Share

Cluster Unit Randomized Trials

3. Statistical Implications

Statistical Implications of Cluster Randomization

A key feature of cluster randomization trials is that while randomization is at the cluster level, statistical analyses are usually conducted at the individual level.

This discordance between the unit of randomization and the unit of analysis, an issue not usually dealt with in standard statistical texts, creates special methodological challenges at every stage of the trial. These challenges arise essentially because individuals in the same cluster tend to respond more similarly than individuals in different clusters, i.e. the assumption of statistical independence required for the application of standard statistical methods is now violated. Thus the outcome measure is now characterized by two separate sources of variation, one within clusters and the other between clusters.

Within-cluster dependencies may arise from several different sources:

  • Subject self-selection is one important factor, as when female patients choose female physicians, or when individuals with respiratory problems choose to live in dry weather communities.
  • External factors may also be relevant, as when differences in temperature among nurseries are related to infection rates, or when differences in smoking bylaws influence the success of smoking cessation programs.
  • Finally, a variety of internal factors may also lead to between-cluster variation, particularly when individuals respond similarly to an intervention that is provided in a group setting.