Cluster Randomized Field Experiment
Also known as: CRFE, cluster-randomized trial in the field, group-randomized field experiment, community-randomized field experiment
A cluster randomized field experiment (CRFE) assigns intact groups — schools, villages, clinics, workplaces — rather than individuals to treatment or control conditions, and the experiment is conducted in real-world settings rather than a laboratory. Randomization at the group level controls for contamination between conditions while preserving the ecological validity of the natural environment. It is the dominant design for evaluating community-level, school-based, or workplace interventions in public health, education policy, and development economics.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a cluster randomized field experiment when: (1) the intervention is delivered at the group level (e.g., a policy, a training program, a mass-media campaign) and individual randomization is logistically or ethically impossible; (2) contamination between treatment and control units would invalidate individual randomization; (3) ecological validity — knowing whether the intervention works under real-world conditions — is the research priority. Do NOT use when: you have fewer than ~20 clusters per arm (power will be critically low); the outcome of interest is primarily an individual-level biological or cognitive measure that could be studied without risk of contamination (a standard RCT is simpler and more powerful); or budget and logistics cannot support field deployment across multiple sites.
Strengths & limitations
- High ecological validity — results reflect real-world implementation rather than artificial lab conditions.
- Prevents treatment contamination when individuals within the same cluster interact or share resources.
- Supports evaluation of group-level or policy-level interventions that cannot be assigned individually.
- Random assignment of clusters (when sufficient clusters are available) supports causal inference about treatment effects.
- Compatible with stratified or matched designs that improve balance on key cluster-level confounders.
- Statistical power is substantially lower than an individually randomized trial of the same total sample size, because of intracluster correlation; more clusters are needed to compensate.
- The cluster — not the individual — is the unit of inference for power calculations, requiring many clusters (typically 20+ per arm), which is expensive.
- Fidelity monitoring across geographically dispersed field sites is resource-intensive and results may vary by site.
- Partial non-compliance and cluster-level attrition can threaten internal validity and complicate intent-to-treat analysis.
Frequently asked
How many clusters do I need?
At minimum 10–15 per arm for any meaningful power, but 20–30 or more per arm is strongly recommended. The required number of clusters is determined by the ICC, mean cluster size, expected effect size, and desired power. Software such as Optimal Design or the R package clusterPower can compute this. When clusters are scarce, consider a stratified or matched-pair design to improve balance.
What is the intracluster correlation coefficient (ICC) and why does it matter?
The ICC measures how similar individuals within the same cluster are relative to individuals across clusters. An ICC of 0 means clustering is irrelevant; even an ICC as small as 0.01–0.05 can dramatically reduce effective sample size. The ICC must be estimated from pilot data or prior literature and used in the power calculation via the design effect formula DEFF = 1 + (m-1)*ICC.
Can I use multilevel modeling or do I need GEE?
Both approaches are valid. Multilevel (mixed-effects) models provide cluster-specific (conditional) estimates and handle unequal cluster sizes naturally. GEE provides population-averaged (marginal) estimates with cluster-robust standard errors. The choice depends on whether you want to describe the average effect across all individuals (GEE) or model variation between clusters (multilevel). Either is preferable to ignoring clustering entirely.
How is this different from a standard cluster-randomized controlled trial (CRT)?
A standard cluster RCT may be conducted in clinical or institutional settings with tight protocol control. A cluster randomized field experiment specifically emphasizes real-world field conditions — the intervention is delivered in naturalistic settings with the attendant variability in implementation, context, and population. The emphasis on ecological validity and external generalizability is the distinguishing feature.
What if clusters drop out entirely after randomization?
Cluster-level dropout is the most serious threat to internal validity in CRFEs because it can reintroduce confounding at the cluster level. Unlike individual-level attrition, it cannot be handled by standard intent-to-treat analysis alone. Sensitivity analyses, inverse-probability weighting at the cluster level, and transparent reporting of which clusters dropped out and why are recommended.
Sources
- Murray, D. M. (1998). Design and Analysis of Group-Randomized Trials. Oxford University Press. ISBN: 978-0195120424
- Hayes, R. J., & Moulton, L. H. (2017). Cluster Randomised Trials (2nd ed.). CRC Press. ISBN: 978-1498728225
How to cite this page
ScholarGate. (2026, June 3). Cluster Randomized Field Experiment. ScholarGate. https://scholargate.app/en/experimental-design/cluster-randomized-field-experiment
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cluster Randomized Controlled TrialExperimental design↔ compare
- Factorial Field ExperimentExperimental design↔ compare
- Field ExperimentExperimental design↔ compare
- Randomized Controlled TrialExperimental design↔ compare