Before sharing a striking health story, compare the claim in three places: the paper, the press release and the headline. This case study turns that comparison into a reading exercise.

THE SHORT READ
  • Trace the wording back to the research question.
  • Replication can support one finding while leaving another uncertain.
  • A useful caveat explains a specific limit, close to the claim it qualifies.
THE STUDY AT A GLANCEUniversity press-release replication
Study material
348 university press releases
Setting
UK, 2014–2015 samples
Design
Observational replication
Research date
2019; historical case study

Original publication: 18 Nov 2019 · The date above refers to this brief.

THE NUMBERS, IN CONTEXT

Causal exaggeration in linked news

Percentage of associated news reports; model-based estimates

Release exaggerated the causal claim82
Release did not exaggerate it16
0100
Replication study: 95% confidence intervals 68–91% and 10–26%, respectively. These are selected news-report groups, not all 348 releases. Bars show an association, not the effect of editing a release.

What the researchers compared

Bratton and colleagues compared health reporting with its underlying papers. They found a link between stronger causal language in press releases and stronger causal language in news. The analogous association for advice did not replicate. Multiple news stories could come from one release, so the analysis accounted for clustering. This is evidence about a communication pathway, not a census of today’s media.

Source 1 ↗

What changed when researchers intervened?

A separate randomised study worked with 312 releases from nine press offices. It tested claim alignment and causal caveats. Assigned-group differences were small; comparisons based on the wording actually used were larger, but those comparisons were observational. The trial found no evidence that careful wording reduced news uptake. This does not establish that every cautious headline will attract equal attention.

Source 2 ↗

Connect the studies without merging them

The earlier 2014 study sampled 462 releases and 668 news stories. The later replication therefore supplies another cohort, while the press-office experiment asks a different question: can changing communication improve what is reported? Treat those as distinct contributions. Do not average their percentages as though their samples, definitions and interventions were interchangeable.

Source 3 ↗Source 1 ↗Source 2 ↗

Try a claim audit on one story

Original exercise: imagine a fictional paper measuring sleep and test performance in volunteers. A release calls the two linked; a headline promises that extra sleep will raise anyone’s score. Write three short entries: who was measured, what was measured, and which word strengthened the claim. Then rewrite the headline using only the information actually available. Keep the population visible, and leave the causal question open. This invented example is not a finding from the studies above.

A useful next research question

Our proposal: test a clearly specified headline change in a new setting and measure both reach and reader understanding. A successful communication intervention should be evaluated against its intended outcome. Clicks, accurate recall and informed interpretation belong in separate columns. That proposal is a future test, not evidence that a particular format already works.

CONNECT THE EVIDENCE

Three roles in the evidence chain

EvidenceWhat it contributesWhat remains open
Original observational workIdentifies a pattern worth checkingCannot assign the effect of changing wording
ReplicationChecks whether the pattern travelsDifferent sampling and coding may matter
Communication experimentTests an editing strategyAdoption of assigned changes must be examined

A reading map, not a pooled estimate. The linked sources describe separate research questions.

READER QUESTIONS

Your questions, answered

Does a strong association prove that the release caused the news wording?

No. For this question, look for the intervention comparison as well as the observational pattern. The larger as-treated contrast should not inherit the protection of random assignment.

Did every kind of exaggeration replicate?

No. The replication’s advice finding differed from its causal-language and human-inference findings.

Can an accurate headline still be interesting?

Try specificity: name the population, outcome and remaining question. That is an editorial exercise, not a promise about readership.

What should I save while checking a story?

Save the paper’s identifier, the release, the headline and the access date. Write your interpretation separately from any reported result.

LIMITATIONS

Limits of this interpretation

  • Historical UK samples do not estimate the rate of exaggeration in all current media.
  • AI checked the cited passages, not the underlying coding datasets.
  • This comparison does not measure whether readers changed treatment or behaviour.
SOURCE NOTES

Sources & transparency

  1. Bratton et al. (2019): exaggeration in university press releases — replication

    Public full-text HTML: sampling, results and discussion checked. Raw coding files were not reanalysed. · Accessed 27 Sep 2026

    DOI: 10.12688/wellcomeopenres.15486.2
  2. Adams et al. (2019): Claims of causality in health news — a randomised trial

    Public full-text HTML: abstract, intervention design and interpretation of intention-to-treat versus as-treated results. · Accessed 27 Sep 2026

    DOI: 10.1186/s12916-019-1324-7
  3. Sumner et al. (2014): academic press releases and health news

    Indexed primary-paper abstract and bibliographic record. The publisher full text could not be retrieved during this check. · Accessed 27 Sep 2026

    DOI: 10.1136/bmj.g7015

Prepared with AI using the source passages described below. No human editorial or clinical review has been completed. Reading exercises and future research questions are our interpretation, not additional study findings. This is an educational case study about research reporting, not a treatment recommendation.

Source check: AI source check — primary-source passages and reported numbers; 27 September 2026

Clinical review: Not applicable to this educational guide

Suggest a correction