Cluster randomized trial
Why this level?
Participants or clusters are randomly allocated to concurrent comparison groups. NHMRC 2009 sourceRandomisation still happens; it is clusters rather than people that are allocated, which does not change what the design can support. What it changes is the analysis: ignore the clustering and the interval comes out too narrow.
What question does this help answer?
It suits interventions that act on a group rather than a person: a ward protocol, a clinic workflow, a staff training programme, anything a single clinician could not deliver to one patient and withhold from the next.
When to use it
Use it when individual allocation is impossible or would contaminate the comparison, because staff cannot easily un-learn a protocol for the next patient. It is also the honest choice when the intervention is genuinely delivered at the group level, since pretending otherwise misstates what was randomized.
When not to use it
Avoid it when individual randomization is feasible, because clustering costs statistical efficiency for nothing in return. It is also a poor fit when only a handful of clusters are available: randomizing six wards leaves baseline imbalance that individual randomization would have smoothed out, and inference on few clusters is unreliable.
How it works in plain language
Clusters are randomized, then outcomes are measured on the people inside them. Because people in the same cluster resemble each other, each additional person contributes less new information than an unrelated person would. The design must be sized for that from the start, and the analysis must account for it, or the confidence intervals will be too narrow and the p-values too small.
What randomizing clusters looks like
Why Mallard may recommend it
Mallard may recommend a cluster randomized trial when the intervention is delivered to a group, when contamination between individuals in the same setting is likely, or when the investigator describes randomizing sites rather than patients.
Key assumptions and design requirements
- The number of CLUSTERS, not the number of people, drives the precision of the comparison.
- An intracluster correlation coefficient is stated and justified, and the sample size accounts for it.
- The analysis models the clustering (a random effect, or cluster-robust standard errors), or explains why it does not.
- With fewer than about 30 clusters, a small-sample correction is named rather than assumed unnecessary.
- Recruitment and consent procedures are described at both the cluster and the individual level.
What to watch out for
Ignoring the clustering inflates false positives: people in the same ward or clinic are more alike than people picked at random, so the effective sample size is closer to the number of clusters than the number of people. Randomizing few clusters also leaves baseline imbalance that individual randomization would have smoothed out.
Clinical example
Question: Does an enhanced hand-hygiene programme reduce hospital-acquired infection? Design: Randomize entire wards to the programme or usual practice, then measure infection among admitted patients, sizing the study on the number of wards and the expected within-ward correlation rather than on the patient count alone.
- Consort 2010 statement: extension to cluster randomised trials (opens in a new tab) · BMJ · doi 10.1136/bmj.e5661
- Overview of clinical study designs (opens in a new tab) · Clin Exp Emerg Med · doi 10.15441/ceem.23.036