regdiag 0.1.2

R-style regression diagnostic plots (plot(lm)) for statsmodels and scikit-learn

regdiag

R's plot(lm_model) for Python — one line, works with both statsmodels and scikit-learn.

R has a famous one-liner for regression diagnostics: plot(lm_model) produces four classic diagnostic plots (Residuals vs Fitted, Normal Q-Q, Scale-Location, Residuals vs Leverage). Python has never had a clean equivalent — the pieces exist scattered across statsmodels (only works on its own OLS objects) and a few small, low-traction packages. regdiag fills that gap.

from sklearn.linear_model import LinearRegression
import regdiag

model = LinearRegression().fit(X, y)
regdiag.diagnose(model, X, y)

Why regdiag

• One line, R-style output. regdiag.diagnose(model, X, y) gives you the full 4-panel figure, same as plot(lm).
• Works with statsmodels and scikit-learn. Any sklearn estimator exposing .coef_/.intercept_ (LinearRegression, Ridge, Lasso, ElasticNet, SGDRegressor, BayesianRidge, ...) or a statsmodels OLS results object.
• statsmodels is fully optional. Core dependencies are just numpy, scipy, matplotlib, and scikit-learn. VIF and the Breusch-Pagan test are computed with plain numpy/scipy — no statsmodels needed at runtime, even though it's the library people usually reach for to get those numbers.
• Honest about black-box models. Pass a RandomForestRegressor or any other model that only exposes .predict() and regdiag switches to a clearly-labeled generic mode (Residuals vs Fitted, Residual Distribution, Actual vs Predicted). It will never compute leverage or Cook's distance for a model where the hat-matrix math doesn't apply — those statistics are only well-defined for linear models.

Install

pip install regdiag

# with statsmodels support (accept statsmodels OLS objects as input):
pip install "regdiag[statsmodels]"

Quickstart

import numpy as np
from sklearn.linear_model import LinearRegression
import regdiag

rng = np.random.default_rng(0)
X = rng.normal(size=(200, 3))
y = 1.0 + X @ np.array([2.0, -1.5, 0.5]) + rng.normal(scale=1.0, size=200)

model = LinearRegression().fit(X, y)

diag = regdiag.diagnose(model, X, y)   # fits nothing, model must already be fit; shows the 4-panel plot
print(diag.summary())                   # Breusch-Pagan test, VIF table, top Cook's-distance points

Also works directly with statsmodels:

import statsmodels.api as sm

X_const = sm.add_constant(X)
results = sm.OLS(y, X_const).fit()
regdiag.diagnose(results, X, y)

And degrades gracefully for black-box models instead of lying about statistics:

from sklearn.ensemble import RandomForestRegressor

rf = RandomForestRegressor().fit(X, y)
diag = regdiag.diagnose(rf, X, y)   # 3-panel generic residual analysis
diag.vif()                           # raises NotLinearModelError — VIF isn't defined here

API

regdiag.diagnose(model, X, y, *, feature_names=None, plot=True,
                  figsize=(10, 8), annotate=True, n_annotate=3, save=None)

Returns a RegressionDiagnostics object:

Attribute / method	Linear mode	Generic mode
.residuals, .fitted_values	✅	✅
.leverage, .cooks_distance, .studentized_residuals	✅	None
.plot()	4-panel	3-panel
.residuals_vs_fitted(), .qq_plot(), .scale_location(), .residuals_vs_leverage()	✅	raises NotLinearModelError
.residual_histogram(), .actual_vs_predicted()	✅	✅
.breusch_pagan(), .vif()	✅	raises NotLinearModelError
.summary()	text report	text report

Comparison

	regdiag	lmdiag	statsmodels (raw)	yellowbrick
Works with sklearn linear models	✅	✅	❌	✅
Works with statsmodels OLS	✅	✅	✅	❌
statsmodels is optional at runtime	✅	❌	n/a	✅
Built-in VIF / Breusch-Pagan	✅	❌	✅ (manual)	❌
Honest black-box fallback (no fake leverage)	✅	❌	❌	partial
One-object API (diagnose(model, X, y))	✅	functions	manual	partial

Statistical notes

• Leverage is computed via QR decomposition of the design matrix (never forms the full n × n hat matrix).
• Cook's distance uses the classic OLS-based formula. It's applied uniformly to any linear-coefficient model (including Ridge/Lasso) for simplicity, but strictly its "influence on fitted coefficients" interpretation assumes an OLS estimator — interpret with some caution for heavily regularized models.
• Rank-deficient design matrices (collinear columns) and n ≤ p datasets raise a clear error rather than emitting NaN/inf silently.
• statsmodels WLS/GLS results are not supported in v1 (OLS only).

See CHANGELOG.md for release notes and CONTRIBUTING.md to get set up for development.

License

MIT — see LICENSE.