Sequential methods and group sequential designs for comparative clinical trials

Comparative clinical trials arc performed to assess whether a new treatment has superior efficacy than a placebo or a standard treatment (one-sided formulation) or whether two active treatments have different efficacies (two-sided formulation) in a given population. The reference approach is the single-stage design and the statistical test is performed after inclusion and evaluation of a predetermined sample size. In practice, the single-stage design is sometimes difficult to implement because of ethical concerns and/or economic reasons. Thus, specific early termination procedures have been developed to allow repeated statistical analyses to be performed on accumulating data and stop the trial as soon as the information is sufficient to conclude. Two main different approaches can be used. The first one is derived from strictly sequential methods and includes the sequential probability ratio test and the triangular test. The second one is derived from group sequential designs and includes Peto, Pocock, and O'Brien and Fleming methods, alpha and beta spending functions, and one-parameter boundaries. We review all these methods and describe the bases on which they rely as well as their statistical properties. We also compare these methods and comment on their advantages and drawbacks. We present software packages which are available for the planning, monitoring and analysis of comparative clinical trials with these methods and discuss the practical problems encountered when using them. The latest versions of all these methods can offer substantial sample size reductions when compared with the single-stage design not only in the case of clear efficacy but also in the case of complete lack of efficacy of the new treatment. The software packages make their use quite simple. However, it has to be stressed that using these methods requires efficient logistics with real-time data monitoring and, apart from survival studies or long-term clinical trials with censored endpoints, is most appropriate when the endpoint is obtained quickly when compared with the recruitment rate.