How do you write a limitations section?

A limitations section names each specific threat to your study’s validity, states the direction and likely size of the bias it would produce, and says what you did about it. Write three to five, ordered by how much each could change your conclusion, then change the conclusion to match. Generic caveats that leave the claims untouched are the failure mode reviewers punish.

That is the whole instruction, and almost nobody follows it. ter Riet and colleagues examined 300 papers from 30 biomedical journals published in 2007 and found that 27% (81/300) mentioned no limitation at all, while among those that did, 183 of 219 — 84%, 95% CI 78 to 88 — did not temper their conclusions in light of the limitations they had just acknowledged (PLoS ONE 2013;8(11):e73623). This page covers what reporting guidelines actually require, the sentence shape that satisfies them, how to find your own limitations, how to quantify direction and magnitude, what belongs elsewhere, and what reviewers write when the section is padding.

What reporting guidelines require of a limitations section

Reporting guidelines ask for more than a list, and the exact wording is worth reading rather than paraphrasing.

CONSORT 2025, item 30, asks for “Trial limitations, addressing sources of potential bias, imprecision, generalisability, and, if relevant, multiplicity of analyses.” The expanded checklist elaborates it as three bullets: any methodological limitations and, if relevant, any methods used to minimise or mitigate them; any imprecision in the results; and generalisability of the results. In CONSORT 2010 the same requirement sat at item 20 and did not name generalisability, which was a separate item 21 — a renumbering worth knowing when a journal’s author instructions still reference the old scheme.

STROBE, item 19, is the strictest sentence in any of these documents: “Discuss limitations of the study, taking into account sources of potential bias or imprecision. Discuss both direction and magnitude of any potential bias.” Direction and magnitude are a specific, checkable demand. A limitations paragraph that says selection bias is possible, without saying which way it pushes the estimate or by roughly how much, does not satisfy STROBE item 19 — and most published observational limitations paragraphs do not.

PRISMA 2020 splits the requirement in two: item 23b, “Discuss any limitations of the evidence included in the review”, and item 23c, “Discuss any limitations of the review processes used.” Systematic reviews routinely answer 23b and skip 23c entirely.

The ICMJE Recommendations instruct authors to “State the limitations of your study, and explore the implications of your findings for future research and for clinical practice or policy”, and separately to “avoid unqualified statements and conclusions not adequately supported by the data.” Those two clauses are one instruction. The second is how an editor decides whether the first was performed honestly.

The three parts of a limitation a reviewer will accept

Every usable limitation has three parts in one or two sentences: the specific mechanism, the direction and size of the distortion it produces, and what you did about it or why nothing could be done.

Compare a limitation missing all three with the same limitation carrying all three.

Weak: “This was a single-centre study, and our sample size was relatively small, so the results should be interpreted with caution.”

Usable: “Recruitment was confined to one tertiary referral centre, so participants had a median Charlson comorbidity index of 5 against 3 in the national registry cohort; because comorbidity predicts the outcome, the absolute event rate reported here is likely higher than in community practice, while the relative effect estimate is less sensitive to that difference. We report both, and the community-based estimate should be regarded as an extrapolation.”

The second version specifies the mechanism (referral-centre case mix), the direction (absolute rate inflated), the differential effect on two quantities (absolute versus relative), and the consequence for the claim. It is longer, and length is not the point — the point is that a reviewer can agree or disagree with it, which they cannot do with “interpreted with caution”.

The same test applies to every limitation you write. Ask: could a reviewer disagree with this sentence? If the sentence is unfalsifiable, it is not a limitation, it is a hedge.

Finding the limitations in your own manuscript

Limitations are found by auditing the manuscript against itself, not by consulting a list of common limitations. Six passes will surface nearly all of them.

Read your own abstract conclusion and list what must be true for it to hold. If the conclusion says an intervention “improves outcomes in older adults”, the required premises include that the sample contained older adults in useful numbers, that the outcome measured is the outcome named, and that follow-up was long enough for it to occur. Each premise you cannot fully support is a limitation, already ranked by importance because it attaches to the main claim.

Diff your registered protocol against your reported outcomes. Outcomes registered but not reported, and outcomes reported but not registered, are the most reliably detected defect in a submitted trial manuscript — the COMPare project made a public exercise of checking published trials in leading journals against their registrations. A silently changed primary endpoint discovered by a reviewer is far more damaging than the same change disclosed in your own limitations paragraph.

Follow the numbers down your flow diagram. Every drop between screened, eligible, enrolled, and analysed is a candidate selection limitation. If 240 were eligible, 180 enrolled and 141 analysed, the limitations section owes the reader an account of the 99, and whether those lost differed on exposure, outcome, or both — the last case being collider bias rather than simple attrition.

Scan Table 1 for imbalance you did not adjust for, and the Methods for every variable you wanted and did not have. Unmeasured confounding is only a serious limitation when you name the confounder; “residual confounding is possible” is boilerplate, whereas “we had no measure of smoking pack-years, which is associated with both the exposure and the outcome and would bias the estimate away from the null” is a limitation.

Find every sentence in the Methods beginning with a concession — “owing to”, “because of the pandemic”, “data were unavailable for”, “assay drift required”. Each was a compromise you already recognised while doing the work. Concessions that appear only in the Methods and never in the Discussion are precisely what a methods-focused reviewer collects.

Check the gap between your measurement and your construct. A surrogate endpoint, a self-reported behaviour, a cell line standing in for a tissue, an in vivo model standing in for a human disease: each is a validity limitation that lives in the distance between what was measured and what is claimed. For animal work the gap between model and human disease is a standard reviewer request, and it has its own machinery in model relevance review.

Ordering limitations by consequence, not by convenience

Order limitations by how much each could move your conclusion, most consequential first. The near-universal ordering — sample size, single centre, short follow-up, then everything else — is convenience ordering, and it puts the two least informative items in the position of greatest emphasis.

Sample size is a weak lead limitation for a specific reason: it affects precision, which the reader can already see in your confidence interval. A limitation that duplicates information already in the results table spends the reader’s attention without adding anything. Precision belongs in the limitations section only when the interval is wide enough that the study cannot distinguish clinically distinct effects, and then the right sentence names those effects: “the 95% CI ranges from a 4% absolute reduction to a 9% absolute increase, so this trial cannot distinguish benefit from harm at magnitudes that would change practice.” That sentence is not about sample size, it is about what the study can and cannot decide.

Never argue sample size retrospectively with observed power. Post hoc power is a deterministic function of the p-value and adds no information beyond it, which is why post hoc power objections recur in review reports as a criticism rather than a defence.

A defensible ordering usually runs: threats to internal validity that could reverse or nullify the effect; threats to construct validity, where the thing measured is not the thing claimed; threats to external validity and generalisability; and finally precision and scope. ter Riet’s audit found the opposite balance in the literature — 62% of acknowledged limitations concerned internal validity, mostly measurement error, and 38% external validity, mostly selected study populations. Those counts describe what authors chose to raise, not what was present in the studies.

Quantifying direction and magnitude

STROBE item 19 asks for magnitude, and magnitude can usually be estimated rather than asserted. Four methods convert a prose limitation into a number a reviewer can check.

E-values state how strong an unmeasured confounder would have to be, on the risk-ratio scale, in association with both exposure and outcome, to explain away the observed effect (VanderWeele and Ding, Annals of Internal Medicine 2017). Reporting an E-value of 1.9 for the point estimate and 1.3 for the confidence limit converts “residual confounding cannot be excluded” into a claim readers can evaluate against known confounder strengths in the field.

Quantitative bias analysis propagates a specified misclassification or selection mechanism through the analysis and reports the corrected estimate under stated assumptions. It requires you to name sensitivity and specificity, or selection probabilities, which is itself disciplining.

Negative control outcomes and exposures test whether a bias you suspect is actually present. An association between the exposure and an outcome it cannot plausibly cause is evidence of confounding or selection, and its magnitude bounds the bias in the outcome you care about.

Tipping-point and pattern-mixture analyses address missing data by asking how different the unobserved values would have to be from the observed ones before the conclusion changes. Reporting that the result holds unless dropouts fared more than twice as badly as completers is a stronger statement than “loss to follow-up may have introduced bias”.

None of these removes the limitation. All of them replace an unbounded worry with a bounded one, which is the difference between a limitation a reviewer accepts and one they escalate.

What each study design is expected to declare

Reviewers arrive with design-specific expectations, and a limitations section that omits the expected item reads as unaware rather than concise.

Registry and electronic-health-record studies attract the most predictable objections of the group, because the data were generated for administration rather than research; the recurring set is catalogued in registry data limitations, and time-allocation errors of this kind are immortal time bias rather than a limitation adjustment can address.

What does not belong in a limitations section

A limitations section is not the place for methodological failures that should have been fixed, for material that belongs in the Methods, or for humility performed for its own sake.

Fixable problems belong fixed. A missing control, an unreported randomisation method, an unadjusted multiplicity — disclosing these as limitations does not neutralise them, and a reviewer who spots a fixable defect declared as a limitation reads the declaration as an attempt to trade disclosure for exemption. When the fix is another experiment, that is a different judgement, and additional-controls requests are usually negotiated in review rather than pre-empted here.

Methods descriptions belong in Methods. “We used a convenience sample” is a methods statement. “Convenience sampling recruited from a single outpatient clinic, which over-represents treatment-adherent patients and would bias the adherence-stratified estimate towards the null” is a limitation. The first states what happened; the second states what it does to the number.

Universal caveats add nothing. “This was an observational study, so causation cannot be inferred” is true of every observational study ever published and carries no information about yours. If your paper needs that sentence, the actual problem is causal language elsewhere in the manuscript, which is a causal language issue to fix in the Results and Discussion rather than to apologise for at the end.

Do not invent limitations to appear balanced. Reviewers notice a section that lists three trivial limitations and omits the obvious one, and the omission becomes the review’s opening sentence. Padding the section also dilutes the limitations that matter.

Do not undo the paper. A limitations section that concedes so much that no claim survives invites the reasonable question of what the manuscript is for. The target is calibration, not self-abolition.

Length, placement, and the abstract

Place the limitations section in the Discussion, after the interpretation of your findings and before the conclusion, under its own subheading where the journal permits — three to five limitations, roughly 200 to 400 words in total for a standard research article.

The subheading matters more than it appears. A labelled “Limitations” subheading makes the section findable by an editor triaging a submission, by a reader scanning the Discussion, and by retrieval systems that chunk documents at heading boundaries. Burying limitations in the fourth paragraph of an unheaded Discussion is a common cause of the reviewer comment that limitations were not addressed when in fact they were.

The abstract is the least-used and most-noticed position. ter Riet’s audit found a limitation mentioned in the abstract of 16 of 300 papers — 5.3%, 95% CI 3.3 to 8.5 — and papers in general medical journals were substantially more likely to do it than papers in specialty journals (odds ratio 3.57, 95% CI 1.27 to 10.0). One clause in the abstract’s conclusion sentence, naming the principal constraint, is a small edit that materially changes how an editor reads the paper’s self-awareness. Many structured abstracts have no limitations field; the clause then attaches to the conclusion sentence.

The conclusion has to move

A limitations section is only credible if the conclusions differ from what they would have been without it. This is the test editors apply, and the one most manuscripts fail: in the ter Riet audit, 84% of papers that acknowledged limitations did not temper their conclusions in response.

Tempering is not adding “however, further research is needed”. It is changing the scope of the claim to match the evidence. Concretely: change “the intervention improves survival” to “the intervention improved survival in this single-centre cohort of patients with preserved performance status”; change “X causes Y” to “X was associated with Y after adjustment for the confounders listed, with an E-value of 1.9”; change “these results support routine use” to “these results support evaluation in a randomised trial powered for the clinical endpoint.”

Then check the title, the abstract conclusion, the final paragraph of the Discussion, and any graphical abstract for consistency with the tempered claim. Overclaiming survives in those four places long after the Discussion body has been made careful, and mismatch between them is what an overclaim check is looking for. Unsupported claims propagating through the title and abstract are among the most common grounds for rejection.

What reviewers say when the limitations section is mishandled

“The limitations section is generic and does not engage with the specific threats to validity in this study.” “The authors acknowledge limitations but do not modify their conclusions accordingly.” “The direction and likely magnitude of the bias introduced by [X] are not discussed.” “The most important limitation of this design is not mentioned.” “Limitations are confined to sample size and single-centre recruitment, while the more serious issue of outcome ascertainment is not addressed.” “The claim in the abstract is not supported once the stated limitations are taken into account.”

Two of these are worth anticipating specifically. “The authors acknowledge limitations but do not modify their conclusions” is the comment that most often converts a major revision into a rejection, because it reads as a judgement about candour rather than about method. And “the most important limitation is not mentioned” is unanswerable in a response letter — you cannot argue the omission away, only add it, which is why the audit passes above are worth running before submission rather than after. When the comment does arrive, the drafting problem becomes a response letter problem.

Honest boundaries of this advice

A limitations section cannot rescue a study whose design cannot answer its question, and no wording makes an underpowered, confounded, or misdirected study publishable in a journal that would otherwise reject it. Disclosure changes how a paper is read; it does not change what the data support.

Conventions also differ by field. Physics, mathematics and theoretical computer science have no equivalent section, and a “Limitations” heading in a theory paper reads as out of register. Machine learning and NLP have moved the other way: ACL Rolling Review desk-rejects submissions that lack a dedicated “Limitations” section, and the NeurIPS paper checklist directs authors to write one. In qualitative research the frame is transferability and reflexivity rather than bias and generalisability, and importing the epidemiological vocabulary can itself draw an objection — the standards there are set by COREQ and SRQR and interrogated in qualitative rigour review. Some journals cap the Discussion tightly enough that three limitations is the practical maximum. Read recent papers in your target journal before deciding the shape.

Finally, the empirical base for this advice is thinner than it should be. The ter Riet survey sampled publications from 2007, its classification of limitations was in the authors’ own words “to some extent subjective”, and the audits published since — in dental, orthodontic, manual-therapy and psychology journals — are narrower in scope, and none has re-measured whether authors now temper their conclusions. Treat the 27% and 84% figures as the best available characterisation of the problem, not as a current measurement.

Related

Confounding · Collider bias · Registry data limitations · Post hoc power · Response to reviewers · Why papers get rejected

Before submission, the PerfectPaper overclaim check compares the claims in your title, abstract and conclusion against the limitations you stated and reports where the two do not agree.

Review my manuscript

Frequently asked questions

How do I write the limitations section of a research paper?

Name three to five specific threats to validity, and for each give the mechanism, the direction and likely size of the bias, and what you did about it. Order them by how much each could change your conclusion, then revise the conclusion, abstract and title to match what survives.

What should a limitations section include?

Sources of potential bias, imprecision in the estimates, and constraints on generalisability, each tied to a specific feature of your study rather than to research in general. CONSORT 2025 item 30 and STROBE item 19 define this scope, and STROBE additionally requires the direction and magnitude of any potential bias.

How should I phrase a study limitation so a reviewer accepts it?

Write a sentence a reviewer could disagree with. “Results should be interpreted with caution” is unfalsifiable and adds nothing; “recruitment from a referral centre inflates the absolute event rate while leaving the relative estimate largely intact” states a mechanism, a direction and a consequence that can be checked.

How do you write limitations without weakening the paper?

Bound each limitation rather than leaving it open. An E-value, a sensitivity analysis, a negative control outcome or a tipping-point analysis converts an unlimited worry into a stated quantity, which strengthens the paper. Weakness comes from vague concession, not from precise disclosure.

How do you write the limitations of a study in the discussion section?

Place the limitations after your interpretation of the findings and before your conclusion, under a labelled subheading, in roughly 200 to 400 words. A labelled subheading prevents the common reviewer comment that limitations were not addressed when they were in fact buried mid-paragraph in an unheaded Discussion.

How do I write limitations for a thesis or dissertation?

Use the three-part structure — mechanism, direction and magnitude, what was done — but expect to write more of them and in more detail, since examiners test whether you understand the boundaries of your own design. Distinguish limitations of the studies from limitations of the methods available in your field.

How many limitations should a paper have?

Three to five for a standard research article. The survey of 300 biomedical papers by ter Riet and colleagues found a median of three among papers that acknowledged any. Fewer than three usually means the obvious one was omitted; more than six usually means trivial items are diluting the ones that matter.