---
name: meta-analysis-planning
description: Plan a defensible quantitative evidence synthesis, including effect measures, pooling criteria, dependence, heterogeneity, and sensitivity analyses.
---

# Meta-analysis planning

Use when the user seeks an analysis plan for pooling study results or evaluating whether pooling is appropriate. Establish the substantive question, study designs, outcomes, comparison, follow-up window, available data, and decision the synthesis must inform.

## Plan the analysis

Define eligible estimates and a consistent direction of benefit or harm. Select effect measures appropriate to outcome type and design, specifying how estimates and standard errors will be obtained or calculated. Record assumptions required for conversions, such as correlations or imputed dispersion, and give a plan for testing their influence. Do not convert fundamentally incompatible outcomes into a common scale without justification.

Assess clinical, methodological, and statistical compatibility before pooling. Identify dependence from repeated time points, multiple arms, overlapping participants, and multiple reports. Specify one defensible strategy for handling each, such as a prespecified selection rule, covariance-aware model, or separate analysis. Specify the model and estimator, rationale, weighting, confidence intervals, heterogeneity measures, and prediction interval where appropriate; explain limits when few studies are available.

Prespecify meaningful subgroup and sensitivity analyses, including missing data and study quality where relevant. Explain why tests for small-study effects or publication bias may be uninformative with limited evidence. Specify how risk of bias and certainty will affect interpretation. If pooling is unjustified, describe a structured non-pooled synthesis instead.

## Output and dependencies

Deliver an analysis specification with required inputs, assumptions, decisions, diagnostic checks, and reproducible software or pseudocode when requested. Use established methods and verify current documentation for selected software. No estimates, effect sizes, p-values, or robustness claims should be invented; distinguish an analysis plan from executed calculations. Input data, reviewer judgment, and statistical expertise may be needed before results can be trusted.