Cluster Unit Randomization Trials

17. Reporting

Reporting of Cluster Randomization Trials

The well-known CONSORT statement for individually randomized trials (Begg et al., 1996; Moher et al., 2001; Altman et al., 2001) has now been extended to cluster randomized trials (Campbell et al., 2004). The principle features of this extension include recommendations to:

• Provide the rationale for adopting a cluster design;
• Specify how the effects of clustering were incorporated into the sample size calculation and the statistical analysis; and
• Present a chart showing the flow of both clusters and individuals through the trial.

An earlier set of guidelines were provided by Donner and Klar, 2000, Chapter 9. Aside from reporting standards that are unique to CRTs, there are some that have become routinely accepted for individually randomized trials, but now need to be reconsidered. This includes the presentation of baseline characteristics, which for CRTs should be provided separately for cluster level characteristics (e.g., geographic area, cluster size) and individual level characteristics (e.g., age, gender). The presentation of baseline cluster level characteristics is straightforward, since the clusters assigned to each group are independently distributed.

Some special caution is required when comparing individual level baseline characteristics.

• Although it is now recognized that the use of significance tests for this purpose is always a logically flawed procedure (e.g., Senn, 1994), this practice can be particularly misleading when applied to CRTs. This is because the test procedures typically used, such as t-tests and chi-square tests, may fail to account for the clustering effects that apply at baseline as well as at outcome. The resulting p-values will be biased downwards, potentially leading to an ill-advised decision to adjust for the characteristic (covariate) in question in the statistical analysis.
• Standard deviations for continuous variables that are used for descriptive purposes will also be biased downwards by clustering effects, but only slightly unless the overall sample size is small and the intracluster correlation coefficient is large (White et al., 2005), conditions unlikely to apply in most CRTs.
• Finally it must be recognized that the effective sample size for the variables involved is no longer the number of individuals n per treatment group but rather n/VIF. Failure to recognize this makes it difficult to accurately compare the amount of information provided by different trials.