03 · What You Need to Know
How the Unit of Randomization Changes Trial Design
What Is the Unit of Randomization?
The unit of randomization is the entity assigned by a random mechanism to an intervention condition. It may be an individual, classroom, hospital, village, institution, or defined period of intervention delivery.
It is important to distinguish three concepts:
Unit of randomization
The entity assigned to a treatment condition by the randomization procedure.
Unit of intervention delivery
The entity or organizational level at which the intervention is implemented.
Unit of observation or analysis
The entity on which outcomes are measured or the level at which statistical comparisons are performed.
These units can differ. A school may be randomized, teachers may deliver the intervention, and student outcomes may be measured individually. The statistical analysis must account for the resulting dependence among observations.
The fundamental decision concerns what is actually assigned by randomization, not simply how many people are measured.
When Should You Randomize Individual Participants?
Individual randomization assigns each eligible participant independently to one of the trial conditions, subject to any prespecified restrictions such as blocking or stratification.
It is generally attractive when participants can receive different interventions without substantially influencing one another's treatment or outcomes.
For example, researchers testing two versions of an individually delivered digital tutoring application might randomly assign students to either version if access can be separated and meaningful spillover is unlikely.
Individual randomization often provides greater statistical efficiency than cluster randomization with the same number of participants because treatment assignment varies at the individual level. Nevertheless, participants may still be nested within classrooms or institutions, and relevant baseline clustering may need to be considered in the analysis.
Its main vulnerability arises when the assigned conditions cannot remain meaningfully separate. Teachers may share instructional materials, students may exchange access credentials, or treatment delivery may change the behavior of people assigned to the control condition.
Such contamination can reduce the contrast between conditions. Its direction and consequences depend on the mechanism, so it should not automatically be assumed to bias every estimate toward zero.
When Should You Randomize Groups or Clusters?
Cluster randomization assigns intact groups, such as classrooms, clinical practices, departments, or communities, rather than separately assigning their members.
It is especially relevant when the intervention operates through group-level practices or shared environments.
Examples include a classroom-wide teaching strategy, a hospital infection-control protocol, a community health campaign, or a departmental management policy.
Randomizing individual members within these settings may be impractical or inconsistent with the intended intervention. A teacher implementing a new classroom discussion method, for instance, may be unable to deliver two genuinely separate instructional environments to students participating in the same discussion.
Cluster randomization can reduce some forms of contamination, but it does not eliminate spillover between clusters. Teachers may collaborate across classes, patients may receive care at multiple facilities, and communities may interact.
When Is Site Randomization Appropriate?
Site randomization is a form of cluster randomization in which entire institutions or implementation locations are assigned to trial conditions. Sites may include schools, universities, hospitals, clinics, or workplaces.
It is useful when the intervention requires organizational authority, infrastructure changes, or site-wide implementation.
Consider a university policy that changes assessment procedures across all departments. Randomizing individual students would not reproduce the organizational intervention. Assigning entire universities might be more defensible, provided enough institutions can participate.
However, site randomization may be statistically challenging because the number of sites is often small and institutions can differ substantially in baseline characteristics. Randomization does not guarantee comparable groups in any particular realized allocation.
Researchers should consider prespecified restrictions, stratification, or covariate adjustment when appropriate, without assuming these procedures can compensate for having too few independently randomized sites.
When Should You Randomize Time Periods?
Time-period randomization assigns intervention conditions or intervention schedules according to randomly determined periods. It can be useful when an intervention is delivered through a shared system and conditions can change over time.
One possibility is a cluster crossover design, in which participating sites receive different interventions during different periods according to randomized sequences. Another is a stepped-wedge cluster randomized trial, in which clusters transition from control to intervention at randomly assigned times.
These designs answer different questions and involve different assumptions. A cluster crossover design generally requires careful attention to carryover effects and the reversibility of the intervention. A stepped-wedge design requires explicit modeling of calendar-time effects because intervention exposure is associated with later periods.
Randomizing time periods is not equivalent to measuring outcomes repeatedly. A simple before-and-after study does not become randomized merely because it contains several measurement occasions.
Time-based designs may be appropriate when implementation must occur in phases, but they are not automatically superior to parallel-group trials. Secular trends, seasonality, implementation learning, and changing participant composition may complicate interpretation.
How Do the Four Options Compare?
| Randomization unit |
Typical application |
Main advantage |
Principal concern |
| Individual |
Separately delivered treatments or learning applications |
Often greater statistical efficiency |
Contamination or interference between participants |
| Group |
Classroom teaching methods or team interventions |
Matches collective delivery |
Within-cluster correlation and limited clusters |
| Site |
Institutional policies or facility-wide interventions |
Allows organization-level implementation |
Site heterogeneity and recruitment feasibility |
| Time period |
Switchable interventions or staggered implementation |
May accommodate operational sequencing |
Time trends, carryover, and complex correlation |
Why Does Cluster Randomization Usually Require More Participants?
Participants within the same cluster often resemble one another. Students in a classroom may share a teacher, curriculum, and peer environment. Patients within a hospital may receive similar clinical practices.
This similarity is represented by the intracluster correlation coefficient, commonly abbreviated ICC. Positive intracluster correlation reduces the amount of independent information contributed by additional participants within the same cluster.
For a simplified design with equally sized clusters, a commonly used approximation is:
This is an approximation, not a complete sample-size calculation. Unequal cluster sizes, the number of clusters, baseline adjustment, repeated measurements, and the intended analysis can change the required sample size. A large number of participants distributed across very few clusters may still provide inadequate precision.
Why Must the Analysis Reflect the Randomization Structure?
If classrooms are randomized, treating every student as though independently assigned can underestimate standard errors and inflate apparent precision. This is a form of unit-of-analysis error.
Appropriate approaches may include mixed-effects models, generalized estimating equations with suitable small-sample corrections, cluster-level analyses, or randomization-based methods. Their suitability depends on the outcome, number of clusters, design, and estimand.
Time-based designs may additionally require period effects, within-site correlation structures, and explicit handling of carryover or time-varying confounding.
The analysis plan should therefore be developed alongside the randomization plan, not reconstructed after the data have been collected.
What Ethical and Recruitment Issues Arise?
Cluster trials can introduce distinctive consent and selection problems. In some settings, an organization may authorize implementation of a policy, but this does not automatically replace individual consent where consent is ethically or legally required.
Recruiting participants after clusters have been allocated may also introduce selection bias if recruiters or participants know the assigned condition and this knowledge influences participation.
Researchers should consult the applicable research ethics committee, clarify which activities occur at the cluster and individual levels, and establish procedures that minimize allocation-related recruitment differences.
04 · A Practical Example
Choosing Randomization for an AI-Assisted Teaching Intervention
Hypothetical Example
Twenty Classes, Six Hundred Students, and One Instructional Intervention
A research team wants to determine whether an AI-assisted formative feedback approach improves student writing performance compared with conventional feedback. Twenty classes, each containing approximately 30 students, are available across several schools.
The intervention includes teacher training, revised feedback procedures, and classroom discussions about AI-generated suggestions.
Option A: Randomize students.
Individual assignment could be efficient, but students in the same classroom would encounter a shared teacher and discussion environment. Separating the instructional approaches might be difficult.
Option B: Randomize classrooms.
Ten classes could be allocated to each condition. This better matches classroom-level implementation, although the analysis must account for clustering and the number of classes may constrain precision.
Option C: Randomize schools.
This might be preferable if teachers routinely share instructional practices across classes or if school-wide training is required. However, the available number of schools must be sufficient to support credible estimation.
Option D: Randomize implementation periods.
A staggered or crossover design might be considered if implementation can be scheduled and the assumptions concerning time trends, learning, and carryover are defensible.
For this hypothetical intervention, classroom randomization is a plausible initial choice because treatment delivery is collective. It is not automatically the final choice. Researchers would still need to assess teacher overlap, school-level spillover, available cluster counts, anticipated ICC, and feasibility.
Suppose the design uses 20 classes with 30 students each and an assumed ICC of 0.05. The approximate design effect is 2.45, yielding an effective sample-size approximation of 245 rather than 600. This illustrates why the number of students alone cannot establish adequate statistical power.
A defensible protocol would report the number of randomized classes, recruitment timing, assignment procedure, cluster-adjusted sample-size assumptions, primary outcome, and analysis model.
