Run a Unit-of-Analysis Check Before Pooling Results

Draw the study's structure in words

Before extracting an effect estimate, identify four things: who or what was allocated, who received the intervention, what was measured, and what the analysis treated as independent. These may differ. A trial can randomize clinics, measure patients, and collect several visits from each patient.

Cochrane's chapter on variants of randomized trials explains why clustering, crossover designs, and multiple intervention groups need appropriate handling. Ignoring dependence can exaggerate precision and give a study too much influence in a synthesis.

Work through a fictional extraction problem

Imagine 12 clinics randomized between two training strategies, with 240 patients contributing outcomes. The participant count does not mean 240 independent randomizations. If each patient has three visits, the 720 recorded rows are not 720 independent participants either.

Create separate fields for clinics, patients, and observations. Then inspect the reported model for how clinic and repeated-patient structure were handled. Keep the model name and adjustment description beside the extracted estimate. Merely seeing the word “regression” does not establish that the relevant correlation was addressed.

Follow a short review sequence

First, inspect methods, participant flow, and table footnotes for words such as cluster, practice, household, paired, bilateral, repeated, or crossover. Second, identify whether the chosen estimate already accounts for the design. Third, confirm that the reported confidence interval and sample-size description belong to that estimate, rather than to a different unadjusted table.

Prefer extracting a suitable design-adjusted estimate with its uncertainty when available. If only an analysis ignoring clustering is reported, document the problem and seek the information needed for an appropriate correction. Do not invent an intracluster correlation or silently reduce the sample size.

Check independence across extracted rows

A three-arm study can contribute two comparisons that share a control group. Two publications can describe the same participants. Several outcomes or time points can also come from one cohort. Label these relationships before arranging the synthesis dataset.

Decide the handling strategy in the analysis plan: selecting an appropriate comparison, combining eligible groups, or using a method that explicitly accounts for dependence. The right choice follows the review question and design; it cannot be chosen from sample size alone.

End with a one-sentence audit conclusion for each study: “The extracted estimate accounts for clinic clustering,” or “Dependence remains unresolved; hold from the planned synthesis.” This gives the next reviewer a concrete decision to verify.

Source checked on 30 October 2026; example is fictional.