Sensitivity Analysis in Systematic Reviews: A Practical Guide

A sensitivity analysis systematically tests how much your pooled result changes when you alter a specific methodological decision -- excluding a study, switching models, changing a definition -- and it is one of the clearest ways to show readers how robust, or fragile, your headline conclusion actually is.

What sensitivity analysis is actually for

Every systematic review makes a series of judgment calls: which risk-of-bias threshold counts as acceptable, whether to include a borderline-eligible study, which statistical model to use. Sensitivity analysis asks a direct, honest question about each of these calls: if we had decided this differently, would our overall conclusion have changed? A conclusion that survives several different reasonable choices is considerably more trustworthy than one that flips depending on a single close judgment call.

Common candidates for sensitivity analysis

Excluding studies at high risk of bias and checking whether the pooled estimate shifts meaningfully is one of the most standard sensitivity checks, directly testing whether your overall result depends on including lower-quality evidence. Switching between a fixed-effect and random-effects model, and comparing the two pooled estimates, tests whether your conclusion is sensitive to this modeling choice specifically. Excluding studies published only as abstracts or preprints, where full peer-reviewed detail isn't available, tests whether your result depends on less rigorously vetted sources. Re-running your analysis with an alternative, equally defensible effect size calculation formula tests whether a specific statistical choice is driving your finding.

Planning sensitivity analyses in advance

Sensitivity analyses should be planned and named in your protocol before you see your pooled results, for the same reason your eligibility criteria and primary analysis plan are pre-specified. Running sensitivity analyses only after seeing an unexpected or unwelcome primary result, then selectively reporting whichever alternative analysis produces a more favorable picture, is a form of undisclosed analytical flexibility that undermines the very purpose sensitivity analysis is meant to serve.

Reporting sensitivity analysis results honestly

Present your sensitivity analyses alongside your primary result, not buried in a supplementary appendix your average reader will never open, particularly when a sensitivity analysis reveals your conclusion is genuinely fragile to a specific decision. If excluding high-risk-of-bias studies substantially changes your pooled estimate, this is a critically important finding about the reliability of your overall conclusion, not an inconvenient result to minimize or relegate to a footnote.

Sensitivity analysis versus subgroup analysis

These two are related but conceptually distinct. Subgroup analysis asks whether an effect genuinely differs across a specific population characteristic -- age, geography, intervention variant. Sensitivity analysis asks whether your overall conclusion depends on a specific methodological decision rather than a substantive characteristic of the studies themselves. Both use similar statistical machinery, re-running an analysis under a modified condition, but they answer genuinely different questions and should be labeled and interpreted according to which one you're actually running.

How many sensitivity analyses is reasonable

There's no fixed universal number, but running so many sensitivity analyses that some become essentially exploratory data dredging, particularly if not pre-specified, undermines their credibility for the same reason excessive unplanned subgroup analysis does. A handful of genuinely important, pre-specified sensitivity checks addressing your review's most significant methodological judgment calls is more valuable and more credible than an exhaustive, unplanned list run simply because the software makes it easy to generate more output.

What a stable result actually tells you

When your pooled estimate remains substantively similar across your planned sensitivity analyses, this is meaningful, positive evidence your conclusion doesn't rest fragile on a single contestable decision, and this stability is worth stating explicitly and with genuine confidence in your discussion section, not left as an implicit inference readers have to draw from a supplementary table themselves.

What an unstable result tells you

If your pooled estimate shifts meaningfully under a reasonable alternative analytical choice, this is equally important information, and it should shape how confidently you state your overall conclusion. A finding that depends heavily on including or excluding a small number of specific studies is genuinely less certain than one holding steady across multiple reasonable analytical paths, and GRADE's imprecision and risk-of-bias domains should reflect this instability where it exists.

A practical checklist before finalizing your review

Confirm your planned sensitivity analyses were specified in your protocol, not decided only after seeing your primary results. Confirm you're presenting sensitivity results transparently alongside your primary finding, not selectively reporting favorable ones while omitting less favorable ones. And confirm your discussion section explicitly addresses what your sensitivity analyses collectively suggest about how much confidence readers should place in your headline conclusion, rather than treating this step as a procedural formality completed and then set aside.

Sensitivity analysis and your GRADE ratings

Where a sensitivity analysis reveals genuine instability in your pooled estimate, this should feed directly into your GRADE assessment, typically contributing to a downgrade on the imprecision or risk-of-bias domain depending on which specific decision the instability traces back to. Treating sensitivity analysis results as disconnected from your certainty ratings, rather than as direct evidence informing them, misses one of the most concrete, practical uses this kind of analysis actually serves.

Communicating sensitivity findings to a non-statistical audience

A sensitivity analysis section written purely in statistical terms -- reporting a series of alternative pooled estimates without explaining what their similarity or difference actually means for a reader's confidence in your conclusion -- misses an opportunity to make this analysis genuinely useful. A short, plain-language summary stating whether your key finding held steady or shifted meaningfully across your planned checks helps readers who won't parse a full sensitivity analysis table on their own. This small translation step, added at the end of an already-completed analysis, is often the single sentence in your entire results section that a busy clinician or policymaker actually reads and remembers. Investing the small additional effort this translation requires is disproportionately valuable relative to the modest time it actually takes to write clearly once the underlying statistical work is already finished. Reviewers who see this kind of clear, considered translation consistently note it favorably, since it demonstrates the authors genuinely understood and thought through the practical implications of their own statistical findings.