meta-analysis

• English guide → README.md · 中文指南 → README_zh-CN.md

<div align="center"> <img src="assets/icon.svg" width="240" height="240" alt="meta-analysis logo"/> </div>

Easy-to-use R-based Meta-Analysis for Clinical Researchers

You don't need to code or memorize commands — just describe your meta-analysis needs in plain language inside a chat, and the skill automatically runs the full analysis (pooling, figures, report) for you. Powered by R and 14 core + 2 optional professional R packages (metafor, meta, netmeta, bayesmeta, dosresmeta, mada, etc.), it returns results in Chinese or English depending on your OS language setting (you can force-switch via a prompt at any time). Once you describe a request, the skill auto-executes and returns results + figures; ask for the full reproducible R code at any time.

---

Who This Is For

meta-analysis is part of the CT-series skill family, built for three groups:

• Clinical-trial practitioners at pharmaceutical companies — sponsors, CROs, and medical / statistical / regulatory roles who need to pool and compare study evidence;
• Clinicians and nurses who design, conduct, or participate in clinical research and need meta-analysis / systematic-review support;
• Medical students who want to learn evidence-based medicine and meta-analysis methodology in a structured way.

---

1. How to Use It in a Chat (the Core)

meta-analysis is a conversational skill: you simply tell the assistant what you want in natural language — no commands, no parameter names to remember. As a WorkBuddy skill it auto-loads with no extra installation.

Below are 7 real conversational examples. Example 1 is the upstream topic-selection gate (choose among candidate directions before running any analysis), followed by Examples 2–7 ordered from simple to advanced. Each shows "You say" and a sketch of "The assistant replies", plus how to get the actual number.

Example 1 · Topic selection: choose among candidate meta-analysis directions

You say:

I'm interested in SGLT2 inhibitors for renoprotection and cardiovascular benefit in chronic kidney disease (CKD), but I haven't decided which specific meta-analysis direction to pursue. Can you help me shortlist candidate directions?

Assistant replies (sketch):

Sure — I'll use a topic-selection approach to lock down the direction. First I run the in-skill dedup probe (Cochrane + PubMed real hit counts), then shortlist — otherwise a saturated direction can be wrongly pushed as top pick.

First, rule out the saturated directions: the generic "SGLT2i on cardiorenal outcomes in CKD" is Cochrane 23 / PubMed 2785 — highly saturated. Counter-intuitively, the seemingly niche non-diabetic CKD (Cochrane 20 / PubMed 2402) and dialysis / advanced CKD (Cochrane 22 / PubMed 1067) are also heavily covered (DAPA-CKD / EMPA-KIDNEY spawned many reviews) — also not recommended.

Re-shortlisting 3 candidates by evidence gap + novelty (I'll produce 1–3 directions; I won't decide for you):

Candidate ① Renoprotection in IgA nephropathy (top pick)

• Cochrane 4 / PubMed 224 (≈1/10 of the broad direction) → real gap
• Most common primary GN; SGLT2i already a guideline adjunct; many RCTs, data available

Candidate ② Net benefit / benefit–risk framework

• Cochrane 5 / PubMed 442 → gap
• Pools renal + cardiac benefit with safety events (DKA / infection / volume depletion) in one framework — strong methodological differentiation

Candidate ③ Specific glomerular diseases (FSGS / membranous / ADPKD / lupus)

• FSGS Cochrane 0, ADPKD 2, membranous 1, lupus 3 → very deep gap
• ⚠️ but sparse RCTs → power risk; better as a scoping review than a hard efficacy MA

Suggestion: Start with ① (real gap, high clinical need); ② as the methodological-differentiation alternative; ③ conditional. Next step: a full topic assessment (PICO + dedup search + PRISMA/AMSTAR-2 pre-check), and run ct-literature first for a full dedup confirmation (the probe is a quick signal; FSGS=0 may include index-phrasing variance, don't rely on it alone).

📌 Note: When you have a direction but are unsure which specific meta to run, the assistant uses the topic-selection framework to produce 1–3 candidate directions + a 4-dim score + meta type, instead of deciding for you or giving a single answer. This is the topic-selection gate before analysis — no R computation is invoked. Candidates are stratified by evidence gap / novelty, each backed by the in-skill dedup probe (Cochrane + PubMed real hit counts) — so a saturated generic direction (including the "intuitively niche" non-diabetic CKD or dialysis) is no longer mis-pushed as the top pick.

Example 2 · Two-group binary meta (most common)

You say:

Combine the OR across these 5 binary studies: Study A: experimental 30/100, control 20/100 Study B: experimental 45/120, control 30/100 ...

Assistant replies (sketch):

✅ 5 studies, random-effects (DL), OR = 1.52 (95%CI: 1.18–1.96), I² = 34% (Ask for the full reproducible R code at any time.)

📌 Auto-execution: Once you describe the request, the assistant auto-executes and returns the real result; ask for the reproducible code by saying "give me the full R code".

Example 3 · Effect size conversion

You say:

Convert Cohen's d = 0.8 to logOR

Assistant replies (sketch):

✅ d = 0.8 → logOR ≈ 1.45 (SE ≈ 0.38), based on logOR = d × π/√3

Example 4 · Heterogeneity + subgroup

You say:

I did a continuous meta with 8 studies, SMD pooled, but I² is high — I want to subgroup by region

Assistant replies (sketch):

✅ 8 studies, SMD = 0.42 (95%CI: 0.15–0.69), I² = 78% Subgroup by region:

• Asia (k=4): SMD = 0.55, I² = 45%
• Euro-US (k=4): SMD = 0.28, I² = 62%
• Between-group Q = 4.12, p = 0.04

Example 5 · Complex: Network meta with multiple decisions (popup menu)

You say:

I want to do a network meta with 4 interventions, but I haven't decided whether to use frequentist or Bayesian, and I'm not sure if I should include dose-response

Assistant replies (sketch):

4 interventions clear. Two key choices to clarify:

① Which analysis framework?

• (a) Frequentist — netmeta, inverse-variance based, good for consistent networks
• (b) Bayesian — gemtc (JAGS) or multinma (Stan) local only (a known coze limitation: the container has no root to install JAGS; on the cloud use (a) netmeta); bayesmeta supports Bayesian pairwise comparisons locally

② Include dose-response?

• (a) No — pure pairwise comparison network
• (b) Yes — needs dose-effect parameters, Bayesian framework preferred

③ Can't decide? → say "explain the differences between these choices in detail", and I'll clarify the clinical/statistical meaning before you choose

📌 Note: When you've named the test type but are stuck on "which option / design family", the assistant pops a routing menu to confirm step by step — instead of deciding for you or dumping the full menu. The menu also offers an "explain the differences" option — when you're unsure, it explains before you decide.

Example 6 · Vague: Not sure which model (grill-me)

You say:

I want to do a meta-analysis but I'm not sure which model to use — help me figure it out

Assistant replies (sketch):

Sure — let's pin down your needs first; a few key questions will locate the right model:

1. What type is your primary endpoint? — continuous (e.g. SMD) / binary (e.g. OR) / survival (HR) / single-group rate / other?
2. Is it a two-group comparison or multi-group network?
3. Any special dependency structure? — multi-arm / multi-outcome / IPD / none

(After you answer, the assistant keeps probing until the model is locked, then computes — no full menu dumped on you.)

📌 Note: When you say "not sure which model", the assistant uses branch-by-branch probing to clarify your needs, instead of dumping the full menu.

Example 7 · Systematic review workflow

You say:

Help me generate a PRISMA flow diagram

Assistant replies (sketch):

Sure, please provide:

• Records identified: ___
• Records screened: ___
• Records excluded: ___
• Full-text assessed: ___
• Full-text excluded (+ reasons): ___
• Studies included: ___

User: 520, 380, 140, 80, 52(20 no data, 18 wrong pop, 14 wrong design), 28

✅ PRISMA flow diagram generated! [SVG output]

💡 Tip: Most analyses need only three things — effect size (or rate / HR) + α + power. Anything you omit is filled with sensible defaults. It's fine to be incomplete — the assistant will tell you what's missing.

---

2. What You Can Do — Scenario Index

Tests are grouped by analysis purpose (7 categories below). Each row gives the typical clinical scenario and a line you can copy verbatim under "Try saying". The same test may be reached from multiple entry points.

The underlying R packages (metafor / meta / netmeta …) are listed in Section 6 "Advanced Reference"; ordinary users don't need to care.

① Pairwise Meta-Analysis

| Scenario | Try saying in chat | |:---|:---| | Binary (OR/RR/RD) | "Combine the OR across these 5 binary studies" | | Continuous (SMD/MD) | "Pool the SMD of these 6 continuous studies" | | Pre-calculated (yi+CI) | "I have effect sizes and CIs for 5 studies — draw the forest plot directly" | | Survival (HR) | "Pool the HR across these 8 studies" | | Correlation (r→Zr) | "Convert these 4 correlations via Fisher z then pool" | | Single-group rate/mean | "Pool the incidence rates across these studies" | | Generic inverse-variance | "I have yi and vi — run the meta directly" |

② Heterogeneity & Bias

| Scenario | Try saying in chat | |:---|:---| | Heterogeneity assessment | "I ran a meta, I² is very high — help me assess heterogeneity" | | Subgroup analysis | "Run subgroup analysis by region" | | Meta-regression | "Run meta-regression on publication year and sample size" | | Egger test | "Check publication bias, run Egger's test" | | Begg test | "Begg rank-correlation test" | | Trim-and-fill | "Correct publication bias with trim-and-fill" | | Selection model | "Assess publication bias with a selection model" | | Sensitivity analysis | "Run leave-one-out sensitivity analysis" | | Cumulative meta | "Run cumulative meta by publication year" | | GOSH plot | "Plot a GOSH graph to see heterogeneity patterns" | | Baujat diagnosis | "Make a Baujat plot to see which study contributes most heterogeneity" | | Drapery plot | "Plot a Drapery graph to assess α robustness" |

③ Advanced Models

| Scenario | Try saying in chat | |:---|:---| | Frequentist NMA | "Run network meta with 4 interventions, use netmeta" | | Bayesian NMA (Stan) | "Run Bayesian network meta, Stan backend" | | Bayesian NMA (JAGS) | "Run Bayesian network meta, JAGS backend" | | Multilevel meta | "Run 3-level meta with multiple effects within studies" | | Multivariate meta | "Pool a meta with multiple correlated outcomes" | | IPD meta | "I have individual patient data — run IPD meta" | | Dose-response | "Run dose-response meta, dosresmeta" | | Survival meta | "Pool survival HR via metafor (survmeta removed)" | | Trial sequential analysis | "Run TSA — see how many more studies are needed" | | Bootstrap meta | "Use Bootstrap for nonparametric DL estimation" |

④ Effect Size & Conversion

| Scenario | Try saying in chat | |:---|:---| | Mean/SD→d | "Convert mean and SD to Cohen's d" | | t/F→d | "Convert a t value to d" | | r→Fisher z | "Convert a correlation to Fisher z" | | d↔logOR | "Convert d to logOR" | | OR↔logOR | "Convert OR to logOR" | | Batch convert | "Batch convert SMD to logOR" | | NNT | "Calculate NNT" |

⑤ Visualization

| Scenario | Try saying in chat | |:---|:---| | Forest plot | "Draw a forest plot, lancet theme" | | Funnel plot | "Draw a funnel plot with contour enhancement" | | Bubble plot | "Draw a meta-regression bubble plot" | | GOSH plot | "Plot a GOSH graph" | | Network plot | "Draw the network meta graph" | | League table | "Draw the NMA league table" | | RoB traffic-light | "Draw a risk-of-bias traffic-light plot" | | Power curve | "Draw a power curve" | | Drapery plot | "Plot a Drapery graph" | | Inconsistency heatmap | "Plot an NMA inconsistency heatmap" |

⑥ Study Quality

| Scenario | Try saying in chat | |:---|:---| | RoB 2.0 | "Assess risk of bias with RoB 2.0" | | RoB 1.0 | "Assess with Cochrane RoB 1.0" | | ROBINS-I | "Non-randomized study — use ROBINS-I" | | GRADE | "Do a GRADE evidence-quality assessment" | | PRISMA checklist | "PRISMA checklist" |

⑦ Systematic Review Workflow

| Scenario | Try saying in chat | |:---|:---| | PRISMA flow | "Help me generate a PRISMA flow diagram" | | Literature screening | "Title/abstract screening, AI-assisted" | | PDF batch download | "Batch download full texts from a DOI list (needs confirmation)" | | Graph digitize | "Extract data from a scatter plot" | | Missing value imputation | "Impute missing standard deviations" |

⚠️ PDF batch download connects to external networks and writes files to your local disk. Run it only on explicit user instruction, and respect copyright and access controls.

---

---

2.1 Supported Figures (23)

The skill renders 23 analysis figures on the cloud coze R engine. Pass the plot type via the plots field. prisma_flow / prisma and rob / rob2 resolve to the same figure.

| # | Plot type | 中文名 | English name | Analysis area | Purpose | |:---:|:---|:---|:---|:---|:---| | 1 | forest | 森林图 | Forest plot | Pairwise / NMA | Pooled effect-size summary | | 2 | funnel | 漏斗图 | Funnel plot | Pairwise | Publication-bias visual | | 3 | prisma_flow | PRISMA 流程图 | PRISMA flow diagram | Systematic review | Four-stage screening flow | | 4 | rob / rob2 | 偏倚风险图 | RoB traffic-light / summary | Study quality | Cochrane RoB 1.0 / 2.0 | | 5 | cumulative | 累积 Meta 图 | Cumulative meta plot | Pairwise | Accumulated by study order | | 6 | baujat | Baujat 图 | Baujat plot | Heterogeneity | Heterogeneity contributor | | 7 | labbe | L'Abbe 图 | L'Abbe plot | Pairwise (binary) | Effect-consistency check | | 8 | radial | Radial 图 | Radial / Galbraith plot | Heterogeneity | Radial heterogeneity view | | 9 | sucra | SUCRA 排名图 | SUCRA ranking plot | NMA | Intervention rank probability | | 10 | egger | Egger 回归散点图 | Egger's regression plot | Bias | Quantitative bias test | | 11 | contribution | NMA 贡献图 | NMA contribution plot | NMA | Design / comparison contribution | | 12 | loo | 留一法影响图 | Leave-one-out plot | Sensitivity | Sensitivity analysis | | 13 | gosh | GOSH 图 | GOSH plot | Heterogeneity | Heterogeneity pattern clusters | | 14 | bubble | 气泡图 | Bubble plot | Meta-regression | Covariate–effect relationship | | 15 | netgraph | 网络关系图 | Network graph | NMA | Evidence-network structure | | 16 | dose_resp | 剂量反应图 | Dose-response plot | Dose-response | Dose–effect relationship | | 17 | drapery | Drapery 图 | Drapery plot | Sensitivity | α robustness | | 18 | sroc | SROC 曲线 | SROC curve | Diagnostic MA | Diagnostic accuracy | | 19 | tsa | 试验序贯分析图 | Trial sequential analysis | TSA | Evidence sufficiency / required N | | 20 | power | 功效曲线 | Power curve | Power | Statistical power | | 21 | influence | 影响诊断图 | Influence diagnostic plot | Sensitivity | Single-study omission impact | | 22 | nodesplit | 节点拆分图 | Node-splitting plot | NMA | Local inconsistency | | 23 | trimfill | 剪补法漏斗图 | Trim-and-fill funnel plot | Bias | Bias-corrected funnel |

Note: netleague (NMA league table) is a tabular output, not a figure, so it is excluded from the count of 23.

3. First-Time FAQ

Q: I only gave effect size and study count, no other parameters — will it still compute? A: Yes. Most analyses need only 3 items — effect size (or rate / HR) + α + power. Omitted parts (two-sided α=0.05, 1:1 randomization, follow-up) are filled with sensible defaults; if something truly required is missing, the assistant will ask.

Q: Is the n in the result per group or total? A: By default it's per group; paired / crossover designs report per-sequence, and survival often reports total events needed. The output always labels this clearly.

Q: Does the analysis run as soon as I describe a request? A: Yes. Once you describe the request, the assistant auto-executes and returns the real numbers + figures — no extra trigger word needed. Computation runs on the cloud coze R engine (data disclosure in Section 5).

Q: I want the reproducible R code for submission or audit — how do I ask? A: Say "give me the full R code". Every analysis returns reproducible R code (with R and package versions), which you can copy, modify, and re-run yourself.

Q: On a Chinese system, is the output in Chinese? A: Yes. By default the output language follows your OS language setting — Chinese on a Chinese-OS, English otherwise. This default requires no extra permission and only affects display language; you can force-switch anytime via a prompt (e.g. "用中文回复" / "switch to English").

Q: My data is in SPSS/Excel/Stata format — what do I do? A: Say "help me convert my SPSS/Excel data to CSV" — the assistant will recommend installing @skill:statdata-transfer for 50+ format conversions.

Q: What if my data must stay confidential? A: Run the whole analysis with simulated / placeholder data, then ask the skill for the full reproducible R code and run it yourself locally with your real data. The skill itself only sends your analysis parameters / summary statistics (event counts, sample sizes, effect sizes) to the cloud coze R engine — it never touches your raw datasets or individual-patient records (unless you explicitly choose to run an IPD analysis through the cloud, in which case sending IPD to the cloud is your decision).

Q: What if I found an error in the result — how do I report it? A: This skill follows the standard bug-report workflow. If you suspect the result is wrong (or the engine errored), just say "report a bug" / "上报问题" / "提交错误报告". The skill also proactively asks whether to report when it detects a likely defect (e.g. the engine errors or retries still fail) — at most once per session, and you can always decline. Either way, the assistant will:

1. Propose a sanitized report (11-field whitelist: skill / skill_version / test / error_type / error_code / engine_status / description / locale / query_origin / session_hash / attempts — no raw input values or personal data, except the description field where you decide what to disclose, e.g. the algorithm/function used and the error message);
2. Show the full report text for your review — you can add a problem description or correct anything before confirming;
3. Send after your explicit confirmation — to the unified endpoint https://ct-bugreport.coze.site/run (if this session called coze) or, if purely local, save the sanitized report locally and show you the author contact so you can email it yourself if you choose (the skill itself does not send it; data never leaves your machine unless you email it);
4. Receive an acknowledgment — including whether a previously submitted report from your source has already been fixed (with the fix note) or is still pending.

You stay in full control: the report is shown to you before anything is sent, and nothing is transmitted without your explicit "send" confirmation.

---

4. Execution Model

• Auto-execution: Once you describe a request, the skill auto-executes the analysis and returns real numbers + figures — no extra trigger word or confirmation needed. Computation runs on the cloud coze R engine by default.
• Default compute path: The skill sends the analysis request to the cloud coze R engine (https://ct-meta.coze.site/run) (data disclosure in Section 5).
• Reproducible code: Every analysis returns reproducible R code (with R + package versions); say "give me the full R code" to obtain it for submission or audit.
• Outbound authorization: The default endpoint is pre-approved and runs automatically; a custom endpoint (COZE_META_ENDPOINT) asks for confirmation on first use (see Section 5).
• Output is for reference only — validate before journal submission or regulatory use.

---

5. Data & Privacy

The skill sends data externally in two situations: ① when you describe an analysis request, the skill auto-sends the analysis request to execute; ② when you confirm sending an error report. Neither sends personal identifiers.

5.1 Analysis request (cloud computation)

• What is sent: your analysis data — summary statistics such as study event counts / sample sizes / effect sizes. No personal identifiers; payloads are sanitized before sending.
• When: the skill auto-sends after you describe a request; before the first outbound call each session, the skill gives you a one-time spoken disclosure of what is sent and to which endpoint (then executes automatically, without per-call confirmation).
• Endpoint: default https://ct-meta.coze.site/run (pre-approved in adapters/config.json auto_approve_endpoints). A custom endpoint (COZE_META_ENDPOINT) asks for confirmation on first use (AUTH-BLOCK), and is persisted to the whitelist after you approve.
• If declined: the skill returns a clear "cloud analysis not used" message.

5.2 Metadata sent with the request Each request also carries two metadata fields (in both the analysis request and the error report):

• query_origin: a SHA-256 hash of your machine hostname, used only for server-side attribution / rate-limiting — not your plaintext hostname;
• locale: your OS language, for bilingual output.

Neither is used to identify you personally.

5.3 Error report

• What is sent: only the 11-key whitelist envelope (skill / skill_version / test / error_type / error_code / engine_status / description / locale / query_origin / session_hash / attempts) — no analysis data and no personal identifiers. description is the only free-text field, and you review it before consent (hard boundary: no identifiable person/institution/subject info).
• Endpoint: unified bug-report endpoint https://ct-bugreport.coze.site/run.
• If declined: nothing is sent; if there is no cloud call this session, the report is saved locally instead (save_local_report, data never leaves the machine).

In one sentence: your analysis summary data is auto-sent to the cloud after you describe a request (with a one-time disclosure before the first outbound call each session); error reports go to the unified endpoint only after your confirmation; the two metadata fields (query_origin hash + locale) are for anonymous attribution. Raw data and individual records never leave your machine.

---

6. Advanced Reference (moved to a separate file)

CLI examples, bidirectional solving, curve mode, core formulas, system requirements, common errors, full file structure, and references for developers have been moved to references/ADVANCED.md. Ordinary users don't need it; see Sections 1-5 for daily use.

---

Version: v2.1.5 | License: MIT | Authors: medstatstar, phoe-zip

For feature requests, bug reports, or other feedback, please contact the author directly at medstatstar@gmail.com (Wintone Zhang / 张文彤).

---

Confidentiality Notice

The CT series consists of 20+ specialized domain skills, organized into two tiers — A, B — by "confidential-data-exfiltration risk + whether external retrieval is needed", providing full coverage of the entire new-drug clinical trial (Clinical Trial) lifecycle.

• Tier A (non-confidential, public): inputs are ordinary data, run fully locally (network=off) or with external public retrieval (network=public-retrieval, e.g. ct-registry / ct-advisor); no confidential information involved. Tier A skills are published openly on GitHub.
• Tier B (confidential, internal): involve strictly confidential clinical-trial data and internal information from pharma sponsors (e.g., ct-analysis, ct-sdtm, ct-eligibility); Tier B is processed locally (egress=none, data never leaves the boundary) or requires approval for outbound (egress=approval-req, e.g. ct-eligibility). Tier B skills are designated for internal enterprise use only and are not publicly released at present.

If you do have a genuine need for these confidential skills, please contact the author to request custom installation.

📧 Contact: medstatstar@gmail.com (Wintone Zhang / 张文彤)