pylmrob 0.5.15

Python port of robustbase::lmrob (MM regression)

pylmrob

Python port of the lmrob MM-estimator from R's robustbase package.

Documentation: pylmrob.readthedocs.io.

Status: v0.5.11 (alpha). End-to-end pipeline matches R's lmrob element-wise on the classical robust regression datasets (stackloss, coleman, delivery, aircraft, phosphor, etc.) within rtol=1e-3 for both coefficients and covariance. Median wall-clock is 0.93x R across the 34-case bench corpus. See plan.md for the full roadmap and docs/numerical-notes.md for known divergences from R.

What works

• All six psi families (bisquare, huber, hampel, optimal, lqq, ggw) match R's Mpsi/Mchi/Mwgt to rtol=1e-12 (1e-7 for ggw rho's polynomial approximation).
• M-scale matches R's find_scale to rtol=1e-9.
• Fast-S resampling + MM iteration produces final coefficients matching R on the validation corpus (10 classical datasets) within per-dataset tolerances ranging 1e-3 (well-conditioned) to 1e-1 (small-n).
• Sandwich covariance estimator (vcov_avar1).
• Control presets matching R's lmrob.control(setting={"KS2014","KS2011","MM"}).
• Affine equivariance, scale equivariance, regression equivariance verified via Hypothesis property tests.
• init="M-S" for factor designs (Maronna-Yohai 2000, 4-phase port of robustbase's R_lmrob_M_S).
• D-scale refinement (Koller & Stahel 2014) for setting="KS2014" and setting="KS2011"; full SMDM pipeline matches R element-wise.
• vcov.w with all five cov.corrfact branches plus the Huber finite-sample correction, matching R within rtol=1e-3.
• summary() (coefficient table, robust R-squared) and anova() (Wald + Deviance) matching R element-wise.
• Per-case weights argument: matches R element-wise on the NumPy path. Non-trivial weights force Control(engine_c=False) for now.
• Default Control() runs the monolithic Cython engine (single nogil C block for fast-S, MM, D-scale, and vcov); see Performance.

What does not work yet

• Bit-identical reproducibility with R's MT RNG. We use NumPy's PCG64 (waived in plan.md §5.2; coefficients agree with R within basin-of-attraction tolerances documented in docs/numerical-notes.md).

Performance

Median wall-clock vs R across the 34-case bench corpus is 0.93x; pylmrob is faster than R on more than half of cases (down to 0.32x R on n>=2000 fits) and within 1-2x R on small classical datasets.

The default Control() runs the fit through a monolithic Cython engine (fast-S, MM, D-scale, and vcov in one nogil C block). At larger n (n*p^2 >= 100k) lmrob() auto-falls-back to a threaded NumPy path because the monolithic kernel is a single non-parallel call. To force the legacy NumPy path explicitly, pass Control(engine_c=False).

See docs/bench-report.md for the full per-case table and docs/engine_c.md for the trade-off discussion.

Install (from source)

Requires Python 3.10+, a C compiler, and (for the validation harness) a local R install with robustbase.

git clone https://github.com/anevolbap/pylmrob
cd pylmrob
uv venv
uv pip install --no-build-isolation -e ".[dev]"
pytest

Quickstart

import pandas as pd
from pylmrob import lmrob, Control

df = pd.read_csv("stackloss.csv")
fit = lmrob(
    "stack.loss ~ Air.Flow + Water.Temp + Acid.Conc.",
    df,
    control=Control(setting="KS2014"),
    seed=42,
)
print(fit.summary())

# Coefficients match R within rtol=1e-3:
# (Intercept) -41.5246
# Air.Flow      0.9388
# Water.Temp    0.5796
# Acid.Conc.   -0.1129
# scale = 1.9123

# Predict on new rows; the formula spec is re-applied automatically.
new_rows = pd.DataFrame({
    "Air.Flow":    [60, 80],
    "Water.Temp":  [20, 25],
    "Acid.Conc.":  [85, 90],
})
fit.predict(new_rows)

# Confidence intervals:
fit.confint(level=0.95)

Documentation

The Sphinx docs are hosted at pylmrob.readthedocs.io and built on every push (config in .readthedocs.yaml). To build locally:

uv pip install -e ".[docs]" --no-build-isolation
cd docs && uv run sphinx-build -b html -W . _build/html
open _build/html/index.html

Pages: quickstart, API reference, R-to-Python porting cheatsheet, and the numerical-notes log of documented divergences from R.

Comparison with statsmodels

statsmodels.robust.RLM implements M-estimators (Huber, Tukey, Hampel) but not MM-estimators. It will not give you the high-breakdown-point + high-efficiency combo that lmrob provides. If you need MM, use pylmrob. If you only need M with no contamination concerns, RLM is fine and has a more mature API.

License

GPL-3.0-or-later, matching robustbase. See LICENSE and NOTICE.

Citing

If you use pylmrob in research, please cite the project alongside robustbase. Citation metadata is in CITATION.cff (GitHub shows a "Cite this repository" button on the sidebar). The references include the key papers: Yohai (1987) on MM-estimators, Salibian-Barrera & Yohai (2006) on the fast-S algorithm, and Koller & Stahel (2011) on the small-sample Wald correction implemented in vcov.w.