正文
获取原文件references/review_sources.md
Reviewed 2026-10-01. This is an evidence-appraisal workflow, not an implementation of formal scoring algorithms, a statistical package, or a clinical decision system. Use the complete current tool documents for a formal assessment and retain version/date, result identity, source locations, signalling answers, and judgment rationale.
Risk of bias, reporting, and certainty
| Source | Verified scope and consequence |
|---|---|
| Cochrane Handbook, chapter 8 | RoB 2 is result-specific; assignment versus adherence and trial variant matter. Overall judgments use domain rules, not an average score. |
| ROBINS-I V2 developers | November 2025 V2 is still labelled draft, for follow-up/cohort intervention studies. It differs from ROBINS-I 2016; name the version. |
| QUADAS-3 developers | Current recommended diagnostic accuracy tool, published February 2026; listed download v1.2. Assessment is per accuracy estimate, with synthesis questions and an ideal test accuracy trial. |
| AMSTAR 2 developers | Appraises systematic reviews of healthcare interventions; critical domains determine confidence, not a summed score. |
| PEDro scale | Physiotherapy-trial appraisal scale; eligibility item is not part of the total. Not equivalent to RoB 2 or a GRADE certainty rating. |
| Newcastle-Ottawa Scale developers | Cohort/case-control forms address selection, comparability, and outcome/exposure ascertainment. Do not invent universal star thresholds. |
| CASP checklists | Separate appraisal checklists by design; use the matching current form. |
| CONSORT/SPIRIT developers and extensions | Core trial-results and protocol guidance is CONSORT 2025/SPIRIT 2025. Some applicable extensions still reference older cores. Reporting standards do not certify conduct. |
| STROBE developers | Cohort, case-control, and cross-sectional reporting; explicitly not a study-quality assessment instrument. |
| PRISMA 2020 developers | Systematic-review reporting checklist, expanded checklist, and flow diagrams; not a bias or certainty score. |
| STARD guidance | STARD 2015 for diagnostic accuracy reports; STARD-AI is a 2025 extension for AI studies. |
| COREQ guidance | Qualitative interviews and focus groups; not all qualitative designs indiscriminately. |
| CARE developers | Case-report guidance developed in 2013 with 2017 explanation; reporting does not identify a comparative treatment effect. |
| GRADE Book principles | Current chapter dated August 2025: outcome-level certainty relative to explicit thresholds/ranges, absolute effects, alternative NRSI starting approaches, and no double counting. |
| GRADE Book overview | Updated May 2026; intervention, diagnostic, prognosis, exposure, and values applications differ. The intervention design hierarchy is not universal. |
| GRADE recommendations | Updated March 2026: recommendation direction/strength requires a separate evidence-to-decision rationale and context. |
| Cochrane Handbook, chapter 14 | Practical Summary of Findings and ROBINS-I/GRADE integration; preserve outcome/comparison scope and explicit domain explanations. |
The GRADE Book is replacing the older handbook progressively. Its three chapters above were read through rendered-page extraction because ordinary HTTP/browser text returned a JavaScript shell. The newer threshold/range framing should not be replaced with the older shortcut that certainty only means closeness to a point estimate.
Statistical and causal appraisal
- ASA p-value statement: p-values do not give hypothesis probabilities, effect importance, or a complete basis for decisions. Interpretation requires the analysis context and full reporting.
- Greenland et al., statistical misinterpretations: test statistics and confidence intervals depend on the full model and sampling/selection assumptions; avoid observed-power, threshold, and interval-overlap shortcuts.
- Hernán and Robins, Causal Inference: What If: define the causal estimand and defend consistency, exchangeability, positivity, and selection assumptions. The authors' current book link is dated August 19, 2026.
- Cochrane Handbook, chapter 13: missing-evidence appraisal cannot be reduced to a funnel plot. Asymmetry has alternative explanations; tests usually require at least ten studies and appropriate effect measures.
- ARRIVE 2.0 experimental-unit explanation: state the independent unit and avoid treating subsamples or repeated measurements as additional independent experimental units.
- FDA multiple-endpoints guidance: supports specifying endpoint families and multiplicity strategies in human drug/biologic trials; its regulatory scope is not a universal rule for all exploratory science.
- WMA Declaration of Helsinki: current 2024 revision concerns medical research involving human participants. Institutional review and applicable local requirements must be checked; this skill certifies no compliance.
Executed checks and limits
- The skill contains no scripts, SDK calls, scientific API endpoints, or statistical computation engine. References to REDCap, Qualtrics, registries, and statistical software identify possible workflow tools; they are not integrations or tested instructions.
- Tiny known-answer arithmetic checks verify the illustrative predictive values and independent multiple-testing probability; they do not establish diagnostic performance in a real population or validate a formal risk-of-bias/GRADE assessment.
- The optional scientific-schematics command was checked at the local CLI/help boundary. No paid generation of that example or private-data transmission was performed. Its provider contract and dependencies belong to that skill, not this appraisal workflow.
- No patient data, clinical study, registry submission, or authenticated scientific service was accessed. Human clinical/policy decisions require the applicable professional process.