|
Abstract:
|
The analysis of mixed environmental exposures on health often rely on regression-based statistical methods making efficient estimation of a joint exposure with complex relationships difficult and results difficult to interpret. Novel nonparametric methods with an interpretable target parameter of interest focusing on interactions are needed to ensure robust estimation of a joint exposure. We approach this by treating decision trees as a data-adaptive target parameter where V-fold cross-validation is used to create a training sample and an estimation sample from the data. Thresholds of a mixed and marginal exposure, estimated from the training sample, are applied to the estimation sample. In the estimation sample, cross-validated targeted minimum loss-based estimation (TMLE) is used to estimate the ATE of the mixed exposure. This method, called CVtreeMLE, guarantees consistency, efficiency, and multiple robustness by using TMLE to update machine learning estimates of the data-adaptive parameter determined by the best fitting decision tree. CVtreeMLE provides researchers with V-fold specific and pooled results for ATEs determined by decision trees with accompanying exposure rules.
|
