How to Read a Cluster Randomized Trial: When Groups, Not People, Are Randomized

A cluster randomized trial assigns whole groups, such as clinics, schools, or villages, to a treatment rather than assigning each person separately. This is the right design when an intervention naturally works at the group level or would otherwise leak between individuals, but it carries a cost: people within the same group are alike, so the trial holds less information than its headcount suggests. Reading one well means checking that the analysis accounts for that clustering and that people were recruited in a way that could not be shaped by knowing the group's assignment.

Randomizing groups instead of people

In a standard trial, each participant is randomized one at a time. In a cluster randomized trial, the unit that gets randomized is a group: a whole general practice, a hospital ward, a school, or an entire village. Everyone inside a chosen cluster receives whatever that cluster was assigned.

There are good reasons to design a trial this way. Some interventions are delivered to a group by nature, like a change to how a clinic runs or a public-health campaign broadcast across a community. Others would contaminate a person-level trial: if you train a doctor in a new technique, you cannot have that doctor use it for some patients and not others without the learning bleeding across. Randomizing the whole practice avoids that leakage.

Why the headcount overstates the evidence

The efficiency of an ordinary trial rests on each participant being an independent piece of information. In a cluster trial that independence is gone. People treated in the same clinic, taught by the same teacher, or living in the same village tend to resemble one another, so each additional person inside a cluster adds less new information than a fresh, independent participant would.

Statisticians measure this similarity with the intracluster correlation coefficient, and its practical consequence is captured by the design effect, the factor by which a cluster trial needs to be larger than an individually randomized trial to carry the same weight. A trial of ten thousand people spread across a handful of clusters can hold far less information than the number ten thousand suggests. When you read a cluster trial, look for the intracluster correlation and a sample size that was inflated to account for it.

The analysis must respect the clustering

The same clustering that shrinks the effective sample size must be built into the analysis. Methods that account for clustering, such as mixed models or generalized estimating equations, give honest confidence intervals. An analysis that ignores clustering and treats every individual as independent will report intervals that are too narrow and p-values that are too small.

This is not a subtle academic quibble. Analyzing a cluster trial as though it were an individually randomized one can turn a genuinely uncertain result into a falsely confident one. If the methods section does not mention how clustering was handled, that omission alone should lower your confidence in the reported precision.

Recruitment bias, the quiet threat

The most easily missed problem in cluster trials happens before any treatment is given. In an individually randomized trial, people are usually enrolled and consented before they are assigned, so their characteristics cannot be influenced by the assignment. In a cluster trial, the cluster is often randomized first, and then individuals inside it are identified and recruited.

That order opens a door. If the people recruiting participants know which arm a cluster is in, they may, even unintentionally, enroll different kinds of people into the intervention clusters than into the control clusters. The result looks like a treatment effect but is really a difference in who was studied. The reporting guidance for cluster trials asks authors to describe exactly how and when participants were identified relative to randomization, precisely so readers can judge this risk.

What to check before you trust the result

Reading a cluster trial well comes down to a few questions. Was randomizing groups the right choice, given the intervention or the risk of contamination? Does the sample size account for clustering through the design effect, and is the intracluster correlation reported? Does the analysis use a method that keeps the clustering intact rather than treating individuals as independent? And is it clear that participants were recruited in a way that could not be steered by knowing the cluster's assignment?

When those pieces line up, a cluster trial can answer questions an individual trial cannot. When they are missing, the same design can quietly overstate both the certainty and the size of the effect.