Causal Machine Learning Approaches for Adjusting Treatment Switching in Clinical Trials

Post-randomization treatment switching is common in randomized trials and can bias treatment-effect estimates when the estimand targets the hypothetical outcome under no switching. Existing adjustments either rely on strong structural assumptions about treatment effects (e.g., on the failure-time scale) or sacrifice effective sample size by censoring switching patients. We propose a causal machine-learning framework that (i) learns an arm-specific outcome model using non-switching patients, (ii) corrects prognosis-driven selection bias via transfer-learning–based reweighting to align the covariate distribution of non-switching patients with that of switching patients, and (iii) predicts counterfactual outcomes for switching patients as if they had remained on their randomized treatment. Treatment effects for the no-switch estimand are then computed on the reconstructed no-switch population using standard estimators. The approach extends naturally to multi-directional switching and multi-arm trials. Comprehensive simulation studies and a real-data application demonstrate the practical performance of the proposed framework.