The use of urn models in response-adaptive randomized designs: a simulation study

Recently, response-adaptive designs have been proposed in randomized trials to achieve ethical and cost advantages by using sequential accrual information collected during the trial to dynamically update the probabilities of treatment assignments. In this context, urn models - where the probability to assign patients to treatments is interpreted as the proportion of balls of different colors available in a virtual urn - have been used as response- adaptive randomization rules [1]. We propose the use of Randomly Reinforced Urn (RRU) models in a simulation study based on a randomized clinical trial on the efficacy of home enteral nutrition in cancer patients after major gastrointestinal surgery. We compare results (number of patients allocated to the inferior treatment and empirical power of the t-test for the treatment coefficient) obtained with the RRU design with those previously published with the non-adaptive approach. In detail, we simulate 10,000 trials based on the RRU model in three set-ups of different total sample sizes. For each sample size, in approximately 75% of the simulation runs, the number of patients allocated to the inferior treatment by the RRU design is lower. The empirical power of the t-test for the treatment effect is similar in the two designs.