Design and analysis of group-randomized trials: A review of recent developments

Abstract

Purpose: This paper will review the design and analysis issues in group-randomized trials and summarize recent developments that affect those issues. Methods: Group-randomized trials involve the allocation of identifiable groups instead of individuals to study conditions; examples include whole communities, worksites, schools, and clinics. Members of those groups are measured to assess the impact of an intervention; examples include residents, employees, students, and patients. Results: The allocation of identifiable groups guarantees some level of correlation among the observations taken from members of the same group. That intraclass correlation violates the assumption of independent errors that underlies most of the analysis methods used in clinical trials. Use of those methods in group-randomized trials can lead the investigators to overstate the significance of their findings, often badly. Much attention has been focused on the group as the proper unit of analysis. However, that is neither a necessary nor a sufficient condition for a valid analysis. Under certain conditions, such an analysis can have a highly inflated Type I error rate. Under other conditions, an analysis that ignores the group can have the nominal 5% Type I error rate. Conclusions: Attention should be focused on careful matching of the analytic model to the underlying structure of the data, as dictated by the design of the trial. The analytic model will need to reflect all measurable sources of variation as well as the pattern of variation both between and within groups and members. Specific recommendations are presented in the paper.