A Flexible Method Using a Parametric Bootstrap for Reducing Bias in Adaptive Designs With Treatment Selection

Abstract

Adaptive designs are being implemented at an increasing rate in an effort to improve the efficiency of the drug development process. One such adaptive design is the treatment-selection design whereby a study begins with k treatment arms, but only a subset is carried forward after an interim analysis. The final analysis of the selected arm(s) is then performed using the data from both stages of the study. One drawback from such studies is that the maximum likelihood estimate at the final analysis is often biased due to the selection method, and several approaches have been previously proposed for reducing this bias. We describe here a novel approach that reduces the bias of the point estimate in these designs by comparing the observed results to what would be expected when the treatment arms had equal means. We show that this estimator provides a reasonable balance between bias and mean squared error across several scenarios. An advantage of this approach is that it can be applied when the endpoint comes from a normal or binomial distribution, and it can be applied to other distributions as well.

Key Words:

• Binomial
• Drop-the-Losers
• Normal
• Point estimation
• Seamless Phase II/III

Acknowledgments

The authors thank Gheorghe Doros, Joseph Massaro, Josee Dupuis, and the two anonymous referees for their helpful comments that improved this work.