Extract a Cluster-Trial Effect and Its Correct Uncertainty

Identify the randomized unit

A cluster trial allocates groups such as clinics, schools or wards. Participants within a cluster may resemble one another, so an individual-level sample size does not fully describe the information available. Extract the number of randomized clusters, participants, cluster sizes, recruitment timing and analysis population separately for each arm.

Cochrane’s guidance on cluster trials discusses design-aware analyses and approximate corrections when clustering was ignored. The first task is to discover what the investigators already did. Applying another design effect to an already adjusted standard error can correct the same issue twice.

Locate the estimand and variance

Find the effect measure, model, adjustment variables and reported standard error or confidence interval. Check for mixed models, generalized estimating equations, cluster-level analyses or another stated method accounting for clustering. A phrase such as “adjusted analysis” may refer only to baseline covariates; it does not necessarily mean cluster dependence was addressed.

Preserve the number of clusters used in the analysis and any small-sample degrees-of-freedom method. An interval based on a t distribution cannot always be inverted using a normal critical value without changing its interpretation.

Audit a hypothetical record

Imagine 20 clinics randomized, ten per arm, with 400 participants overall. The paper gives an adjusted mean difference of −3 units and a standard error of 1.2 units from a stated cluster-aware model. Its inverse-variance contribution is 1/(1.2²) = 0.6944, before any synthesis-specific weighting beyond inverse variance.

The extraction should show 20 randomized clusters and 400 participants, not “n = 400” alone. If two clinics withdrew, establish whether the model includes 18 or 20 clinics and whether their participant outcomes remain available. Do not reduce the standard error simply because 400 appears to be a large sample.

Handle missing design information transparently

When clustering was ignored, collect the reported individual-level variance, cluster sizes and any intracluster correlation estimate needed for an approximate correction. Record the source and applicability of an externally borrowed correlation. Vary that assumption when it materially affects conclusions; the design-effect guide explains the simplified arithmetic and its limits.

Finalize a design-aware extraction row containing both unit counts and the provenance of the uncertainty estimate. Flag unexplained cluster losses, recruitment after allocation and inconsistent cluster sizes for appraisal. Correct standard errors address dependence; they do not repair recruitment bias or make unlike intervention settings interchangeable.

Source verification: 6 November 2026. Examples are hypothetical.