Sensitivity analysis in meta-analysis tests whether the pooled result holds up when reasonable analytic choices change. A sensitivity analysis re-runs the meta-analysis under different assumptions, for example excluding high risk-of-bias studies or switching the statistical model, and checks whether the conclusion stays the same. If the pooled effect barely moves, the finding is robust; if it flips or shrinks, the original estimate depended on a fragile decision that readers deserve to know about.
Robustness, not a fishing expedition
A sensitivity analysis answers a single question: would a sensible reviewer reach the same conclusion if they had made a defensible choice differently? That is distinct from subgroup analysis and meta-regression, which try to explain why studies differ. Sensitivity work tests stability. Like every analytic decision, the planned sensitivity checks belong in the protocol, so they read as principled rather than as a hunt for a result you prefer.
Common sensitivity analyses
Excluding studies at high risk of bias
The most common check removes studies judged at high risk of bias and re-pools the rest. If the trustworthy studies alone tell the same story, confidence rises. If the effect depends on the weak studies, that is a serious caveat and feeds directly into the certainty of evidence rating.
Switching the statistical model
Re-running the pool under a fixed-effect versus random-effects model shows how sensitive the result is to the weighting assumption. The two usually agree on direction; large disagreement signals influential heterogeneity worth investigating. You can compare both models quickly in the meta-analysis calculator.
Leave-one-out analysis
A leave-one-out analysis removes each study in turn and re-pools the remainder, exposing any single trial that drives the result. If dropping one study changes the conclusion, that study is influential and warrants a closer look at its size, weight, and quality. This pairs naturally with reading the weights on a forest plot, where a dominant study is visually obvious.
Varying assumptions for missing data
When studies omit standard deviations or use different definitions, you often impute or harmonise values to pool them. A sensitivity analysis repeats the synthesis under alternative, plausible assumptions for those missing values to confirm the imputation did not manufacture the effect. A common pattern is to pool first with an imputed correlation coefficient of 0.5 for change-from-baseline data, then repeat at 0.2 and 0.8 to show the pooled effect does not hinge on a number that was guessed rather than reported.
Switching the heterogeneity estimator and the effect measure
Under a random-effects model the between-study variance, tau-squared, can be estimated several ways, and the choice changes the weights. Re-running with the DerSimonian-Laird estimator and then the restricted maximum-likelihood or Paule-Mandel estimator shows whether the result depends on that technical decision, which matters most when studies are few and heterogeneity is high. A parallel check switches the effect measure itself, for example pooling a risk ratio and then a risk difference for a binary outcome, to confirm the conclusion is not an artefact of the metric.
A worked leave-one-out example
Imagine six trials pooling to a standardised mean difference of 0.45 in favour of treatment, with a confidence interval from 0.20 to 0.70. One large early trial carries 55 percent of the weight. Removing it in a leave-one-out pass drops the pooled estimate to 0.18 with an interval that now includes zero, while removing any other single study barely moves it. That single result tells you the headline depends almost entirely on one influential study, and the discussion must say so, rather than presenting 0.45 as a settled finding. Spotting which study to suspect is far easier once you can read the box sizes on a forest plot.
How to report a sensitivity analysis
State clearly which assumption you varied, show the alternative pooled effect size with its confidence interval, and say plainly whether the conclusion changed. Reporting only the analyses that confirmed your headline, while hiding those that did not, is a form of selective reporting that the PRISMA 2020 guideline is designed to surface. Transparency about an unstable result is far stronger than a confident claim that cannot survive a reasonable challenge.
A practical order for running the checks
Sensitivity analyses are most convincing when they follow a planned sequence rather than appearing piecemeal:
- List in the protocol every assumption you consider debatable but defensible, such as the risk-of-bias cut-off, the model, the estimator, and any imputed value.
- Run the primary analysis first and record its pooled effect, interval, and heterogeneity as the benchmark.
- Re-run once per debatable assumption, changing only one thing at a time so you can attribute any movement to a single cause.
- Tabulate each alternative pooled estimate beside the primary one and judge whether the direction and significance hold, not merely whether the point estimate shifted a little.
- Carry the verdict, stable or fragile, into the certainty rating and state it plainly in the discussion.
Common mistakes that undermine a sensitivity analysis
The recurring errors are easy to name. The first is post-hoc cherry-picking: deciding which studies to exclude after seeing how exclusion changes the answer, which inverts the whole purpose and is why the checks belong in the pre-specified protocol. The second is changing several assumptions at once, so a shifted result cannot be traced to any one decision. The third is reporting only the confirming runs and quietly dropping the ones that destabilised the headline, a form of selective reporting that the PRISMA 2020 guideline exists to surface. The fourth is confusing a sensitivity analysis with a subgroup analysis: one tests whether the same conclusion survives a different choice, the other asks what explains differences between studies, and treating an unstable subgroup finding as a robustness result misreads both.
Where it sits in the analysis
Sensitivity analysis comes after the main meta-analysis and runs alongside checks for publication bias. Together they answer the questions a careful reader will ask: is the effect real, is it driven by weak or missing studies, and does it survive a change of method? A review that can answer all three earns a high certainty rating and a much smoother peer review. If you would rather have these checks designed and run to a standard that holds up in review, our meta-analysis service builds them into the synthesis from the protocol stage.