pymmeans 0.1.0__tar.gz

pymmeans-0.1.0/.gitignore

@@ -0,0 +1,62 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+dist/
+*.egg-info/
+*.egg
+.eggs/
+
+# Virtual envs
+.venv/
+venv/
+env/
+
+# Test / coverage
+.pytest_cache/
+.coverage
+.coverage.*
+htmlcov/
+.tox/
+.mypy_cache/
+.ruff_cache/
+
+# Editors / OS
+.idea/
+.vscode/
+.DS_Store
+*.swp
+*~
+.Trash-*
+
+# Jupyter checkpoint dirs
+.ipynb_checkpoints/
+*/.ipynb_checkpoints/
+
+# Docs build (mkdocs writes to ./site, sphinx to docs/_build)
+docs/_build/
+site/
+.mkdocs_cache/
+
+# Benchmarks output
+benchmarks/r_results.csv
+benchmarks/PERFORMANCE_REPORT.md
+
+# Internal dev docs (kept locally; not published)
+docs/pymmeans_project_plan.md
+docs/v0_1_limits.md
+
+# JSS manuscript (separate publication track)
+paper/
+
+# Internal-only test file (regression guards from a long iterative
+# development cycle; kept locally for the maintainer, omitted from
+# the public surface)
+tests/test_audit_regressions.py
+
+# Package-manager lockfile (this is a library, not an application;
+# end-users install via pip with their own resolver)
+uv.lock

pymmeans-0.1.0/CHANGELOG.md

@@ -0,0 +1,53 @@
+# Changelog
+
+All notable changes to `pymmeans` will be documented in this file.
+
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/);
+this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.1.0] — 2026-05-25
+
+Initial release.
+
+### Added
+
+- Reference-grid construction (`ref_grid`) and estimated marginal means
+  extraction (`emmeans`, `lsmeans`) for fitted `statsmodels` models,
+  `linearmodels` panel / IV results, and any user-supplied
+  `predict_fn(data) -> ndarray` callable.
+- Pairwise contrasts (`pairs`), generic linear contrasts (`contrast`),
+  compact letter displays (`cld`), and pairwise p-value matrices
+  (`pwpm`).
+- Multiplicity adjustments: Tukey HSD (exact studentised-range
+  integral), Dunnett (exact via the multivariate-_t_ CDF), Šidák,
+  Bonferroni, Holm, Benjamini–Hochberg FDR, and the generic `mvt`
+  integral.
+- Small-sample mixed-model inference: `apply_satterthwaite`,
+  `apply_kenward_roger`, plus the six headline `pbkrtest`
+  equivalents (`kenward_roger_vcov`, `get_kr`, `ddf_lb`, `krmodcomp`,
+  `satmodcomp`, `pbmodcomp`).
+- Response-scale back-transformation (`type="response"`) for
+  GLM-style links, including bias-adjusted estimates for log links.
+- Bootstrap (`bootstrap_ci`, both parametric and case-resampling)
+  and permutation tests (`permutation_test`).
+- Bayesian-posterior EMMs via `from_pymc` / `posterior_emmeans` /
+  `posterior_emm_summary` (optional `arviz` / `PyMC` integration).
+- Survey-weighted EMMs via `from_survey` (Lumley-style designs).
+- ML adapter (`from_predict`, `ml_emmeans`, `ml_pairs`,
+  `ml_contrast`) for tree ensembles, gradient-boosted models, neural
+  networks, and any other `.predict()`-capable estimator.
+- `emtrends` for derivatives of the regression surface at focal points.
+- `OrderedModel` (cumulative-link ordinal) and `MNLogit` (multinomial
+  logit) support.
+
+### Tested
+
+- 858 unit tests against `statsmodels`, `linearmodels`,
+  `scikit-learn`, `xgboost`, `lightgbm`, `PyTorch`, `survey`, and
+  R-side reference values from `emmeans`, `lme4` + `lmerTest`,
+  `pbkrtest`, `marginaleffects`, and `survey`.
+- R parity tolerances range from `atol < 1e-7` (survey-weighted
+  Gaussian) to `atol < 1e-3` (finite-difference Satterthwaite df);
+  the six `pbkrtest` equivalents match the R reference at
+  `atol < 1e-5` on identical $\hat\theta$ inputs.
+

pymmeans-0.1.0/CITATION.cff

@@ -0,0 +1,62 @@
+cff-version: 1.2.0
+message: "If you use pymmeans in your research, please cite it as below."
+title: "pymmeans: Estimated marginal means for Python"
+abstract: >
+  pymmeans is a native-Python implementation of R's emmeans package
+  (Lenth 2024) for estimated marginal means (EMMs), pairwise and
+  custom contrasts, and multiplicity adjustments on fitted statistical
+  models. It supports statsmodels OLS / GLM / MixedLM / GEE / Cox /
+  BetaModel, the linearmodels panel and IV result classes, and any
+  prediction-only model with a ``.predict()`` method via a
+  beyond-R-parity ML adapter (g-computation / prediction-surface
+  averaging). The implementation uses an analytic Kronecker-product
+  marginalisation (no grid materialisation) so EMMs on designs with
+  500K-row to 46M-row reference grids run in milliseconds where R
+  emmeans refuses or runs out of memory. Inference paths include
+  Wald, Satterthwaite, Kenward-Roger (Kackar-Harville), parametric
+  bootstrap, non-parametric case bootstrap with refit, permutation
+  testing with Phipson-Smyth correction, and Bayesian / posterior
+  draws.
+authors:
+  - family-names: Turner
+    given-names: Jason
+    email: jason.s.turner@gmail.com
+version: 0.1.0
+date-released: 2026-05-25
+license: GPL-3.0-or-later
+repository-code: "https://github.com/jturner-uofl/pymmeans"
+url: "https://github.com/jturner-uofl/pymmeans"
+type: software
+keywords:
+  - estimated marginal means
+  - emmeans
+  - least-squares means
+  - lsmeans
+  - linear models
+  - mixed-effects models
+  - generalized linear models
+  - multiplicity adjustment
+  - Satterthwaite degrees of freedom
+  - Kenward-Roger correction
+  - bootstrap inference
+  - g-computation
+  - statsmodels
+  - Python
+references:
+  - type: article
+    title: "Least-squares means: the R package lsmeans"
+    authors:
+      - family-names: Lenth
+        given-names: Russell V.
+    year: 2016
+    journal: "Journal of Statistical Software"
+    volume: 69
+    issue: 1
+    doi: 10.18637/jss.v069.i01
+  - type: software
+    title: "emmeans: Estimated marginal means, aka least-squares means"
+    authors:
+      - family-names: Lenth
+        given-names: Russell V.
+    year: 2024
+    url: "https://cran.r-project.org/package=emmeans"