Sequential Designs

Sequential designs are applied in situations where the objective consists in early termination of a trial, either for futility or overwhelming efficacy. In planning such trials, it is critical to specify the purpose, frequency of analysis, and procedures to be applied to control Type I error probabilities. Some of the strategies require prospective definitions of both the number of interim analyses and the amount of information to be observed. The choice of a particular strategy also depends on the intended purpose and operational considerations. For example, the O’Brien-Fleming approach tends to require strong evidence for early stopping (O’Brien and Fleming 1979); while Pocock’s method tends to lead to more frequent early stopping

(Pocock 1977). On the other hand, the general alpha-spending approach provides flexibility about the number and timing of interim looks, since it mainly depends on specification of a function for how the Type I error probability is spent (Lan and DeMets 1983). However, even this approach may result in inflation of Type I error if the timing of interim analysis is based on knowledge of the observed treatment effects.

As mentioned earlier, when using group sequential designs, estimates of treatment effects are known to be biased, and confidence intervals do not have the desired nominal levels. Therefore, appropriate adjustment should be made to the estimates when reporting such results (Jennison and Turnbull 1999; Wassmer and Brannath 2016). Further, independent monitoring of the study is essential to safeguard the integrity of the trial and credibility of the results (see, e.g., Ellenberg et al. 2002).

Group sequential designs may be more efficient than sample-size reestimation designs; however, depending on the circumstances, samplesize reestimation could still be the preferred design (Levin et al. 2013). The choice of the sequential design may depend on the stage of development and prior knowledge and expectation of the treatment effect.

Adaptive Designs for Dose and Treatment Selection

Modifications could also be made to treatment arms, including doses of the same agent, based on interim results. In early-phase trials, adaptive dose-ranging trials may be conducted to select an optimal dose or doses for further evaluation. An example is the Continual Reassessment Method (CRM), a model-based approach discussed in Chapter 1, which involves fitting a dose-toxicity curve to estimate the maximum tolerated dose for a new drug (Le Tourneau et al. 2009).

In late-phase drug development, an adaptive dose-modification design may allow interim selection of candidate doses from two or more competing candidates, thereby permitting assignment of future patients to the selected treatment arms. Seamless Phase II/III trials are sometimes used as a viable option when there is interest to shorten time-to-market of a new medicine, by addressing two objectives in the same study; namely, dose selection, at an interim analysis, and efficacy determination, at the end of the study. In such designs, suitable procedures should be used to pool information from patients enrolled before and after the adaptation to perform inference at the final analysis, while controlling the Type I error. Examples of approaches include the combination test techniques described earlier, or the adaptive Dunnett design (see, e.g., Stallard and Todd 2010).

In general, the choice of a specific approach requires a careful evaluation of several factors, including trial objective, outcome measure, and other therapeutic and practical considerations.

Adaptive Enrichment Designs

Adaptive enrichment designs allow modifications to the patient population based on comparative interim results. The final analysis may involve multiple hypotheses; thereby requiring appropriate procedures to adjust for multiplicity (see, e.g., Wassmer and Brannath 2016).

Enrichment designs have especially become an attractive option in oncology, as the focus of drug development shifted from those centered on cytotoxic agents (i.e., those agents that may stop cancer cells from dividing and growing or cause tumors to shrink) to those using molecularly targeted agents (which target specific molecular markers) or use a patients immune response to attack cancer cells. To implement such designs, it is essential to ensure that the molecular marker is well established (i.e., is strongly correlated with the outcome measure) and that a diagnostic tool for evaluating it is available. Further, it should be ascertained that the study drug is not effective in the marker-negative patients. If the latter is not known, the randomization should be stratified by molecular marker positivity or negativity or use more complex designs with a common protocol (Simon 2017a). When enrichment designs are used in seamless Phase II/III trials, the required sample size and analysis for each phase may be evaluated using the same or different endpoints. For example, assuming progression-free survival for end of Phase II and overall survival for Phase III, the corresponding analyses may be performed after the following numbers of events E_p (p = I or II) have been observed:

( Z_{a +Zp

J ^{1 u}p Pp k ^log^vp

where two-sided significance level a and higher power 1-p are typically used for Phase III, for detecting the respective hazard ratio Vp (Schoenfeld 1983).