Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Epidemiology / Evaluating interventions

Cluster, stepped-wedge and pragmatic trials

13 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Explain when randomising groups rather than individuals is appropriate

  • Describe parallel cluster, stepped-wedge and cluster-crossover designs

  • Distinguish explanatory from pragmatic trials

Some interventions act on groups, not individuals: a community health worker programme, school feeding, water systems, mass drug administration, a clinic-level protocol. Or individual randomisation risks contamination: if neighbours in the same village are in different arms, those in the control arm may learn about the intervention. In such cases the cluster (village, school, clinic) is randomised.

Parallel-group cluster randomised trials

  • Entire clusters are allocated to intervention or control.
  • The analysis must respect clustering: summarise at cluster level or use mixed models/GEE.
  • Sample size needs the design effect (1 + (m − 1)ρ; see biostatistics chapter 9).
  • Few clusters (under about 15 per arm) can give poor balance; use restriction or stratified randomisation.
  • Identification and recruitment bias: if participants are recruited after clusters know their assignment, the people enrolled may differ by arm. Prefer recruiting everyone in the cluster, or recruiting before allocation.
  • Consent may be needed at cluster level (gatekeepers) and individual level.

Stepped-wedge designs

All clusters start in the control condition and cross to the intervention at randomly assigned times, until all receive it. Measurements are taken in each period. Useful when the intervention is believed to be beneficial and withholding it permanently is unacceptable, or when it can only be rolled out in stages.

A stepped-wedge schedule

Six clinics are randomised to start a new TB screening protocol in different quarters. 'C' = control, 'I' = intervention.

Clinic pairQ1Q2Q3Q4
1 and 2CIII
3 and 4CCII
5 and 6CCCI

Each clinic acts as its own control. The comparison is between clinic-quarters under intervention and control, adjusting for calendar time (because other changes happen over time and are confounded with the rollout). Weaknesses: longer duration, more complex analysis, carry-over of learning, risk that secular trends are mistaken for effects.

Other designs

  • Cluster-crossover: each cluster receives both conditions in sequence, with a washout.
  • Factorial: test two or more interventions at once (A, B, A+B, neither) to examine each and their interaction.
  • Adaptive: pre-planned changes (dropping arms, re-estimating sample size) based on accumulating data.
  • Non-inferiority: show a new option is not worse than standard care by more than a margin.

Explanatory versus pragmatic trials

Explanatory ('Can it work?')Pragmatic ('Does it work in practice?')
ParticipantsHighly selectedTypical patients
SettingSpecialist centreRoutine care
DeliveryClosely supervisedBy usual staff
ComparatorPlaceboUsual care
Follow-upIntensiveRoutine records
AnalysisEfficacyEffectiveness (ITT)

The PRECIS-2 tool scores a trial on nine domains from explanatory to pragmatic. Neither is better; they answer different questions. Policy-makers need pragmatic evidence; mechanism research needs explanatory trials.

Individually randomisedParallel clusterStepped wedgeGreatest statisticalefficiencyRisk of contamination forgroup interventionsPrevents contaminationNeeds more participants(design effect)Analysis by clusterEveryone eventually getsthe interventionEfficient with fewclustersTime effects must bemodelled
Match the randomisation unit and schedule to the intervention.

Reporting

Use the CONSORT extension for cluster randomised trials and for stepped-wedge designs. Report the number of clusters and participants per cluster, the intracluster correlation, and how clustering was handled in the analysis and sample size.

Typical mistakes

Calculating sample size as if individuals were independent; analysing individual-level data without accounting for clusters; forgetting to account for time in a stepped wedge; comparing only a few clusters in each arm and then attributing a difference to the intervention when it is due to cluster-level differences.

Summary

  • Randomise clusters when the intervention acts on groups or contamination is likely.

  • Cluster trials need larger samples and analysis that accounts for clustering.

  • Stepped-wedge trials roll out the intervention to all clusters in random order; time must be modelled.

  • Explanatory trials test efficacy under ideal conditions; pragmatic trials test effectiveness in routine care.

  • Follow the CONSORT extensions for cluster and stepped-wedge trials.

Check your understanding

Question 1
1

A health education programme is delivered at schools; the best unit of randomisation is probably:

Question 2
2

In a stepped wedge, all clusters:

Question 3
3

A key analytic requirement of stepped-wedge trials is to:

Question 4
4

A trial in routine clinics using usual staff and usual care as comparator is:

Question 5
5

Why can few clusters be a problem?

Question 6
6

Contamination in an individually randomised trial means:

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

The randomised controlled trialNatural experiments and quasi-experimental evaluations