eden-kramer-lab/regularized_glm

Tools for fitting a generalized linear model with a L2 or difference penalty

regularized_glm

A simple python package for fitting L2- and smoothing-penalized generalized linear models.

Built primarily because the statsmodels GLM fit_regularized method is built to do elastic net (combination of L1 and L2 penalities), but if you just want to do an L2 or a smoothing penalty (like in generalized additive models), using a penalized iteratively reweighted least squares (p-IRLS) is much faster.

Installation

pip install regularized_glm

OR

conda install -c edeno regularized_glm

Other packages

• Statsmodels
• pyGAM

References

Wood, S. (2006). Generalized additive models: an introduction with R (CRC press).

regularized_glm's Issues

Use scipy.linalg instead of numpy.linalg

Weighted GLM does not work

Deal with exponential families with estimated scale

Currently code does not handle families with estimated scale. Need to adjust the covariance and AIC for the normal and gamma distributions where the scale is estimated.

Add 2 * estimated_scale for AIC.

For covariance multiply by sfit ** 2:

ssr = ssr(predicted_response, response, prior_weights)
sfit = sqrt(ssr / (n_observations - effective_degrees_of_freedom))