Increasingly, epidemiologists are faced with the need to evaluate the impact of an intervention that is delivered at the level of a community or cluster of individuals, rather than at the individual level. This has profound implications for the design and interpretation of a study to evaluate its impact. We start by discussing the issues arising in the extension of the randomized double-blind controlled trial methodology to the evaluation of interventions delivered to clusters of individuals, or to whole communities, where the unit of randomization is a cluster of individuals rather than an individual. We then consider alternative approaches to design, discuss their relative strengths and weaknesses and present a framework of design options. Finally we propose a pragmatic approach to evaluation design in this setting. We believe that the answer lies in the judicious selection of different design elements, combined in such a way that when the evidence from each is presented together, a clear picture of the impact of the intervention emerges. We illustrate this using an example from the recent literature.