To the Editor:

We read with interest the article by Traini et al1 concerning the effects of air pollutants on all-cause mortality and consider it an important contribution to the topic. In particular, the authors acknowledge that air pollutants are better considered as a mixture and accordingly apply statistical methods to try to characterize the effect of this mixture on the outcome. By additionally using propensity scores to control for confounding, authors conclude that their methodology produces results with a causal interpretation—albeit acknowledging some limitations of their study. However, we think additional considerations are necessary for the causal interpretation of these results.

First, we argue that the use of data-driven methods to choose confounders for multivariate models is not necessarily correct and can result in residual confounding and/or collider bias.2 Instead, expert knowledge should drive the choice of confounders, and the assumptions made by the authors to make this choice can be expressed using causal graphs. Although the authors did choose potential confounders based on previous literature, they excluded those that did not cause a relevant change in the coefficient of multivariate regression models. It is not clear if this negligible impact on the regression coefficients would remain true in the multivariable (mixture) analysis or in the calculation of the propensity scores.

Second, we applaud the novel use of inverse probability weighting (IPW) from multivariate generalized propensity scores in the context of a mixture of environmental exposures and consider they set an important precedent for future studies. However, IPW and other propensity score methods assume that all relevant confounders are being considered, and best practices recommend empirical assessment of whether covariate balance is indeed achieved by this method.3 Although the causal claim made by the authors seems to rely heavily on the use of this method, it is not clear to what extent they assessed its validity.

Last, we observe the implicit idea expressed by the authors that the use of causal inference methods leads to causal conclusions and that the lack of use of these methods does not. We argue that most epidemiologic studies about the association between air pollution and health have an implicit or explicit causal objective. Furthermore, virtually any statistical method can lead to causal conclusions, given that a set of assumptions—usually considered under the potential outcomes framework—are met.4 It is the careful consideration of whether these assumptions are met that increases confidence in causal conclusions, not necessarily the use of causal inference statistical methods.

ACKNOWLEDGMENTS

I thank Marc Weisskopf for his insightful comments.

REFERENCES 1. Traini E, Huss A, Portengen L, et al. A multipollutant approach to estimating causal effects of air pollution mixtures on overall mortality in a large, prospective cohort. Epidemiology. 2022;33:514–522. 2. Hernán MA, Hernández-Díaz S, Werler MM, Mitchell AA. Causal knowledge as a prerequisite for confounding evaluation: an application to birth defects epidemiology. Am J Epidemiol. 2002;155:176–184. 3. Austin PC, Stuart EA. Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies. Stat Med. 2015;34:3661–3679. 4. Westreich D. Epidemiology by Design: A Causal Approach to the Health Sciences. Oxford University Press; 2020.