Combining Ensemble Learning Techniques and G-Computation to Investigate Chemical Mixtures in Environmental Epidemiology Studies

Background Although biomonitoring studies demonstrate that the general population experiences exposure to multiple chemicals, most environmental epidemiology studies consider each chemical separately when assessing adverse effects of environmental exposures. Hence, the critical need for novel approaches to handle multiple correlated exposures. Methods We propose a novel approach using the G-formula, a maximum likelihood-based substitution estimator, combined with an ensemble learning technique (i.e. SuperLearner) to infer causal effect estimates for a multi-pollutant mixture. We simulated four continuous outcomes from real data on 5 correlated exposures under four exposure-response relationships with increasing complexity and 500 replications. The first simulated exposure-response was generated as a linear function depending on two exposures; the second was based on a univariate nonlinear exposure-response relationship; the third was generated as a linear exposure-response relationship depending on two exposures and their interaction; the fourth simulation was based on a non-linear exposure-response relationship with an effect modification by sex and a linear relationship with a second exposure. We assessed the method based on its predictive performance (Minimum Square error [MSE]), its ability to detect the true predictors and interactions (i.e. false discovery proportion, sensitivity), and its bias. We compared the method with generalized linear and additive models, elastic net, random forests, and Extreme gradient boosting. Finally, we reconstructed the exposure-response relationships and developed a toolbox for interactions visualization using individual conditional expectations. Results The proposed method yielded the best average MSE across all the scenarios, and was therefore able to adapt to the true underlying structure of the data. The method succeeded to detect the true predictors and interactions, and was less biased in all the scenarios. Finally, we could correctly reconstruct the exposure-response relationships in all the simulations. Conclusions This is the first approach combining ensemble learning techniques and causal inference to unravel the effects of chemical mixtures and their interactions in epidemiological studies. Additional developments including high dimensional exposure data, and testing for detection of low to moderate associations will be carried out in future developments.