Skip to contents

Scope and limitations

Source: vignettes/scope-and-limitations.Rmd
scope-and-limitations.Rmd

rtransparency is a pattern-based detector. It is designed for high precision on the statements it targets, and its predictions come with the exact text that triggered them so they can be audited. This vignette describes what each indicator does and does not capture, so results are interpreted correctly.

What the indicators mean

Indicator Detects Does not mean
Conflicts of interest A COI disclosure is present (including “the authors declare no competing interests”) That a conflict exists
Funding A statement that funding was received Presence of a funding section (a “no funding” section is read as absence)
Registration A protocol/trial registration identifier or statement Ethics/IRB approval numbers
Novelty The article claims its own work is novel or first That the work is objectively novel
Replication A replication or external/independent validation was performed An internal train/test split, or future/recommended validation
Data sharing The authors’ own data are made available (repository, accession, or in-article) Data merely reused, cited, or available “upon request”
Code sharing The authors’ own analysis code is shared Use of third-party software/tools
AI disclosure A statement discloses generative-AI use in manuscript preparation (including “no AI was used”) Use of AI as a research method
Open-access license The article is openly licensed (Creative Commons / CC0), and which license That every reader has full reuse rights (NC / ND licenses restrict reuse)
Reporting guideline The authors state they followed a reporting guideline, and which one That the study fully complied, or that a cited/animal-welfare/clinical guideline counts

Conflicts of interest and AI disclosure are disclosure-based: a statement addressing the topic counts as present, whether the disclosure is positive or negative. This mirrors how these are reported and counted in the literature.

Known limitations

  • Language. Detection is strongest in English. Conflict-of-interest and funding statements are also detected in Spanish, Portuguese, French, German and Italian; other indicators and other languages are not yet covered.
  • Data availability “upon request”. Data offered only on request are not counted as shared, reflecting the modern open-data standard. This is stricter than some earlier definitions and will report lower data-sharing prevalence than tools that count availability-on-request.
  • Novelty and replication are claim detection. They identify what authors state, not whether a study is truly novel or a replication succeeded. The replication indicator in particular is precision-limited because validation language (“validation cohort”, “independent”) is heavily overloaded; see the replication-enriched benchmark in inst/benchmark/.
  • Plain text vs XML. The plain-text detectors share the same logic as the PMC XML detectors but cannot use XML-structural cues (tagged funding groups, conflict footnotes, section types), so a few statements detectable in XML are not detectable in plain text. The plain-text AI detector rt_ai() is a special case: with no publication date and no section structure available, it applies no 2023 year gate (it never returns NA) and scans the whole document, so the caller must restrict it to 2023-or-later articles and tolerate a higher false-positive rate on AI-method papers than rt_ai_pmc().
  • Reporting guideline is claim detection. It identifies a stated adherence to a reporting guideline, not whether the study fully complied. It is tuned for precision (validated 93.8% sensitivity / 99.0% specificity on a hand-labeled 1000-article sample); the residual misses are mostly non-English statements, and residual false positives are guidelines named but not actually followed by the article.
  • Open-access licensing is XML-structural. It reads the JATS <license> element, so it is near-deterministic for PMC XML but, in plain text, depends on a license statement being present in the extracted text.
  • Accuracy correction. rt_summary() can correct apparent prevalence using bundled sensitivity/specificity estimates (rt_accuracy). These derive from the validation benchmarks; supply your own via rt_summary(accuracy = ) when you have study-specific estimates. AI disclosure is reported uncorrected (its prevalence is too low in unselected literature for a stable estimate).

Output schema

Every per-article detector returns the prediction columns is_coi_pred, is_fund_pred, is_register_pred, is_novelty_pred, is_replication_pred, is_open_data, is_open_code, the year-gated is_ai_pred (NA before 2023), is_open_access (with oa_license) and is_reporting_pred (with reporting_guideline), each paired with the extracted text or value. rt_all_pmc() returns all ten for one file; rt_all_pmc_dir() runs a whole directory.

library(rtransparency)

res <- rt_all_pmc("article.xml", remove_ns = TRUE)
res[, c("is_coi_pred", "is_fund_pred", "is_open_data", "is_open_code")]

Linking to FAIR assessment

The data- and code-availability links the detector extracts (open_data_links, open_code_links) can be passed to FAIR-assessment tooling such as rfair, a native R implementation of FAIR data and software assessment, to score the findability and accessibility of the shared resources.

res <- rt_all_pmc("article.xml", remove_ns = TRUE)
links <- strsplit(res$open_data_links, " ; ")[[1]]
# rfair::assess_fair(links)