PyMetaAnalysis is an open-source Python library for conventional study-level meta-analysis. I built it around a simple principle: a statistical result should be easy to use, but never opaque.
Why I built it#
Meta-analysis often begins with data already living in a pandas DataFrame, yet many workflows require reshaping that data around a tool’s assumptions. Just as importantly, a pooled estimate alone is not enough: analysts need to understand which studies were included, how edge cases were handled, which methods were selected, and how the result can be reproduced.
PyMetaAnalysis keeps that full trail attached to the result. It accepts DataFrames, NumPy arrays, and ordinary Python sequences, and returns immutable result objects containing study effects, exclusions, weights, diagnostics, method choices, provenance, and structured reports.
What it supports#
- Generic effects supplied with a sampling variance or standard error
- Binary outcomes using odds ratios, risk ratios, and risk differences
- Continuous outcomes using mean differences and Hedges’ g
- Common-effect, random-effects, and Mantel–Haenszel models
- REML, Paule–Mandel, and DerSimonian–Laird heterogeneity estimators
- Subgroup analysis, leave-one-out analysis, and cumulative meta-analysis
- Forest and funnel plots through optional Matplotlib integration
- JSON and Markdown reports designed for downstream use
Designed for auditability#
Rows excluded by missing-data or sparse-data policies remain visible, with stable identifiers and explicit exclusion reasons. Transformations—such as converting a standard error to a variance—are recorded in provenance. For ratio measures, model-scale values remain available alongside display-scale estimates.
The test suite combines hand calculations, statistical invariants, numerical edge cases, and committed reference results from R’s metafor package. Continuous integration covers Python 3.10–3.13, strict typing and linting, documentation, and distribution builds.
A small example#
import meta_analyze as ma
result = ma.meta_analysis(
effect=[0.12, 0.35, -0.08, 0.21],
variance=[0.04, 0.06, 0.03, 0.05],
study=["Trial A", "Trial B", "Trial C", "Trial D"],
model="random",
tau2_method="REML",
)
print(result.summary())
print(result.study_results)The distribution is named PyMetaAnalysis, while the import name is meta_analyze. The project is released under the MIT License and is currently in its early 0.x stage.