03 · What You Need to Know
What Determines the Appropriate Unit of Randomization?
Individual and cluster randomization assign interventions at different levels
In an individually randomized trial, each eligible participant is independently assigned to an intervention or comparison condition, subject to the chosen allocation procedure. Two students attending the same school, for example, might receive different interventions.
In a cluster-randomized trial, intact groups are allocated to conditions. All eligible participants within a randomized classroom, clinic, school, or other defined group are ordinarily exposed to the condition assigned to that cluster, although actual participation and adherence may vary.
The cluster serving as the unit of randomization must be distinguished from the unit at which outcomes are measured. A school may be randomized while student achievement remains the outcome of interest.
| Design consideration |
Individual randomization |
Cluster randomization |
| Allocation unit |
Individual participant |
Intact group or cluster |
| Typical intervention |
Individually delivered treatment or support |
Classroom practice, institutional policy, or group-wide program |
| Contamination |
May occur between differently allocated participants |
May reduce within-group contamination but does not eliminate spillover between groups |
| Statistical efficiency |
Often greater for a comparable total number of participants |
Usually reduced by within-cluster correlation |
| Analysis |
Accounts for the individual allocation design |
Must account for clustering and the number of independent randomized units |
Group-level interventions may make individual randomization inappropriate
Some interventions cannot meaningfully be delivered to selected individuals without changing their nature. Examples include a school-wide attendance policy, a hospital infection-control protocol, a community sanitation program, or a classroom teaching approach that reorganizes whole-class instruction.
Consider a professional development program intended to change how teachers plan and deliver lessons. If the intervention requires the same teacher to use one instructional approach with selected students and another with their classmates, individual allocation could create an artificial implementation arrangement. Randomizing teachers or classrooms may better preserve the intervention as intended.
However, the appropriate cluster level still requires justification. If teachers independently deliver instruction to separate classes, classroom randomization might be feasible. If teachers routinely share materials and jointly implement the intervention, school-level randomization may be more defensible.
Contamination can justify cluster allocation, but its likelihood matters
Contamination occurs when participants or providers assigned to one condition are exposed to elements intended for another condition. In educational trials, students may share intervention materials, teachers may apply newly learned practices to comparison students, or staff may adopt procedures introduced in another treatment arm.
Cluster allocation can reduce some forms of contamination by keeping participants who interact closely under the same assigned condition. Yet contamination does not automatically justify cluster randomization. Researchers should examine how likely spillover is, whether it could materially change the estimated effect, and whether practical safeguards could reduce it.
Furthermore, spillover is not always merely an implementation failure. It may be a substantive feature of the intervention. A program designed to spread teaching practices through professional networks, for example, may require an evaluation that explicitly considers indirect effects.
Interference between participants can change the causal question
Conventional individually randomized analyses often rely on assumptions that one participant's assigned treatment does not affect another participant's outcome. This is commonly discussed under the broader framework of interference and the stable unit treatment value assumption.
In social and educational settings, that assumption can be difficult to maintain. A peer-learning intervention may influence classroom interactions, including those involving students who did not directly receive the intervention.
Cluster randomization can help when interference is largely contained within groups, but it does not guarantee independence between clusters. Researchers must still consider relationships across classrooms, schools, organizations, or geographic areas.
Administrative convenience is not sufficient justification by itself
Institutions may prefer to assign whole schools because distributing interventions individually requires additional coordination. Convenience can be relevant to feasibility, but it should not be mistaken for methodological necessity.
If an intervention can be delivered independently, contamination is limited, and individual allocation is ethically and operationally acceptable, individual randomization may provide a more efficient design. Researchers should not adopt cluster randomization merely because participants are already organized into groups.
Cluster randomization changes sample-size requirements
Participants within a cluster may share teachers, organizational practices, resources, environmental conditions, or demographic characteristics. Their outcomes can therefore be correlated.
This correlation is commonly summarized by the intracluster correlation coefficient (ICC). A positive ICC indicates that outcomes from members of the same cluster are more similar, on average, than outcomes from members of different clusters under the relevant model.
For a simplified equal-size cluster design, a frequently used approximation to the design effect is:
This is a planning approximation, not a universal sample-size formula. Unequal cluster sizes, covariate adjustment, baseline measurements, outcome type, allocation ratio, and the number of clusters can change the calculation. A design effect of 2.20 does not mean that 2.20 times as many clusters or participants will necessarily guarantee adequate power.
Consequently, sample-size planning for cluster-randomized trials must explicitly address both the number of clusters and the number of participants within them.
Recruitment and allocation timing can introduce bias
Randomizing clusters before identifying or recruiting individual participants may be unavoidable in some settings. However, recruiters who know a cluster's allocation may consciously or unconsciously influence who enters the trial. Participants may also be more willing to enroll in one condition than another.
Where feasible, researchers should identify eligible participants before allocation, use standardized recruitment procedures, and limit the influence of allocation knowledge on enrollment and outcome assessment.
When recruitment must occur afterward, post-randomization recruitment procedures need explicit safeguards and transparent reporting.
The causal estimand should inform the choice
An estimand specifies the treatment effect the study intends to estimate, including the population, intervention conditions, outcome, and handling of relevant events or differences in exposure.
Researchers should distinguish the effect of offering an intervention to an entire institution from the effect of delivering it to a selected individual. These may be substantively different questions.
For example, evaluating the institutional adoption of an AI-supported teaching policy may require measuring its effects on the school environment, teacher practices, and student outcomes. Randomizing individual students might answer a narrower question about personal access to the technology, but not the effect of institutional adoption.
Cluster randomization should therefore follow the intervention's causal structure rather than merely the administrative structure of the available sample.
04 · A Practical Example
Evaluating a Classroom-Wide Teaching Intervention
Hypothetical Example
A collaborative learning program across secondary schools
A research team wants to evaluate a structured collaborative learning approach in 24 secondary schools. The program requires teachers to reorganize classroom activities, establish shared group tasks, and assess students through collaborative exercises. The primary outcome is individual student achievement.
The researchers initially consider randomly assigning students within each classroom. They realize that the instructional approach necessarily changes interactions among all students in the class. Individual allocation would be difficult to implement without changing the intervention.
Identify the intervention unit. The teaching approach operates at the classroom level because instructional routines and peer interactions are shared.
Assess contamination. Teachers may exchange lesson materials and instructional practices within schools, potentially contaminating comparison classrooms.
Select the cluster level. The team chooses school-level randomization after determining that within-school contamination would be difficult to control. Twelve schools are allocated to each condition.
Plan recruitment and analysis. Student eligibility criteria are established before allocation. The analysis will account for students nested within schools and adjust for prespecified baseline characteristics.
Evaluate feasibility. The team conducts a power assessment using plausible ICC values and the available number of schools. It does not assume that a large student count compensates for having only 24 randomized schools.
Interpretation: School-level randomization is defensible because the intervention affects shared teaching practices and contamination between classrooms is plausible. Nevertheless, classroom-level randomization might have been more efficient if teachers could implement conditions independently and spillover were limited.
The decision requires evidence about how schools actually operate, not merely the assumption that schools are natural clusters.
06 · What This Means for You
Choose the Smallest Defensible Randomization Unit
When designing a randomized evaluation, begin by identifying how the intervention operates and whether independently allocated participants could realistically receive different conditions.
Then assess whether contamination, interference, organizational requirements, or ethical considerations make individual allocation unsuitable. If several clustering levels are feasible, compare them in terms of implementation fidelity, statistical efficiency, recruitment bias, and interpretability.
A practical decision framework
If the intervention can be delivered independently and spillover is limited
Prefer evaluating individual randomization unless another substantive reason favors group allocation.
If the intervention changes shared practices or institutional conditions
Consider randomizing the group at which those practices or conditions operate.
If contamination would seriously compromise an individually randomized comparison
Evaluate cluster randomization and alternative contamination-control strategies.
If only a small number of clusters is available
Assess power, baseline imbalance, and analytical limitations before committing to the design.
For educational technology studies, this distinction is especially consequential. Individual access to a learning application and institution-wide adoption of an AI-supported instructional policy are different interventions. Their appropriate randomization units may therefore differ even when the same student-level outcome is measured.
07 · A Quick Checklist
Before Deciding to Randomize Groups, Check These Conditions
Before finalizing the randomization unit, check:
Specify the intervention, comparison condition, target population, outcome, and causal estimand.
Identify the level at which the intervention is delivered and the level at which its effects operate.
Determine whether individual allocation is operationally and ethically feasible.
Assess plausible contamination and interference pathways within and between groups.
Compare feasible cluster levels rather than assuming the largest organizational unit is best.
Estimate the available number of clusters, expected cluster sizes, and plausible ICC values.
Plan eligibility assessment, recruitment, and consent procedures to minimize allocation-related selection bias.
Specify an analysis that accounts for clustering and prespecified covariates.
Document the methodological justification for the chosen randomization unit in the protocol.
