STARD (Standards for Reporting of Diagnostic Accuracy Studies)

What it is

— STARD (Standards for Reporting of Diagnostic Accuracy Studies) is a reporting guideline whose current version, STARD 2015, is a 30-item checklist plus a recommended participant flow diagram for studies that estimate the accuracy of a test (index test) against a reference standard ("gold standard"). Its purpose is to make a diagnostic accuracy study fully and transparently reportable so a reader can judge internal validity (risk of bias), applicability, and the estimates themselves (sensitivity, specificity, predictive values, likelihood ratios). It is maintained as part of the EQUATOR Network library of reporting guidelines; the checklist was first published in 2003 (Bossuyt et al., a 25-item list) and substantially revised in 2015 (Bossuyt et al.) with a companion explanation-and-elaboration paper (Cohen et al., 2016). STARD is a reporting standard — it prescribes what to write, not how to design or appraise the study. In real-world evidence, the natural "diagnostic test" is a computable phenotype / claims or EHR algorithm (e.g., a 1-inpatient-or-2- outpatient code rule, an NLP outcome classifier, a registry case definition) and the "reference standard" is the validated truth (adjudicated chart review, linked gold-standard registry); STARD is the guideline for reporting the validation study that estimates how well that algorithm performs.

When to use

— Apply STARD whenever the study object is the diagnostic/classification accuracy of a test or algorithm against a reference standard, and you are reporting it for a peer-reviewed journal, a regulatory submission, or an HTA/payer dossier. In RWE this is precisely the algorithm/phenotype validation study: estimating the positive predictive value, sensitivity, specificity, and (where the design allows) negative predictive value of a claims or EHR case-finding algorithm against adjudicated chart review or a linked gold standard. Decision rule for choosing the right guideline: use STARD when the estimand is test accuracy (PPV/Se/Sp of an algorithm or biomarker); use STROBE or its pharmacoepidemiology extension RECORD/RECORD-PE when the study is an etiologic/comparative observational study whose outcome happens to be algorithm-defined; use HARPER or the ENCePP checklist for the protocol of a non-interventional comparative study; use TRIPOD+AI (not STARD) when the index test is a multivariable prediction model that outputs a risk score rather than a binary classification validated against a reference standard; and use STARD-AI when the index test is an AI/ML diagnostic system. A study can need more than one: a comparative-effectiveness paper that embeds an algorithm validation sub-study reports the main study under STROBE/RECORD-PE and the validation sub-study under STARD.

What it requires

— STARD 2015's 30 items compel reporting of the elements that determine whether an accuracy estimate is trustworthy and transferable, organized across title/abstract, introduction, methods, results, and discussion. The substantive methods and results items that matter most for RWE validation are: an explicit study question and intended use / target population (item 1–4); the index test (algorithm) definition in enough operational detail to reproduce it — exact codes, code positions (inpatient vs outpatient), diagnosis fields, time windows, and the version/date of the code set (items 10a, 11–13); the reference standard and its rationale, including how the truth was ascertained (adjudication process, blinding) and why it is credible (items 10b, 12b); participant flow with a diagram and the eligibility, sampling, and time interval between index test and reference standard (items 5–9, 19, 21); how indeterminate/missing results and patients lost to the reference standard were handled (items 23, 24, 25); a 2×2 cross-tabulation of index-test results by reference-standard status (item 23); accuracy estimates with precision (PPV, sensitivity, specificity, and confidence intervals, item 24); and any subgroup, threshold, or sensitivity analyses (item 25). For claims/EHR work the high-leverage requirements are: report the verification scheme (was the reference standard applied to a sample of test-positives only, or to test-positives and a sample of test-negatives — i.e., is this a PPV-only design or a full Se/Sp design, and is there partial-verification / spectrum bias); report the time-zero alignment between the algorithm date and the reference-standard date; and report how the validation sample's spectrum and prevalence relate to the target database so readers can judge transportability of PPV (which is prevalence-dependent).

When NOT to use — limitations and common misapplications

— STARD is a reporting checklist, not a risk-of-bias instrument and not a quality score; the appraisal tool for diagnostic accuracy studies is QUADAS-2, and completing STARD says nothing about whether the study was well designed — only whether it was fully described. Concrete failure modes: (1) Wrong guideline for the question — using STARD to report a comparative observational drug study because its outcome is algorithm-defined; that study is STROBE/RECORD-PE, and STARD covers only the embedded validation sub-study, if any. (2) Using STARD for a prediction model — a multivariable risk score validated by discrimination and calibration belongs under TRIPOD+AI, not STARD. (3) Checklist-as-theater — page-numbering all 30 items while leaving the algorithm codes, the reference-standard adjudication process, or the verification scheme vague defeats the purpose; the value is reproducibility, not the completed grid. (4) Reporting PPV as if it were portable — PPV is prevalence-dependent, so a PPV validated in one database or era does not transfer to a target cohort with different case mix; STARD requires the spectrum and prevalence information that makes this judgeable, and omitting it is a common, consequential lapse. (5) Ignoring partial verification — validating only test-positives yields PPV but cannot yield sensitivity or specificity; reporting Se/Sp from a test-positive-only sample is a misapplication STARD's flow and 2×2 items are designed to expose. (6) Confusing "STARD-compliant report" with "valid algorithm" — a fully STARD-compliant paper can still describe an algorithm too inaccurate to use; the checklist surfaces that, it does not prevent it.

How it maps to this catalog

— In this repo, STARD is the reporting standard for the validation concepts; each requirement is implemented by a concept a methodologist can build and pre-specify against: - The core RWE "index test" / algorithm definitions STARD asks you to report reproducibly: diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe (code rules, positions, time windows), ehr-phenotyping-algorithms-rwe and outcome-algorithm-construction-rwe (algorithm construction), and procedure-identification-and-measurement-in-claims-ehr. - The validation study itself and its accuracy estimands: algorithm-validation and claims-outcome-algorithm-ppv-sensitivity-rwe operationalize the PPV/Se/Sp/2×2 results items; endpoint-adjudication-chart-review-rwe implements the reference-standard / adjudication items (10b, 12b) STARD requires you to describe. - The consequences of imperfect accuracy STARD's estimates feed into: misclassification-bias-correction-rwe and external-adjustment-validation-substudy-bias- correction-rwe use the validated Se/Sp/PPV to correct downstream effect estimates, and quantitative-bias-analysis-toolkit-rwe propagates that uncertainty. - Applicability/transportability of a prevalence-dependent PPV across databases: medicare-ffs-ma-commercial-claims-differences-rwe (case mix and coding-intensity differences), fit-for-purpose-data-assessment-rwe, and database-feasibility-attrition-funnel-rwe (the participant-flow item in a data context). General data handling lives in claims-analysis.

Applied note (claims/EHR/registry RWE)

A STARD-compliant validation of a claims case-finding algorithm should state the intended use and target database; give the exact algorithm (codes, positions, diagnosis fields, time windows, code-set version) so it is reproducible; describe the reference standard (e.g., two-physician adjudicated chart review with disagreement resolution) and blinding; show a flow diagram and the index-to-reference time interval; present the 2×2 table; report PPV with a confidence interval and — only if test-negatives were also verified — sensitivity and specificity; declare any partial-verification design explicitly; and characterize how the validation sample's prevalence and spectrum compare to the analytic cohort, because the PPV you report will not transfer to a database with a different case mix without that context.