Learning Optimal Distributionally Robust Individualized Treatment Rules.

Area of Science:

• Data-driven decision science
• Machine learning
• Causal inference

Background:

• Individualized treatment rules (ITRs) aim to maximize expected outcomes based on individual covariates.
• Existing ITR methods often assume identical training and testing data distributions, leading to poor generalizability.
• Covariate shifts between training and testing data pose a significant challenge for robust ITR estimation.

Purpose of the Study:

• To develop a novel framework for finding optimal ITRs under unknown covariate changes between training and testing distributions.
• To enhance the generalizability and robustness of ITRs in data-driven decision-making.
• To propose a method that guarantees reasonable performance across a set of plausible underlying distributions.

Main Methods:

• Introduced a distributionally robust ITR (DR-ITR) framework.
• The DR-ITR maximizes the worst-case value function across distributions close to the training distribution.
• Proposed a calibrating procedure to adapt the DR-ITR to a target population using calibration data.

Main Results:

• The proposed DR-ITR framework demonstrates improved generalizability compared to standard ITR methods.
• Numerical studies confirmed that the calibrated DR-ITR maintains performance across distributions with covariate shifts.
• The method provides performance guarantees across a neighborhood of distributions around the training data.

Conclusions:

• The DR-ITR framework offers a robust approach to individualized decision-making when data distributions are not stationary.
• The proposed calibrating procedure further enhances the practical applicability and generalizability of ITRs.
• This work advances the field of data-driven decision science by addressing distribution shifts in ITR estimation.