Adaptive designs in the time to event setting: The potential for benefit and risk

Thesis (Ph.D.)--University of Washington, 2016-06

Group sequential designs (GSDs) have been the standard sequential approach to maintain scientific, ethical, and efficiency goals in any confirmatory Phase III studies. Over the past two decades, adaptive extensions to group sequential designs have been proposed to allow more flexible modification of aspects of the trial. However, such use of unblinded estimates computed from accruing clinical trial data has been viewed extremely cautiously by regulatory agencies such as the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA). In their Guidance to Industry on adaptive designs, the FDA has distinguished adaptive designs that are ``less well understood'' from the ``well understood'' GSDs. There is thus much interest in characterizing potential benefits (e.g., efficiency, flexibility), as well as the potential for harm (e.g., inflation of statistical error rates, introduction of operational bias), when using adaptive designs. Randomized clinical trials (RCTs) involving censored time to event data are of particular interest, as the aspects of adaptive designs that are ``less well-understood'' in that setting may have as much to do with how well we understand standard censored data models as with the general properties of unblinded adaptation. A major focus of any sequential procedure is the appropriate control of statistical operating characteristics, including the type 1 error. A common requirement of all commonly used sequential methods is the proper characterization of the growth of statistical information about the parameter of greatest interest. When sequentially analyzing time to event data, the censoring distribution can have great impact. It can affect both the choice of the distributional parameter used to summarize treatment effect (e.g., 5 year survival, median survival, hazard ratios) and the growth of statistical information over time. Further, efficiency of inference might need to consider not only the number of subjects accrued to the study, but also factors related to time as measured by both the typical time of patients on study (``study time''), as well as the calendar time needed from the start of accrual until the final analysis is performed. In this research, we investigate how these issues of information growth and time may impact (a) scientific interpretation, and (b) statistical credibility (control of type 1 error and study precision). In the first part of the research, we focus on the proportional hazards setting wherein issues of (1) calendar time and (2) information growth are separable. We first investigate the efficiency of the adaptive weighting scheme as a consequence of changing the timing of the adaptation in prevention trials with potentially low background rates (either as a consequence of overestimating the event rate and/or high treatment efficacy). Noting that GSDs are better able to avoid any operational bias that might be introduced by the more flexible forms of adaptive designs, we compare our ability to preserve study precision solely through the use of blinded adaptations within prespecified GSDs versus the use of an unblinded adaptation which might better distinguish between low event rates versus extreme treatment effects. We next investigate how poor understanding of information growth (in the weighted logrank statistics) can impact the ability of adaptive procedures to preserve the overall Type 1 error. We examined scenarios whereby simply changing the censoring distribution can directly impact the ability of adaptive procedures to preserve the overall Type 1 error. We provide some recommendations from our findings. In the second part of our research we consider settings in which we cannot presume a parametric or strongly semiparametric probability model, for instance, when crossing survival curves are plausible.Under the weak null/non proportional hazards setting, calendar time and information growth are no longer separable. We investigate the degree to which the use of ``less well-understood'' statistics in presence of time varying treatment effect and censoring as induced either sequentially or accrual affects the degree to which we can control the probability of rejecting the null hypothesis when we may be concerned with the weak null.