Covariate-adjusted response-adaptive designs for semiparametric survival models.

This article focuses on the development and evaluation of covariate-adjusted response-adaptive (CARA) randomization designs for two-arm right-censored survival clinical trials, based on the semiparametric Cox proportional hazards (PH) model. These CARA designs aim to ethically allocate more patients to superior treatments while accounting for patient heterogeneity through covariates, without relying on strict parametric assumptions about survival distributions. The authors propose optimal allocation targets derived from the Cox model and implement randomization procedures such as the covariate-adjusted doubly adaptive biased coin design (CADBCD) and the covariate-adjusted efficient-randomized adaptive design (CAERADE), validating their performance via extensive simulations. Results demonstrate that the proposed CARA designs maintain controlled Type I error rates and comparable power to traditional balanced randomization while ethically skewing allocations toward better treatments, even under model misspecifications and varying recruitment patterns. The methodology is further illustrated by re-designing a real-life colorectal cancer trial, showing practical applicability and advantages over conventional designs in handling treatment–covariate interactions and improving patient benefit during the trial.