Sensitivity Analysis in Meta-Analysis: When and How

A sensitivity analysis asks: does my pooled result hold up if I change one methodological decision at a time? It's one of the fastest ways to show reviewers your findings aren't an artifact of a single arbitrary choice.

What sensitivity analysis is (and isn't)

Sensitivity analysis re-runs the same meta-analysis with a specific decision changed, to see whether the conclusion changes — it's not the same as subgroup analysis, which splits studies by a clinical or methodological characteristic to explore why effects differ. Sensitivity analysis asks whether your result is robust; subgroup analysis asks what's driving variation.

Common sensitivity analyses

• Excluding high risk-of-bias studies — re-run the pooled estimate with only low/moderate risk-of-bias studies included, and compare.
• Fixed-effect vs random-effects model — confirm your conclusion doesn't depend entirely on which model you chose. See our guide to fixed-effect vs random-effects meta-analysis.
• Excluding outlier or highly influential studies — a single large or extreme-effect study can dominate a pooled estimate; removing it one at a time (leave-one-out analysis) shows whether any single study is driving your result.
• Different imputation or estimation choices — e.g., comparing results with and without studies where summary statistics had to be estimated. See our guide to handling missing data.
• Published vs unpublished/grey literature — re-running with only published studies can reveal how much a result depends on grey literature inclusion.

How to report it

State the analysis was pre-specified in your protocol (if it was — sensitivity analyses decided after seeing results are far less convincing), report the original and adjusted pooled estimates side by side, and interpret plainly: does the conclusion hold, or does it change meaningfully? A conclusion that shifts substantially under a reasonable sensitivity analysis is important to report honestly, not something to omit.

Sensitivity analysis vs GRADE certainty

A result that's robust across multiple sensitivity analyses supports higher certainty in your GRADE rating for that outcome; a result that shifts substantially under minor changes is a legitimate reason to downgrade certainty, separate from the individual study risk-of-bias ratings.