Bayesian Outcome-Adaptive Randomization Trial Designs: A Promise Not Without Perils

Bayesian Outcome-Adaptive Randomization (OAR) designs for clinical trials are becoming popular. While traditional designs consider a fixed or equal randomization probability during the trial, OAR designs make use of the outcome information obtained for patients already included in the trial to continuously update the probability. As this generally results in more patients being assigned to the ‘more promising’ treatment, based on all current information, the adaptation is suggested to increase patient-specific benefits in clinical trials.

Another extension of the traditional randomized clinical trial designs is targeted designs. In these designs, patients are pre-screened by using, e.g., biomarkers, before being randomized to treatments which would be deemed the ‘most promising’ based on patients’ screening results.

Designs combining OAR with the idea of targeted designs have been proposed to avoid the need to pre‑select patients based on their biomarker status. By using OAR, it is possible to assign patients within a particular biomarker stratum to the ‘most promising’ treatment arms during the course of the trial. Additionally, it is also possible to stop accrual to ‘non-promising’ treatments during the trial.

Implementation of the Bayesian (biomarker-driven) OAR designs is not trivial. Elements such as the selection of the prior distributions, early-stopping criteria, or biomarker­-assay accuracy strongly influence operational characteristics of the designs. Also, while the designs may offer advantages such as reduced total target sample size or decreased variation of the accrued sample size, several issues have been identified. These include statistical inefficiency due to an imbalance in the number of patients assigned to different treatment arms and a non-trivial probability of ending up with a substantially larger number of patients assigned to the less-efficient treatment arm.

In this webinar, the speakers will discuss the crucial elements of implementing OAR designs and illustrate their impact on the operational characteristics. They will then review the advantages and disadvantages of the designs, including some new results related to the use of imperfect biomarker assays.

Speakers

Tomasz Burzykowski, PhD, Professor of Biostatistics, Hasselt University (Belgium), Vice-President of Research, International Drug Development Institute (IDDI)

Tomasz Burzykowski has over 25 years of experience in clinical trial design, with a particular focus in oncology. His research interests include clinical trial design, meta-analysis, validation of biomarkers and surrogate endpoints, statistical methods in oncology and survival analysis. A full list of his publications can be found here.

Message Presenter

Leandro Garcia Barrado, PhD, Post-Doctoral Researcher, Hasselt University (Belgium)

Leandro Garcia Barrado has eight years of experience in optimizing methods for development and validation of biomarker-based diagnostic tests and Bayesian outcome-adaptive designs. His research interests include Bayesian statistics, development and validation of diagnostic tests, adaptive clinical trial designs and cross-over designs.

Message Presenter