Comparison of traditional Cox regression and causal modeling to investigate the association between long-term air pollution exposure and natural-cause mortality within European cohorts

Most studies investigating the health effects of long-term exposure to air pollution used traditional regression models, although causal inference approaches have been proposed as alternative. However, few studies have applied causal models and comparisons with traditional methods are sparse. We therefore compared the asso-ciations between natural-cause mortality and exposure to fine particulate matter (PM2.5) and nitrogen dioxide (NO2) using traditional Cox and causal models in a large multicenter cohort setting. We analysed data from eight well-characterized cohorts (pooled cohort) and seven administrative cohorts from eleven European countries. Annual mean PM2.5 and NO2 from Europe-wide models were assigned to baseline residential addresses and dichotomized at selected cut-off values (PM2.5: 10, 12, 15 mu g/m3; NO2: 20, 40 mu g/m3). For each pollutant, we estimated the propensity score as the conditional likelihood of exposure given available covariates, and derived corresponding inverse-probability weights (IPW). We applied Cox proportional hazards models i) adjusting for all covariates (traditional Cox) and ii) weighting by IPW (causal model). Of 325,367 and 28,063,809 par-ticipants in the pooled and administrative cohorts, 47,131 and 3,580,264 died from natural causes, respectively. For PM2.5 above vs. below 12 mu g/m3, the hazard ratios (HRs) of natural-cause mortality were 1.17 (95% CI 1.13-1.21) and 1.15 (1.11-1.19) for the traditional and causal models in the pooled cohort, and 1.03 (1.01-1.06) and 1.02 (0.97-1.09) in the administrative cohorts. For NO2 above vs below 20 mu g/m3, the HRs were 1.12 (1.09-1.14) and 1.07 (1.05-1.09) for the pooled and 1.06 (95% CI 1.03-1.08) and 1.05 (1.02-1.07) for the administrative cohorts. In conclusion, we observed mostly consistent associations between long-term air pollu-tion exposure and natural-cause mortality with both approaches, though estimates partly differed in individual cohorts with no systematic pattern. The application of multiple modelling methods might help to improve causal inference.299 of 300 words.

Automated summary

Most studies on the health effects of long-term exposure to air pollution have used traditional regression models, but few have applied causal inference approaches. In this study, we compared the associations between exposure to PM2.5 and NO2 and natural-cause mortality using both traditional Cox and causal models in a large cohort setting. The results showed consistent associations between air pollution exposure and natural-cause mortality using both approaches, although there were some differences in estimates among individual cohorts. Multiple modeling methods may help improve causal inference.