| name | evidence-appraisal |
| description | Rates how much certainty a body of evidence deserves, outcome by outcome, on the GRADE scale (Grading of Recommendations Assessment, Development and Evaluation — high, moderate, low, very low), naming the reason behind every downgrade or upgrade and writing the summary-of-findings row and a calibrated certainty statement. Use when asked "how strong is the evidence for X?", "GRADE this evidence", "is this proven or just one study?", or "the studies conflict — what is actually known?". Not for one study's risk of bias (use `assess-study-bias`) or for grading a non-study source such as an article or dataset (use `rate-source-admiralty`). |
| license | MIT |
| metadata | {"category":"evidence-verification","method":"GRADE certainty of evidence (Grading of Recommendations Assessment, Development and Evaluation)","origin":"GRADE Working Group — G. H. Guyatt et al., 2008 (BMJ series); GRADE Handbook, 2013","version":"2.0.0"} |
Evidence Appraisal (GRADE)
GRADE — from the GRADE Working Group (Guyatt et al., 2008; GRADE Handbook, 2013) — rates the certainty of a body of evidence per outcome, not per study: randomized bodies start High, observational bodies Low, five named factors rate down and three rate up, and the result is stated as high / moderate / low / very low. Its core principle: certainty is a judged, argued property of the whole evidence base for one outcome, every move traceable to a reason. It prevents two symmetrical failures — treating one exciting trial as a fact, and a pile of weak studies as one.
When to invoke
Invoke when:
- The question is how far a body of studies deserves trust: "how strong is the evidence for {X}?", "GRADE this".
- The literature is mixed and a decision must still be made: "the studies conflict — what is known?".
- A
systematic-review or meta-analysis result needs a certainty rating.
Do NOT invoke when:
- One study's quality is the question — use
assess-study-bias (RoB 2 / ROBINS-I); its ratings are an input here.
- The study set is not assembled yet — use
systematic-review; GRADE starts once the body is known.
- The source is not a study (article, dataset, vendor claim) — use
rate-source-admiralty.
- Several studies need pooling first — use
meta-analysis; its results feed the factors.
Procedure
1 — Frame the question and the outcomes
GRADE is per question, per outcome; "does X work?" is unappraisable as stated. Write the PICO (population, intervention, comparator, outcomes) and rank outcomes by importance to the decision — the outcome list, not the study list, sets how many ratings to produce. Artifact: PICO plus ranked outcomes.
2 — Assemble the evidence body
Take the study set from a systematic-review where one exists; otherwise document the search (sources, date window, inclusion rules) so the body is auditable. Record each design and sample size. A randomized body starts High (4); an observational body Low (2). Artifact: study table.
3 — Appraise each study, routed by design
Route each study to its per-study tool: randomized → RoB 2, non-randomized → ROBINS-I (via assess-study-bias); other designs → the matching CASP checklist; single-arm series → hypothesis-generating. Every judgment carries a quote (≤ 20 words) or is marked NOT REPORTED — unreported methods are never assumed done. Artifact: appraisal table.
4 — GRADE the body, one outcome at a time
Start from the design level and apply the five rate-down factors, each 0 / −1 (serious) / −2 (very serious), with a one-line reason grounded in the appraisal table: risk of bias; inconsistency (unexplained heterogeneity, I²); indirectness (population, intervention, comparator or outcome differ from the PICO); imprecision (the CI crosses the minimal important difference); publication bias (small-study asymmetry, unpublished trials). Then the three rate-up factors: large effect (RR > 2 or < 0.5, +1; RR > 5 or < 0.2, +2) with no plausible confounder; dose–response gradient (+1); residual confounding working against the observed effect (+1). Rating up is "mostly applicable to observational studies" and comes only after rating down; the Handbook has "yet to find a compelling example" of rating up randomized evidence, so such an upgrade is exceptional and must be argued. State the strongest case for every factor not applied. Artifact: evidence profile with reasons.
5 — Communicate in certainty language and set the surveillance trigger
Write one summary-of-findings row per outcome and a plain statement ("moderate certainty that X probably reduces Y"). Banned words: "proven", "conclusive", "no evidence" (write "no eligible studies found" if true). Name the new evidence — trials in progress, replications, an explained heterogeneity — that would change the grade. Artifact: summary-of-findings table and surveillance note.
Output template
## Evidence Appraisal — {question}
**PICO:** P={population} | I={intervention} | C={comparator} | O={outcomes, ranked}
**Evidence body:** {N studies} — {designs, total n, source}
| Outcome | Studies (n) | Effect (95% CI) | Certainty | Factors applied (reason) |
|---|---|---|---|---|
| {outcome 1} | {N RCTs, n=…} | {MD/RR …} | ⊕⊕⊕⊕ High | none |
| {outcome 2} | {…} | {…} | ⊕⊕◯◯ Low | −1 inconsistency ({reason}); −1 imprecision ({reason}) |
**Certainty statement:** "{high/moderate/low/very-low} certainty that {X} {will/probably/may} {effect}"
**Surveillance trigger:** {trials, replications or mechanisms that would change the grade}
Mandatory: the PICO, one row per outcome with factor-by-factor reasons (including "none"), the certainty statement in GRADE language, and the surveillance trigger. A level without reasons is not a grade.
Worked example
Question (illustrative body): in adults with chronic insomnia (P), does CBT-I (I) versus sleep-hygiene education (C) improve sleep-onset latency (O)? Body: 5 RCTs, n = 412, documented search — start High. Appraisal: four studies at low RoB 2 risk, one "some concerns" (unblinded outcome logging) at about 8 % of pooled weight. python3 scripts/grade.py rate --demo reproduces the rating (abridged: risk of bias, indirectness, publication bias and all three rate-up factors print 0 with reasons):
GRADE certainty of evidence — Sleep-onset latency (min)
Body: 5 RCTs, n = 412, effect MD −12 min (95% CI −18 to −3)
Starting level: High (4) — randomized trials [Handbook §5.1.1]
Rate down (Handbook Table 5.2; 0 / −1 serious / −2 very serious):
inconsistency −1 serious point estimates range −4 to −22 min, I² = 68%
imprecision −1 serious 95% CI −18 to −3 spans the pre-set 10-min MID
Arithmetic: 4 (High) −1 inconsistency −1 imprecision = 2 → Low
Certainty: ⊕⊕◯◯ Low
Summary-of-findings row: Sleep-onset latency (min) | 5 RCTs (n = 412) | MD −12 (−18 to −3) | ⊕⊕◯◯ Low | −1 inconsistency (I² = 68 %); −1 imprecision (CI spans the 10-min MID). Certainty statement: "Low certainty that CBT-I may shorten sleep-onset latency by about 12 minutes versus sleep-hygiene education." Surveillance trigger: the 300-patient SLEEPIQ-2 trial — a CI wholly beyond the 10-min MID removes the imprecision downgrade (→ Moderate); the inconsistency downgrade stands until the heterogeneity is explained.
Verification
Companion tool
scripts/grade.py (stdlib only) does the step-4 bookkeeping: start level by design (RCT → High, observational → Low; --start high for ROBINS-I-appraised bodies), five rate-down factors (0/−1/−2), three rate-up factors (+1/+2 large effect, +1 dose–response, +1 opposing confounding), the clamp to Very low…High, and the Table 5.1 interpretation. Judgements stay with the analyst; the tool totals them and warns when rating up meets rating down or randomized evidence.
python3 scripts/grade.py rate --file body.json [--json] — one outcome; rate --demo reproduces the worked example
python3 scripts/grade.py sof --file bodies.json — Summary-of-Findings rows
python3 scripts/grade.py factors — the 5+3 factor definitions, cited
python3 scripts/grade.py --selftest
Sample output: the worked example above. Usable without the tool; the rules are those of step 4.
Pair with adjacent skills
assess-study-bias — the per-study sibling; its RoB 2 / ROBINS-I ratings feed the risk-of-bias factor.
systematic-review — assembles the study set (PRISMA flow) this skill grades.
meta-analysis — pools the body first; its I² and Egger results feed two of the factors.
rate-source-admiralty — grades non-study sources before they enter an evidence body.
bayesian-update — a GRADE level plus effect size disciplines the likelihoods.
- Methodology counterpart: methodologies/scientific-methods/evidence-appraisal.md — hierarchies, CASP, causation viewpoints.
Anti-patterns
- Do not grade one study and call it "the certainty of the evidence"; GRADE rates the body per outcome.
- Do not apply the evidence hierarchy mechanically; it ranks bias-resistance for therapy questions, other question types need their own frames, and a well-run cohort can outrank a badly run RCT.
- Do not move a factor without a reason from the appraisal table; "feels weak" is not a downgrade, "3 of 5 studies have high attrition risk" is.
- Do not rate up casually: the upgrades are mostly for observational bodies not already rated down, and a randomized-body upgrade is exceptional.
- Do not average certainty across outcomes or report only the best; "probably helps A, uncertain for B" is the deliverable.
- Do not write "proven", "conclusive" or "no evidence"; decisions rest on imperfect evidence, graded honestly.
- Do not let checklist scores replace judgment; careful teams can differ by one level.
Reference
- G. H. Guyatt et al., "GRADE: an emerging consensus on rating quality of evidence and strength of recommendations," BMJ, vol. 336, pp. 924–926, 2008. doi:10.1136/bmj.39489.470347.AD
- H. Balshem et al., "GRADE guidelines: 3. Rating the quality of evidence," J Clin Epidemiol, vol. 64, no. 4, pp. 401–406, 2011. doi:10.1016/j.jclinepi.2010.07.015
- H. Schünemann, J. Brożek, G. Guyatt, A. Oxman (eds.), GRADE Handbook, GRADE Working Group, 2013, §5 (Tables 5.1–5.3, 5.9). https://gdt.gradepro.org/app/handbook/handbook.html
- H. J. Schünemann et al., "GRADE guidelines: 18. How ROBINS-I and other tools … should be used to rate the certainty of a body of evidence," J Clin Epidemiol, vol. 111, pp. 105–114, 2019. doi:10.1016/j.jclinepi.2018.01.012
- J. A. C. Sterne et al., "ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions," BMJ, vol. 355, i4919, 2016. doi:10.1136/bmj.i4919