Sensitivity Analysis Systematic Review: A Practical Guide

A sensitivity analysis systematic review checks whether a review’s estimate or interpretation changes when uncertain decisions or assumptions are varied in plausible ways. It is a robustness check, not a method for choosing the result that looks most favorable or becomes statistically significant. [1]

The useful question is not simply whether the result changes, but how much it changes, why, and whether the review’s interpretation depends on the choice being tested. A transparent analysis makes those choices visible so readers can judge how robust the conclusions are.

What question does a sensitivity analysis answer?

A sensitivity analysis asks whether the review’s result is robust to plausible alternatives for uncertain methodological decisions. [1] It compares the main analysis with an alternative that is defensible, such as applying a different reasonable rule for handling an uncertain study or assumption.

The analysis should keep its target question clear. If changing an assumption materially changes the estimate, its direction, or the interpretation, the conclusion depends on that assumption. If the result remains stable, that supports confidence only for the alternatives examined; it does not establish that every source of uncertainty has been addressed. [1][2]

Sensitivity analysis is not a search for a preferred answer. Running multiple versions and reporting only the one that supports a desired conclusion can mislead. The purpose is to reveal how analysis choices affect the result, including when the conclusion is less certain than the main analysis alone suggests. [1]

Which review decisions are worth testing?

Test decisions that are genuinely uncertain and could plausibly affect the result, rather than changing choices arbitrarily. Useful candidates include borderline study eligibility, risk-of-bias decisions, handling of missing or imputed data, outcome time points, effect measures, and reasonable pooling models. [1]

For each candidate, identify the uncertainty and define a defensible alternative. For example, if study eligibility is borderline, specify which decision rule would change and which studies would be affected. If the concern is missing data, make clear what alternative assumption is being tested; do not present an unsupported assumption as though it were known to be true.

It is not necessary to test every imaginable choice. Prioritize alternatives that are methodologically plausible and relevant to the review’s conclusions. When several decisions could be examined, explain why the selected analyses address important uncertainties. The available guidance does not set a universal ranking rule, so prioritization requires reviewer judgment.

How do you perform a sensitivity analysis in a systematic review?

Perform the analysis by defining the uncertainty, specifying a defensible alternative, rerunning the relevant analysis, and comparing the result with the main analysis. [1] Where possible, decide on the alternatives in advance and describe them in the protocol; if an analysis is added later, explain why. [1]

Keep the comparison interpretable: change the decision or assumption under examination while holding other analysis choices constant where feasible. Record which studies or data are affected, and report the resulting estimate and confidence interval. If the alternative changes more than one choice at once, readers may not be able to tell what drove the difference.

The statistical approach depends on the uncertainty being examined. A sensitivity analysis might use a different reasonable pooling model, but model choice should not be selected solely in response to a heterogeneity test. Such tests can have low power when there are few or small studies, and I² ranges are rough interpretive guides—not automatic triggers for a model choice or a downgrade in certainty. [1][2]

Automation can support this workflow by helping reviewers track assumptions, rerun pre-specified analyses, and document outputs. Those functions should support human governance: reviewers still need to decide whether an alternative is defensible, whether a result is meaningful, and how it should be interpreted.

How should you interpret changes in estimates and uncertainty?

Interpret changes by comparing effect size and direction, uncertainty, and the implications for the review—not just whether statistical significance changes. [1][2] A result crossing a significance threshold does not, by itself, show that the underlying conclusion has changed.

Consider whether the estimate moves meaningfully, whether its direction changes, and whether the uncertainty becomes wider or narrower. Then ask whether those changes affect the interpretation. If the conclusion shifts under a plausible alternative, report that it is sensitive to the tested assumption. If it remains similar, state that it was stable under those alternatives, without implying that it is robust to untested uncertainties. [1][2]

Excluding studies judged at high risk of bias illustrates why interpretation should include precision. Exclusion may reduce bias but can also reduce precision; a non-significant difference between risk-of-bias groups does not establish that bias is absent, because such comparisons may have low power. [3][1]

Keep sensitivity results separate from GRADE certainty judgments, while using them where relevant to assess the evidence. GRADE assesses certainty for each important outcome across risk of bias, inconsistency, indirectness, imprecision, and publication bias. These are evidence-level judgments, not a study-quality score or a mechanical tally. [2]

How is sensitivity analysis different from subgroup analysis?

Sensitivity analysis tests whether an estimate for the same target remains robust under alternative decisions or assumptions; subgroup analysis estimates effects in separate groups and formally compares them. [1] Although both may involve more than one analysis, they answer different questions.

A sensitivity analysis asks, “Would this estimate or interpretation change if we handled this uncertainty differently?” A subgroup analysis asks whether effects differ across groups. Do not describe a subgroup comparison as a robustness check merely because the results are presented separately. Being explicit about the question helps readers distinguish uncertainty about analytic decisions from possible differences between groups.

How should results be reported—and what can automation support?

Report why each sensitivity analysis was performed, what changed, how many studies were affected, the resulting estimate and confidence interval, and whether the interpretation changed. [1] A concise summary table can make these comparisons clearer than many separate forest plots. [1]

A practical table can include the uncertainty tested, the main-analysis decision, the alternative, the studies affected, the estimate and confidence interval under the alternative, and a brief account of whether the interpretation changed. These fields are a reporting aid; the key requirement is that the rationale, changes, results, and interpretation are transparent. [1]

For analyses added after the protocol, state that they were post-protocol and explain the reason. Treat post hoc explorations as hypothesis-generating rather than as planned confirmation. [1] In a governed workflow, automation may help preserve the record of choices and outputs, but it cannot decide whether the analysis was justified or what its findings mean.

Key takeaways

• Sensitivity analysis tests whether results change under plausible alternatives to uncertain review decisions or assumptions. [1]
• Test methodologically defensible alternatives, and specify them in advance where possible. [1]
• Interpret effect size, direction, uncertainty, and implications rather than relying only on statistical significance. [1][2]
• A stable result supports confidence only for the alternatives tested; a material change signals an assumption-sensitive conclusion. [1][2]
• Report the rationale, affected studies, resulting estimate and confidence interval, and any change in interpretation. [1]

FAQ

Should I report a sensitivity analysis that does not change the result?

Yes. A stable result is informative if you say which alternative was tested and what remained stable. Limit the conclusion to those alternatives; stability does not show that untested uncertainties have no effect. [1][2]

Does a non-significant result mean the sensitivity analysis found no important difference?

Not necessarily. Statistical significance alone does not show whether the estimate, its uncertainty, or its interpretation changed. Comparisons between risk-of-bias groups may also have low power, so a non-significant difference does not establish the absence of bias. [1][3]

Can sensitivity analysis fix missing studies or poor review methods?

No. It cannot recover unknown missing studies, correct poor review methods, or remove bias in primary studies. Implausible alternatives and arbitrary exclusions can mislead rather than clarify. [1]

What are the GRADE downgrading domains?

GRADE assesses certainty by outcome across risk of bias, inconsistency, indirectness, imprecision, and publication bias. These domains support evidence-level judgments; they are not a mechanical tally or a score of individual study quality. [2]