Meta Analysis R

R Meta-Analysis Workflow

Use this skill when the user wants to actually run a meta-analysis (not just search literature). Inputs are a CSV of extracted study data; outputs are plots and a Markdown report under ~/Desktop/meta-analysis//.

This is the natural downstream step from [[referencepubmedskill]] (pubmed-search-casp) — but data extraction is a manual step in between and this skill must not synthesize numeric data from PDFs.

Core Rules

1. Never invent data. If a cell in the input CSV is empty or non-numeric, stop and ask. Do not impute means, SDs, or event counts from study abstracts unless the user explicitly says "estimate from abstract" and the estimate is recorded with that caveat in the report.
2. Validate before running. Check column types, n > 0, SD > 0, no negative counts, ≥2 studies for pairwise (≥3 + connected network for netmeta).
3. Show the R script. Always write the R code to a file in the output directory so the user can re-run / modify. Do not hide it inside a one-shot heredoc.
4. Random effects by default (REML estimator). Rehab/exercise data is almost always heterogeneous — fixed-effect is misleading. Override only if user asks.
5. Hedges' g (SMD) by default for continuous outcomes when scales differ; MD when all studies use the same scale.

Environment Check (run once at start)

R --version
R -e 'pkgs /dev/null | tail -3

If anything missing, tell the user the install command (install.packages(c("meta","metafor","netmeta"))) — don't silently install.

Input Modes

Mode A — Template CSV (primary)

User fills one of these templates depending on outcome type:

Binary outcome (binary_template.csv):

study,year,event_t,n_t,event_c,n_c,subgroup
Smith2020,2020,12,45,20,44,older

Continuous outcome (continuous_template.csv):

study,year,mean_t,sd_t,n_t,mean_c,sd_c,n_c,scale,subgroup
Smith2020,2020,42.1,5.3,30,38.4,6.1,30,6MWT,older

Generic / pre-calculated (generic_template.csv):

study,year,TE,seTE,subgroup
Smith2020,2020,0.34,0.12,older

Network meta (long format) (network_template.csv):

study,treatment,event,n
Smith2020,strength,12,30
Smith2020,aerobic,8,30
Smith2020,control,4,30

(or study,treatment,mean,sd,n for continuous)

Save templates inside the skill directory and copy them into the user's output folder when they ask "give me the template".

Mode B — From CASP output folder

If the user points to a folder produced by [[referencepubmedskill]] (typically ~/Desktop//included/):

1. Read the included-articles list (filenames, PMIDs, or a manifest file if present).
2. Generate an empty template populated with study column = first-author + year extracted from filenames. All numeric columns left blank.
3. Save as _to_fill.csv in the output folder.
4. Tell the user: "I've created the template with N studies pre-filled. Please open it and fill the numeric columns from each study's full text. Re-run me when done."
5. Do not proceed to analysis until the numeric columns are non-empty.

Workflow

1. Confirm scope. Ask: outcome type (binary / continuous / generic / network), effect measure (RR/OR/MD/SMD), topic/title for the output folder. If the user already named the file binary_*.csv etc., skip the question.

1. Validate the CSV.

• Required columns present
• No NA in required numeric columns
• n, sd, event counts ≥ 0; sd, n strictly > 0
• ≥2 studies (pairwise) or ≥3 studies forming connected network (netmeta)
• For binary: warn (don't block) on zero cells — meta handles continuity correction automatically

1. Create output folder. ~/Desktop/meta-analysis// with subfolders plots/ and data/ (copy input CSV into data/).

1. Generate the R script (analysis.R) into the output folder. Template structure below.

1. Run the script. R --no-save analysis.log 2>&1. Check exit code; if it fails, surface the R error verbatim and stop.

1. Write the Markdown report (report.md) — see structure below.

1. Tell the user the output path and list the three artifacts (analysis.R, plots/*.png, report.md).

Required Plot Set

Every run produces a standard plot bundle so the user gets the full diagnostic picture, not just the headline forest plot. Conditional plots are gated by what makes sense given the data.

| Plot | Always? | Skip condition | What it shows | |---|---|---|---| | Forest (main) | yes | — | Pooled effect + per-study effects | | Funnel | yes (k≥3) | k.csv", stringsAsFactors = FALSE)

m =3; contour-enhanced) if (nrow(dat) >= 3) { png("plots/funnel.png", width = 1000, height = 1000, res = 150) funnel(m, studlab = TRUE, contour = c(0.9, 0.95, 0.99), col.contour = c("grey80","grey60","grey40")) legend("topright", legend = c("p>0.1","0.05=10 if (nrow(dat) >= 10) { eg = 2) { ms / ├── analysis.R ├── analysis.log ├── report.md ├── data/ │ ├── .csv (copy of user's input) │ ├── summary.txt │ ├── leaveoneout.txt │ ├── egger.txt (if k≥10) │ ├── subgroup.txt (if subgroups) │ ├── metareg.txt (if meta-regression) │ └── notes.txt (any skips/caveats) └── plots/ ├── forest.png ├── funnel.png (k≥3) ├── baujat.png (k≥3) ├── radial.png (k≥3) ├── drapery.png (k≥3) ├── forestloo.png (k≥3) ├── forestsubgroup.png (if subgroups) ├── labbe.png (binary only) ├── bubble.png (meta-regression only) ├── netgraph.png (network mode only) └── netrank_forest.png (network mode only)

## When to refuse / push back

- User asks for a meta-analysis with 1 study → refuse, suggest narrative synthesis instead
- User asks to "estimate the missing SDs" → push back; offer SD-from-SE / SD-from-CI / SD-from-IQR formulas (Cochrane Handbook 6.5.2) and ask which the source paper actually reports
- User asks to combine clearly heterogeneous outcomes (e.g. 6MWT seconds with gait speed m/s) without standardization → push back, recommend SMD or splitting into separate analyses
- Network has disconnected components → run pairwise on the connected piece, tell the user which arms are disconnected

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [Lokenqaq111](https://github.com/Lokenqaq111)
- **Source:** [Lokenqaq111/meta-analysis-skill](https://github.com/Lokenqaq111/meta-analysis-skill)
- **License:** MIT

Install and usage instructions live in the source repository linked above.