How do I run a sensitivity analysis to test a pooled winter-health result?

Run a sensitivity analysis by removing the trials most likely to distort the pool, such as high-risk or outlying studies, recomputing the pooled estimate, and comparing the new result with the original in a side-by-side table. Cochrane's meta-analysis guidance treats sensitivity analysis as a way to check how robust a pooled result is to analytic choices and to the inclusion of particular studies.

Steps

1. Define the removal rule before you recompute. Use a criterion tied to your review question, such as excluding trials judged high risk of bias with RoB 2 or excluding trials whose effect estimate sits far outside the rest of the pool. Record the rule and the list of removed studies so the comparison is reproducible.
2. Recompute the pooled estimate with the same effect measure and model you used in the primary analysis, changing only the included studies. Report the new estimate alongside the original, with both confidence intervals.
3. Build the comparison table with one row per scenario: original pool, high-risk trials removed, and outliers removed. Add columns for the number of studies, the pooled effect, the confidence interval, and the heterogeneity statistic.
4. State your rule for a material change before you look at the results. For example, treat the conclusion as sensitive if removing a set of trials moves the confidence interval across the null value or reverses the direction of the pooled effect.
5. Interpret each row against that rule and report which removed studies drove any shift. If the conclusion holds across scenarios, describe the pooled result as robust to those removals; if it shifts, report the sensitivity finding and identify the influential trials.

What to watch

Use RoB 2 to justify which trials count as high risk, since it assesses a specific randomized-trial result across bias domains and produces low risk, some concerns, or high risk judgements with written justifications. Keep the removal rule narrow enough that each scenario answers a question a reader would recognize, and show the excluded studies by name in the table so readers can trace the change.