DOUBLY ROBUST OPTIMAL TREATMENT REGIME VIA NEURAL NETWORKS

Abstract

The optimal treatment regime (OTR) is a recent development in personalized medicine. It assigns the best treatment to an incoming patient according to his/her covariate profile. Construction of the optimal regime hinges on unbiased estimates of the treatment effects. In observational studies, it is known that traditional estimates of treatment effects are biased, and causal inference methods are needed to construct unbiased estimates. The doubly robust estimation (DRE) is a recent popular causal inference procedure for this purpose. It requires the specification of a propensity score model and an outcome model, yielding unbiased estimates as long as one of the models is correctly specified. However, when both models deviate from the true ones even slightly, the DRE is still biased, which can be the case if both models are specified parametrically. Here, we consider observational studies with large-enough data for the more robust nonparametric neural network model toworkwell but not large enough for deep learning. We developOTR using DRE from neural networks. Extensive simulation studies are carried out to evaluate the performance of the proposed method, and compare it against its parametric counterpart. Our findings show that the resulting OTR outperforms that under the parametric model when it is incorrectly specified.