科研 Skill

效应量元分析

alterlab-meta-analysis

Pools effect sizes across studies with fixed/random-effects models, heterogeneity stats, forest/funnel plots, and publication-bias tests. Use for a systematic review’s quantitative synthesis.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Meta-Analysis — Pool Honestly, Then Interrogate Heterogeneity and Bias

Skill type: ANALYSIS MODULE. Synthesizes effect sizes across studies. The discipline is not the pooled point estimate — it is the model choice (fixed vs random effects), the heterogeneity you must characterize, and the publication-bias diagnostics that decide whether the pooled estimate is trustworthy at all.

Core Mission

THE POOLED EFFECT IS ONLY AS GOOD AS ITS HETEROGENEITY STORY AND ITS PUBLICATION-BIAS CHECK.

When to Use This Skill

  • "Pool these effect sizes / run a meta-analysis across N studies."
  • "Compute I² / τ² — how heterogeneous are my studies?"
  • "Is there publication bias? Draw a funnel plot / run Egger's test."
  • "Convert these means and SDs (or 2×2 tables) into effect sizes and combine them."

Does NOT Trigger

The request is really about…Route toWhy not this skill
Finding / screening the literature (search, PRISMA flow)alterlab-deep-researchDiscovery & screening, upstream of pooling.
A single study's descriptive/inferential statisticsalterlab-statistical-analysisOne dataset, not cross-study synthesis.
Fitting one regression / GLMalterlab-statsmodelsNot effect-size pooling.
Whether the review question/design is soundalterlab-ssci-design-gateDesign routing.

Hand-off rule. When the request matches a row above, open your answer by naming that skill by its exact id (for example alterlab-deep-research) and saying in one line why it fits better, then stop applying this skill's procedure. Name it even when you also need to ask for missing input. If only part of the request is out of scope, name the skill for that part and do only the in-scope part. A brief general pointer is fine; the named skill does the work.

Running the code. In Claude Code, run analysis code in a throwaway environment instead of installing packages: uv run --python ">=3.11" --with "statsmodels>=0.14" --with pandas python meta.py. This needs no install approval and leaves the user's environment untouched. If uv or the network is unavailable, say so and hand over the script; never report results you did not compute.

Fixed-effect vs random-effects (the first decision)

  • Fixed-effect — assumes one true effect; studies differ only by sampling error. Rarely defensible across heterogeneous social-science studies.
  • Random-effects — assumes a distribution of true effects (between-study variance τ²); the default when studies vary in population, measure, or design. Report the model and why.

Verified calls (pinned)

Python — statsmodels:

from statsmodels.stats.meta_analysis import effectsize_smd, combine_effects
eff, var = effectsize_smd(m1, sd1, n1, m2, sd2, n2)          # Hedges' g + variance
res = combine_effects(eff, var, method_re="dl")              # DerSimonian-Laird random effects
res.summary_frame()                                          # fixed + random rows, CIs, weights
res.tau2, res.i2, res.q                                      # heterogeneity
res.plot_forest()
# binary outcomes: effectsize_2proportions(c1, n1, c2, n2, statistic="odds-ratio")  # log OR + var

R — metafor (field standard):

library(metafor)
dat <- escalc(measure = "SMD", m1i=, sd1i=, n1i=, m2i=, sd2i=, n2i=, data = studies)  # or "OR"/"RR"
res <- rma(yi, vi, data = dat, method = "REML")              # random-effects
summary(res)                                                 # I2, tau2, Q
forest(res); funnel(res)
regtest(res)                                                 # Egger's test for funnel asymmetry
trimfill(res)                                                # trim-and-fill sensitivity

Heterogeneity — characterize, don't hide

  • Cochran's Q (test of homogeneity; low power with few studies),
  • I² (% of variation due to heterogeneity, not chance),
  • τ² (between-study variance, on the effect-size scale). High I²/τ² means the pooled mean summarizes a distribution — report a prediction interval, and explore moderators (meta-regression / subgroups) rather than over-interpreting the point estimate.

Publication bias

  • Funnel plot — asymmetry suggests small-study effects / missing null results.
  • Egger's regression test (regtest) — a formal asymmetry test.
  • Trim-and-fill (trimfill) — imputes "missing" studies as a sensitivity analysis. No single test is definitive; report the funnel plot plus at least one test, and treat them as sensitivity analyses, not proof.

Reporting standard (PRISMA)

Report per PRISMA 2020: the search/screening flow (hand off discovery to alterlab-deep-research), inclusion criteria, per-study effect sizes and weights, the pooling model, I²/τ²/Q, the publication-bias diagnostics, and risk-of-bias assessment.

References

  • references/pooling_and_bias.md — effect-size formulas, RE estimators, heterogeneity, bias diagnostics, PRISMA.

Part of the AlterLab Academic Skills suite.

示例提示词

帮我汇总多篇研究的效应量,画出森林图并检验发表偏倚。

所属插件

社会学、政治学与城市研究面向社会科学研究:问卷与访谈设计、质性编码、文本即数据、社会网络、ABM、比较政治与选举数据、RCT 与 DID、城市与区域分析。

安装该插件会一并获得同领域的 58 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

效应量元分析能做什么?

Pools effect sizes across studies with fixed/random-effects models, heterogeneity stats, forest/funnel plots, and publication-bias tests. Use for a systematic review’s quantitative synthesis。

如何安装效应量元分析?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「效应量元分析」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

效应量元分析需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

效应量元分析经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。