Can I combine arms that received different doses?

You can, and whether you should depends on the question. If dose is plausibly an effect modifier and you have a prespecified subgroup analysis on it, combining destroys the information you planned to examine. In that case use the arm-selection approach or move to a network meta-analysis.

What do I do when a trial has two control arms, placebo and no treatment?

Same problem in reverse. Prespecify which control matches your comparator definition, or combine them if your review treats both as usual care. Note the decision, since placebo and no treatment differ meaningfully for subjective outcomes.

Does this apply to a cluster trial with only two arms?

Yes, though the issue is clustering rather than double-counting. An unadjusted cluster trial carries too much weight. Report the intracluster correlation used, whether it was reported or imputed, and the source if imputed.

Multi-Arm Trials and the Unit-of-Analysis Error in Meta-analysis

12 September 20264 min

A trial randomises participants to high-intensity training, moderate-intensity training, or usual care. Your meta-analysis compares exercise against usual care.

The intuitive move is to enter two comparisons: high versus control, and moderate versus control. That double-counts the control group. Those participants contribute to the analysis twice, the trial receives more weight than its sample size justifies, and the confidence interval around the pooled estimate is too narrow because the two comparisons are correlated and the model assumes they are independent.

This is the unit-of-analysis error. It appears in a substantial share of published meta-analyses and it is straightforward to avoid.

Four legitimate approaches

Combine the intervention arms. Pool high and moderate intensity into a single group using the formulae in the Cochrane Handbook, which combine means and standard deviations across arms, then compare against the full control group. This is usually the right answer when the arms represent variants of the same intervention and your research question is about the intervention as a class.

Split the control group. Divide the control participants between the two comparisons, halving the sample size in each. This avoids double-counting but understates precision and produces awkward non-integer counts. Acceptable for dichotomous outcomes; less satisfactory for continuous ones.

Select one arm. Prespecify a rule such as the most intensive arm, or the arm most closely matching the review's intervention definition, and use only that. Simple and defensible, and it discards data, so say what was discarded and why.

Use network meta-analysis. If the distinction between arms is the point, a network handles multi-arm trials correctly by accounting for the correlation between comparisons from the same study. This is the right choice when comparing intensities is your question rather than a nuisance.

Choose in the protocol

Each approach gives a different pooled estimate, which means choosing after seeing the results is choosing a result.

A protocol sentence that works: "Where trials include multiple eligible intervention arms, arms will be combined using the formulae in Chapter 6 of the Cochrane Handbook where the interventions represent variants of the same modality. Where the arms differ in a way relevant to a prespecified subgroup analysis, the arm matching the subgroup will be used and the alternative reported in a sensitivity analysis."

The related error: multiple outcomes and time points

The same problem arises in three other forms.

Multiple time points from one trial entered as separate rows. Prespecify which time point is primary and group the rest by follow-up window.

Multiple instruments measuring the same construct entered separately. Prespecify a hierarchy of instruments and use one per study.

Cluster-randomised trials entered as if individually randomised. This ignores clustering, inflates the effective sample size and narrows the interval. The correction is to adjust the sample size by the design effect, which requires the intracluster correlation coefficient. If it is not reported, impute it from similar trials and test the assumption across a range.

Crossover trials need their own handling too: paired analysis where the within-participant correlation is reported, and a stated decision about whether to use first-period data alone where carryover is a concern.

How reviewers detect it

The clue in a forest plot is the same study name appearing on more than one row without an explanatory label. A careful statistical reviewer will check the total participant count in the forest plot against the sum across included trials, and a mismatch is the tell.

Label multi-arm handling explicitly on the plot, for instance "Smith 2021 (combined arms)". It takes a footnote and prevents a query.

References

Higgins, J. P. T., Thomas, J., Chandler, J., Cumpston, M., Li, T., Page, M. J., & Welch, V. A. (Eds.). (2024). Cochrane handbook for systematic reviews of interventions (Version 6.5). Cochrane. https://training.cochrane.org/handbook

Common questions

Can I combine arms that received different doses?: You can, and whether you should depends on the question. If dose is plausibly an effect modifier and you have a prespecified subgroup analysis on it, combining destroys the information you planned to examine. In that case use the arm-selection approach or move to a network meta-analysis.
What do I do when a trial has two control arms, placebo and no treatment?: Same problem in reverse. Prespecify which control matches your comparator definition, or combine them if your review treats both as usual care. Note the decision, since placebo and no treatment differ meaningfully for subjective outcomes.
Does this apply to a cluster trial with only two arms?: Yes, though the issue is clustering rather than double-counting. An unadjusted cluster trial carries too much weight. Report the intracluster correlation used, whether it was reported or imputed, and the source if imputed.