Should Clusters Be Matched or Stratified Before Randomization?

Should Similar Schools or Institutions Be Grouped Before Random Assignment?

Suppose you are evaluating a digital learning intervention across 16 schools. Some schools are large and well-resourced, while others are small and have limited digital infrastructure. Their baseline achievement levels also differ substantially.

If you randomly assign eight schools to the intervention and eight to the comparison condition, there is a possibility that one arm will contain disproportionately more high-performing or well-resourced schools.

You could first match schools with similar characteristics and randomly assign one school from each pair to each condition. Alternatively, you could divide schools into broader categories, such as high- and low-baseline achievement, and randomize within those categories.

Both approaches aim to improve baseline comparability, but they impose different restrictions on randomization and have different implications for analysis, statistical efficiency, and implementation.

Use Matching or Stratification When Important Baseline Differences Could Reduce Precision

How Matching and Stratification Affect Cluster-Randomized Trial Design

Why baseline balance deserves particular attention in cluster trials

Cluster-randomized trials often involve substantially fewer randomized units than individually randomized trials. A study might randomize 20 schools while measuring outcomes for several thousand students.

With relatively few schools, chance differences in institutional characteristics can be substantial. One arm may contain schools with higher baseline achievement, better infrastructure, or different student populations.

Randomization still provides a valid allocation mechanism, but a particular allocation may be poorly balanced on characteristics that strongly predict the outcome.

Such imbalance can reduce precision and complicate interpretation. Restricted randomization methods attempt to reduce this risk by using information collected before allocation.

The importance of baseline balance is especially evident when many participants are recruited from relatively few clusters.

What matched-pair cluster randomization means

In matched-pair cluster randomization, researchers identify pairs of clusters that are similar on prespecified baseline characteristics. Within each pair, one cluster is randomly assigned to the intervention and the other to the comparison condition.

For example, 16 schools may be organized into eight matched pairs based on baseline achievement and school size. Random allocation then determines which school in each pair receives the intervention.

Matching can ensure that the two arms contain one school from each pair. It may improve precision when the characteristics used for matching strongly predict the outcome and the resulting pairs are meaningfully comparable.

However, matching schools on superficial similarities may provide little statistical benefit. With very few pairs, the analysis can also have limited degrees of freedom, and withdrawal of a cluster may complicate matched comparisons.

What stratified cluster randomization means

Stratified randomization divides clusters into categories defined by one or more prespecified baseline characteristics. Randomization then occurs separately within each stratum.

Suppose schools are classified into two baseline achievement categories: higher and lower achievement. Schools within each category can be randomly allocated to intervention and comparison conditions, using an allocation procedure intended to maintain balance within the strata.

Stratification is particularly practical when a small number of categorical characteristics are strongly prognostic.

Unlike pair matching, a stratum can contain more than two clusters. This may provide greater flexibility in allocation and analysis.

However, creating too many strata can leave too few clusters within each category to maintain balanced allocation. The number of strata must be appropriate to the available cluster count.

How matching and stratification differ

Pair matching can be understood as a particularly restrictive form of stratification in which each stratum contains exactly two clusters.

The appropriate choice depends on the number of clusters, the strength of available prognostic variables, and whether meaningful matches or strata can be constructed.

Which baseline characteristics should be used?

Variables selected for matching or stratification should generally be measured before randomization and have a defensible relationship with the primary outcome.

In educational trials, relevant candidates may include baseline achievement, school size, geographic region, institutional type, or characteristics associated with intervention implementation.

Baseline outcome measurements can be especially informative when they strongly predict follow-up outcomes.

However, adding more variables is not necessarily better. Excessive restrictions can reduce the number of permissible allocations and complicate both implementation and analysis.

Variables should be chosen based on substantive knowledge and empirical evidence where available, rather than selecting numerous characteristics merely because they can be measured.

Why matching does not automatically increase statistical power

Matching is sometimes assumed to improve a trial because it produces visibly comparable intervention and comparison clusters. Yet improved appearance of balance does not necessarily translate into greater precision.

The statistical benefit depends partly on how strongly outcomes are correlated within matched pairs. If matched schools have similar baseline characteristics but their follow-up outcomes remain weakly related, the expected efficiency gain may be limited.

Pair matching also constrains the allocation structure. With a small number of pairs, matched analyses may have relatively few degrees of freedom.

Methodological investigations have consequently reached nuanced conclusions about matching. It can improve efficiency in suitable settings, but stratification or covariate-constrained randomization may perform as well or better under other conditions.

Researchers should not select matching solely because the cluster count is small.

How matching changes the randomization procedure

Consider eight matched pairs of schools. Within each pair, one school is assigned to the intervention and the other to the comparison condition.

Under independent, equally probable assignment within each pair, the number of possible allocations is:

The restricted allocation space is intentional. It ensures that each pair contributes one cluster to each arm.

However, fewer possible allocations can affect randomization-based inference and make allocation procedures more predictable if assignment information is disclosed prematurely.

The allocation sequence should therefore be generated and implemented using procedures that preserve allocation concealment until assignment.

Why excessive stratification can become counterproductive

Suppose a trial has 16 schools and researchers want to stratify by region, school size, baseline achievement, and institutional type.

If these variables create many combinations, some strata may contain only one school or an odd number of schools. Balanced allocation within every stratum may become impossible.

Researchers may need to simplify categories, select fewer stratification factors, or adopt another restricted randomization approach.

Stratification should be designed around the available number of clusters rather than allowing a large collection of baseline variables to dictate an impractical allocation structure.

Covariate-constrained randomization is another option

Matching and stratification are not the only approaches for improving baseline balance.

Covariate-constrained randomization evaluates a set of possible allocations according to prespecified balance criteria. Allocations meeting an acceptable balance requirement are retained, and the final allocation is randomly selected from the acceptable set.

This approach can accommodate several baseline covariates, including continuous characteristics, without necessarily dividing every variable into categories.

However, the procedure requires reliable baseline data, explicit balance criteria, and sufficient allowable allocations. Excessively restrictive criteria may leave too few allocations or produce undesirable dependencies between cluster assignments.

The analysis should reflect the restricted randomization procedure and adjust for relevant allocation covariates.

Covariate-constrained randomization may be attractive when multiple important prognostic factors must be balanced and the research team has appropriate statistical support.

Matching and stratification must be reflected in the analysis

Restricted randomization is not merely an administrative step completed before data collection. It forms part of the experimental design and has implications for statistical inference.

Current methodological guidance generally recommends adjusting for covariates used in restricted randomization when conducting the primary analysis. Such adjustment can improve precision and help ensure that inference appropriately reflects the design.

For matched-pair trials, the analysis should account for the pairing structure. Depending on the outcome, number of pairs, and estimand, appropriate approaches may include paired cluster-level analyses, regression adjustment, or randomization-based procedures.

For stratified trials, relevant stratification factors should ordinarily be incorporated into the analysis plan.

Ignoring the allocation restrictions can produce inefficient or inappropriate inference, particularly when the number of clusters is small.

Cluster withdrawal can be especially consequential in matched designs

If one cluster withdraws from a matched pair, the planned paired comparison is disrupted.

Some conventional matched analyses may no longer be able to use the pair in its original form. Alternative methods may retain information from the remaining cluster, but their validity depends on the missing-data mechanism, estimand, and analytical assumptions.

Researchers should not automatically discard both clusters from every disrupted pair. Nor should they simply ignore the original matching without evaluating the consequences.

The protocol should anticipate possible cluster withdrawal and specify how incomplete pairs will be handled.

Restricted randomization does not eliminate selection bias

Matching or stratification can improve balance on measured baseline characteristics, but it does not prevent selection bias arising from recruitment after allocation.

If recruiters know which schools received the intervention, they may enroll different types of students across trial arms. Matching schools on baseline achievement does not necessarily correct such differences in subsequent participant selection.

Where possible, eligible participants should be identified before randomization. When recruitment occurs afterward, standardized procedures and masking of recruiters to allocation may reduce bias.

The consequences of recruiting participants after cluster randomization should be considered separately from baseline balance.

Randomization restrictions should be planned before allocation

Researchers should specify the matching variables, stratification factors, balance criteria, allocation ratio, and randomization procedure in advance.

Matching clusters after observing treatment assignments does not reproduce a properly planned matched-randomization design.

Similarly, changing stratification criteria after examining outcomes can introduce analytical flexibility that undermines transparency.

Prespecification helps distinguish legitimate design decisions from post hoc attempts to obtain favorable comparisons.

Matching or Stratifying Schools in a Digital Learning Trial

Interpretation: Both matching and stratification can improve balance on measured baseline characteristics. Neither is automatically superior.

If schools can be paired meaningfully and within-pair outcome similarity is expected to be strong, matching may improve precision. If suitable pairs are difficult to form, stratification may offer a simpler and more flexible alternative.

The researchers should not infer that visually similar baseline groups guarantee an unbiased treatment-effect estimate. Random allocation, appropriate recruitment, intervention fidelity, and analysis remain necessary.

Misconceptions About Matching and Stratifying Clusters

Choose a Balance Strategy That Matches the Number and Characteristics of Your Clusters

Before selecting an allocation method, identify baseline characteristics that are likely to predict the primary outcome and can be measured reliably before randomization.

Then examine whether the available clusters form meaningful pairs or broader categories. Consider how each approach affects the number of allowable allocations, statistical analysis, and vulnerability to cluster withdrawal.

When several prognostic variables require balancing, covariate-constrained randomization may be worth evaluating rather than forcing numerous variables into a complicated matching or stratification structure.

For educational technology interventions, baseline achievement and institutional digital readiness may be relevant allocation factors, but their value depends on whether they predict the outcome or influence intervention implementation.

The objective is not to make the intervention and comparison groups appear identical. It is to improve the efficiency and credibility of a genuinely randomized comparison without introducing unnecessary restrictions.

Before Matching or Stratifying Clusters

Frequently Asked Questions About Matching and Stratified Randomization

Balance Important Prognostic Characteristics Without Overrestricting Randomization

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.