Citation Persistence

What it is

Tatsioni, Bonitsis and Ioannidis published the founding study in JAMA in 2007, on the persistence of contradicted claims in the literature. The strongest recent instance is Tom Hardwicke and colleagues in 2021, who examined citation patterns after strongly contradictory replications across four case studies in psychology. They found a small post-replication decline in favourable citations, a small rise in unfavourable ones, and that most articles citing the originals did not cite the replications. In the best-known of their cases, between thirteen and forty-one per cent of articles citing the original also cited the replication, depending on the year. About half of the articles citing both favourably offered an explicit defence of the original.

Independent work agrees in direction. A 2021 study in Psychological Science and a 2023 analysis in PNAS both report citation declines after failed replications that are real but modest.

Four case studies is a small base and this page says so.

In effect

This is the finding that makes retraction and replication weaker instruments than they sound. A study can be publicly overturned and remain, in practice, the thing a field keeps repeating, because citation is a social habit rather than a verification step.

It is also the reason a library like this one exists. If the scientific record self-corrected on contact with evidence, checking a claim would mean checking whether it had been challenged, and the literature would tell you. Instead the overturning paper and the overturned one sit side by side, with the overturned one usually better cited, and the only thing that separates them is somebody deliberately going and looking.

The instances in this vault are not coincidences. A retracted study still doing work in one book, a collapsed finding defended in a later afterword in another, an author's own study surviving its own correction in a third, and ego depletion running across four sources in four different broken ways: that is what citation persistence looks like from inside a reading list.

The mechanism has variants worth naming. A finding can persist after its citation has been stripped off altogether, as with the layered triune brain model, which is obsolete and is taught in one book twice as current brain science without its originator being named, so that a reader who wanted to check could not even begin. A misattribution persists the same way and travels faster, because nothing about it looks wrong and there is no replication attempt to fail: an external pass on one eight-hundred-page history confirmed ten of ten sampled errors, seven of them simple misattributions of a name, a date or a concept.

And the mechanism runs behind the origin of a claim as well as after it. A 2018 gift book with no apparatus names nine sources in its whole running text, and four of its most memorable stories descend from four nineteenth-century fabricators, concentrated in the small part of the book that looks sourced. A reader who did the thing this publication normally asks for, and followed the named source, would land on the forgery. A citation is not a verification.

What it does not say

It does not say that correction does nothing. The honest statement is that it reduces citation slightly, and the evidence for that is consistent across several independent analyses.

It does not say the effect size is known here. Figures from the primary papers are given in the sources below and nowhere else, and the full text of the 2021 paper has not been opened by this publication.

It does not follow from a book reporting the pattern. The chapter this vault entered it through states the finding and names nobody, so the identification of the primary papers here is ours rather than the book's.

It does not have a fix that operates downstream. Preregistration and replication govern how a study is made, and corrections applied afterwards do not propagate reliably. The only instrument that works is somebody going and looking.

A standing rule follows for this publication. Treat an attribution from a secondary synthesis exactly as a finding from one: as a pointer to be checked, never as the thing itself.

---