|
Abstract:
|
The Bayesian uncertainty-directed (BUD) design is an attractive adaptive design for multi-arm trials that randomizes patients to arms relative to how much new information the assignment is expected to generate about the key trial goal, given currently observed data. However, standard BUD designs assume that when new patients are enrolled all previous patients' outcomes have already been observed, and in the common trial setting where this assumption is false they can make poor patient assignments. In this work we extend the BUD information metric to account for previous patients' pending outcomes so a much broader range of multi-arm clinical trials can benefit from the BUD approach. This more general information metric, which integrates over the posterior predictive distribution of the pending outcomes, can be substantially more computationally demanding so we develop an improved Sequential Monte Carlo strategy using a simple identity from information theory to keep these designs practical.
|
