What STARD is and who maintains it
STARD, the Standards for Reporting Diagnostic Accuracy Studies, is a checklist for reporting studies that evaluate how well one or more index tests identify a target condition against a reference standard. The current version, STARD 2015, has 30 items and was developed by an international group of methodologists building on the original 2003 statement (Bossuyt et al., 2015).
The checklist covers the title and abstract, introduction, methods (study design, participants, test methods, the reference standard, analysis), results (participant flow, test results, estimates of accuracy) and discussion. A participant flow diagram, similar in spirit to the CONSORT diagram, shows how many participants were enrolled, tested, and included in the final accuracy analysis.
Why it exists
A diagnostic accuracy study is easy to report in a way that looks more favourable than the test actually is. If the reference standard used to confirm the true diagnosis is itself imperfect, if patients with indeterminate results are quietly excluded, or if the study population is enriched with clear-cut cases and obvious non-cases, the reported sensitivity and specificity can overstate how the test will perform in the mixed population it is meant to be used on. This last problem, sometimes called spectrum bias, is one of the most common and least visible distortions in diagnostic research: a test that looks excellent when evaluated on clearly diseased patients and clearly healthy controls can perform far less well once it meets the harder, borderline cases that make up most of real clinical practice.
STARD exists to make these choices visible rather than hidden inside a headline accuracy figure. A reader who can see the reference standard, how indeterminate results were handled, and who was included and excluded can judge for themselves how much the reported accuracy is likely to generalise to their own patients.
STARD is published through the EQUATOR Network's coordinated library of reporting guidelines, which was itself established because incomplete reporting was recognised as a recurring problem across many types of health research, not only diagnostic studies (Moher et al., 2008).
Which studies it fits
STARD applies to any study that evaluates the accuracy of one or more index tests, which may be laboratory assays, imaging findings, clinical signs, symptoms, questionnaires or algorithms, against an independent reference standard taken to represent the true presence or absence of a target condition.
- A single test evaluated against a reference standard in a defined patient group: core STARD applies.
- Two or more tests compared head-to-head, or combined in sequence: STARD still applies, with items on how each test and any combination rule was interpreted.
- A test whose results are continuous, such as a biomarker level, where a cut-off is chosen to classify patients: STARD expects the cut-off and how it was chosen to be reported, including whether it was pre-specified or selected to optimise performance in the same data.
- A test interpreted by more than one reader, such as an imaging finding: STARD expects the number of readers, their level of experience and how disagreements between them were resolved to be reported.
Across all of these variants, the common requirement is the same: name the index test and the reference standard precisely enough that another researcher could repeat the comparison, and report every patient who entered the study, not only those whose results were straightforward to classify. Reporting the test's technical details, such as the assay manufacturer, software version or imaging protocol, also matters, because accuracy can shift between versions of the same nominal test.
Where STARD does not apply
STARD is specifically about index tests measured against a reference standard for a single diagnosis at a point in time. It is not the right checklist for a multivariable model that combines several predictors to estimate the probability of a diagnosis or a future outcome; such studies should use TRIPOD+AI, which has items on model development, validation and performance that STARD does not cover.
STARD also does not apply to studies of prognosis without a diagnostic test, to observational studies of risk factors, which should use STROBE (von Elm et al., 2007), or to trials that compare clinical outcomes between patients managed with versus without a test, which are themselves usually reported with CONSORT (Hopewell et al., 2025) because the comparison being made is a trial of a management strategy, not a measurement of accuracy.
A systematic review that pools several diagnostic accuracy studies should be reported with PRISMA, using its diagnostic test accuracy extension, rather than with STARD itself; STARD operates at the level of a single primary study, and PRISMA at the level of the review that brings several such studies together (Page et al., 2021).
Why it matters
Diagnostic accuracy estimates feed directly into clinical decisions, screening policy and, increasingly, the evidence base that systematic reviews and guideline panels draw on. Because accuracy estimates are sensitive to the study population and the reference standard used, GRADE and similar frameworks explicitly ask reviewers to consider indirectness, the match between the study population and the population a test will actually be used on, when rating how much confidence to place in a diagnostic accuracy estimate (Guyatt et al., 2008). STARD reporting gives a reviewer the information that judgement depends on.
This is also why systematic reviews of diagnostic accuracy pay close attention to differences between studies in how the reference standard was applied and what patient group was studied; two studies of the "same" test can report very different accuracy simply because one recruited from a specialist referral clinic and the other from a general screening programme.
How it is applied in practice
Journals in radiology, laboratory medicine and clinical pathology, among others, commonly require a completed STARD checklist at submission, and several have adopted STARD-based editorial checks as a condition of peer review for diagnostic accuracy manuscripts. Systematic reviews of diagnostic test accuracy, reported with the PRISMA-DTA extension, routinely use STARD items to assess the quality of reporting in the primary studies they include, which gives authors a further incentive to report completely even outside journals that formally require it.
For a doctoral researcher running a diagnostic accuracy study, the most useful time to consult STARD is while designing the study, particularly to decide the reference standard, the interval between the index test and the reference standard, and the rule for classifying indeterminate results, since these decisions shape the data that will later be available to report.
It is also worth deciding, before recruitment, whether the people interpreting the index test will be blinded to the reference standard result and vice versa, since unblinded interpretation is one of the commonest sources of inflated accuracy in the published literature and is very hard to correct for retrospectively once data collection is finished.
An illustrative example
The passage below is a hypothetical illustration, not a quotation from a published study.
| Element | Without STARD | With STARD |
|---|---|---|
| Reference standard | "Diagnosis was confirmed clinically." | "The reference standard was histopathological confirmation by two independent pathologists blinded to the index test result, with discrepancies resolved by a third pathologist." |
| Participant flow | "120 patients were studied." | "150 patients were eligible; 12 declined and 8 had indeterminate reference standard results and were excluded from the primary analysis but reported separately; 130 were included in the accuracy analysis." |
| Cut-off | "A positive result was recorded above the threshold." | "The biomarker cut-off of 4.2 ng/mL was pre-specified from a prior cohort and applied without modification in this study." |
| Accuracy estimates | "Sensitivity was 92% and specificity was 88%." | "Sensitivity was 92% (95% CI 86–96%) and specificity was 88% (95% CI 81–93%), with a 2x2 table of index test versus reference standard results reported in full." |
| Population | Not described | "Participants were adults referred for suspected disease at a single tertiary centre; prevalence of the target condition in the study sample was 34%, which may differ from primary care settings." |
What to check before you submit
- Is the reference standard named and justified, with how and by whom it was interpreted?
- Is there a participant flow diagram showing eligible, tested, excluded and analysed participants, including indeterminate results?
- Is the index test cut-off, if any, stated as pre-specified or chosen from the study data?
- Are sensitivity, specificity and other accuracy measures reported with confidence intervals and the full 2x2 table?
- Is the study population described well enough for a reader to judge whether the estimates will generalise to their own setting?
- Official checklist and source: STARD 2015 (EQUATOR Network) .
Key takeaways
- STARD 2015 is a 30-item checklist for reporting studies of diagnostic test accuracy against a reference standard (Bossuyt et al., 2015).
- It exists because accuracy estimates can look better than they are if the reference standard, indeterminate results or study population are not disclosed.
- It fits single or combined index tests against a reference standard; multivariable prediction models need TRIPOD+AI instead.
- It does not cover risk-factor studies (STROBE (von Elm et al., 2007)) or management trials (CONSORT (Hopewell et al., 2025)).
- How well a study reports its population and reference standard shapes how indirectness is judged under frameworks such as GRADE (Guyatt et al., 2008).
- A systematic review of diagnostic accuracy studies uses the PRISMA-DTA extension, not STARD, at the review level (Page et al., 2021).
