Cluster trials in critical care: half do not say why they clustered

Design

systematic review of cluster randomised trials, MEDLINE and Embase, 2004-2022

Population

102 cluster randomised trials conducted in critical care settings

Primary outcome

ethical and methodological reporting features

Effect

46% justified clustering; 23% reported provider consent where providers were participants; 13% of trials with under 40 clusters used a small-sample correction

Cluster randomisation is how system-level questions in critical care get tested - a protocol, a checklist, a staffing model - because randomising individual patients to a unit-wide practice is often impossible. A systematic review of 102 such trials published between 2004 and 2022 examined how well they handled the ethical and statistical problems the design creates.

The gaps are specific. A justification for using cluster randomisation at all appeared in 46%. Patients or caregivers were identified as research participants in 97%, and of those 75% reported on consent, most commonly waived. Healthcare providers were identified as participants in 47% of trials, but only 23% of those reported anything about provider consent. On the statistical side, 80% reported a power calculation and 82% of those accounted for intracluster correlation, but among the 85 trials with fewer than 40 clusters only 13% applied a small-sample correction.

That last figure is the one that should affect how you read a paper. With few clusters, standard analyses understate uncertainty, and confidence intervals come out narrower than they should be. Eighty-seven per cent of small cluster trials in this field did not correct for it - so when a stepped-wedge trial of a new ICU bundle reports a tight interval around a modest benefit, the interval is probably not that tight.

• When reading a cluster trial, check the number of clusters before the number of patients.
• With fewer than 40 clusters, look for a small-sample correction; assume the interval is too narrow without one.
• Check whether intracluster correlation was accounted for in both the power calculation and the analysis.
• Note who was treated as a participant: provider consent is rarely addressed even when providers are the ones being randomised.
• A missing justification for clustering is a reason to ask whether individual randomisation was actually feasible.

Why it matters

Much of what changes ICU practice arrives as a cluster trial, and the commonest defect makes those trials look more certain than they are.

Don't overread it

This describes reporting quality, not whether the trials' conclusions were wrong - poor reporting of a method is not proof it was not done.

The statistics, in plain English

In a cluster trial the unit of randomisation is the ward, not the patient, so the effective sample size is closer to the number of wards than the number of patients. Intracluster correlation quantifies how similar patients within a cluster are; ignoring it inflates apparent precision. With fewer than about 40 clusters, even correct methods need a small-sample adjustment, and without one the reported interval is systematically too narrow and the p value too small.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.