Causal inference methods in environmental epidemiology: different approaches to evaluate the health effects of industrial air pollution.

This study aims at implementing different causal inference approaches for the first time in a longitudinal cohort analysis with a continuous exposure, to assess the causal effect of industrial air pollution on health. A first review of the literature on the addressed causal inference methods is conducted, focusing on the main assumptions and suggested applications. Then the main longitudinal study, from which the causal inference methods originate, is described. A standard time-to-event analysis is performed to assess the relationship between exposure to air pollution (PM10 and SO2 from industrial origin) and mortality, as well as morbidity, in the cohort of residents around a large steel plant in the Taranto area (Apulia region, Italy). The Difference-in difference (DID) approach as well as three methods using the generalized propensity score (Propensity Function-PF of Imai and van Dyk, the Dose-response Function DRF by Hirano and Imbens, and the Robins’ Importance sampling-RIS using the GPS) were implemented in a Cox Proportional Hazard model for mortality. The main study demonstrated a negative effect of exposure to industrial air pollution on mortality and morbidity, after controlling for occupation, age, time period, and socioeconomic position index. The health effects were confirmed in all the causal approaches applied to the cohort, and the concentration-response curves showed increasing risk of natural and cause-specific mortality for higher levels of PM10 and SO2. We conclude that the health effects estimated are causal and that the adjustment for socioeconomic index already takes into account other, not measured, individual factors.