The monitoring of interim results in clinical trials should be planned in advance, preferably with a limited number of interim analyses focusing on only one or two specified endpoints. In this regard, various group sequential designs developed in recent years have helped considerably in the design and interpretation of randomized trials. All such designs require more stringent p-values in order to stop the trial, and this article will discuss the relative merits of various stopping rules. Unfortunately, many trials have interim analyses which were not planned in advance. Such unplanned analyses may be “data provoked, ”for example, investigators request formal analysis after informal inspection already suggests a treatment difference. This can greatly increase the type-I error, and recommendations will be made on how to cope with unplanned interim analyses. For trials that stop early, even with a formal stopping rule, the magnitude of the treatment difference (and its confidence limits) will tend to be exaggerated. This bias in the reporting of trial results needs wider recognition, and authors should routinely report on whether publication is prompted by the significance of an interim analysis. Other problems with interim analyses concern trials with more than two treatments, how to cope with multiple endpoints, and the extent to which prognostic factor adustments should be made. There is a continuing need to relate statistical methodology to the reality of monitoring clinical trials, so that clearly defined analysis policies can lead to objective unbiased decision-making and reporting of trial findings.
