Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · Sensitivity Analysis

The Use of Sensitivity Analysis in Meta-Analysis: Purpose, Methods and Reporting

Sensitivity analysis in meta-analysis repeats the pooled calculation under alternative, defensible methodological choices to determine whether the principal result depends on a single analytic decision, study or assumption, or whether it is robust across them.

Pubrica AcademyWritten by a meta-analysis methodologist Reviewed bya biostatistics reviewer Published10 October 2026 Last checked10 October 2026 Reading time7 minutes Sources10 peer-reviewed

Sensitivity analysis in meta-analysis is the practice of re-running the pooled effect calculation under alternative, equally defensible methodological choices (excluding one study at a time, switching between fixed-effect and random-effects models, restricting to studies above a quality threshold) to determine whether the overall conclusion changes. It is necessary because a single pooled estimate, calculated in one way, can appear precise without revealing how dependent that precision is on a particular analytic decision or a particular influential study.

Learning objectives

  • The main forms sensitivity analysis takes in a meta-analysis, including leave-one-out and model-choice checks
  • How sensitivity analysis differs from heterogeneity assessment, which is a related but distinct concept
  • Why an influential single study can change a pooled result even when overall heterogeneity looks low

Common misconceptions

  • “Sensitivity analysis and heterogeneity analysis are the same procedure.” Heterogeneity analysis quantifies how much true effects vary across studies, typically via the I-squared statistic, while sensitivity analysis tests whether the pooled result is robust to specific analytic decisions; a meta-analysis can show low heterogeneity and still have a result that depends heavily on one influential study, which only a sensitivity analysis would reveal.

Purpose of sensitivity analysis

Sensitivity analysis tests whether a meta-analysis's pooled conclusion survives reasonable changes to the choices made in calculating it, rather than testing the underlying clinical or scientific question itself1. The Cochrane Handbook frames this explicitly as distinct from the primary analysis: the primary analysis produces the headline pooled estimate using the pre-specified method, and sensitivity analyses then ask how far that estimate would shift had specific, justifiable alternative choices been made instead1. A result that changes direction or loses statistical significance under small, defensible changes in method is a materially weaker finding than one that remains stable across them, even if both produce the same headline point estimate.

This framing also clarifies what a sensitivity analysis is not: it is not a search for the most favourable version of the result among several possible analyses, and reporting only the sensitivity analysis that confirms the primary finding while omitting others that weaken it subverts the purpose of the technique. The purpose of the exercise is to give readers an accurate indication of how much confidence the pooled estimate merits, which requires reporting the full, pre-specified set of checks regardless of the direction in which each one points.

Common forms of sensitivity analysis

The most widely used form is the leave-one-out analysis, recalculating the pooled estimate with each study removed in turn to check whether any single study is driving the result, a method formalised in diagnostic tools such as those described by Viechtbauer and Cheung for identifying outlier and influential studies in meta-analysis2. Other common forms include comparing fixed-effect against random-effects pooling, since the two models weight studies differently and can produce different pooled estimates when heterogeneity is present3, restricting the analysis to studies above a defined risk-of-bias or quality threshold, and testing alternative choices for how missing data or unpublished results were handled1. Each of these targets a different analytic decision point rather than a single generic robustness check.

A further common form restricts the analysis to a specific subset of studies defined by a clinical or methodological characteristic decided in advance, for instance comparing the pooled estimate using only studies with long-term follow-up against the estimate using all available studies, to determine whether the effect appears to persist over time or is concentrated in shorter-term results. This differs from an exploratory subgroup analysis in that it is framed and reported specifically as a robustness check on the primary pooled estimate, not as a search for a separate, subgroup-specific effect to be highlighted in its own right.

A worked illustration of its practical importance

In a hypothetical meta-analysis of eight trials of a treatment, seven small trials show a modest, non-significant benefit and one unusually large trial shows a strong, significant benefit; because the large trial’s precision gives it substantial weight in an inverse-variance-weighted pooled estimate, the primary analysis can readily reach significance overall even though seven of the eight individual trials did not reach significance on their own. A leave-one-out sensitivity analysis removing that large trial would show the pooled estimate losing significance, revealing that the headline finding is substantially a function of one study rather than a consistent signal across the evidence base, information that the primary analysis alone does not reveal.

This is not, by itself, a reason to discard the large trial or to distrust the overall finding; large, well-conducted trials are not inherently suspect simply because they are large and precise. It is, however, a reason to report the leave-one-out result transparently, so that readers, including guideline committees deciding how much weight to place on this evidence, can see that the statistical strength of the finding depends heavily on a single, identifiable study, which is precisely the kind of information a sensitivity analysis is intended to reveal and a headline pooled estimate alone would conceal.

Distinction from heterogeneity assessment

Heterogeneity assessment, most commonly reported through the I-squared statistic developed by Higgins and Thompson, quantifies the proportion of variability across study estimates attributable to genuine differences between studies rather than chance4. This is a different question from sensitivity analysis: heterogeneity describes how much studies disagree, while sensitivity analysis tests whether the pooled conclusion itself is fragile to specific methodological choices5. A body of evidence can show low heterogeneity by the I-squared statistic and still have its pooled estimate determined disproportionately by one large or unusual study, a pattern that a leave-one-out sensitivity analysis, not a heterogeneity statistic, reveals.

Patsopoulos and colleagues’ empirical evaluation of heterogeneity metrics specifically cautions against treating I-squared as a fixed, universally interpretable threshold, since its value is also sensitive to the number and precision of the studies included, independently of the extent to which the underlying true effects vary5. This partly explains why sensitivity analysis and heterogeneity assessment are best treated as complementary checks reported together: a high I-squared value signals that pooling may be masking real differences between studies, while a sensitivity analysis shows explicitly how much, and in what direction, the pooled estimate shifts depending on how those differences are handled.

Planning and reporting sensitivity analyses

Good practice, reflected in both PRISMA 2020's reporting items and AMSTAR 2's appraisal criteria, is to pre-specify which sensitivity analyses will be conducted in the review protocol, rather than running many post-hoc comparisons and selectively reporting the ones that support the primary finding67. The results should be reported alongside the primary analysis, not only when they confirm it, since a reader's confidence in a pooled estimate should be informed by how it behaved under alternative, defensible choices, not solely by the single number the primary model produced.

One practical way to present this in a manuscript is a short table listing each pre-specified sensitivity analysis alongside the primary pooled estimate and showing how the point estimate and confidence interval shift under each one, which allows readers to review the full set of checks at a glance rather than searching the prose for scattered references to robustness. Where every sensitivity analysis moves the estimate only modestly and in no consistent direction away from the primary finding, that pattern should itself be stated explicitly, since demonstrated stability under scrutiny is as informative, and as worth reporting, as instability would be. Journals and guideline bodies increasingly expect such a table as a matter of course for a pooled estimate that will inform clinical or policy decisions, which makes its construction a routine part of drafting the results section rather than an optional addition requested only after peer review. Archiving the raw output from each sensitivity run alongside the main analysis, and not only the summary table, also makes it considerably easier to answer a peer reviewer’s follow-up question about a specific check without rebuilding the entire analysis months after it was first run. This archival practice requires little effort at the time of analysis and typically saves considerably more effort later, whether the request comes from a reviewer, an editor or the review team itself when updating the analysis with newer evidence. None of this requires specialised software beyond that already used for the primary meta-analysis; it is a matter of consistent record-keeping rather than additional statistical complexity.

Key points

Key points

Sensitivity analysis examines whether a pooled meta-analytic result is robust to specific, defensible methodological choices, a question distinct from heterogeneity; both should be pre-specified and reported irrespective of whether they confirm the primary finding.

Practical application

Practical application

  1. Sensitivity analyses should be pre-specified in the review protocol. The analytic choices to be tested (model type, quality threshold, missing-data handling) should be decided before the pooled result is seen.
  2. A leave-one-out analysis should be performed whenever a meta-analysis includes a markedly larger or more influential study. It should be established specifically whether removing that study changes the direction or significance of the pooled estimate.
  3. Sensitivity analysis results should be reported even when they weaken the primary finding. The full set of pre-specified checks should be presented in the manuscript, not only those that reinforce the principal result.

Frequently asked questions

It is not universally mandatory, but current reporting and appraisal standards such as PRISMA 2020 and AMSTAR 2 expect it to be considered and reported where relevant, particularly when heterogeneity or an influential study is present.

It means the pooled conclusion changes materially, in direction, magnitude, or statistical significance, under one or more defensible alternative analytic choices, which should lower confidence in the primary finding.

No single sensitivity check determines which model is ‘correct’; the aim is to understand how far the conclusion depends on the choice, which then informs how cautiously the overall finding should be interpreted.

No. They answer different questions and are typically reported together: heterogeneity quantifies disagreement between studies, while sensitivity analysis tests the stability of the pooled conclusion under alternative choices.

There is no fixed number, but each one should test a specific, pre-specified, defensible methodological alternative; running many undirected post-hoc comparisons and reporting only favourable ones undermines the purpose of the exercise.

References

  1. Higgins JPT, Thomas J, Chandler J, et al. (editors). Cochrane Handbook for Systematic Reviews of Interventions, version 6.4, Chapter 10: Analysing data and undertaking meta-analyses. read 10 October 2026. View Chapter 10
  2. Viechtbauer W, Cheung MW. Outlier and influence diagnostics for meta-analysis. Research Synthesis Methods. 2010;1(2):112-125. read 10 October 2026. View on Research Synthesis Methods
  3. DerSimonian R, Laird N. Meta-analysis in clinical trials. Controlled Clinical Trials. 1986;7(3):177-188. read 10 October 2026. View on Controlled Clinical Trials
  4. Higgins JPT, Thompson SG. Quantifying heterogeneity in a meta-analysis. Statistics in Medicine. 2002;21(11):1539-1558. read 10 October 2026. View on Statistics in Medicine
  5. Patsopoulos NA, Evangelou E, Ioannidis JPA. Sensitivity of between-study heterogeneity in meta-analysis: proposed metrics and empirical evaluation. International Journal of Epidemiology. 2008;37(5):1148-1157. read 10 October 2026. View on IJE
  6. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. read 10 October 2026. View on The BMJ
  7. Shea BJ, Reeves BC, Wells G, et al. AMSTAR 2: a critical appraisal tool for systematic reviews. BMJ. 2017;358:j4008. read 10 October 2026. View on The BMJ
  8. Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-analysis detected by a simple, graphical test. BMJ. 1997;315(7109):629-634. read 10 October 2026. View on The BMJ
  9. Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Introduction to Meta-Analysis. Chichester: John Wiley & Sons; 2009. read 10 October 2026. View the publisher listing
  10. Thompson SG, Higgins JPT. How should meta-regression analyses be undertaken and interpreted? Statistics in Medicine. 2002;21(11):1559-1573. read 10 October 2026. View on Statistics in Medicine

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). The Use of Sensitivity Analysis in Meta-Analysis: Purpose, Methods and Reporting. Pubrica. https://pubrica.com/academy/concepts-definitions/why-is-sensitivity-analysis-used-in-meta-analyses/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Meta-Analysis Sensitivity Check Plan

These analyses should be pre-specified before the primary meta-analysis is run.

[ ] Leave-one-out analysis planned for any disproportionately large or unusual study
[ ] Fixed-effect vs. random-effects comparison planned where heterogeneity is expected
[ ] Quality/risk-of-bias threshold restriction planned, with the threshold defined in advance
[ ] Missing-data or unpublished-result handling alternatives specified
[ ] Plan to report all pre-specified sensitivity analyses regardless of outcome

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s biostatistics team pre-specifies and performs sensitivity analyses alongside the primary meta-analytic model, so that the robustness of a pooled result is documented rather than assumed.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Robustness of the pooled result to leave-one-out analysis

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.