Is privacy invoked to withhold data that involve no people?
Written against the design standard. Related work of the author’s examined the phrase “data not shown” in dental research (PMID 35944050), another form of withholding in plain sight.
1. The question
“Data cannot be shared publicly because of ethical and privacy restrictions” is a legitimate sentence when a study holds identifiable data about people. In a study of mice, cell lines, or published articles it is a sentence with no referent. If such statements appear in studies with no human data, the justification is being used as a formula rather than a reason, and the rate at which it happens is a direct measure of pretext.
2. Prior work, and the search behind the novelty claim
Log key P27.
corpus regex: (data availability statement|data sharing statement).{0,120}
(privacy|confidential|ethic|restrict)
PubMed: ("data availability statement*"[tiab] OR "data sharing statement*"[tiab])
AND (privacy[tiab] OR confidentiality[tiab] OR restrict*[tiab]
OR justification*[tiab])Twenty PubMed records, three in the corpus. The relevant ones find trial authors unwilling to share raw data (PMID 30738169), describe sharing practices among Canadian funded publications (PMID 42116839), and describe data statements in medRxiv preprints (PMID 33983972). None tests whether a privacy justification fits the data it is attached to.
3. Hypothesis
Primary. Privacy or ethics is invoked to restrict access in studies with no human participant data, at a rate whose 95 percent lower confidence bound exceeds 1 percent of restricting statements.
Secondary, relational. Within journal, privacy-justified statements in studies without human data are more similar to the journal’s most common statement text than privacy-justified statements in studies with human data are, which is what boilerplate predicts.
What would falsify them. A lower bound at or below 1 percent for the primary. No difference in similarity for the secondary.
4. Design and unit
Cross-sectional. One article with a restricting statement is one unit.
5. Sample
PMC open-access research articles 2019 to 2025 whose data availability statement, located by rt_data_code_pmc, restricts access (on request, controlled access, or not available). Statements that share data openly are excluded because they restrict nothing.
6. Classification of the study
No human participant data: animal, in vitro, in silico, and studies of published literature. Human data: any identifiable or de-identified individual-level human data. Classified from methods text by a model gated on the author’s labels, with MeSH Humans as a sensitivity check rather than the primary rule, because MeSH tags human relevance, not human data.
7. Outcome
The statement invokes privacy, confidentiality, ethics approval, or consent as a reason. Coded from the statement text by a rule published with the protocol, and gated.
8. Validation gate
The author labels 200 articles for human data and 200 statements for privacy justification, each against a written rubric. Kappa on each binary contrast; below 0.80 the author labels every article in the no-human-data group, which is expected to be small and is where the primary estimate lives.
9. Analysis
In R. The primary is a proportion with a Wilson interval, reported with its numerator and denominator. The secondary compares cosine similarity to the journal modal statement between groups with a linear mixed model:
lmer(sim_to_modal ~ human_data + (1 | journal), data = subset(d, privacy_reason))10. Threats
| threat | handling | removed or disclosed | standard |
|---|---|---|---|
| Some non-human studies do hold sensitive data, such as location data for endangered species | The rubric records these as a legitimate-restriction class | Disclosed | S8 |
| Journal templates may state privacy for every article | That is what the secondary tests | Disclosed | |
| Statement detection misses some statements | Detector sensitivity 0.765 for data sharing; missed statements reduce the denominator, not the rate, unless missing differs by class | Disclosed | S3 |
| Human-data classification | Gated, with full labeling of the key group on failure | Disclosed | S8 |
11. Probes
| probe | computes | goes to | what would change the design | status |
|---|---|---|---|---|
| P0 | whether rt_data_code_pmc returns statement text for restricting statements, not only a sharing flag, on 50 known restricting articles |
5, 7 | if it does not, statements are read from JATS data-availability sections in <front>, <back> and <body> by markup, with the P01v2 false-negative audit applied to that route |
not run |
| PA | count of restricting statements, marginally | 5 | fewer than 2000 makes the no-human group too small | not run |
| PB | share of studies classed as no human data on an author-labeled pilot of 100 removed from the sample | 6 | under 5 percent makes the primary denominator tiny | not run |
| PC | the author’s two sets of 200 labels | 8 | failing either gate triggers full labeling | needs the author |
12. What this cannot show
- Why a statement was written as it was. A template, an editor, and an author reluctant to share produce the same text.
- Whether the data, if requested, would be provided.