Covariate-Adaptive Optimization in Online Clinical Trials

The decision of how to allocate subjects to treatment groups is of great importance in experimental clinical trials for novel investigational drugs, a multibillion-dollar industry. Statistical power, the ability of an experiment to detect a positive treatment effect when one exists, depends in part on the similarity of the groups in terms of measurable covariates that affect the treatment response. We present a novel algorithm for online allocation that leverages robust mixed-integer optimization. In all tested scenarios, the proposed method yields statistical power at least as high as, and sometimes significantly higher than, state-of-the-art covariate-adaptive randomization approaches. We present a setting in which our algorithm achieves a desired level of power at a sample size 25%-50% smaller than that required with randomization-based approaches. Correspondingly, we expect that our covariate-adaptive optimization approach could substantially reduce both the duration and operating costs of clinical trials in many commonly observed settings while maintaining computational efficiency and protection against experimental bias.