03 · What You Need to Know
How Randomization and Measurement Units Affect Experimental Design
What Is the Unit of Randomization?
The unit of randomization is the entity independently allocated to an intervention or comparison condition through a random assignment procedure.
In an individually randomized trial, participants are assigned separately. In a cluster-randomized trial, intact groups such as classrooms, schools, hospitals, or communities are assigned.
For example, if 30 classrooms are randomly assigned to two instructional conditions, the study contains 30 independently randomized classrooms, regardless of how many students subsequently provide outcome measurements.
The randomization unit establishes the level of independent treatment allocation. It is particularly important for understanding experimental replication and the uncertainty associated with treatment comparisons.
What Is the Unit of Measurement?
The unit of measurement is the entity or occurrence for which a particular outcome, exposure, or characteristic is measured.
For example, a school-level intervention may be evaluated using individual student examination scores. The students are the outcome measurement units, while schools are the randomization units.
Measurement may also occur repeatedly. A patient in a hospital-randomized trial might provide several outcome assessments, creating repeated measurements nested within patients and patients nested within hospitals.
Researchers should distinguish observation and analysis units from the unit of randomization because these concepts address different aspects of the study.
How Do the Different Units Relate?
| Research design |
Randomization unit |
Measurement unit |
| Individual instructional intervention |
Student |
Student |
| Classroom instructional intervention |
Classroom |
Student |
| School-wide educational policy trial |
School |
Student or teacher |
| Hospital quality improvement trial |
Hospital |
Patient |
| Community public health intervention |
Community |
Household or individual |
| Workplace training intervention |
Organization or team |
Employee |
| Within-patient bilateral treatment trial |
Eye, when independently allocated within patient |
Eye |
These examples illustrate possible arrangements. The actual assignment procedure, rather than the study setting alone, determines the randomization unit.
Why Would Researchers Randomize Groups Instead of Individuals?
Group randomization may be appropriate when interventions are delivered collectively, individual allocation is impractical, or participants assigned to different conditions could influence one another.
For example, a teacher may implement an instructional strategy for an entire classroom. Randomizing individual students within that classroom could be operationally difficult and may introduce contamination if students share instructional activities.
Similarly, a hospital-wide safety protocol cannot always be implemented independently for randomly selected patients without affecting the broader clinical environment.
Cluster randomization may therefore align treatment assignment with the actual intervention delivery process.
However, it can require more participants or groups than individual randomization to achieve comparable precision, particularly when outcomes are positively correlated within clusters.
Why Does Group Randomization Create Statistical Dependence?
Individuals within a randomized group share treatment assignment and may experience common contextual influences.
Students within the same classroom may have similar outcomes because of shared teachers, instructional resources, peer interactions, or prior classroom characteristics.
Patients within a hospital may share clinical protocols and organizational conditions.
This similarity is often described using the intracluster correlation coefficient, or ICC.
Ignoring relevant clustering can produce misleading standard errors and confidence intervals, especially when the analysis treats individually measured outcomes as though their treatment assignments were independent.
Researchers should therefore consider clustering during the design stage.
Can Individual-Level Outcomes Still Be Analyzed?
Yes. Cluster randomization does not require researchers to discard individual outcomes or calculate only group averages.
Individual-level analyses can preserve information about participant characteristics and outcome variation.
Appropriate approaches may include mixed-effects models, generalized estimating equations, or other cluster-aware methods.
The choice depends on the estimand, number of randomized clusters, outcome distribution, and assumptions.
For example, a mixed-effects model might include student-level covariates and a classroom-level random intercept. A marginal model might estimate a population-averaged intervention association while accounting for classroom clustering.
Neither approach automatically eliminates confounding introduced by post-randomization selection or other design problems.
What Is the Difference Between the Experimental Unit and the Outcome Unit?
The experimental unit is the smallest entity independently assigned to a treatment condition under the experimental design.
The outcome unit is the entity for which the outcome is defined and measured.
In a classroom-randomized study, the classroom is the experimental unit, while student achievement may be the outcome.
Individual students are legitimate outcome observations, but they are not independent treatment replications.
Confusing these levels can produce pseudoreplication.
Correct identification of the experimental unit is essential when determining how much independent information the randomization provides about the intervention contrast.
How Does the Number of Randomized Clusters Affect Precision?
The number of independent clusters is an important determinant of precision in cluster-randomized trials.
Suppose two studies each measure 600 students. One randomizes 20 classrooms containing 30 students each, while the other randomizes four classrooms containing 150 students each.
Although both contain 600 students, they provide different levels of independent classroom replication.
With positive intracluster correlation, adding participants to existing clusters often produces diminishing gains in precision compared with recruiting additional clusters.
Very small numbers of randomized clusters also complicate statistical inference, even when individual outcome counts are large.
Researchers should evaluate whether sample size should be planned around participants, clusters, or both.
How Is the Cluster Design Effect Calculated?
For equal-size clusters and an exchangeable correlation structure, a common approximation is:
The formula is illustrative rather than universal. Cluster-size variation, covariate adjustment, baseline measurements, intervention allocation, and analytical methods can alter sample size requirements.
Sample size planning should reflect the specific treatment-effect estimand and proposed model.
What Happens When Researchers Ignore the Randomization Unit?
Ignoring the level of randomization may lead to standard errors that are too small, confidence intervals that are too narrow, and unsupported conclusions about treatment effects.
For example, if four schools are randomized but 800 students are measured, an analysis that treats all 800 students as independently randomized exaggerates the independent assignment information.
The consequences depend on the correlation structure and analytical model, but the underlying design problem remains.
This is related to inflating sample size through nonindependent observations.
Correcting variance estimates may help when sufficient independent clusters are available. It cannot create additional randomized clusters or remove confounding when treatment condition is inseparable from a particular group.
Can Randomization and Measurement Occur at the Same Level?
Yes. In individually randomized trials, each participant may be independently assigned and subsequently measured.
However, repeated measurements from the same participant still require appropriate treatment of within-person dependence.
For example, individually randomized students measured at four occasions contribute repeated outcome observations. Their randomization unit is the student, but the measurement occasions are not independent participants.
Researchers should distinguish repeated measurements from independent observations even when assignment occurs individually.
What If Participants Are Recruited After Clusters Are Randomized?
Recruitment after randomization can introduce selection bias if knowledge of allocation affects which individuals enter the study.
For example, teachers who know their classroom received an innovative instructional program may encourage certain students to participate more actively in outcome assessment.
Alternatively, students may be more willing to enroll when they know they will receive the intervention.
These processes can undermine comparability between treatment conditions despite valid initial cluster randomization.
Possible safeguards include identifying eligible participants before randomization, standardized recruitment, and masking recruiters to allocation where feasible.
The distinction between recruitment and analysis units is therefore relevant to the validity of cluster-randomized trials.
What If Cluster Sizes Differ Substantially?
Unequal cluster sizes can affect precision, weighting, and the interpretation of the estimated intervention effect.
For example, one school may contribute 50 students while another contributes 500.
An analysis giving every student equal weight may target a different population quantity from one giving every school equal weight.
Researchers should determine whether the intended estimand concerns the average individual or the average cluster.
When cluster size is related to outcomes or treatment effects, additional analytical considerations may be necessary.
Cluster-size variation should be incorporated into sample size planning and reported transparently.
What If There Are Very Few Randomized Clusters?
Few-cluster trials require particular caution because conventional large-sample statistical approximations may be unreliable.
For example, a trial randomizing only three schools per condition may have limited information for estimating between-school variation.
Cluster-robust standard errors may require small-sample corrections, while some random-effects models may rely heavily on distributional assumptions.
Randomization-based inference or cluster-level analyses may be considered when suitable for the assignment design, but they do not eliminate the limitations created by very few clusters.
Researchers should avoid interpreting a large student count as evidence that the number of randomized schools is sufficient.
Can Treatment Be Assigned Within Participants?
Yes. Some experimental designs assign different treatments to paired organs, anatomical sites, or measurement conditions within the same participant.
For example, one eye may be randomly assigned to one treatment and the fellow eye to another.
The eye may then serve as the treatment assignment and outcome unit, while patient identity creates a paired dependence structure.
Such designs may improve precision by controlling for stable participant characteristics.
However, researchers must consider potential interference between treated sites, including systemic treatment effects.
The relevant experimental unit depends on the actual assignment procedure rather than a universal assumption that the participant must always be randomized as a whole.
How Does the Assignment Level Affect Causal Inference?
The intervention level helps define the causal effect being estimated.
A classroom-level instructional program may operate through teaching practices, peer interactions, and shared learning environments.
Its effect is not necessarily equivalent to the effect of independently assigning the same instructional activity to individual students.
Cluster-level interventions may also produce spillovers among individuals within the same group. Such effects may be part of the intervention rather than a violation of its intended implementation.
Researchers should define whether the estimand concerns the effect of assigning a group to an intervention, the effect of individual exposure, or another explicitly specified causal contrast.
Correct statistical adjustment for clustering does not automatically establish the validity of all causal assumptions.
How Should Randomization and Measurement Units Be Reported?
Methods sections should identify the entity randomized, the allocation procedure, the number of randomized units, and the entities on which outcomes were measured.
For example, a study might report that 24 classrooms were randomized, with student-level achievement measured among 720 students.
Researchers should describe cluster sizes, outcome assessment procedures, and how clustering was addressed analytically.
Where relevant, explain whether participants were identified before or after randomization and whether recruitment could have been influenced by allocation.
The CONSORT extension for cluster-randomized trials provides reporting recommendations concerning cluster allocation, participant recruitment, outcome measurement, and statistical analysis.
