Confirm both effect measures come from the same underlying data and model, then recompute one from the other to locate the discrepancy. Cochrane Handbook Chapter 10 treats effect measures as properties of the analysis, so two measures for one outcome should agree when they describe the same comparison and dataset.
- Trace each measure to its source: record the included studies, the comparison, the outcome definition, and the model (fixed or random effects) behind each figure. If the study sets or models differ, the conflict is a difference in analyses, and you resolve it by deciding which analysis answers the review question.
- Recompute one measure from the other using the same data and model. A ratio measure such as a risk ratio and a difference measure such as a risk difference can be derived from shared event counts and group sizes. If the recomputed value matches one report and not the other, the mismatch sits in the second report's inputs or arithmetic.
- Run a sensitivity analysis on the point where the two diverge, such as a changed model or a different handling of missing data, and compare the resulting estimates and confidence intervals. Cochrane Handbook Chapter 10 describes sensitivity analyses for testing how robust pooled results are to such choices.
- Document the reconciliation: state the source data, the model, the recomputation, and the reason one figure was retained. Then state which measure the review will report and use it consistently in the abstract, results, and summary tables.
When reporting, pair the chosen measure with its confidence interval and exact P value, and describe absolute and relative effects where both help readers judge applicability, as Cochrane Handbook Chapter 15 advises. If recomputation still leaves the two figures apart, report the discrepancy as unresolved and describe what each source can verify.