A decision that could reasonably go another way
Imagine a fictional study with some missing follow-up outcomes. The main analysis uses one set of assumptions about those outcomes. A sensitivity analysis explores a plausible less favorable scenario for the intervention group.
If the conclusion changes substantially, the assumptions about missing data are important to the interpretation. If it does not, the result appears less sensitive to that particular alternative.
Choose alternatives for a reason
An analysis is more informative when it explains why each alternative is plausible. Trying many methods and presenting only those that preserve a preferred conclusion defeats the purpose.
Cochrane's meta-analysis guidance describes sensitivity analysis as examining the effect of decisions made in a review. Examples can include excluding studies with particular methodological concerns or changing an analytical assumption.
A compact reporting example
Main estimate: a small benefit. After excluding studies with a prespecified concern: a similar point estimate, with wider uncertainty because fewer studies remained.
This invented summary distinguishes the estimate from its precision. “The result stayed the same” would be too vague: the direction, magnitude, interval, and decision implication may behave differently.
What a stable result cannot prove
Several analyses can share the same unmeasured confounding, measurement error, or missing evidence. Agreement among them does not remove a flaw they all inherit.
Conversely, a change across analyses does not automatically make the study useless. It can identify the assumption that future research needs to resolve.
Read the alternatives beside the main result
A useful report gives enough detail to understand what changed and why. The sensitivity analysis is not merely a ceremonial appendix. It shows where the conclusion is firm, where it depends on judgment, and which uncertainty remains despite the extra calculations.
