Meta Analysis Forge

Meta-Analysis Forge

Use this skill when evidence synthesis requires statistical pooling of primary-study effects.

Core Principle

A meta-analysis is valid only when the effect sizes being combined are conceptually and statistically comparable enough for the target inference.

Separate:

• effect-size extraction;
• effect-size conversion;
• dependence among effects;
• model choice;
• heterogeneity interpretation;
• publication-bias diagnostics;
• substantive conclusion.

Intake

Identify:

• outcome construct;
• effect-size metric;
• standard error, confidence interval, p-value, or sample size availability;
• number of studies;
• multiple effects per study;
• study designs;
• expected heterogeneity;
• moderators;
• field norms.
• whether raw, participant-level, sample-level, or harmonized derived data are available.

Load:

• references/effect-sizes.md for effect metrics and extraction.
• references/soil-fauna-carbon-meta.md when the project pools ecological effects on both carbon stocks and carbon fluxes and needs trait or climate moderators without collapsing incompatible outcome families.
• references/ecological-meta-ml-path-model-paradigm.md when the project combines meta-analysis, mixed-effects meta-regression, random forest variable ranking, and PLS-PM/SEM-family path modeling.
• references/high-value-paper-reproducibility-audit.md when a strong published meta-analysis should become a reusable template and the task requires checking code, data-table structure, rma.mv, random forest, PLS-PM/SEM-family modeling, and reproducibility.
• references/ipd-and-mega-analysis.md when the task involves individual participant data, multi-site raw/derived data harmonization, small-sample dataset integration, or mega-analysis.
• references/synthesis-models.md for model choice and diagnostics.
• references/network-meta-analysis.md when comparing three or more treatments across a connected evidence network (frequentist NMA, P-score ranking, and node-splitting inconsistency with an explicit trust check).
• references/ml-moderator-analysis.md for exploratory interpretable-ML moderator analysis (MetaForest/SHAP/EBM/GAM) with the small-k honesty guardrails — complements, never replaces, pre-specified meta-regression.
• references/meta-analysis-quality-gates.md for pre-pooling checks.
• templates/coding-schema.csv and templates/validation-rules.md for machine-readable coding-sheet structure and validation.
• scripts/validate_coding_sheet.py before statistical execution.
• scripts/effect_size_helpers.R for transparent mechanical conversions during extraction.
• scripts/run_meta_analysis.R only after coding validity and pooling appropriateness have been checked.
• scripts/install_r_packages.R when setting up the minimal R environment.

Workflow

1. Define the effect-size family.
2. Build the coding sheet.
3. Convert or preserve metrics with justification.
4. Identify dependence: multiple outcomes, time points, samples, or models per study.
5. Pass the quality gates before pooling.
6. Choose model: fixed, random, multilevel, robust variance, Bayesian, or narrative synthesis.
7. Report heterogeneity: tau2, I2, prediction interval.
8. Assess small-study effects or publication bias when feasible.
9. Run sensitivity checks.
10. Write interpretation with limits.

For IPD or mega-analysis, first build a dataset inventory, harmonization plan, quality-control ledger, and study/site heterogeneity model before any pooled interpretation.

Output Modes

Coding Sheet

Use:

• templates/coding-sheet.md for a human-readable table.
• templates/coding-schema.csv for field definitions.
• templates/example-coding-sheet.csv for a minimal machine-readable example.

Analysis Plan

Effect-size metric:
Inclusion for pooling:
Model:
Dependence handling:
Heterogeneity:
Bias diagnostics:
Sensitivity checks:
Software:
Interpretation limits:

Minimal R Run

Use scripts/run_meta_analysis.R for a small reproducible demonstration when the coding sheet has one harmonized effect metric and valid standard errors.

Input CSV:
Output directory:
Effect metric:
Pre-pooling checks passed:
Known limits:

Validation and Conversion

Use scripts/validate_coding_sheet.py to check required fields, numeric estimates, positive standard errors, duplicate effect IDs, and mixed effect metrics.

Use scripts/effect_size_helpers.R only for transparent mechanical helpers such as CI-to-SE, log-ratio transforms, Fisher z, approximate SMD SE, and lnROM. Record formulas and assumptions in the coding sheet notes.

IPD / Mega-Analysis

Use:

• references/ipd-and-mega-analysis.md for workflow and guardrails.
• templates/mega-analysis-dataset-inventory.csv for data access and harmonization.
• templates/mega-analysis-audit-report.md for audit output.

Audit

Flag:

• incompatible outcomes;
• mixed effect metrics without conversion;
• missing uncertainty;
• multiple effects treated as independent;
• overuse of I2 without prediction interval;
• meta-regression overclaiming;
• publication-bias tests with too few studies.

Ecological Meta + ML + Path Model

Use templates/ecological-meta-ml-path-model-audit.md when a meta-analysis combines pooled effects, moderator testing, machine-learning driver ranking, and a path model or SEM-family diagram.

Meta-analysis layer:
ML layer:
Path-model layer:
Effect-size families:
Dependence plan:
Main reuse lesson:
Main overclaim risk:

High-Value Paper Reproducibility Audit

Use templates/high-value-paper-reproducibility-audit.md when the user wants to learn from a strong article, especially a Nature Communications or similar paper with public data/code. Do not stop at a paper summary.

Extract:

• file and repository inventory;
• data table structure;
• effect-size and uncertainty logic;
• metafor::rma.mv() implementation;
• shared-control VCV or other dependence handling;
• random forest or machine-learning layer;
• plspm, PLS-PM, PLS-SEM, or other path-model layer;
• peer-review lessons;
• reproducibility gaps;
• reusable skill rules.

Article logic:
Data table structure:
Effect-size logic:
rma.mv / dependence implementation:
Random forest layer:
PLS-PM / path-model layer:
Reproducibility verdict:
Reusable rule:

Guardrails

• Do not invent effect sizes.
• Do not pool effects solely because they are numerically available.
• Do not interpret meta-regression causally unless design supports it.
• Do not interpret random-forest importance or PLS-PM paths causally unless the design supports it.
• Do not ignore within-study dependence.
• Do not treat a high pooled N as proof of high evidence quality.
• Do not use vote-counting as a substitute for effect-size synthesis.
• Do not treat the minimal R script as a full meta-analysis pipeline; it does not solve effect conversion, dependence, or certainty assessment.
• Do not run effect-size helper conversions without preserving original reported values and source anchors.
• Do not call a project a mega-analysis unless raw, participant-level, sample-level, or harmonized derived data are reprocessed or remodeled under a common framework.
• Do not call a high-value paper reproducible until its public code/data files, data schema, package versions, and model scripts have been inspected.