Comparing “causal” and “traditional” approaches in the association of long-term exposure to ambient air pollution on mortality: How sensitive are the results?

Abstract

Background: Few comparisons between causal inference and traditional approaches have been performed. We applied “causal” and “traditional” methods to investigate the association between long-term air pollution exposure (PM2.5 and NO2) and mortality. Methods: We analyzed pooled data from eight well-characterized cohorts and one administrative cohort. We defined the generalized propensity score (GPS) as the conditional likelihood of exposure given confounders, and derived corresponding inverse-probability weights (IPW). We applied Cox-proportional hazard models weighted by IPW, adjusted for GPS, and directly adjusting for all confounders. Results: In IPW models, PM2.5 5 µg/m3 increases were associated with hazard ratios (HR) = 1.141 (95% confidence interval (CI): 1.107, 1.176) and 1.050 (1.014, 1.088) in the pooled and administrative cohorts. Corresponding estimates for traditional Cox models were 1.132 (1.107, 1.158) and 1.057 (1.025, 1.089). Almost identical results were found for all approaches and both pollutants, when unbalanced covariates were adjusted for in causal models. Conclusions: Traditional and causal approaches provided consistent associations between long-term exposure to air pollution and mortality.

Keywords

• Air pollution
• Causal inference
• Generalized propensity score
• Inverse probability weighting
• Mortality