03 · What You Need to Know
How Clusters Are Defined and Distinguished From Other Research Units
The unit of randomization determines what counts as a cluster
The defining feature of a cluster-randomized trial is that allocation occurs at a level above the individual participant, or at a defined grouping level whose members receive a common assigned condition.
For example, if 30 classrooms are randomly allocated to two teaching methods, the classrooms are the randomized clusters. Students within those classrooms are participants, not additional randomized clusters.
If the same study instead randomly allocates ten schools, the schools become the clusters, even if outcomes are measured for students in multiple classrooms within each school.
This distinction is central to the CONSORT extension for cluster-randomized trials, which emphasizes identifying and reporting the unit of randomization and the multilevel structure of the study.
Unit of randomization
The unit assigned to an intervention condition through a random allocation procedure.
Unit of observation
The unit from which a particular measurement is obtained, such as a student, teacher, classroom, or school.
Unit of analysis
The level at which observations are represented or summarized in the statistical analysis. Individual-level and cluster-level analyses may both be possible, provided the randomization and dependence structure are handled correctly.
What kinds of groups can qualify as clusters?
There is no requirement that a cluster must be a formal institution or a geographically separate organization. Its suitability depends on whether it can be clearly identified, randomized, and meaningfully assigned to the intended condition.
| Possible cluster |
Illustrative intervention |
Possible outcome measurement |
| Classroom |
Collaborative teaching strategy |
Individual student achievement |
| School |
School-wide attendance initiative |
Student attendance records |
| Teacher |
Teacher-delivered instructional intervention |
Learning outcomes among students taught by each teacher |
| Hospital ward |
Ward-level patient safety protocol |
Patient-level adverse events |
| Clinic |
Clinical practice improvement program |
Patient care outcomes |
| Workplace |
Organization-wide occupational program |
Employee-level outcomes |
| Household |
Household-level behavioral intervention |
Individual household member outcomes |
| Community |
Community-wide public health campaign |
Resident-level outcomes |
These are possible examples rather than automatic classifications. A school is not necessarily a randomized cluster merely because its students participate in a trial. The study must actually allocate schools as units.
Cluster boundaries must be explicit
Researchers should specify which individuals, organizational units, or locations belong to each cluster. Boundaries can be administrative, geographical, relational, or operational, depending on the intervention.
A classroom usually has a defined student roster, but membership can become less straightforward when students attend different classes for different subjects. A teacher may teach several classes, while students may receive instruction from multiple teachers. Similarly, patients may move between hospital wards during a trial.
These circumstances do not necessarily invalidate cluster randomization. They do, however, require clear rules for determining cluster membership, intervention exposure, outcome attribution, and possible cross-cluster movement.
For a school-based intervention, a protocol might define the cluster as an entire school operating under one administration during a specified academic year. For a classroom-level intervention, it might define the cluster as a particular scheduled class section taught by a designated teacher.
The definition should be established before randomization whenever feasible, rather than adjusted afterward to accommodate unexpected findings.
A cluster can contain multiple levels of participants
Cluster-randomized trials often have hierarchical data structures. Students may be nested within classrooms, which are nested within schools. Patients may be nested within wards, which are nested within hospitals.
Consider a trial that randomizes schools while collecting outcomes from students in different classrooms. The school is the randomized cluster, but classroom-level similarities may also affect student outcomes.
The statistical model may need to account for more than one level of dependence. Simply identifying schools as clusters does not establish that every other source of correlation can be ignored.
In some studies, the data structure is cross-classified rather than strictly nested. For example, students may interact with multiple teachers, and teachers may work across different classes or schools. Researchers should distinguish the assignment structure from the structure generating outcome dependence.
Cluster membership does not require every member to receive the intervention
Randomization assigns a condition to the cluster. It does not guarantee that every eligible person within that cluster will participate, adhere to the intervention, or receive the same exposure.
A school may be assigned to implement a professional development program, but some teachers may not attend every session. A clinic may adopt a new care protocol, while some patients receive care that differs from the intended procedure.
The cluster remains the unit of randomization. Differences in participation, adherence, and exposure must be addressed through the study's estimand, implementation assessment, and analysis plan.
Researchers should not redefine clusters according to who ultimately complied with the intervention. Doing so could compromise the protection provided by random allocation.
Clusters are not the same as sampling clusters
Cluster sampling and cluster randomization are distinct procedures, although both involve groups.
In cluster sampling, researchers select groups from a population as part of obtaining a sample. In cluster randomization, researchers assign groups to intervention conditions through random allocation.
A study can use cluster sampling without being a cluster-randomized trial. For example, researchers might randomly select schools for a survey and observe students without assigning any intervention.
Conversely, researchers might recruit a convenience sample of eligible schools and then randomly assign those schools to intervention and comparison conditions. That would involve cluster randomization, although the school sample itself would not be a probability sample of all schools.
Clusters are not statistical clusters discovered through analysis
Statistical cluster analysis groups observations according to similarities in measured characteristics. Such clusters are produced through an analytical procedure rather than necessarily existing as units of intervention allocation.
For example, an educational researcher might identify groups of students with similar learning behaviors using a clustering algorithm. These analytical groups do not become randomized clusters unless the trial actually uses them as the units of random allocation.
The distinction matters because clustering algorithms, cluster sampling, and cluster-randomized trials address different methodological questions despite sharing similar terminology.
Time periods are not automatically clusters
Suppose a hospital uses its conventional protocol in January and a new protocol in February. Alternating conditions by calendar month does not automatically create a cluster-randomized trial.
The defining issue is whether identifiable clusters were randomly allocated under a valid randomization scheme. A design that randomly assigns hospitals to different treatment sequences may qualify as a cluster-randomized crossover trial. Merely alternating interventions over time without random allocation does not.
Some designs also randomize the timing at which clusters switch to an intervention. A stepped-wedge cluster-randomized design is one example. In that arrangement, the clusters remain identifiable groups, while their assigned intervention timing changes across study periods.
Why the number of clusters matters more than the label
Participants within the same cluster often share conditions that make their outcomes statistically dependent. This dependence is commonly summarized using the intracluster correlation coefficient (ICC).
When the ICC is positive, adding more participants to an existing cluster usually provides less independent information than adding participants through additional independently randomized clusters.
This is why determining how many clusters are needed is a separate design problem from determining the total number of participants.
The number of randomized clusters also influences the reliability of variance estimation and statistical inference. Trials with very few clusters can face serious limitations even when the total number of individual observations is large.
Watch Out
Do not report the number of individual participants as though it were the number of independently randomized units. If ten schools are randomized and 2,000 students provide outcome data, the trial has ten randomized clusters, not 2,000 independent allocation units.
Unequal cluster sizes do not change what the clusters are
Clusters do not need identical numbers of members. One school may have 150 eligible students while another has 500. Both can be randomized clusters if schools are the allocation units.
Nevertheless, substantial differences in cluster size may affect statistical efficiency, allocation balance, and the interpretation of treatment effects.
For instance, a participant-average treatment effect gives greater aggregate weight to larger clusters than a cluster-average treatment effect. These estimands can differ when intervention effects vary across clusters of different sizes.
Researchers should therefore specify the target estimand and address unequal cluster sizes in the design and analysis rather than assuming that every cluster contributes identically to every statistical question.
Cluster definitions should reflect the intervention mechanism
Choosing between classrooms, teachers, departments, and schools is not merely a matter of selecting the most convenient organizational label.
The appropriate unit depends on how the intervention operates and how exposure might spread between participants. If teachers routinely share instructional materials across classrooms, randomizing classrooms may create substantial contamination. School-level allocation might then be preferable.
On the other hand, randomizing entire schools when an intervention could be delivered independently to individual students may introduce unnecessary statistical inefficiency.
The central design principle is to justify group randomization according to the intervention and causal question, not simply according to the organizational structure in which participants happen to be found.
04 · A Practical Example
Identifying the Clusters in a School-Based Randomized Trial
Hypothetical Example
An AI-supported instructional intervention
A researcher evaluates an AI-supported formative assessment approach involving 600 students in 20 classrooms distributed across five schools.
Each classroom contains 30 participating students. Ten classrooms are randomly allocated to use the intervention, while ten continue with existing assessment practices.
Student assessment scores are collected at baseline and after implementation. The researcher wants to determine whether the intervention improves student performance.
Step 1: Identify what was randomized.
The random allocation procedure assigned classrooms, not individual students or schools, to the two conditions.
Step 2: Count the randomized clusters.
The trial contains 20 clusters, with ten allocated to each condition.
Step 3: Identify the observational units.
Individual students provide achievement measurements. There are 600 participating students, assuming complete data.
Step 4: Examine the hierarchical structure.
Students are nested within classrooms, and classrooms are located within five schools. The analysis must account for classroom clustering and consider additional school-level dependence.
Step 5: Assess the consequences of the design.
With only five schools, school-level characteristics and within-school relationships may complicate inference. The researchers must evaluate whether classroom-level allocation provides adequate independence and whether contamination between classrooms is manageable.
Interpretation: The trial has 20 randomized classroom clusters and 600 student participants. It does not have 600 independently randomized units, nor does it have five randomized school clusters.
However, the presence of 20 classroom clusters does not by itself establish that the trial has 20 statistically independent sources of information. Shared school-level influences and possible spillover between classrooms may require additional design and analytical safeguards.
If the researchers instead randomly assigned the five schools, the trial would contain five randomized clusters, regardless of how many classrooms or students participated. Such a small number of randomized units would create substantial challenges for balance, power, and reliable inference.
06 · What This Means for You
Define the Cluster Before Designing the Randomization Procedure
When planning a cluster-randomized trial, write an operational definition of the cluster that another researcher could apply consistently. Identify its boundaries, membership rules, allocation process, and relationship to the intervention.
For educational research, do not assume that the school must be the cluster merely because the intervention takes place in a school. Determine whether allocation should occur at the student, teacher, classroom, department, or school level.
The smallest feasible unit may provide greater statistical efficiency, but only if the intervention can be delivered without compromising implementation or introducing unacceptable interference.
A practical framework for defining clusters
If the intervention is delivered to entire classrooms
Consider classrooms as clusters, while assessing teacher overlap and cross-classroom contamination.
If the intervention changes a school-wide policy
Consider schools as clusters because the intervention may not be separable within the institution.
If individuals belong to multiple overlapping groups
Define allocation and exposure rules explicitly, and plan analysis for the resulting dependence structure.
If cluster membership changes during follow-up
Prespecify how entrants, departures, transfers, and repeated measurements will be handled.
If very few clusters are available
Reassess the feasibility of the design before relying on a large participant count to provide statistical power.
When reporting the trial, distinguish the number of clusters randomized, the number of participants recruited or included, and the number contributing outcome data. These counts answer different questions and should not be combined into one ambiguous sample-size statement.
Finally, the chosen cluster definition should remain consistent across the protocol, registration, allocation records, statistical analysis plan, and published report. Any changes should be documented with their timing and justification.
