In cluster-randomized trials, groups of individuals rather than individuals are randomized to different interventions. Cluster-randomized trials are also known as group-randomized trials. We say the ‘unit of allocation’ is the cluster, or the group. The groups may be, for example, schools, villages, medical practices or families. Such trials may be done for one of several reasons. It may be to evaluate the group effect of an intervention, for example herd-immunity of a vaccine. It may be to avoid ‘contamination’ across interventions when trial participants are managed within the same setting, for example in a trial evaluating a dietary intervention, families rather than individuals may be randomized. A cluster-randomized design may be used simply for convenience.
One of the main consequences of a cluster design is that participants within any one cluster often tend to respond in a similar manner, and thus their data can no longer be assumed to be independent of one another. Many of these studies, however, are incorrectly analysed as though the unit of allocation had been the individual participants. This is often referred to as a ‘unit-of-analysis error’ (Whiting-O'Keefe 1984) because the unit of analysis is different from the unit of allocation. If the clustering is ignored and cluster trials are analysed as if individuals had been randomized, resulting P values will be artificially small. This can result in false positive conclusions that the intervention had an effect. In the context of a meta-analysis, studies in which clustering has been ignored will have overly narrow confidence intervals and will receive more weight than is appropriate in a meta-analysis. This situation can also arise if participants are allocated to interventions that are then applied to parts of them (for example, to both eyes or to several teeth), or if repeated observations are made on a participant. If the analysis is by the individual units (for example, each tooth or each observation) without taking into account that the data are clustered within participants, then a unit-of-analysis error can occur.
There are several useful sources of information on cluster-randomized trials (Murray 1995, Donner 2000). A detailed discussion of incorporating cluster-randomized trials in a meta-analysis is available (Donner 2002), as is a more technical treatment of the problem (Donner 2001). Special considerations for analysis of standardized mean differences from cluster-randomized trials are discussed by White and Thomas (White 2005).