Skip to content
Back to search results

Cluster Randomized Controlled Trial — Group-Level Randomization

Introduced by
Cornfield (1978); systematised by Donner and colleagues (1980s)
Year
1978–1980s
Type
Experimental design
Family
Experimental design
Font size

What does it do?

A cluster randomized controlled trial (cluster RCT) is an experimental design in which intact social or organisational groups — such as schools, clinics, villages, or workplaces — are randomly assigned to treatment conditions rather than individual participants.

Outcomes are still measured at the individual level, but the unit of randomization is the cluster. This design is essential when an intervention is delivered to whole groups, when there is a risk of contamination between participants in the same setting, or when individual randomization is logistically or ethically impractical.

The intuition behind the method

Imagine testing a new teaching method in schools. You cannot randomize individual pupils within the same classroom — the teacher would inevitably mix methods, and children would share strategies. Instead you randomly assign whole schools to receive the new method or continue with the usual curriculum. Pupils within the same school tend to be more similar to each other than to pupils in other schools (they share the same teacher, culture, and resources), so analyses must account for this clustering. The design sacrifices some statistical power compared to individual randomization, but it reflects how educational and public-health interventions actually work in the real world.

History

The first formal statistical treatment of intra-cluster correlation in group-randomized designs was published by Jerome Cornfield in 1978, who warned that trials randomizing groups but analysing individuals would be systematically anti-conservative. Allan Donner and colleagues developed the analytical framework through the 1980s and published the field's foundational textbook in 2000. The CONSORT extension for cluster RCTs, first published by Campbell and colleagues in 2004 and updated in 2012, established reporting standards that are now required by major journals. Cluster RCTs became especially prominent in global health and education research during the 1990s and 2000s as large-scale community intervention trials demonstrated that contamination control was essential for valid evaluation of public-health programmes.

When to use it

Use a cluster RCT when the intervention is inherently delivered to a group (e.g., a policy, a training programme, a community campaign), when contamination between individually randomized participants in the same setting is likely and would bias the treatment effect estimate, or when individual randomization would be logistically impossible or ethically unacceptable. The design requires a sufficient number of clusters — at minimum 6 per arm, and ideally 20 or more per arm — to provide adequate power and support robust inference. Do NOT use a cluster RCT when individual randomization is feasible and there is no contamination risk, because the cluster design reduces power and complicates analysis. Also avoid it when the number of available clusters is very small (fewer than 6 per arm), as estimation of the ICC and cluster-level variance will be unreliable and standard errors will be underestimated even with clustering-corrected analyses.

Limitations and considerations

  • Reduced statistical power compared to individual randomization for the same total number of participants; more participants are needed to compensate for the design effect.
  • Requires a meaningful number of clusters (typically 20+ per arm) for adequate power and reliable ICC estimation; trials with few clusters are underpowered and analytically fragile.
  • Logistically complex: coordinating randomization, intervention delivery, and data collection across multiple sites requires substantial infrastructure.
  • Cluster dropout (an entire school or clinic withdrawing) can severely compromise power and introduce bias that individual dropout does not.
  • The ICC must be estimated or assumed a priori for sample size calculations; an incorrect ICC assumption leads to an over- or under-powered study.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Prevents contamination between conditions when the intervention operates at a group level, preserving the integrity of the comparison.
  • Enables evaluation of interventions that are logistically or ethically impossible to deliver at the individual level.
  • Supports generalisation to real-world delivery contexts — interventions are tested in the settings where they will ultimately be deployed.
  • Stratified or matched cluster randomization can achieve good covariate balance even with a modest number of clusters.
  • When properly analysed, provides unbiased estimates of intervention effects that account for within-cluster correlation.

Applications

  • Evaluating school-based interventions such as anti-bullying programmes or new curricula, where the classroom or school is the unit of delivery.
  • Testing community-level public-health campaigns (e.g., HIV prevention, sanitation, vaccination outreach) randomized at the village or district level.
  • Assessing quality-improvement initiatives in healthcare settings, with GP practices, clinics, or hospital wards as the randomized clusters.
  • Evaluating workplace wellness or safety programmes delivered to whole teams or departments.
  • Testing policy or governance reforms in social services where the administrative unit (e.g., a local authority) is the natural cluster.

How it works

How it works

1

Identify the naturally occurring or administratively defined groups that will serve as the unit of randomization — for example, schools, GP practices, villages, or hospital wards. Clusters should be as similar as possible in size and relevant baseline characteristics. Stratification or pair-matching of clusters before randomization is strongly recommended to improve balance and power.

2

Standard sample-size formulas for individually randomized trials systematically underestimate the required number of participants because observations within the same cluster are correlated. The design effect (DEFF) inflates the required sample size: DEFF = 1 + (m − 1) × ICC, where m is the average cluster size and ICC (intraclass correlation coefficient) is the proportion of total variance attributable to between-cluster differences. Both the number of clusters and the number of participants per cluster must be planned explicitly.

Formula
n = n_{ind} \times (1 + (m-1)\rho)
3

Clusters — not individuals — are randomly allocated to treatment or control arms. Randomization may be simple, stratified by cluster-level covariates (e.g., urban/rural, baseline performance), or pair-matched. Because the number of clusters is typically small (10–100), stratified or matched randomization substantially reduces the chance of imbalance on key confounders.

4

The intervention is administered at the cluster level (e.g., a training programme delivered to all staff in a clinic). Procedures must prevent cross-contamination — leakage of the intervention from treated to control clusters — which is often the primary rationale for choosing this design over individual randomization.

5

Individual-level outcomes are collected from participants nested within each cluster. Baseline characteristics and outcomes should be recorded for all consenting individuals within sampled clusters. Loss to follow-up at both the cluster and the individual level must be tracked and reported separately, as cluster dropout is especially damaging to power and validity.

6

Standard regression or t-test analyses that ignore clustering produce standard errors that are too small, inflating Type I error. Valid analyses include: (1) cluster-level summaries analysed as the unit (analysis of cluster means); (2) individual-level mixed-effects models with a random intercept for cluster; (3) generalised estimating equations (GEE) with a cluster-level working correlation structure. The ICC estimated from the data should be reported to inform future trials.

Formula
\text{SE}_{adj} = \text{SE}_{naive} \times \sqrt{1 + (m-1)\rho}

Common pitfalls

  • Analysing as if participants were individually randomized — ignoring the clustering structure produces standard errors that are far too small and inflates Type I error rates, sometimes dramatically. Always use a clustering-aware analytic method.
  • Too few clusters — trials with only 4–6 clusters per arm have very low power and may produce misleading ICC estimates; consider whether the study is worth conducting at all without more clusters.
  • Unequal cluster sizes without adjustment — large variation in cluster size reduces efficiency and complicates analysis; account for this in sample-size planning and use weighted or mixed-model analyses.
  • Failing to report the ICC — without the observed ICC, subsequent researchers cannot use the trial to inform sample-size calculations for future cluster RCTs in the same domain.
  • Randomizing after recruitment — selecting participants before cluster randomization and allowing them to know their assignment creates allocation concealment failure and selection bias.

Frequently asked questions

How is a cluster RCT different from a standard RCT?

In a standard RCT, individual participants are randomly assigned to treatment or control. In a cluster RCT, entire groups (clusters) — such as schools or clinics — are randomly assigned. Outcomes are still measured on individuals, but the unit of randomization and the primary source of variance for power calculations is the cluster, not the person. This distinction fundamentally changes sample size requirements and the correct analytic approach.

What is the intraclass correlation coefficient (ICC) and why does it matter?

The ICC (rho) measures the proportion of total outcome variance that is attributable to differences between clusters rather than differences between individuals within clusters. Even a small ICC (e.g., 0.05) can greatly inflate the required sample size. The design effect formula DEFF = 1 + (m − 1) × ICC shows that with 50 participants per cluster and ICC = 0.05, you need 3.45 times as many participants as an individually randomized trial would require.

How many clusters do I need?

A common minimum is around 6 clusters per arm for a valid trial, but this provides very low power for most realistic ICCs and effect sizes. Twenty or more clusters per arm is a widely cited practical target for reliable inference. Sample size should be calculated formally using the expected ICC, the average cluster size, the desired power, and the anticipated effect size. Software such as PASS, Stata's clustersampsi, or R packages clusterPower or CRTpowerPack can assist.

What analysis method should I use?

The two main approaches are: (1) cluster-level analysis — compute a summary statistic (e.g., mean outcome) for each cluster and analyse these summaries with a t-test or ANCOVA, treating the cluster as the unit; and (2) individual-level mixed-effects regression with a random intercept for cluster, or GEE with an exchangeable working correlation. Mixed-effects models and GEE are preferred when covariates need adjustment or when cluster sizes are unequal. All approaches must account for clustering — analyses that ignore it are invalid.

When is contamination a real concern?

Contamination is most serious when participants in control clusters could realistically access or adopt the treatment — for example, if intervention participants share the same physical space, social network, or information channel as control participants. Classic examples: a motivational intervention delivered to some nurses in a ward, where other nurses observe and imitate it; or a health-education leaflet trial where leaflets circulate between households in the same neighbourhood. If contamination is unlikely (e.g., a surgical technique randomized across hospitals that do not share staff), individual randomization is usually preferable.

Related methods

Sources

  1. Donner, A., & Klar, N. (2000). Design and Analysis of Cluster Randomization Trials in Health Research. Arnold.

    ISBN: 978-0340652978
    Copied.
  2. Hayes, R. J., & Moulton, L. H. (2017). Cluster Randomised Trials (2nd ed.). CRC Press / Chapman & Hall.

    ISBN: 978-1498728225
    Copied.