Bias and precision in early phase adaptive oncology studies and its consequences for confirmatory trials

The need for a more efficient drug development process led to migration from the traditional fixed-sample clinical trial designs to group-sequential and adaptive designs, especially in early phases of clinical drug development. This, however, came with challenges in inference, since many of these newly proposed designs come without respective methods for statistical inference. In this dissertation, we study the estimation methods for oncology phase II group-sequential and adaptive designs in terms of bias and precision, and we propose new estimation methods for a new class of adaptive designs. We then evaluate the consequences, in terms of power, of using estimates from these designs to plan phase III trial. We also study and propose new approaches to adjust these estimates, based on the observed data, before employing them in planning of phase III sample size, in order to reach the desired power. Literature review showed that many estimation methods have been proposed for the classical single-arm two-stage group-sequential designs with a binary endpoint, which are the most commonly used designs in oncology trials of phase II. Simulation studies showed that the uniformly minimum variance unbiased estimator is the best amongst them in terms of bias and mean square error. However, for the adaptive group-sequential designs, these estimation methods have poor performance. Our proposed estimation methods in oncology phase II adaptive designs showed better performance as compared to the naive maximum likelihood estimator. A direct use of estimates from phase II adaptive designs to plan phase III results in underpowered phase III trials. Therefore, adjusting (discounting) these estimates beforehand is necessary. The amount of discounting, however, depends on the estimator, with our proposed estimators requiring less discounting as compared to the naive maximum likelihood estimator. Our proposed adjustment approaches show power improvements, which are similar across different estimators and design scenarios.