What a PMID Establishes, and What It Does Not

A citation that resolves to the wrong paper

Two papers appeared in the British Medical Journal in 1985. Black's commentary on rates of grommet surgery in children sits at 290(6486):1963-5 under PMID 3924327. De Lacey, Record and Wade's study of quotation accuracy in medical journals sits at 291(6499):884-6 under PMID 3931753. A reference list that pairs the second title with the first number is not broken in any way a reader would notice. The link opens. A real record appears, in the right journal, in the right year. The record is about ear surgery.

That is the ordinary shape of citation drift. Nothing is fabricated, the identifier is valid, and the only thing wrong is the correspondence between the claim and the document. Checking a citation for existence catches none of it, because existence was never in doubt. The check that catches it is slower: the record is opened, what the paper measured is read, and the number in the sentence is confirmed against the number in the paper, for the same species, by the same route, at the same endpoint.

What a PMID actually is

A PMID is an accession number. The National Library of Medicine assigns one to each record as it enters PubMed, and the PubMed User Guide states that PMIDs do not change over time or during processing and are never reused. That is the whole of the guarantee. It is a promise about the stability of a pointer, not about what the pointer aims at. No step in assigning the number involves reading the argument, checking the statistics, or judging whether the conclusions follow from the data.

Nor does a PMID mean the record was indexed for MEDLINE. MEDLINE is the curated subset, selected against published scientific-quality and editorial-practice criteria and marked by the assignment of MeSH headings. PubMed is larger. Counts retrieved from the E-utilities interface on 17 August 2026 return 41,022,049 PubMed records, of which 33,590,005 are indexed for MEDLINE and 6,768,851 sit in the pubmednotmedline subset, with 389,535 at publisher status and 273,559 in process. Those four subsets sum to within ninety-nine records of the total. Roughly one PubMed record in six carries a PMID and will never be indexed for MEDLINE.

The PubMed DTD documentation defines the last of those categories. Its Status attribute gives PubMed-not-MEDLINE as meaning the citation will not be processed for MEDLINE, alongside Publisher for ahead-of-print submissions and In-Process for records awaiting MeSH assignment. These are processing states. They describe where a record sits in a workflow, and none of them is a verdict on the article.

Four identifiers, four different things

Reference lists carry four kinds of identifier, and each answers a different question. The PMID indexes a record in PubMed. A PMCID indexes a full-text deposit in PubMed Central, an archive whose acceptance criteria are separate from MEDLINE's. A DOI is a registrant-purchased handle in a resolution system; the DOI Foundation describes it as a digital identifier of an object, any object, and notes that persistence is a function of organisations rather than of technology. A DOI resolving cleanly establishes that a publisher registered a record and kept the redirect working. An accession number in a domain database, such as a PubChem CID or a ClinicalTrials.gov registration, establishes that something was deposited under that string.

The chemical databases record the two layers separately. The 2025 PubChem update by Kim and colleagues reports 322 million substance records against 119 million compound records, drawn from more than a thousand sources. Substances are what depositors supplied; compounds are what standardisation produced from them. A CID therefore certifies that a structure was normalised into the collection, not that any property attached to it downstream was measured well or measured at all. Each of these identifiers answers whether a record exists. None answers whether the record says what the citing sentence says.

What the index does not screen out

Manca and colleagues, writing in CMAJ in 2018, named the route: PubMed Central runs its own journal selection, and material accepted into that archive is searchable in PubMed with a PMID attached. MEDLINE selection passes through the Literature Selection Technical Review Committee, and MEDLINE indexed roughly 5,600 journals against approximately 30,000 in PubMed; the PubMed Central pre-application asked instead for an ISSN, a two-year publication history and twenty-five peer-reviewed articles, and the authors recorded that the National Library of Medicine may waive the two-year history where a publisher can show comparable editorial experience elsewhere. A record admitted by that door carries a PMID identical in form to one carried by a paper in a journal that passed committee review.

Retraction is recorded rather than enforced. A search of the Retracted Publication publication type on 17 August 2026 returns 33,937 records, flagged and still resolving under their original identifiers. The flag appears on the PubMed record as a Retracted Publication publication type with a linked retraction notice. A title prefix is added by some publishers and not by the National Library of Medicine, so its wording varies and it is absent from a large share of flagged records altogether. Neither device appears in a bibliography compiled before the retraction, and neither appears in a downstream sentence that inherited the figure from a review.

Two studies put numbers on what follows. Budd, Sievert and Schultz searched MEDLINE from 1966 to August 1997, found 235 retracted articles, and counted 2,034 citations to them after the notices appeared; of 299 citations examined, 19 acknowledged the retraction and 280 treated the work as valid. Bolland, Grey and Avenell reported in 2022 on a research group with 113 published concerns regarding publication integrity, where 376 source publications drew 6,926 references from 5,577 citing articles, and 198 of those references appeared after the relevant concern was published.

How often the cited paper fails to contain the claim

Measurements of quotation accuracy begin in the mid-1980s and run through at least 2020. De Lacey, Record and Wade assessed six medical journals published in January 1984 and reported that the original author was misquoted in 15 per cent of references, and that citation errors occurred in 24 per cent, 8 per cent of them major enough to prevent identification of the source. Their stated conclusion was that inaccurate quotation is one route by which untruths enter the record as accepted fact. Evans, Nadjari and Burchell repeated the exercise for three surgical journals in JAMA in 1990, reporting thirteen major and forty-one minor citation errors alongside thirty-seven major quotation errors.

One detail cannot be settled from the 1990 record. Its abstract states that fifty randomly selected references were drawn from a single monthly issue of the three named journals, without making clear whether fifty is the total or the count per journal, so the denominator behind those error counts is underdetermined without the full text. Schulmeister's 1998 study of three nursing journals is unambiguous by comparison: 60 references per journal, 180 in total, 58 carrying citation errors, and 12 containing a major quotation error, including four where the cited article contradicted or was unrelated to the assertion drawn from it.

The two systematic syntheses do not agree. Jergas and Baethge screened 559 studies, included 28, and estimated total quotation errors at 25.4 per cent, 95 per cent CI 19.5 to 32.4, splitting into 11.9 per cent major and 11.5 per cent minor. Mogull, reviewing 15 studies in 2017, arrived at 14.5 per cent, 95 per cent CI 10.5 to 18.6. The gap is a denominator: Mogull counted errors against quotations examined, where the earlier convention counted them against references selected. Both are defensible and they are not interchangeable. Mogull also found 64.8 per cent of content errors to be major, meaning the cited source failed to substantiate, was unrelated to, or contradicted the assertion.

Baethge's 2020 case study addressed the objection that such errors might be trivial. Across 72 psychiatric index articles from five German-language journals, 51 were quoted 235 times in 109 quoting articles; 96 per cent of those quotations were rated of medium or high importance to the quoting paper's message, and 22 per cent were rated major errors against 19 per cent minor. The errors sat in load-bearing sentences.

Citation as a social mechanism

Greenberg's 2009 network analysis in the BMJ moved the unit of analysis from the single citation to the pattern it sits in. He assembled the complete citation network of English-language PubMed-indexed papers addressing one claim — that beta amyloid was produced by, and injured, skeletal muscle in inclusion body myositis — and found 242 papers linked by 675 citations, generating 220,553 citation paths in support of a proposition none of them had tested. The mechanisms he identified were citation bias against papers that refuted or weakened the claim, amplification by papers presenting no data on it, and invention, including the conversion of hypothesis into fact through citation alone.

Invention is the mechanism behind most of the untraced figures logged on this site. A cautious sentence in a primary paper becomes a flatter sentence in a review, the review is cited in place of the paper, the hedge is dropped, and the number arrives on an aggregator page behind a citation trail that is technically continuous and substantively empty. Every link in that chain resolves.

What the ledger does with this

The claim ledger on each compound page records species, route, sample size and source for every figure, and it does so for a mechanical reason. Those four fields are the ones citation drift alters silently. A number survives transplantation from rat to human, from intravenous to oral, from a cell assay to a whole animal, and from an n of six to an implied general finding, without any visible break in the citation. Writing the four fields beside the number makes the substitution visible, because the transplanted version cannot be filled in.

Verification means the claim and the source are checked against each other, not each against the other's existence. A PMID is confirmed to resolve, and then the retrieved record is read to confirm that it contains the figure attached to it, for the stated population and route. Where a widely repeated number traces to no primary source, it is listed and marked as untraced rather than deleted, with a note on what was searched. A claim with no origin is a finding about the record, and removing it would leave a reader to meet it elsewhere with nothing attached.