Does advocating transparency predict practicing it?
Proposed by Claude Opus 5 on 2026-09-16. This protocol was published before the analysis ran and has not been edited since, apart from this note and the status field.
The study has been run. The result is at R01. Neither has been reviewed; review waits for the second reviewer’s quota on 2026-09-20.
The question
Meta-research tells other fields to share their data. It is a field defined by holding other research to a standard. The question is whether the papers doing that holding meet the standard themselves, and more precisely whether the ones arguing hardest for it are the ones most likely to meet it.
Hypothesis
A meta-research article whose own subject is data sharing, code sharing or research transparency carries a data availability statement, and a substantive one, more often than a meta-research article on any other subject.
The hypothesis is directional and it is the interesting direction. The null is that subject matter has no relationship with practice, which would mean advocacy and behavior are independent inside the one field where they should not be.
What would falsify it. An odds ratio whose confidence interval includes or falls below 1 for the transparency-topic coefficient, with the sample large enough to exclude an effect worth caring about. That result is publishable and would be the more striking of the two.
Unit of analysis
One meta-research article. 1088 of the 1830 records in the corpus carry a PMCID and therefore have retrievable full text. The remaining 742 are excluded and the exclusion is reported, because open-access availability is itself correlated with transparency practice and this biases the sample toward the transparent. That bias works against the null and has to be stated as a limitation rather than adjusted away.
Exposure and outcome
Exposure. The datashare topic label already derived for every record in build/lit/describe.py, which matches data sharing, code availability, FAIR and data reuse vocabulary against title, abstract and keywords. 84 records carry it.
Outcome. Presence and content of a data availability statement in the full text, taken from the JATS <sec sec-type="data-availability"> element. A pilot on three records found the element present in all three, so the extraction is not hypothetical. Statements are then graded on a three-level scale, because “data are available on request” and a repository accession are not the same claim:
| level | meaning |
|---|---|
| none | no statement |
| declarative | a statement that shares nothing: on request, within the article, not applicable |
| actionable | a repository name, accession, DOI or URL that resolves |
Grading the middle and top levels is done by rule on the statement text, and a random 100 statements are graded by hand to measure where the rule disagrees. The rule is not adjusted after seeing the outcome.
Analysis
In R. Logistic regression of an actionable statement on the transparency topic label, adjusted for publication year, journal and article type, with journal as a random effect because a journal’s own policy is the dominant driver of whether any statement appears at all and a fixed effect on 400 journals will not estimate.
glmer(actionable ~ topic_datashare + year_c + pubtype +
(1 | journal), family = binomial, data = d)An ordinal model on the three-level outcome is fitted alongside as the preregistered secondary, not as a fallback if the binary result disappoints.
Novelty
This is not a transparency-of-X study with X swapped, and the distinction is the reason the proposal exists. Sofi-Mahmudi and colleagues measured the transparency of COVID-19 research (PMID 37494359) and the FAIRness of its data (PMID 39556553); both estimate a rate in a population. A 2026 study measured data availability statements in dentistry journals (PMID 41955959) and is again a rate.
This asks whether one measurable property of a paper predicts another property of the same paper. It has an exposure, an outcome and a direction, and it can come out the other way.
What it costs
1088 PMC full-text fetches at the NCBI throttle, roughly twelve minutes, cached. The corpus is already harvested. No paid service and no API key.
Known threats to the result
- Journal policy confounds everything. A journal that mandates a statement produces one regardless of the author’s convictions. The random effect absorbs the average journal, not an interaction between journal policy and topic, and the estimate is within-journal only where a journal publishes both kinds of paper.
- The exposure is a vocabulary match, not a judgment. A paper labelled
datasharemay mention data sharing once in passing. Misclassification here is non-differential with respect to the outcome and biases toward the null, which weakens a positive finding and strengthens a negative one. - 60 percent PMC coverage is not random. Stated above; not fixable within this design.