Estimating the causal effects of exposure mixtures: a generalized propensity score method

In environmental epidemiology and many other fields, estimating the causal effects of multiple concurrent exposures holds great promise for driving public health interventions and policy changes. Given the predominant reliance on observational data, confounding remains a key consideration, and generalized propensity score (GPS) methods are widely used as causal models to control measured confounders. However, current GPS methods for multiple continuous exposures remain scarce.

We proposed a novel causal model for exposure mixtures, called nonparametric multivariate covariate balancing generalized propensity score (npmvCBGPS). A simulation study examined whether npmvCBGPS, an existing multivariate GPS (mvGPS) method, and a linear regression model for the outcome can accurately and precisely estimate the effects of exposure mixtures in a variety of common scenarios. An application study illustrated the analysis of the causal role of per- and polyfluoroalkyl substances (PFASs) on BMI.

The npmvCBGPS achieved acceptable covariate balance in all scenarios. The estimates were close to the true value as long as either the exposure or the outcome model was correctly specified, and the results were less impacted by correlations among mixture components. The accuracy and precision of mvGPS and the linear regression model relied on the correctly specified exposure model and outcome model, respectively. The npmvCBGPS outperformed mvGPS in all scenarios. The npmvCBGPS achieved better covariate balance than mvGPS and provided an overall inverse trend between the PFAS mixtures with BMI.

In this study, we proposed npmvCBGPS to accurately estimate the causal effects of multiple exposure mixtures on health outcomes. Our approach is applicable across various domains, with a particular emphasis on environmental epidemiology.