coreSR.AIcoreSR
Usage Guide

CLARITY (Cohort)

How to assess risk of bias in cohort studies

1

What CLARITY is and when to use it

  1. 1CLARITY — "Tool to Assess Risk of Bias in Cohort Studies", CLARITY Group at McMaster University (the re-modified Newcastle-Ottawa instrument shipped with DistillerSR)
  2. 2Use for COHORT studies in your review — exposed vs non-exposed groups followed over time for an outcome
  3. 3NOT for randomised trials (use RoB 2.0 / ROBUST-RCT), non-randomised intervention studies (use ROBINS-I), prediction models (use PROBAST), or prognostic-factor studies (use QUIPS)
  4. 4Unit of assessment: one CLARITY assessment per included cohort study
Tip: CLARITY collapses the classic Newcastle-Ottawa star items into 8 plain-language questions. If your review protocol says "Newcastle-Ottawa for cohort studies", this is the instrument it almost always means.
2

The 8 domains and the 4-point scale

  1. 1D1 Selection — were exposed and non-exposed cohorts drawn from the same population?
  2. 2D2 Exposure assessment — can we be confident in the assessment of exposure?
  3. 3D3 Outcome not present at start — can we be confident the outcome of interest was absent at baseline?
  4. 4D4 Control of confounders — did the study match or adjust for all plausible prognostic variables?
  5. 5D5 Assessment of confounders — can we be confident in the assessment of prognostic factors?
  6. 6D6 Outcome assessment — can we be confident in the assessment of outcome?
  7. 7D7 Adequate follow-up — was follow-up of cohorts adequate?
  8. 8D8 Similar co-interventions — were co-interventions similar between groups?
  9. 9Every domain is rated Definitely yes / Probably yes / Probably no / Definitely no — "yes" means LOW risk of bias, "definitely vs probably" is a confidence gradation, not a severity one
3

Guidance examples are part of the instrument

  1. 1Each domain card carries the published guidance lists (low / higher / high risk examples) in a collapsible section — read them before rating
  2. 2Example: for D2, secure surgical/pharmacy records or repeated interviews are low risk; a single structured interview or written self-report is higher; "uncertain how exposure was obtained" is high
  3. 3Example: for D6, independent blind assessment or record linkage is low risk; for hard outcomes (fractured hip, implant failure) the medical record is sufficient — mirroring the instrument's fractured-hip example
  4. 4The AI sees the same guidance lists plus calibration notes distilled from reviewer-worked examples, so its suggestions are anchored to how the instrument is actually applied
4

AI assessment with quoted evidence

  1. 1Upload the study PDF on the assessment page (text is extracted automatically), then click "Run AI assessment"
  2. 2One whole-study GLASS pass rates all 8 domains with a rationale, a confidence, and a VERBATIM evidence quote per domain
  3. 3Every quote is verified server-side against the extracted text — a quote that is not fully locatable is dropped (the judgement stays, the run is flagged) so fabricated citations never reach the worksheet
  4. 4Click any quote to jump to and highlight that passage in the rendered PDF beside the worksheet
  5. 5AI also suggests the holistic overall judgement (Low / High / Very high) — the worked examples show overall is NOT a mechanical rollup, so accept it only after reading the rationale
5

Reviewer override and the summary matrix

  1. 1Your pick always wins: final = your judgement if set, otherwise the AI suggestion. Clearing your pick restores the AI answer
  2. 2Add a support-for-judgement note per domain — it is yours and AI runs never overwrite it
  3. 3The project page shows the study-by-domain matrix (the same table format as the instrument's published worked examples), colour-coded green → lime → orange → red, with an AI badge on AI-assisted studies
  4. 4Re-running AI is idempotent (keyed on files + prompt version) — use "force re-run" after uploading a new PDF version
6

Reviewer best practice

  1. 1Read the AI rationale and click the evidence quote BEFORE accepting a domain rating — the quote shows you exactly what the model saw
  2. 2D2 (exposure) and D5 (confounder assessment) are where AI and reviewers disagree most — retrospective record review without documented abstraction quality should usually be Probably no or worse
  3. 3D8 (co-interventions) is almost always a "no" in practice — unless the paper explicitly documents similar co-interventions, do not let an optimistic suggestion stand
  4. 4Decide the overall yourself: several "no" domains usually means High overall, but the CLARITY convention allows judgement (isolated severe problems need not sink an otherwise sound study)
coreSR.AICoreSR

AI-powered evidence synthesis