Hypothesis Testing and Estimation in Adaptive Designs with Blinded and Unblinded Interim Analysis

The design of clinical trials requires to make assumptions on numerous parameters that may have an impact on the chance of success. Examples are the variability of outcomes, the effect sizes and safety profiles of investigated treatments, or the influence of covariates. Often, little information on these parameters is available a-priori and it is difficult to obtain reliable estimates. Clinical trial designs that allow to estimate these unknown parameters from interim data and to adapt the remainder of the trial accordingly, have, therefore, raised increasing interest in recent years. A wide range of adaptations, including sample size reassessment, and selection of treatments and subgroups, has been considered so far. For the application of adaptive designs in confirmative clinical trials, preservation of the integrity of the trial is pivotal. Of special concern is the control of the type I error rate and the impact of adaptations on effect estimates. An essential tool to maintain the integrity of controlled randomized clinical trials is blinding: if the treatment allocations of individual patients remain unknown to patients and investigators, potential biases, as, e.g., a biased evaluation of endpoints, can be avoided. Trial adaptations based on blinded interim data (i.e., without revealing the individual treatment allocations) are, therefore, perceived to have much less impact on the integrity of a trial, compared to unblinded interim analyses. Blinded interim analyses prohibit a direct estimation of treatment effects. Nuisance parameters like the variability of endpoints or overall event rates, however, can be easily estimated and can be used as basis for sample size reassessment. The project has two main goals. On the one hand, we will investigate to what extent the common perception, that adaptations based on blinded data reviews have only marginal impact on the properties of standard single stage testing and estimation procedures, is justified. We want to extend previous research on blinded sample size reassessment to a wider range of clinical trial settings. This will include multiarmed trials and trials with several endpoints. Furthermore, we will consider unrestricted adaptation rules based on blinded data to assess their potential impact on statistical inference procedures that do not explicitly account for the adaptations. Besides adaptive clinical trials, we will investigate trials where the choice of the analysis strategy is based on a blinded review of the data. The second major goal of the project is to develop novel testing procedures and confidence intervals that strictly control the type I error rate and coverage probability, respectively. In a first step we will consider adaptive designs with unblinded interim analysis. Such tests are equally applicable to designs with blinded interim analysis. We will improve these tests for the special setting where the adaptations are based only on blinded data and derive corresponding confidence intervals. In a second step we will extend on our previous work on single stage multiple testing procedures based on weighted directed graphs and generalize it to adaptive multiple tests controlling the family-wise error rate. This methodology allows to derive adaptive gatekeeping procedures that have applications in clinical trials with multiple treatment arms, subgroups and/or endpoints where, e.g., treatment groups are selected at interim. Additionally, we plan to derive simultaneous confidence intervals corresponding to the investigated adaptive multiple testing procedures.

In traditional clinical trials with the objective to demonstrate efficacy of new therapies, patients are often recruited for several years but the results are analysed only at the end. Adaptive trial designs in contrast, allow one to perform interim analyses and to adapt the remainder of the trial accordingly. For example, the number of patients included in a trial can be adapted or treatment arms with apparently little or no treatment effect can be dropped. An important field of application is also the development of personalized therapies. Here, by a modification of the inclusion criteria, adaptive designs allow one to exclude groups of patients from a clinical trial that apparently do not benefit from a therapy. For the application of adaptive designs in confirmative clinical trials, the preservation of the integrity of the trial is pivotal. Of special concern is the control of the false positive rate of statistical hypotheses tests and the impact of adaptations on the unbiasedness of effect estimates. In the first part of this project we assessed the robustness of adaptive designs and explored scenarios where things can go wrong, leading to biased results and unfounded claims of treatment effects. We investigated to what extent the common perception that adaptations based on blinded data reviews, where the individual treatment assignments are not revealed, have only marginal impact on the properties of standard single stage testing and estimation procedures, is indeed justified. Especially, we gave examples where blinded adaptations may lead to substantially biased results and derived conditions when adaptations based on blinded interim data have no negative influence on the validity of the statistical analysis. In a second step, we developed novel, robust analysis methods for adaptive clinical trial designs with blinded and unblinded interim analyses that allow for a valid interpretation of study results in spite of data dependent trial adaptations. We especially focused on clinical trials with multiple objectives, where several treatment arms or patient populations are investigated. Another emphasis was clinical trials in oncology, where the survival times of patients are analysed. The project results extend the possibilities for designing, analysing and assessing adaptive clinical trials and have a high importance for their efficient implementation for the authorisation of new therapies by regulatory authorities.