Practical Bayesian Guidelines for Small Randomized Oncology Trials

Randomization is a well-established statistical tool for obtaining fair treatment comparisons in clinical trials. Despite this, most investigators conducting small early-phase oncology trials of different experimental treatments or doses of a single agent do not randomize patients. This may be due to convention, physicians' desire to choose personalized treatments for their patients, or the belief that randomization is of little value in small trials. We argue that, when it is feasible and ethical, randomization is very desirable in early-phase trials because it gives fair treatment comparisons despite the small sample sizes. Illustrations are provided of how confounding and bias may arise when comparing treatments using data from separate single-arm trials. By eliminating confounding treatment effects with between-study differences in known or unknown prognostic variables, randomization provides unbiased treatment comparisons. To facilitate the planning and analysis of small randomized trials, Bayesian criteria for comparing treatments based on response and toxicity rates are provided. Practical guidelines are given for determining sample sizes, specifying Bayesian safety and futility monitoring rules, and constructing a balanced randomization scheme. The methods are illustrated by a trial of engineered cells for treating steroid-refractory graft-versus-host disease.