Bayesian dose selection design for a binary outcome using restricted response adaptive randomization

In phase II trials, the most efficacious dose is usually not known. Moreover, given limited resources, it is difficult to robustly identify a dose while also testing for a signal of efficacy that would support a phase III trial. Recent designs have sought to be more efficient by exploring multiple doses through the use of adaptive strategies. However, the added flexibility may potentially increase the risk of making incorrect assumptions and reduce the total amount of information available across the dose range as a function of imbalanced sample size.

To balance these challenges, a novel placebo-controlled design is presented in which a restricted Bayesian response adaptive randomization (RAR) is used to allocate a majority of subjects to the optimal dose of active drug, defined as the dose with the lowest probability of poor outcome. However, the allocation between subjects who receive active drug or placebo is held constant to retain the maximum possible power for a hypothesis test of overall efficacy comparing the optimal dose to placebo. The design properties and optimization of the design are presented in the context of a phase II trial for subarachnoid hemorrhage.

For a fixed total sample size, a trade-off exists between the ability to select the optimal dose and the probability of rejecting the null hypothesis. This relationship is modified by the allocation ratio between active and control subjects, the choice of RAR algorithm, and the number of subjects allocated to an initial fixed allocation period. While a responsive RAR algorithm improves the ability to select the correct dose, there is an increased risk of assigning more subjects to a worse arm as a function of ephemeral trends in the data. A subarachnoid treatment trial is used to illustrate how this design can be customized for specific objectives and available data.

Bayesian adaptive designs are a flexible approach to addressing multiple questions surrounding the optimal dose for treatment efficacy within the context of limited resources. While the design is general enough to apply to many situations, future work is needed to address interim analyses and the incorporation of models for dose response.