Should Sample Sizes Be Balanced Across Sites in Multisite Research?

Should Every Research Site Contribute the Same Number of Participants?

You are planning a study across several universities, hospitals, schools, or communities. One institution can recruit 300 eligible participants, another can recruit 150, and a smaller site may contribute only 50. Should you impose equal enrollment targets to make the study statistically balanced?

Equal sample sizes can simplify comparisons and improve efficiency in certain designs. Yet enforcing them may require unnecessary recruitment at smaller institutions or prevent larger sites from contributing useful information.

The decision is consequential because the distribution of participants across sites can affect statistical precision, power, institutional representation, and the meaning of pooled estimates. The appropriate allocation depends on what the study is trying to estimate and how its data will be analyzed.

Balanced Site Samples Can Help, but They Are Not Universally Required

How Sample Size Balance Affects Multisite Research

What Does Balanced Enrollment Mean?

A balanced multisite sample has approximately equal numbers of participants across participating sites. An unbalanced sample has unequal site contributions.

For example, a four-university study recruiting 100 students per university has balanced site sizes. A study recruiting 200, 100, 60, and 40 students has unbalanced site sizes, although both studies include 400 participants.

Balance can refer to several different features of a design. These should not be confused.

A study may have equal numbers across sites but unequal treatment allocation within them. It may also have balanced treatment groups while allowing substantially different site sizes.

Neither arrangement automatically establishes representativeness.

Why Can Balanced Site Sizes Improve Statistical Efficiency?

When participants within sites are correlated, observations from the same institution may contribute less independent information than the same number of independently sampled participants.

In a simplified clustered design, distributing participants more evenly across clusters can improve efficiency because very large clusters may provide diminishing additional information.

For a fixed number of clusters, a fixed total sample size, and common positive intracluster correlation, substantial cluster-size variation generally reduces efficiency relative to equal cluster sizes under conventional assumptions.

This does not mean equal site sizes always maximize efficiency in every model. The optimal allocation can change when sites have different outcome variances, recruitment costs, treatment assignments, or scientific importance.

The distinction is especially important when site effects must be considered during study design.

What Is Intracluster Correlation?

The intracluster correlation coefficient, commonly abbreviated ICC, describes the similarity of observations within the same cluster relative to observations from different clusters under a specified statistical model.

For example, students within one university may share teaching practices, institutional resources, or other influences that make their outcomes more similar than those of students at different universities.

In a simple random-intercept model, the ICC can be expressed as the between-cluster variance divided by the sum of between-cluster and within-cluster variances.

Its magnitude depends on the outcome, population, clustering level, and model. An ICC estimated for one outcome or institutional context should not automatically be applied to another.

ICC is also not equivalent to treatment-effect heterogeneity. Sites can differ in baseline outcomes without necessarily having different intervention effects.

How Does Clustering Affect the Effective Sample Size?

Under a simplified equal-cluster-size design with a common positive ICC, a conventional design effect illustrates the loss of information associated with clustering.

The effective sample size is an approximation for the specified variance comparison, not a literal count of independent participants. It should not be used as a universal substitute for a design-specific power calculation.

When cluster sizes vary, the equal-size formula no longer captures the full effect of imbalance. Researchers may use methods incorporating the distribution or coefficient of variation of cluster sizes, or preferably a power calculation or simulation matching the planned design and analysis.

How Does Unequal Site Size Affect Different Research Designs?

The implications of imbalance differ substantially between cluster-randomized, individually randomized, and observational multisite studies.

In individually randomized trials, a site may contain both treatment groups, whereas in cluster-randomized trials the entire site may receive one allocation. These arrangements have different implications for identifying treatment effects.

Researchers should avoid applying a cluster-randomized design effect indiscriminately to every study involving multiple institutions.

When Is Equal Enrollment Across Sites Particularly Useful?

Balanced enrollment may be advantageous when researchers intend to compare the participating sites directly and want similar precision for each site-specific estimate.

It may also simplify certain factorial or stratified designs and improve efficiency under assumptions involving equal variances and comparable recruitment costs.

For example, a study comparing learning outcomes across four universities might seek similar numbers of students at each institution so that one site's estimate is not substantially less precise than the others.

However, equal sample sizes cannot compensate for having too few independent institutions to support a site-level hypothesis.

If the objective is to estimate how institutional policies relate to outcomes across universities, the number and selection of universities may matter more than recruiting hundreds of additional students within a small set of institutions.

When Can Unequal Site Sizes Be Appropriate?

Unequal site sizes may be defensible when institutions serve populations of different sizes, recruitment capacity varies, or the target estimand places different weights on participating populations.

For example, if the aim is to describe a population of students distributed across universities of substantially different enrollment sizes, proportional allocation may be appropriate under a suitable sampling design.

By contrast, if the aim is to compare the average experience of institutions regardless of their size, equal site weighting may be more closely aligned with the question.

These are different inferential targets. Neither weighting strategy is automatically correct.

Practical recruitment limitations also matter. The capacity of individual research sites may make exact equality unrealistic or ethically undesirable if staff feel pressured to recruit beyond the accessible eligible population.

Does the Largest Site Always Dominate the Analysis?

Not necessarily. The influence of a site depends on the estimator, statistical model, outcome variability, weighting, and correlation structure.

In a simple participant-weighted mean, a site contributing more participants receives greater weight. In an equally weighted average of site means, each site contributes the same weight regardless of enrollment.

Multilevel models and other estimators may produce more complex weighting patterns.

Researchers should distinguish the number of observations from their statistical influence. A large site may contribute substantial information about its own population without contributing an equivalent amount of information about variation across institutions.

How Do You Choose Between Participant-Weighted and Site-Weighted Estimates?

The choice should follow the target estimand.

A participant-weighted estimate may be relevant when the objective concerns the average individual in a defined population and the sampling and weighting procedures support that inference.

An equally site-weighted estimate may be relevant when the objective concerns the average participating institution, with each institution contributing equally to the summary.

Neither approach automatically generalizes beyond the selected sites.

Where site sizes are associated with outcomes or treatment effects, different weighting choices can produce substantively different results. Researchers should define the target quantity before examining the outcome data.

Should Researchers Set Minimum Sample Sizes for Every Site?

Minimum site-specific targets may be appropriate when the study requires interpretable institutional comparisons, site-specific estimates, or adequate representation of particular settings.

However, a universal minimum such as 30 or 50 participants per site has no general methodological justification.

The necessary number depends on the outcome, variability, model, desired precision, and intended inference.

For cluster-randomized research, minimum numbers of clusters may be more consequential than minimum numbers of individuals within each cluster.

Researchers should distinguish requirements for estimating within-site quantities from requirements for estimating between-site variation.

What If Enrollment Becomes More Unequal Than Planned?

Recruitment rarely follows projections perfectly. Some sites may enroll more slowly, experience staffing interruptions, or have fewer eligible participants than anticipated.

When imbalance becomes substantial, investigators should update the projected sample distribution and reassess the statistical consequences.

Possible responses include extending recruitment at selected sites, using approved additional recruitment channels, activating reserve sites, or revising the allocation of remaining recruitment targets.

Such decisions should preserve the intended population and should not be driven by knowledge of favorable or unfavorable outcome results.

Where necessary, a statistician can evaluate revised power or precision using the actual or projected site-size distribution.

Changes should be documented and reflected in the protocol or analysis plan when appropriate.

Can Statistical Methods Compensate for Imbalance?

Appropriate models can accommodate unequal site sizes, and some designs can use weighting or stratification to target specific population quantities.

However, statistical adjustment cannot automatically recover information from sites that contribute too few observations or replace independent sites that were never recruited.

Small numbers of clusters can also create difficulties for conventional standard-error estimators and random-effects models.

Researchers should plan analytical methods with the expected imbalance in mind rather than treat model flexibility as a reason to ignore site allocation during design.

Differences in participant composition and institutional conditions may additionally create heterogeneity that affects interpretation.

How Equal and Unequal Site Sizes Change the Meaning of a Study

Neither result is automatically biased or correct for a wider population. The equally weighted mean describes the average of the four university means, while the participant-weighted mean describes the average across the sampled students.

If the researcher intends to estimate the mean for all students across the four universities, the appropriate weighting depends on the actual university populations and sampling probabilities, not simply the observed sample counts.

The example demonstrates that sample allocation can change the quantity being summarized even when the same institutions and total number of participants are involved.

Common Misconceptions About Balanced Multisite Samples

How to Choose a Defensible Distribution of Participants Across Sites

Begin by defining whether your primary inference concerns individuals, the participating institutions, or a broader population of sites. Then determine how many participants and sites are needed to estimate that quantity with acceptable precision.

Consider recruitment feasibility without allowing convenience alone to determine the final sample distribution.

Sample-size balance should be treated as a design choice with identifiable statistical consequences. Exact equality is rarely a scientific objective in itself.

Before Setting Sample-Size Targets for Research Sites

Frequently Asked Questions About Multisite Sample-Size Balance

Balance Site Samples When It Serves the Research 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.