Confirm that the sensitivity analysis applied your prespecified exclusion rule, rerun it like-for-like, and compare the revised estimate with the main result against a material-change threshold you set in advance.
- Compare the records removed in the sensitivity analysis with the exclusion criteria you specified before the main analysis. Cochrane guidance describes sensitivity analyses as checks on how robust pooled results are to assumptions and analytic choices, so the rule should match the one you prespecified. If the removed set differs from the rule, the contradiction may come from the analysis rather than the data.
- Rerun the analysis with the prespecified rule, keeping the same effect measure and model as the main analysis, and record the revised estimate with its confidence interval. A like-for-like comparison isolates the effect of the exclusion itself.
- Compare the revised estimate with the main result against your threshold, such as a shift in the direction of effect or a change in whether the confidence interval crosses the null. Cochrane interpretation guidance emphasizes confidence intervals and exact P values, so compare intervals and direction alongside point estimates. If the revised estimate stays within your threshold, the main result holds and the sensitivity analysis supports it.
- Document the rule applied, the records removed, both estimates with intervals, and whether the threshold was crossed. This record shows readers why the conclusion stands or changed.
- Revise the conclusion if a correctly applied rule moves the estimate past your threshold. GRADE rates certainty at the outcome level using explicit criteria, so a material shift in an outcome's estimate can lower certainty for that outcome and should be reflected in how strongly you state the finding.
If the contradiction persists after these steps, report both analyses and describe what each one shows.