How do you write a research paper abstract?

A research paper abstract is a self-contained summary of one study, usually 150 to 400 words, written last from the finished manuscript. The abstract makes six moves: context, objective, design and setting, the primary result with its effect size and precision, one conclusion scoped to that design, and the identifiers — trial registration, data accession, funding — the journal requires.

Those six moves are easy to list and hard to execute, because the abstract is the one part of a paper that must survive being read with nothing around it. This page covers when to draft it, how to allocate the words, what the ICMJE and the reporting-guideline checklists require it to contain, how to write results sentences that carry numbers rather than adjectives, how to audit a finished abstract against the manuscript, and what reviewers say when the abstract and the paper disagree.

Write the abstract last, from the final tables

The abstract is drafted after the analysis is frozen and the tables are final, using the numbers exactly as they appear in those tables. An abstract written early — during the analysis, or from the grant proposal — is the largest single source of abstract–body inconsistency, because the estimates move and the abstract does not move with them.

Pitkin, Branagan and Burmeister (JAMA 1999;281:1110–1111) audited abstracts in six general medical journals, counting an abstract as deficient when it contained data “either inconsistent with corresponding data in the article’s body (including tables and figures) or not found in the body at all.” The proportion of deficient abstracts ranged from 18% to 68% depending on the journal. That range is not a measure of carelessness in obscure venues; it comes from Annals of Internal Medicine, the BMJ, JAMA, the Lancet, the New England Journal of Medicine and CMAJ.

The practical consequence is a drafting order, not an attitude: finish the analysis, lock the tables, then transcribe from them.

Common questions about when to write the abstract

Should I write my abstract before or after I finish the rest of the paper?

Write it after, from the frozen tables. An abstract drafted early — during data collection, or from the grant proposal — is the largest documented source of abstract-body inconsistency, because estimates move during analysis and revision and the abstract does not move with them unless you deliberately rebuild it from the final results each time you revise.

Is it OK to draft the abstract from my grant proposal or thesis proposal instead of the finished results?

No. A proposal states what you planned to test, not what you found, and an editor sees the gap immediately. Draft the objective sentence from the proposal if the study still matches it, but take every number in the results sentence from the final, frozen tables — never from a power calculation or a preliminary run done months earlier.

The six moves an abstract makes

An abstract makes the same six moves whether or not the journal labels them with headings.

Context, in one or two sentences, ending at the specific gap this study addresses. A field-level truism (“cancer is a leading cause of death”) spends words a reader already has.

Objective, in the grammar of the design: to estimate a quantity, to test whether an intervention changes an outcome, to characterise a population. The verb tells the reader what kind of answer to expect.

Design, setting, participants and measures, naming the design by its accepted label — randomised, parallel-group, double-blind; prospective cohort; nested case-control; cross-sectional — and naming the primary outcome and how it was measured.

Results, giving the primary outcome in both groups with denominators, the effect estimate, and its precision.

Conclusion, one sentence, answering the stated objective with the primary outcome and with the verb the design licenses.

Identifiers, at the end: registration number, data repository and accession, and funding source.

How to allocate 250 words

Word allocation in an abstract is where most drafts go wrong, and the error is consistent in direction: background expands and results contract. In a 250-word abstract, the results block should be the longest of the six, and the background the shortest thing that still frames the question.

Counting the words in your own draft against this allocation takes two minutes and is the fastest structural diagnostic available. A draft with 90 words of background and 40 words of results has a reporting problem the prose cannot disguise, and it will read to an editor as a paper whose findings are thin. Broader drafting questions belong on academic writing; this page stays with the abstract.

Structured and unstructured abstracts

A structured abstract carries explicit labelled sections — commonly Background, Methods, Results and Conclusions, sometimes expanded to Objective, Design, Setting, Participants, Interventions, Main outcome measures, Results and Conclusions. The ICMJE Recommendations state the requirement plainly: “Original research, systematic reviews, and meta-analyses require structured abstracts.”

The format dates to a specific proposal: the Ad Hoc Working Group for Critical Appraisal of the Medical Literature, “A proposal for more informative abstracts of clinical articles,” Annals of Internal Medicine 1987;106:598–604. Hartley’s study in Applied Ergonomics (1999) found that structured abstracts were longer, more informative, and judged clearer by readers than the traditional versions they replaced.

The honest limitation is worth stating. The evidence for structured abstracts comes largely from within-journal before-and-after comparisons and from reader-judgement studies, not from randomised comparisons of formats. The claim that structure increases the amount of information reported is well supported; the claim that it improves reader comprehension or editorial decisions is not established to the same standard.

An unstructured abstract — the norm at Nature, Science, and across physics, mathematics and computer science — removes the labels, not the content. The same six moves appear as continuous prose, and the reader still expects to find the design and the effect size.

The heading set itself is not standardised even among journals that require structure. The 1987 four-part proposal names Background, Methods, Results and Conclusions, but individual journals diverge from it in both directions: some split Methods into separate Design, Setting and Participants headings to match a clinical-trial or observational template, others substitute “Purpose” for “Background” or add a stand-alone “Limitations” heading the 1987 proposal never specified. Copying a generic structured template rather than the target journal’s own recently published articles is a frequent, avoidable source of a returned-for-formatting notice.

Questions about structured abstracts

Do I have to write a structured abstract, or is that only for medical journals?

It depends on the journal and the article type, not the field. The ICMJE Recommendations require structured abstracts for original research, systematic reviews and meta-analyses; many biomedical journals extend that to case reports and brief reports, while physics, mathematics and most computer-science venues keep unstructured, continuous-prose abstracts regardless of study type.

What headings should I use in a structured abstract if the journal doesn’t specify any?

Default to Background, Methods, Results and Conclusions — the four-part set the Ad Hoc Working Group proposed in 1987 and that most journals still use unchanged. Check the target journal’s own recently published articles first, though: some substitute “Purpose” for “Background”, or split Methods into separate Design, Setting and Participants headings, and matching their exact label set avoids an editorial query before review even starts.

What the ICMJE Recommendations require in an abstract

The ICMJE Recommendations set four obligations that drafts routinely miss, and each is quotable verbatim from the Preparing for Submission section.

On content: the abstract “should provide the context or background for the study and should state the study’s purpose, basic procedures (selection of study participants, settings, measurements, analytical methods), main findings (giving specific effect sizes and their statistical and clinical significance, if possible), and principal conclusions.”

On restraint: it “should emphasize new and important aspects of the study or observations, note important limitations, and not overinterpret findings.” Naming a limitation inside the abstract is unusual in practice and is explicitly asked for here.

On registration: “The ICMJE recommends that journals publish the clinical trial registration number at the end of the abstract.” Endpoint and registration consistency is checked by trial registration and endpoints.

On data: “If the data have been deposited in a public repository and/or are being used in a secondary analysis, authors should state at the end of the abstract the unique, persistent data set identifier, repository name and number” — the requirement behind data availability and accessions.

The reporting-guideline checklists that govern abstracts

Three reporting guidelines specify abstract content directly, and a submission is scored against whichever one matches the design.

CONSORT for Abstracts (Hopewell, Clarke, Moher, Wager, Middleton, Altman, Schulz; Lancet 2008, with explanation and elaboration in PLoS Medicine 2008;5(1):e20) covers journal and conference abstracts of randomised trials. It asks for identification of the study as randomised in the title, trial design, eligibility criteria and setting, the interventions, the specific objective, a clearly defined primary outcome, how participants were allocated, whether blinding was used, the number randomised and the number analysed in each group, the result for the primary outcome with its effect size, important adverse events, a general interpretation, and the registration number and funding source. The reason the extension was written is that the reviews behind it found trial abstracts routinely missing those items — the number of participants actually analysed, the size of the effect rather than its significance, the harms, and how allocation was concealed — with conference abstracts among the weakest — in one review of 494 oncology conference abstracts, only nine (2%) reported the method of allocation concealment. Treat the checklist as a presence test: read your own abstract with the item list beside it and mark each item found or absent.

PRISMA 2020 for Abstracts (from Page et al., BMJ 2021;372:n71) is a 12-item checklist for systematic reviews. Item 9 is easy to leave out: “Provide a brief summary of the limitations of the evidence included in the review (e.g. study risk of bias, inconsistency and imprecision).” Items 11 and 12 require the funding source and the register name with registration number.

STROBE has no published abstract extension paper — its website posts only a draft checklist for conference abstracts — so a journal abstract is judged against item 1(b) of the main checklist, which asks authors to provide in the abstract an informative and balanced summary of what was done and what was found. Balanced is the operative word: an observational abstract that reports the adjusted estimate and omits the unadjusted one fails item 1(b) without stating anything false.

Writing the results sentences

A results sentence in an abstract needs four elements: the comparison, both groups’ values with their denominators, the effect estimate, and the precision of that estimate. A p value supplies none of the four, and an adjective supplies none of them either.

Compare a common draft with its repaired form.

Draft: “Treatment significantly reduced 12-month mortality (p = 0.03).”

Repaired: “Mortality at 12 months was 12.4% (31/250) in the treatment group and 18.8% (47/250) in the placebo group (absolute risk difference −6.4 percentage points, 95% CI −12.7 to −0.1).”

The repaired sentence is 29 words against 8, and it is what the reader needs in order to judge whether the effect matters — the question behind is the effect size meaningful. Three further habits belong here. Report the number analysed, not only the number randomised, because the gap between them is the attrition a reader must weigh. Report harms in the abstract of any interventional study, even in one clause. And give the denominator with every percentage, since a percentage without its denominator cannot be checked against the body.

A ratio measure carries its own, separate trap. “Risk reduced by 35% (relative risk 0.65)” reads identically whether the absolute risk fell from 40% to 26% or from 2% to 1.3% — the relative figure is the same in both cases, and the clinical or practical importance is not. Report the relative measure alongside the absolute one, or state the absolute measure alone; a relative-risk figure with no absolute risk beside it is a true number that still produces a false impression of the underlying difference.

The conclusion sentence, and the spin that lives in it

The conclusion sentence of an abstract answers the stated objective, using the primary outcome, with the verb the design supports. The distortion documented in abstracts concentrates in this one sentence.

Boutron, Dutton, Ravaud and Altman (JAMA 2010;303:2058–2064) examined 72 randomised trials whose primary outcome was statistically non-significant and coded them for spin — reporting that distracts from that result. Spin appeared in the abstract conclusions of 58.3% (95% CI 46.1–69.8%), in the abstract results sections of 37.5%, and in the titles of 18.0%. The abstract conclusion was the most affected part of the whole paper.

The devices are recognisable and few: promoting a significant secondary outcome to the conclusion; reporting a significant within-group change instead of the between-group comparison; leading with a subgroup; presenting a per-protocol analysis as the result; and declaring equivalence or safety from a non-significant superiority test, which the test cannot establish. The verb is the tell — “reduces”, “improves” and “causes” require randomisation, while an observational design supports “was associated with”. That boundary is the subject of causal language discipline and of the reviewer objection at correlative, not causal; overreach beyond the data is what overclaim checking reports.

Non-inferiority and equivalence trials need a different conclusion grammar entirely, and the ordinary spin checklist above does not cover the failure mode specific to them. A superiority-test p-value greater than 0.05 means the study failed to detect a difference; it does not mean the two treatments are equivalent, and “no significant difference was found, so the treatments are equivalent” is a distinct error from the five spin devices already listed. The conclusion of a non-inferiority abstract has to state instead whether the observed effect and its confidence interval fell entirely within the margin the protocol pre-specified — a narrower and more specific claim than “no difference.”

Common questions about abstract results and conclusions

Can I say my result was “significant” in the abstract without giving a p-value or a confidence interval?

No — “significant” with no number attached is exactly the sentence reviewers flag, because it hides whether the effect is large enough to matter to anyone. Report the effect estimate with its precision, a confidence interval rather than only a p-value: “6.4 percentage points, 95% CI −12.7 to −0.1” tells a reader something “significantly reduced” never does.

My primary result wasn’t statistically significant — how do I write the conclusion without it looking like I’m hiding that?

State the result and its confidence interval plainly, then draw only the conclusion that interval supports — usually “this study did not detect a difference” rather than “no difference exists”, since a non-significant result is not proof of equivalence. Absence of evidence for an effect is not evidence of its absence; report what the data show, not what you had hoped they would show.

Should my abstract say a treatment “reduces” an outcome, or that it “is associated with” a change in it?

Use “reduces”, “improves” or “causes” only when the design is randomised. An observational study — cohort, case-control, cross-sectional — supports “was associated with” and nothing stronger. Reviewers read this verb as a checkable claim about your design, not a style choice, and a causal verb sitting over observational data is one of the most common single-sentence objections an abstract draws.

Title, keywords and what actually gets indexed

The title and abstract are the portion of a paper indexed in bibliographic databases and, for most readers, the only portion ever read. Every term a searcher would type must therefore appear as a literal string in the title or the abstract — the design phrase included, spelled the way filters expect it (“randomised controlled trial”, “retrospective cohort study”, “genome-wide association study”).

Two mechanical constraints follow. Undefined abbreviations are unsearchable and are usually prohibited: the PLOS ONE submission guidelines direct that the abstract should not include citations, and should avoid abbreviations if possible. And an abstract that names the intervention only by a trade name, or the population only by a cohort acronym, disappears from searches that use the generic term.

The keywords field is indexed separately from the title and the abstract text, and a term that never appears in either still needs to land there. PubMed maps author-supplied keywords onto its own MeSH vocabulary during indexing rather than treating them as literal search strings, while Scopus and Web of Science index author-supplied keywords as their own distinct, filterable field, separate from a full-text search of the abstract. A term absent from both the prose and the keyword list is invisible to a database search restricted to that field, even where a human reader would recognise the same concept from context alone.

The abstract is also the screening surface. Editors decide scope from it, reviewers accept or decline invitations on it, and readers decide whether to open the paper on it — which is why an abstract that hides the design is a common route to being desk rejected without review.

How to audit an abstract against the finished manuscript

Auditing an abstract is a mechanical procedure, not a matter of judgement, and it takes about twenty minutes on a finished paper. Seven passes, in this order.

Number audit. List every numeral in the abstract in one column, and in the next column the table, figure or line of the body where that same value appears. Any numeral with no location is a defect by the Pitkin definition, whether it is wrong or merely unsupported.

Outcome audit. Confirm that the primary outcome named in the abstract is the primary outcome named in the methods and the primary outcome in the registry record. A discrepancy between these three is outcome switching, and reviewers of registered studies check it directly.

Denominator audit. Confirm that every percentage carries its n/N, and that the Ns sum to the number randomised or enrolled.

Attrition audit. Confirm that the number analysed appears, not only the number randomised, and that the difference between them is explicable from the body.

Verb audit. Circle the main verb of the conclusion sentence and ask which design would license it. Cross-sectional and retrospective designs support association verbs only.

Cover test. Read the abstract alone, write down every claim a reader would take from it, then open the paper and mark each claim as supported, partially supported, or absent. Claims in the third category are the ones that generate reviewer comments.

Presence check. Confirm the design label, the primary outcome, the effect size with precision, harms where applicable, the limitation sentence the ICMJE asks for, and the registration and data identifiers.

PerfectPaper runs this audit on the uploaded .docx and reports each abstract figure alongside the location in the body where the same value appears, or its absence. Everything it reports is a pointer for the author to verify against the tables, not a substitute for that check.

How abstracts differ outside clinical medicine

Abstract conventions outside clinical medicine keep the six moves and change their weights.

In computer science and engineering, abstracts run roughly 150 to 250 unstructured words and are built around a contribution claim plus the evaluation that supports it: what was built, what it was compared against, on which benchmark, and by how much it differed. A contribution claim with no evaluation named in the abstract reads as a position paper.

In mathematics, the abstract states the theorem, the hypotheses it requires, and its relation to the previously known result. Methods description is minimal because the proof is the method.

In qualitative and mixed-methods research, the abstract names the approach (grounded theory, framework analysis, interpretive phenomenological analysis), the corpus and how participants were recruited, the number of participants or documents, and the analytic claim — the reporting expectations behind qualitative rigour. Reporting a qualitative study’s findings as percentages in the abstract is a frequent and avoidable objection.

In preclinical work, the abstract should state the species, strain and sex of the animals, the number of independent biological replicates, and whether randomisation and blinding were used — the same four facts a reader of clinical trial abstracts expects for human studies.

A registered report’s abstract is a different object depending on the stage, and reusing one stage’s abstract unchanged at the other is a distinct and avoidable error. The Stage 1 abstract — reviewed and accepted in principle before data collection begins — states the hypothesis, the planned design and the pre-registered analysis plan, and contains no results, because none exist yet. Only the Stage 2 abstract, submitted after the study has run according to that plan, adds the results and the conclusion; a Stage 2 abstract that still reads like a proposal is the tell that the author copied it forward instead of rewriting it from the completed study.

Word limits, stated honestly

Abstract word limits are set per journal and per article type, and no standard exists across journals. The PLOS ONE submission guidelines specify that the abstract should not exceed 300 words; other journals set limits above and below that, and the same journal often sets different limits for research articles, brief reports and reviews. The only authoritative number is the one in the target journal’s author instructions on the day of submission, which is a reason to choose the journal before finalising the abstract — see journal submission.

Two practical points follow. Many submission systems enforce the limit programmatically, so an over-length abstract can be blocked before an editor reads a word of it. And no optimal abstract length has been established within a journal’s permitted range: the evidence that exists concerns discoverability rather than acceptance, and specific word-count targets circulate far more confidently than any study supports. Write to the number of facts the six moves require, then cut to the cap.

A separate, growing requirement sits beside the technical abstract rather than inside its word count: a lay or plain-language summary, written for a reader outside the field, with its own vocabulary rules and its own limit set independently of the abstract’s. Treat it as an independent piece of writing rather than a simplified copy of the technical abstract — a plain-language summary that still carries an undefined abbreviation or a statistical term has not actually been translated.

What reviewers and editors say when the abstract is wrong

Reviewers comment on abstracts in a small and repetitive set of phrasings, and recognising them in advance is the cheapest revision available.

“The effect size reported in the abstract does not match Table 2.” “The abstract reports a benefit that the primary analysis does not support.” “The abstract states the primary outcome as X, whereas the registration lists Y.” “The conclusion is stated causally; the design is cross-sectional.” “Please report the primary outcome with a confidence interval rather than a p value alone.” “The number of participants analysed is not given.” “No adverse events are mentioned in the abstract.” “The study design is not stated in the abstract or the title.” “The abstract’s conclusion concerns a secondary outcome.”

Editors add a shorter and more consequential set at the desk stage: “On the basis of the abstract, the study appears to fall outside the journal’s scope,” and “the abstract does not make the contribution clear.” Those two sentences end papers without review, which is why the abstract deserves a disproportionate share of the final week — see why papers get rejected.

Frequently asked questions

Why do I keep getting desk rejected — could it be my abstract?

Often, yes. Editors screen scope and contribution from the abstract alone before a paper is ever assigned to reviewers, and “the abstract does not make the contribution clear” or “this appears outside our scope, based on the abstract” are two of the most common desk-rejection phrasings on record. An abstract that buries the design or the actual finding reads as out of scope even when the study is not.

A reviewer said my abstract doesn’t match my results — what do they mean?

It means a number, an outcome or a claim in your abstract doesn’t reappear in the same form in your tables, figures or results text — the specific inconsistency Pitkin and colleagues found in 18% to 68% of abstracts across six major medical journals in 1999. Run a number audit before resubmitting: list every abstract numeral and find its exact match in the body.

Is a research abstract supposed to be 250 words or 300 words?

Neither is universal. PLOS ONE caps a research-article abstract at 300 words; many journals set 250; some allow up to 350 for a review article, and the same journal often sets a different limit for a brief report than for a full paper. There is no cross-journal standard, so 250 is a reasonable default to draft against, but the binding number is whatever your target journal’s current author instructions state.

Do conference abstracts need the same things as a journal abstract?

Mostly yes, but a conference abstract is shorter and more often skips items a reader assumes are covered. CONSORT for Abstracts covers conference and journal abstracts of randomised trials under the same checklist and asks both for the number randomised, the primary result with its effect size, and how allocation was concealed — items reviews found conference abstracts reporting the least often of the two.

Related

Academic writing · Journal submission · Overclaim checking · Responding to reviewers · AI peer review

Review my manuscript