Primary Submission Category: Machine Learning and Causal Inference
Doubly Robust Policy Learning for Multi-dimensional Stochastic Interventions through Auto-debiased Neural Networks
Authors: Sylvia Cheng, Alejandro Schuler,
Presenting Author: Sylvia Cheng*
Designing a policy for assignment mechanisms of multiple continuous treatments (e.g. varying dosages of different drugs) creates high-dimensional complexity and remains challenging. Trial and real-world data often lack enough variability due to the exponential growth of the treatment-covariate space, which leads to positivity violations due to little support in rare covariate profiles.
As an alternative, we propose a method of learning an optimal multi-dimensional shift intervention policy by integrating Targeted Maximum Likelihood Estimation with neural nets. It learns a shift under the stochastic treatment regime to modify treatment values based on their natural observed values. Our framework is statistically rigorous to learn an optimal multi-dimensional shift and evaluate its debiased effect with semiparametrical efficiency and doubly robustness. Using a two-step data-splitting process and one compact neural net, it achieves promising run speed. Furthermore, it handles multi-dimensional shift learning via an influence-curve-based loss, optimizing the expected outcome while penalizing variance when deviating from the natural policy. It offers great flexibility for treatment designs and mitigates positivity violations in static regimes, as learned treatment–covariate combinations are supported and less likely to be sparse. We evaluated the method in simulations and showed that the learned shift parameters and causal effect converge to their population truth.