glmm

Linear-regression family — OLS → GLM → LMM → GLMM — in pure Rust, with an eight-name Python API.

Fits fixed-effect and mixed (random-intercept / random-slope) models for Gaussian, Binomial (logit/probit), Poisson, Gamma, and Negative-Binomial outcomes. The numerics are the glmm Rust crate — no BLAS/LAPACK system dependency, no unsafe, validated against R/lme4 and Julia/MixedModels.jl goldens.

Install

pip install glmm

Requires Python 3.10+. The only runtime dependency is NumPy.

One call

You hand fit a data table and an R-style formula. It parses the formula against the table's columns, builds the design matrix, fits, and returns a Fit. There is no model object to construct.

import glmm

data = {
    "y":     [1.02, 1.05, 1.11, 1.13, 1.21, 1.24, 1.30, 1.33, 1.42, 1.44, 1.51, 1.53],
    "x1":    [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1],
    "group": ["a", "a", "b", "b", "c", "c", "d", "d", "e", "e", "f", "f"],
}

# y ~ x1 + (1 | group) — Gaussian, random intercept.
fit = glmm.fit(data, "y ~ x1 + (1 | group)")

assert fit.converged
fit.summary()   # prints the coefficient table (and returns it as a string)

The public surface is eight names: glmm.fit, glmm.Fit, and the six warning categories the diagnostics channel raises — glmm.DiagnosticWarning (the base) plus glmm.IllConditionedWarning, glmm.PirlsExhaustedWarning, glmm.UnusedGroupingLevelsWarning, glmm.ReDesignScaleWarning and glmm.HessianSeFallbackWarning. Families, links, and knobs are string or scalar arguments, not types.

data is documented as a dict[str, array-like], but anything column-addressable works — pandas / polars DataFrame, pyarrow Table — by duck-typing, so there is no dataframe dependency.

Model routing

Routing is inferred from the formula and family — fixed-only means OLS/GLM, adding a (… | g) term makes it LMM/GLMM.

fit = glmm.fit(data, "s ~ x1 + (1 | group)", "binomial", link="probit", nagq=7)

The formula follows R conventions: * desugars to main effects + interaction, A/B to nesting, (1 + x | g) is a correlated random intercept + slope, and additional (… | g2) terms add crossed or nested grouping factors. Treatment contrasts (R's default) code the factors against the column's first level; the full syntax, rejected constructs and workarounds, and factor-coding rules are in formula.md.

Not yet implemented

Two combinations have an approved design but no kernel support, and raise a clean NotImplementedError: quasi-likelihood dispersion= on binomial/poisson, and a float init_theta= seed (only the default init_theta=None cold start is supported). See troubleshooting.md for fixes to common errors.

Documentation

Full walkthrough — every knob, the Fit object, and warm starts: tutorial-python.md.

The algorithm map (dispatch graph, solver paths, tuning knobs) is traced to code in documentation/algorithms.md.

License

LGPL-3.0-or-later.

Paweł Lenartowicz — Freestyler Scientist · GitHub · ORCID