gms | German Medical Science

The scientific shift towards a more patient-individualized treatment approach has led to the implementation of so-called basket trials. These novel clinical trial designs incorporate several subgroups denoted as baskets, each defined by a common characteristic. Additionally, all patients in the trial share one specific characteristic across all baskets which is used as a pathway by the investigated treatment. In practice, the prevalence of the disease and the common characteristic are often low and, hence, the recruitment of patients is challenging. Consequently, an efficient use of the available resources is required to gain as much information as possible. Basket trials address this issue by allowing to share information across baskets and to inactivate baskets midcourse during the trial.

Several different basket designs have been proposed in the literature applying different methods for the assessment of early futility and basket efficacy at the end of the trial. The outcome of interest for each patient is usually a binary response which is aggregated to basket-wise response rates. Based on the observed data, there are several options how to make a decision for stopping enrolment of a particular basket due to futility. Cunanan´s frequentist design [1] requires a pre-specified number of responses at the interim futility assessment in order to continue enrolment to the basket. This rule can be considered as a simplification of a one-sided binomial test, which is used to assess efficacy. The designs using Bayesian hierarchical models base their futility assessments and efficacy evaluations on the posterior distribution of the response rate [2]. The different approaches raise the question whether they come to the same or different conclusions regarding the assessment of futility and efficacy.

We present a general comparison of decision-making based on one-sided binomial tests and on posterior distributions of the beta-binomial model. The performance of the different decision rules is evaluated in terms of type I error rate and power. We show that the Bayesian decision rule ensures type I error rate control for any given significance level, sample size, and treatment effect, and also attains the power achieved by the classical frequentist test. Furthermore, it can be shown analytically that complete agreement between the two approaches can be induced by appropriate choice of design parameters. Therefore, it can be concluded that frequentist and Bayesian decision rules are two sides of the same coin and can be embedded into a joint decision-theoretic framework.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.