Response adaptive randomization

When two or more treatments are being compared in a clinical trial, information accrues on the treatments as the trial proceeds. However, there may be evidence suggesting that some treatments are more promising than others, and consequently, there may be interest in applying response adaptive randomization, which uses the current data to assign more patients to the treatments that are performing better thus far in the trial. In this way, patients have a higher probability of being assigned to these treatments. Examples of this type of randomization are given, such as urn models and sequential maximum likelihood estimation rules, and some of their properties are described. The issue of inference after trials that use response adaptive randomization is then discussed. For example, this approach can lead to a loss in power when using standard tests and the problem of estimation can be considerably more complicated than in a trial that uses complete randomization. The use of stopping rules is also addressed, since, in practice, there may be interest in stopping a trial early if there is convincing evidence of a treatment difference or less promising treatments may be dropped from additional consideration. An indication is given of possible future work.