Causal inference and ensemble learning as new important tools for investigations of health effects of chemical mixtures.

Abstract

OPS 05: Statistical methods to analyze mixtures, Room 114, Floor 1, August 27, 2019, 1:30 PM - 3:00 PM Background: Studies in environmental epidemiology consider each chemical separately when assessing the adverse health effects of environmental exposures. This single pollutant approach suffers from several pitfalls including: 1) risk of false positives; 2) confounding from correlated exposures; and 3) lack of insights on the cumulative or synergistic effects of exposures. Proposed method: We propose a flexible method that leverages an ensemble learning technique – SuperLearner - that offers greater flexibility in approximating the data generating mechanism, to estimate a robust model for the outcome, in combination with the G-computation - a maximum likelihood based substitution estimator – to infer valid estimates of individual and joint effects of mixtures. We extend the method to reconstruct dose-response relationships with no assumptions on the functional form, and to detect potential interactive effects. We end by proposing to augment the method to doubly robust estimation. Simulation study and application: We ran multiple simulations based on real scenarios and compared the method with available approaches handling complex mixtures (e.g. elastic net, weighted quantile regressions, gradient boosting, random forests, and environmental wide association studies). We subsequently applied the method to investigate the potential effects of a mixture of pollutants on child cognitive outcomes in a Faroese cohort. Main results: The proposed method had the lowest false discovery rate in terms of main effects and interactive effects and was able to adapt to the true underlying structure of the data. In the application part, the method confirmed previously reported associations between mercury and lower cognitive function. Additionally, some perfluorinated compounds showed detrimental effects on cognitive function. Conclusions: By combining concepts from data sciences and causal inference, the proposed approach opens new perspectives for estimating causal effects of environmental chemicals in high-dimensional settings, and will allow to correctly address environmental health causal questions that have policy-relevant consequences.