What Exactly Should Be Randomized in a Randomized Controlled Trial?

Are You Supposed to Randomize Participants, Interventions, or Something Else?

You are designing a randomized controlled trial and have decided which intervention to evaluate. The next question seems straightforward: randomly assign participants to groups. But what if the intervention is delivered to entire classrooms, schools, hospitals, or departments?

Should individual students be randomized, or should their classes be assigned to different instructional conditions? If every participant will receive both interventions, should you randomize the order instead? What if the intervention is introduced at different times across institutions?

These are not interchangeable allocation decisions. What you randomize determines how the intervention is delivered, which comparisons are possible, and how the resulting data should be analyzed.

Randomize the Unit That Can Meaningfully Receive the Assigned Condition

How to Identify the Correct Unit of Randomization

Randomization is the use of a chance-based mechanism to assign eligible units to specified intervention conditions. It is a design procedure, not merely the use of random numbers somewhere in the research process.

Before generating an allocation sequence, researchers should identify what is being assigned, what intervention each assignment represents, and how outcomes will be compared.

Distinguish the Unit of Randomization, Intervention, Measurement, and Analysis

Several different units may appear in the same trial. Confusing them can lead to inappropriate allocation or statistical analysis.

These units may coincide in an individually randomized trial, but they often differ in educational and organizational research.

For example, assigning classrooms while measuring individual students creates clustered outcome data. Treating every student as an independently randomized unit would misrepresent the design.

When Should Individual Participants Be Randomized?

Individual randomization is appropriate when participants can meaningfully receive different assigned conditions and when interference between participants is sufficiently limited or appropriately addressed.

Suppose a university evaluates two versions of an independently delivered online tutorial. Eligible students may be randomly assigned to either version, provided their access and outcomes can be managed according to the protocol.

Individual randomization can be efficient because the number of independently allocated units is relatively large compared with many cluster-randomized designs.

It may also help balance pre-assignment characteristics across conditions in expectation.

However, individual randomization may become problematic when participants share intervention materials, receive instruction from the same teacher, or influence one another's outcomes.

The key question is whether assigning different individuals within the same environment produces meaningful and interpretable treatment contrasts.

When Should Groups or Clusters Be Randomized?

Cluster randomization assigns intact groups rather than individual members to intervention conditions.

Clusters may include classrooms, schools, hospitals, clinics, departments, workplaces, or communities.

For example, a professional development program may train teachers to use a particular instructional strategy. If teachers deliver that strategy to entire classes, assigning individual students within the same classroom may be impractical.

Randomizing classrooms or schools may better match the intervention's delivery structure.

Cluster randomization may also reduce contamination when participants within a group interact closely. However, it does not guarantee that information or practices will remain isolated across clusters.

Researchers should consider the choice between individual, cluster, site, and time-based randomization in relation to the intervention mechanism and available units.

Why Does Cluster Randomization Affect Sample Size?

Participants within the same cluster often have correlated outcomes because they share teachers, environments, institutional resources, or other contextual characteristics.

Consequently, 200 students distributed across a small number of classrooms generally provide less independent information about a cluster-assigned intervention than 200 independently randomized students would under otherwise comparable conditions.

A common approximation for equal-sized clusters uses a design effect to describe the inflation in variance associated with intracluster correlation.

Under these assumptions, the variance of a relevant estimate may be approximately 1.95 times what would be expected under an otherwise comparable independent-observation calculation.

The design effect is not a universal multiplier for every cluster-trial analysis. Unequal cluster sizes, baseline adjustment, stratification, repeated measurements, and the number of clusters can materially affect precision.

Researchers should plan sample size using methods appropriate to the actual cluster design.

Should Sites Be Randomized Instead of Participants?

Site randomization is a form of cluster randomization in which institutions or service locations are assigned to conditions.

For example, universities may be randomized to different academic advising strategies, or hospitals may be assigned to different implementation procedures.

This approach can be useful when the intervention changes organizational practice or requires site-wide implementation.

However, the number of participating sites becomes especially important. A trial involving hundreds of individuals but only a few randomized sites may have limited precision for estimating the intervention effect.

Site-level differences, recruitment after allocation, and variation in implementation should also be considered.

Can Time Periods Be Randomized?

Yes. Some trials randomize intervention conditions across time periods rather than assigning each participant or site permanently to one condition.

For example, a service may alternate between two operational strategies across prespecified periods, with the sequence determined randomly.

Time-based randomization can be appropriate when the intervention is reversible, the periods can be meaningfully compared, and carryover or temporal trends can be addressed.

Researchers must consider whether outcomes depend on previous intervention exposure, whether participants appear in several periods, and whether external conditions change over time.

Randomly assigning time periods does not automatically eliminate period effects or interference between successive interventions.

What Should Be Randomized in a Crossover Trial?

In a crossover trial, participants or clusters receive multiple intervention conditions in a specified sequence.

Rather than assigning each participant permanently to one intervention, researchers commonly randomize the order in which interventions are received.

For example, participants may be allocated to receive intervention A followed by intervention B, or intervention B followed by intervention A.

The sequence matters because outcomes may be influenced by treatment order, period effects, or carryover from an earlier intervention.

Crossover designs are most appropriate when the intervention effects are sufficiently reversible, the outcome can be measured repeatedly, and the assumptions needed for within-participant comparison are credible.

Randomizing the sequence does not automatically eliminate carryover effects.

What Should Be Randomized in a Factorial Trial?

A factorial trial evaluates combinations of intervention factors.

For example, researchers may want to evaluate two instructional features: adaptive feedback and peer discussion.

In a two-by-two factorial design, eligible units may be randomly assigned to one of four combinations: neither feature, adaptive feedback only, peer discussion only, or both features.

Randomization therefore assigns units to combinations of intervention-factor levels.

The design may support estimation of main effects and interactions, depending on the research questions, allocation structure, and statistical power.

Researchers should not assume that the effects of two interventions are necessarily additive or independent.

What Is Randomized in a Stepped-Wedge Trial?

In a stepped-wedge cluster-randomized trial, clusters typically begin in a control condition and transition to the intervention at different periods according to a randomized implementation schedule.

The randomized element is commonly the sequence or timing of cluster crossover.

This design may be considered when an intervention is intended to reach all participating clusters eventually.

However, stepped-wedge trials require careful treatment of calendar-time effects, within-cluster correlation, intervention rollout, and possible delayed responses.

The fact that all clusters eventually receive the intervention does not automatically make the design ethically superior or statistically efficient.

Researchers should justify why randomized rollout is preferable to a conventional parallel cluster trial or another evaluation strategy.

Can Intervention Intensity or Dose Be Randomized?

Yes, when the research question concerns the comparative effects of different intervention intensities or doses and the assigned conditions are scientifically and ethically appropriate.

For example, researchers might randomize participants to receive different frequencies of tutoring sessions.

The intervention conditions must be clearly defined, and the intended comparison should reflect the research question.

Random assignment to different doses does not necessarily establish the causal effect of the amount actually received if adherence varies after assignment.

The effect of assigned dose and the effect of received dose are different estimands and may require different analytical assumptions.

Should Participants Be Randomized Before or After Recruitment?

In conventional individually randomized trials, eligibility and enrollment are ordinarily established before assignment, with procedures designed to prevent knowledge of the upcoming allocation from influencing recruitment.

In cluster trials, clusters may be randomized before individual participants are identified or recruited. This can create additional risks when recruiters or participants know the cluster's assignment.

For example, knowledge that a school received a desirable intervention might influence which students are encouraged to enroll or agree to participate.

Researchers should plan recruitment and allocation procedures to minimize selection bias and comply with applicable ethical requirements.

The timing of enrollment relative to randomization should be documented explicitly.

What Is the Difference Between Randomization and Allocation Concealment?

Randomization concerns how the assignment sequence is generated. Allocation concealment concerns preventing those responsible for enrollment or assignment from knowing the upcoming allocation before it is implemented.

A genuinely random sequence can still be compromised if future assignments are predictable or disclosed.

For example, a researcher who knows the next assignment may influence which participant is enrolled at that point.

Proper allocation concealment helps preserve the intended protection against selection bias.

Researchers should also distinguish allocation concealment from blinding, which concerns awareness of assigned conditions after allocation.

Should Randomization Be Simple, Blocked, or Stratified?

Once the unit of randomization has been identified, researchers must choose a suitable allocation procedure.

Simple randomization assigns units according to a random mechanism without requiring exact balance in group sizes at every stage.

Block randomization can help maintain allocation balance within defined blocks. Stratified randomization can help balance assignments within levels of selected baseline characteristics.

The choice depends on the number of randomized units, prognostic characteristics, allocation ratio, and operational requirements.

Researchers should consider simple, block, and stratified randomization only after deciding what unit should receive the random assignment.

Using a sophisticated sequence-generation method cannot correct a fundamentally inappropriate unit of randomization.

Why the Unit of Analysis Must Reflect Randomization

Statistical analysis should respect the dependence structure created by the study design.

When individual participants are randomized independently, participant-level analyses may be appropriate, subject to the outcome and model.

When clusters are randomized, outcomes from individuals within the same cluster are generally correlated. Ignoring this structure can underestimate standard errors and produce misleading confidence intervals.

Researchers may use mixed-effects models, generalized estimating equations, cluster-level summaries, or other suitable approaches depending on the design and number of clusters.

For crossover and repeated-period designs, analyses must address within-unit dependence and relevant period or sequence effects.

The statistical model should follow the allocation and measurement structure, not merely the format in which observations appear in the dataset.

Does Randomization Guarantee That Groups Are Comparable?

No. Proper randomization helps prevent systematic allocation based on participant characteristics, but chance differences can remain.

These differences may be particularly noticeable when the number of randomized units is small.

Researchers should distinguish chance imbalance from evidence that the randomization process was compromised.

Baseline comparisons may be useful descriptively, but significance testing of baseline differences is generally not an appropriate method for determining whether randomization was successful.

The trial's integrity should be evaluated through its allocation procedures, concealment, implementation, and adherence to the protocol.

What If Participants Do Not Receive Their Assigned Intervention?

Randomization assigns intervention conditions, but participants may not always receive or adhere to those conditions.

For example, students assigned to an online tutoring platform may not use it, while students in the comparison group may access similar resources elsewhere.

An intention-to-treat analysis generally compares participants according to their original randomized assignment, preserving the allocation-based comparison for an appropriate treatment-policy estimand.

Estimating the causal effect of actually receiving the intervention may require additional methods and assumptions.

Researchers should therefore distinguish what was randomized from what participants ultimately received.

Why the Randomization Unit Should Be Chosen Before Sample-Size Planning

The number of independently randomized units strongly influences statistical precision.

A study with 300 students randomized individually differs substantially from a study with 300 students distributed across six randomized classrooms.

Both studies may contain the same number of student observations, but the second design has only six independently assigned clusters.

Sample-size calculations must therefore reflect the allocation unit, clustering, expected outcome variation, and intended analysis.

Researchers should not determine a sample size under individual randomization assumptions and then switch to cluster allocation without reassessing the design.

Should an Educational Trial Randomize Students or Classrooms?

The researchers decide to randomize classrooms rather than individual students. Five classrooms are assigned to AI-assisted instruction, and five continue with established instruction.

All 300 students complete the same outcome assessment, but only ten classrooms were independently randomized.

Suppose the researchers anticipate an intracluster correlation of 0.05 for the primary outcome.

This calculation illustrates the loss of statistical information associated with clustering. It does not mean that the trial literally contains 122 independent participants or that the effective sample size alone determines the trial's power.

With only ten randomized classrooms, the number of clusters may impose substantial limitations on statistical inference. Appropriate small-sample methods and more detailed power calculations would be necessary.

The researchers should also consider teacher effects, baseline achievement, classroom differences, and whether the intervention is implemented consistently.

Randomizing classrooms is methodologically coherent because the allocation unit matches the intervention's delivery structure. However, the trial's adequacy still depends on whether the number of classrooms and the analysis can support the intended effect estimate.

The central distinction is that 300 students provide outcome data, but the intervention assignment was randomized across ten classrooms. Treating all 300 students as independently randomized would misrepresent the experimental design.

Common Mistakes When Choosing What to Randomize

How to Select the Appropriate Randomization Unit

Begin by identifying the intervention's delivery mechanism. Determine whether individual participants can receive different assigned conditions without substantial interference or whether assignment must occur at a group or institutional level.

Then identify the intended causal contrast and the level at which outcomes will be measured.

Once the randomization unit is chosen, specify the allocation ratio, sequence-generation method, concealment procedures, sample-size assumptions, and primary analysis.

The responsibility for generating the randomization sequence should be separated appropriately from recruitment and assignment activities where necessary to protect allocation integrity.

Researchers should also document the actual number of randomized units and distinguish it from the number of individuals contributing outcome data.

The appropriate randomization unit should be selected before recruitment and sample-size planning whenever feasible. Changing the allocation structure later may require substantial revisions to the protocol and statistical analysis.

Before Finalizing What to Randomize, Check:

Frequently Asked Questions About What to Randomize

Randomize at the Level That Matches the Intervention and Causal Question

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.