Cluster randomised trial
- Study design
Whole groups (schools, clinics, neighbourhoods) are randomised instead of people, because the intervention is delivered to the group.
What it is
In a cluster randomised trial whole groups of people are randomly allocated to intervention or control. It is used when the intervention acts on the group (training a clinic's staff, changing a school's water supply) or when randomising individuals would contaminate the control group. Because people in the same group resemble each other, each adds less information, and the sample size must be inflated by the design effect.
How it is built
- Clusters (schools, clinics, municipalities) are chosen and their size measured.
- The individual sample size is computed and multiplied by the design effect.
- Clusters are randomised, often by strata or in matched pairs.
- People within each cluster are recruited and measured.
- The analysis uses methods for clustered data (multilevel models or generalised estimating equations).
Worked example
- School handwashing programme; outcome: absence due to diarrhoea. An individually randomised trial would need 600 pupils.
- 50 pupils per school are measured (m = 50) and an ICC of 0.02 is expected.
- Design effect = 1 + (50 − 1) × 0.02 = 1 + 0.98 = 1.98.
- Pupils needed = 600 × 1.98 = 1 188; schools = 1 188 / 50 = 23.8, rounded up to 24 (12 per arm).
- Reading: because of clustering, almost twice as many pupils are needed as in an individually randomised trial.
Randomising people or clusters
| Aspect | Individual | Cluster |
|---|---|---|
| Randomised unit | Person | Group (school, clinic, neighbourhood) |
| Contamination of controls | Possible if they share a place | Minimal |
| Sample size | Smaller | Larger (design effect) |
| Balance between arms | Good with many participants | Doubtful with few clusters |
| Analysis | Standard | Must account for clustering |
Markers
- ICC
- Design effect
- Stepped wedge
When to use it
- Educational interventions in schools, health-worker training and changes to how a service is organised.
- Vaccination, safe water, bed nets or other measures that also protect neighbours.
- The stepped-wedge design brings the intervention to every cluster, in random order and in stages.
Common pitfalls
- Analysing people as if each had been randomised individually: confidence intervals too narrow.
- Using very few clusters (say, 2 per arm): chance cannot balance the differences between them.
- Recruiting people after knowing their centre's arm, without concealment: selection bias.
Test yourself
Think of the answer, then tap the question to see it.
With 20 people per cluster and an ICC of 0.05, what is the design effect?
1 + (20 − 1) × 0.05 = 1 + 0.95 = 1.95: almost twice as many participants are needed.
Why randomise clusters rather than people?
Because the intervention acts on the group, or because within the same place the control group would be contaminated by the intervention.
What happens if clustering is ignored in the analysis?
The standard error is underestimated: confidence intervals come out too narrow and p values too small.
To gain power, is it better to add clusters or people per cluster?
Add clusters: with an ICC above zero, each extra person in a cluster adds less and less.
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: Chapter 23: Including variants on randomized trials (Cochrane, 2024); Consort 2010 statement: extension to cluster randomised trials (EQUATOR Network, 2012); CONSORT 2025 Statement: updated guideline for reporting randomised trials (EQUATOR Network, 2025).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.