Estimating long-term pollution exposure effects through inverse probability weighting methods with Cox proportional hazards models

Abstract

Background: Fine particulate matter (PM2.5) is associated with negative health outcomes in both the short and long term. However, the cohort studies that have produced many of the estimates of long-term exposure associations may fail to account for selection bias in pollution exposure as well as covariate imbalance in the study population; therefore, causal modeling techniques may be beneficial. Methods: Twenty-nine years of data from the National Health Interview Survey (NHIS) was compiled and linked to modeled annual average outdoor PM2.5 concentration and restricted-use mortality data. A series of Cox proportional hazards models, adjusted using inverse probability weights, yielded causal risk estimates of long-term exposure to ambient PM2.5 on all-cause and cardiopulmonary mortality. Results: Covariate-adjusted estimated relative risks per 10 μg/m3 increase in PM2.5 exposure were estimated to be 1.117 (1.083, 1.152) for all-cause mortality and 1.232 (1.174, 1.292) for cardiopulmonary mortality. Inverse probability weighted Cox models provide relatively consistent and robust estimates similar to those in the unweighted baseline multivariate Cox model, though they have marginally lower point estimates and higher standard errors. Conclusions: These results provide evidence that long-term exposure to PM2.5 contributes to increased mortality risk in US adults and that the estimated effects are generally robust to modeling choices. The size and robustness of estimated associations highlight the importance of clean air as a matter of public health. Estimated confounding due to measured covariates appears minimal in the NHIS cohort, and various distributional assumptions have little bearing on the magnitude or standard errors of estimated causal associations.

Parental relationship status and age at autism spectrum disorder diagnosis of their child

“…IPEWs are widely employed across a variety of fields. For instance, IPEWs have been used to evaluate food policies (Edwards et al, 2016), pollution exposure (Higbee et al, 2020), home eviction (Tsai et al, 2020), and substance use (Abdollahpour et al, 2018). Like propensity scores, IPEWs seek to balance exposed and unexposed populations with the goal of "mimicking" experimental methods (Austin, 2011).…”

Mortality Attributable to Long-Term Exposure to Ambient Fine Particulate Matter: Insights from the Epidemiologic Evidence for Understudied Locations

“…Although some remain unconvinced, − the body of evidence, i.e., the consistent epidemiological findings, ,,− the linked biological mechanisms, ,, and results from real-world interventions, overwhelmingly supports the conclusion that long-term exposure to ambient PM 2.5 causes mortality. The World Health Organization (WHO) and USEPA concluded that there was enough evidence to infer a causal link between ambient PM 2.5 and adverse health outcomes before studies using causal methods were published. , These latest study results only serve to strengthen the evidence base for such a conclusion.…”

“…In recent reanalyses of well-established cohorts, authors have started to implement causal inference methods to provide stronger conclusions. Over the past half-decade, many studies were published that utilize these methods. ,− These studies have all concluded that ambient PM 2.5 is causally associated with mortality, even at levels below national standards. ,− …”

“…Kioumourtzoglou et al, Wang et al, Yitshak-Sade et al, and Schwartz et al used difference-in-differences (DID) to analyze this relationship within the Medicare population. ,,, Kioumourtzoglu et al utilized city-specific analyses across time, which eliminated confounding by factors that do not vary across time or that vary across cities, and then, the city-specific health effect estimates were combined using a random effects meta-analysis . Wang et al, Yitshak-Sade et al, and Schwartz et al implemented similar DID methods using census tracts or Zone Improvement Plan (ZIP) codes within their study populations. ,, Wang et al, Awad et al, Danesh Yazdi et al, Higbee et al, and Dominici et al utilized inverse probability weighting (IPW) to adjust their models. − , Wang et al applied two additive hazards models with IPW applied to propensity scores, another method that is commonly employed for causal inference . Awad et al utilized IPW in a subpopulation of Medicare enrollees who moved their residence from one ZIP code to another, which they claim essentially randomized exposure as the new ZIP code is unlikely to be related with any confounders .…”

“…Awad et al utilized IPW in a subpopulation of Medicare enrollees who moved their residence from one ZIP code to another, which they claim essentially randomized exposure as the new ZIP code is unlikely to be related with any confounders . Danesh Yazdi et al and Higbee et al applied IPW within their marginal structural models, yet another method for causal inference. , Dominici et al applied matching, weighting and adjustment to their estimated generalized propensity scores . More about these methods is detailed in Hernan and Robins…”