A simulation-based comparison of estimation methods for adaptive and classical group sequential clinical trials. Pharmaceutical Statistics. 2022; 21( 3): 599– 611. doi:10.1002/pst.2188, , .
Traditional randomized clinical trials enroll the planned number of participants and then evaluate efficacy at the end of the study. In contrast, adaptive group sequential clinical trial designs allow for interim efficacy and safety evaluations as well as sample size adjustments during the course of study follow-up. The use of adaptive designs may allow for early stopping of a trial when a drug is ineffective or overwhelmingly effective or may allow for sample size changes mid-study. In order to obtain valid inference when using adaptive designs, parameter estimation (e.g., estimated treatment effect and confidence interval (CI)) and hypothesis testing may require adjustment. Statistical methods for controlling the type-I error of hypothesis tests in adaptive group sequential clinical trials are well-established and well-understood. However, methods for obtaining statistically valid point estimates and CIs for adaptive designs are not as well-established or as well-understood.
At the end of an adaptive trial, one may calculate the Repeated Confidence Interval (RCI), which provides conservative coverage (e.g. >95% CI when we want a 95% CI) of the true treatment effect, or the Backward Image Confidence Interval (BWCI), which provides exact coverage (e.g., 95% CI when we want a 95% CI) of the true treatment effect. The BWCI can also provide a median unbiased estimate (MUE) of the true treatment effect (e.g., the MUE over-estimates the true treatment effect just as often as it under-estimates it).
We conducted a simulation study to evaluate the performance of the RCI, BWCI, and naïve (e.g, un-adjusted point estimate and Wald confidence interval) methods following adaptation at an interim look of a group sequential clinical trial. Generally, the BWCI provided exact coverage, the naïve CI provided inconsistent coverage, and the RCI provided conservative coverage. Additionally, we note considerable asymmetry in the coverage from above/from below for the RCI (e.g., approximately 2.5% of RCI had an upper limit below the true treatment effect but far less than 2.5% of RCI had a lower limit above the true treatment effect).
At the end of an adaptive group sequential trial, we strongly recommend the use of the BWCI (and associated MUE), with the RCI computed during interim looks; the naïve CI should be avoided.