(work in progress!)

BayesGLM

This package aims to bring convenient Bayesian GLMs (and other GLM-like models) to Python, with the help of Patsy, Stan, and PyStan.

Stan is a very flexible modeling language with a state-of-the-art sampler. This package aims to use Stan's power to make Bayesian GLMs (and other GLM-like models) available to users, without the need to write Stan code, by generating appropriate Stan code for each model and fitting the model using Stan.

Here are some examples:

from bayesglm import *
from bayesglm.example import df

# linear regression
glm("y ~ x1 + x2", df, RegressionModels.linear)

# linear regression with no intercept
glm("y ~ 0 + x1 + x2", df, RegressionModels.linear)

# linear regression w/ priors
glm("y ~ x1 + x2", df, RegressionModels.linear, priors = {"Intercept": NormalPrior(0, 10), "x2": NormalPrior(0,5)})

# flat priors for main effects, normal prior for interaction term
# (in patsy's notation, "x1*x2" includes the interaction and main effects. "x1:x2" refers to the interaction term.)
glm("y ~ x1*x2", df, RegressionModels.linear, priors = {"x1:x2": NormalPrior(0,5)})

# logistic regression
glm("y_binary ~ x1 + x2", df, RegressionModels.logistic)

# specify the outcome distribution and link function by hand:
glm("y_binary ~ x1 + x2", df, RegressionModel(OutcomeDistributions.bernoulli, "Phi"))

Todo

Before alpha release

• format output more nicely
  • (At the moment, glm gives you raw output from PyStan. I should do something to translate that back into the terms of the variables in the original formula / dataframe.)
• allow making predictions on new data
• test suite
• more examples!

Later

• ability to specify prior for parameters of the outcome distribution (e.g. sigma for normal distributions)
• switch the way I specify betas priors in the Stan code so that we don't have to compile so many models. (e.g. the parameters of the prior distribution should be data sent to Stan, not in the Stan code itself.)
• prepackage compiled Stan models so that user doesn't have to wait for compilation even once

Much later

• hierarchical models!