on administrative data have addressed the role of long-term exposure to air pollution on mortality. However, individual lifestyle factors are usually missing in such studies and a causal modeling approach to investigate the association has not been used. The propensity scores approach, based on a potential outcomes framework, allows the use of a substantial number of correlated covariates that might overcome these limitations. Methods A cohort of more than 1 million adults was followed from 2001 to 2010 with 144,441 deaths for non-accidental causes. Residential annual average exposures to NO2, PM10, Coarse, PM2.5 and Absorbance of PM2.5 were assessed at the baseline using land use regression models from ESCAPE. Data included individual socio-demographic variables (sex, age, level of education, occupation, marital status, place of birth), several area-based indicators of socioeconomic position (e.g. income, unemployment rates), individual data on previous hospitalizations for diseases associated with potential confounders (i.e. hypertension, COPD, and diabetes), and area-based lung cancer mortality rates. We used propensity scores to evaluate the association between pollutants and mortality within strata of deciles of the propensity scores in Cox regression models. Within strata defined by the propensity scores the expected exposure is essentially the same while the true exposure is likely random with respect to confounders thus yielding likely causal effect estimates. Results We found a statistically significant 4% increased risk for 10 µg/m3 PM2.5, a 3% higher risk for 10-5/m PM2.5 absorbance, and a 2% higher risk for 10 µg/m3 NO2. The results were not changed under different model specifications. Discussion This study shows an association of long-term exposure to particles and nitrogen oxides with non-accidental mortality using causal modeling methods that were robust to several model specifications.