Cluster RCT

Intact groups (clusters) are randomised to intervention or control, and everyone within a cluster gets the same allocation. Used when the intervention acts at group level (a ward protocol, a staff-training package) or to stop contamination between arms. Because people within a cluster are alike, the effective sample size is smaller than the headcount, and the analysis must account for clustering.

A trial flow where the unit of randomisation is the cluster, not the patient: hospitals are split into arms, and all the patients within a hospital carry that hospital’s allocation through to an analysis that explicitly allows for clustering.

Notice that randomisation happens one level up. Patients in the same cluster resemble each other (same staff, case-mix, protocols), so they supply less independent information than the same number of unrelated individuals. This shrinkage is the design effect, driven by the intracluster correlation coefficient (ICC).

A cluster design avoids contamination and can test group-level interventions, but it buys that at a cost: it needs more participants, the analysis must be clustered, and with only a few clusters, chance imbalance between arms is hard to exclude. Get any of these wrong and the trial over-states its certainty.

• Unit of randomisation = the cluster, not the patient
• Design effect = 1 + (m − 1) × ICC inflates sample size
• Analysis MUST account for clustering (else CIs too narrow)

REDUCE MRSA (NEJM 2013) — a pragmatic cluster RCT randomising 43 hospitals (74 adult ICUs, 74,256 patients) to screening-and-isolation, targeted decolonisation, or universal decolonisation. Hospitals, not patients, were the unit of randomisation. Universal decolonisation cut all-pathogen bloodstream infections most (HR 0.56, about a 44% reduction). A ward-level protocol like ICU-wide chlorhexidine bathing cannot be randomised patient-by-patient without contamination — everyone in the unit shares the environment — which is exactly why the cluster was the hospital/ICU.

• Ignoring the design effect / ICC — powering or analysing as if patients were independent.
• Recruitment (identification) bias — enrolling patients after the cluster’s allocation is known.
• Too few clusters — baseline imbalance between arms cannot be excluded by chance.