How do I check a cluster-randomized trial for unit-of-analysis issues before pooling?

Check whether the trial analyzed at the cluster or participant level, whether the analysis adjusted for clustering, and record the design effect or intracluster correlation (ICC) used in a unit-of-analysis log before pooling.

What to extract

• The reported sample size: confirm whether it counts clusters, participants, or both, and note the average cluster size.
• The analysis method: look for cluster-level summaries, a multilevel or mixed model, or participant-level analysis with a variance correction.
• The adjustment parameters: record the ICC or design effect the authors report, and note when neither is given.
• The outcome data: check whether the effect estimate and its standard error come from an analysis that accounts for clustering.

How to run the check

1. Read the methods and results for the unit of randomization and the unit of analysis, then write both in the log. If they differ and no adjustment is described, flag the result as an unadjusted cluster analysis.
2. Locate the ICC or design effect the authors used and copy the value and source into the log. If the trial reports neither, mark the adjustment as unreported and record the average cluster size so the gap is visible.
3. Compare the reported sample size against the number of clusters and participants in the flow diagram. A participant count used as if it were independent observations signals a unit-of-analysis issue.
4. Carry the flag into your risk-of-bias assessment of that specific result, using the Cochrane RoB 2 domains and signalling questions, and justify the judgement in writing.
5. Decide the pooling step: keep adjusted results in the main analysis, and handle flagged unadjusted results through a sensitivity analysis or exclusion, documenting the choice in the log.

The Cochrane Handbook's meta-analysis chapter covers effect measures, heterogeneity, sensitivity analyses, and interpretation of pooled results, which is where the log entry feeds in. The risk-of-bias chapter describes how RoB 2 judges a specific trial result across domains as low risk, some concerns, or high risk.