Interacting reinforced urn models for the design and analysis of Basket clinical trials
In Basket trials, one or more treatments are tested across multiple types of diseases that share a common characteristic (biomarker/genetic feature/molecular alteration) linked to the therapeutic effect of the treatment. For example, in oncological trials, often multiple types of cancer share a common biomarker and the idea is to identify a therapy that can treat cancer based on the genetic/molecular alteration, regardless of the tumor type. In this kind of study, the sample size is limited and there is a strong ethical imperative of assigning more patients to the better treatment. The basket trial we consider in this work is a randomized sequential experiment for comparing two treatments with binary responses. According to the disease type, when patients are enrolled, they are naturally divided into subgroups. The objective is to randomly force the assignments toward the best-performing treatment based on the already accrued data on its performance in (i) the patient-specific subgroup and (ii) the other subgroups, as treatments’ behavior may be similar across the subgroups. The idea is to exploit this borrowing of information across the groups at the early stage of the study, when just a few patients have been enrolled and, as the trial evolves, give more weight to information coming from the patient-specific group. The design of this trial can be modeled with an interacting reinforced urn model where at each treatment assignment, the urns’ composition is updated based on the patient subgroup, treatment assigned, and observed response. The resulting procedure falls into the class of Covariate-Adjusted Response-Adaptive randomization.
Area: CS40 - Recent developments for urn models (Andrea Ghiglietti and Giacomo Aletti)
Keywords: clinical trials, adaptive experiment, ethics, Covariate-Adjusted Response-Adaptive
Il paper è coperto da copyright.