技能 数据科学 综述全面性与平衡性评估

综述全面性与平衡性评估

v20260724
revedres-comprehensiveness-and-balance
本工具用于系统性综述或元分析的严格质控,指导用户进行深度检验,确保研究结论的全面性和稳健性。它要求进行饱和度证据验证、偏倚风险评估,以及异质性、发表偏差等高级统计分析,防止报告结果出现遗漏或方法学上的缺陷。
获取技能
497 次下载
概览

Comprehensiveness & Balance (revedres-comprehensiveness-and-balance)

When to trigger

  • The corpus and framework exist, but you have not stress-tested coverage or robustness
  • A pooled effect is reported without heterogeneity, bias, or sensitivity analysis
  • You worry a reviewer will name an omitted study, school, or contradictory finding
  • Conflicting studies are being tallied rather than weighed by design quality

Comprehensiveness: prove there are no holes

RER asks for comprehensive reviews; the burden is on you to make exhaustiveness provable, not asserted.

  1. Saturation evidence. Show that the search reached the point where new searches stopped yielding eligible studies (from revedres-literature-synthesis).
  2. Named-omission test. Could an informed reviewer name an important study, research group, or adjacent literature you left out? If yes, include it or justify the boundary explicitly.
  3. Grey-literature & language reach. Excluding dissertations, reports, or non-English work narrows scope and biases effects — acknowledge and, where feasible, widen.

Balance: weigh evidence, never vote-count

The amateur move is vote-counting — tallying significant vs. null studies. The RER standard is to weigh evidence by what each study measures and how credibly.

  1. Risk-of-bias appraisal. Apply the a-priori tool to every included study; let credibility, not count, drive emphasis. A handful of well-identified studies can outweigh many weak ones.
  2. Estimand discipline. Reconcile conflicting findings by asking whether studies estimate the same object (population, construct, horizon, comparison). Apparent contradictions often dissolve.
  3. Steelman rival perspectives. State each theoretical or methodological camp at its strongest before its limits; flag — and bracket — your own program so emphasis is identity-blind.

Robustness (meta-analysis): make the number survive scrutiny

A pooled effect is a claim, not a fact, until you show it is not an artifact.

Probe What it guards against
Heterogeneity (Q, I², τ²) reporting one number for a mix of different effects
Moderator analysis masking real variation the framework should explain
Publication-bias diagnostics (funnel, Egger, trim-and-fill, p-curve/selection models) an inflated effect from missing null results
Sensitivity analysis (leave-one-out, influence, alternative models) a result driven by one study or one modeling choice
Dependent-effects handling (multilevel / robust variance) false precision from multiple effects per sample

For a narrative synthesis, the analogues are: confidence in the body of evidence (e.g. a GRADE-style judgment), explicit handling of conflicting findings, and a sensitivity check on which conclusions survive dropping the weakest studies.

Checklist

  • Saturation documented; no eligible study/database a reviewer could name as missing
  • Grey-literature/language exclusions acknowledged with their bias implications
  • Risk-of-bias appraised for every study; emphasis tracks credibility, not count
  • Conflicting findings reconciled by estimand + design, not tallied
  • Rival camps steelmanned; author's own work bracketed for identity-blind emphasis
  • (Meta) heterogeneity, moderators, publication-bias, and sensitivity all reported
  • (Meta) dependent effects modeled (multilevel / robust variance), not ignored
  • (Narrative) strength-of-evidence judged and a drop-the-weakest sensitivity check run

Anti-patterns

  • Vote-counting significant vs. null studies as if each carries equal weight
  • A single pooled effect with no I²/τ², no moderators, and no publication-bias check
  • Ignoring dependent effect sizes, manufacturing false precision
  • Excluding grey literature silently, then reporting an upward-biased effect
  • Caricaturing the camp the author disagrees with instead of steelmanning it
  • Treating "I found a lot of studies" as proof of comprehensiveness without saturation evidence

Output format

【Saturation】documented? Y/N — named-omission test passed? Y/N
【Grey lit / language】exclusions + bias implication stated? Y/N
【Risk of bias】appraised for all studies; emphasis credibility-weighted? Y/N
【Conflict handling】reconciled by estimand/design (not vote-count)? Y/N
【Heterogeneity】I²/τ² + moderators reported? Y/N (meta) | strength-of-evidence judged (narrative)
【Publication bias】funnel/Egger/trim-fill/p-curve run? Y/N
【Sensitivity】leave-one-out / alt models / dependent-effects model? Y/N
【Next step】→ revedres-tables-figures (PRISMA flow, forest/funnel, coding tables)
信息
Category 数据科学
Name revedres-comprehensiveness-and-balance
版本 v20260724
大小 5.04KB
更新时间 2026-07-29
语言