Abstract

Adaptive clinical trials have design features that adapt to the accumulating data, meaning that the design itself is informative about the treatment effect. Consequently, the overall information from an adaptive clinical trial is a combination of information from two sources, the realised design and the observed outcomes. I will present a general framework for the analysis of adaptive clinical trials, based on the decomposition of overall information into design information and outcome information. The framework provides transparent delineation and comparison of unconditional and conditional approaches to treatment effect estimation. Conditional estimation involves conditioning on the observed design so that the treatment effect is interpreted with reference to the particular design that actually occurred. Unconditional estimation involves interpreting the treatment effect with reference to all designs that could possibly have occurred. Unconditional estimation may be more efficient because it uses more information, but it is potentially subject to bias for some designs. Identifying such bias in a given clinical trial is a motivation of the proposed framework and we show that this is most likely to occur when the outcome information and the design information are inconsistent. Thus, we can detect the presence of bias by assessing heterogeneity between the information provided by the design and the information provided by the outcomes. When such heterogeneity is detected, conditional inference may be more appropriate. Various examples will be considered including response-adaptive randomization, multi-stage phase II studies, multi-arm treatment selection and Bayesian adaptive trials.