Bug Fixes

• PoissonGAM no longer produces -inf log-likelihoods when using non-integer exposure.
• PoissonGAM checks exposure, weight, and y array shapes before fitting.

Bug Fixes

• datasets are loadable like:

from pygam.datasets load cake
X, y = cake(return_X_y=True)

• better model initializations for complex models by using the solution to linear unpenalized problem. This makes the second order PIRLS optimizer less likely to diverge by overshooting the maximum likelihood estimate.
• ReadMe call for collaboration, examples reference dataset loaders, fix typos

Bug Fixes

• bug fix in p-value for models with unknown variance. f-statistic was sensitive to estimated variance when it should be invariant.
• typos

New Features:

• p-values!
• you can now see p-values in the model summary. each feature function will have a p-value, and a code describing it's level of significance.

Bug Fixes

• improving documentation

Bug Fixes

• use scipy stats log-pdfs for computing log-likelihoods
• disable progress bars in gridsearch setting progress=False
• add verbosity attribute to GAMs to control warnings

Bug Fixes:

• alow for changing SVD shapes during PIRLS iterations due to changing mask shapes
• change coefficient initialization to constant model
• change GammaGAM and InvGaussGAM to use non-canonical log-links by default.

New Features

• all GAMs have a sample() method that samples:

  • response variables,
  • model coefficients,
  • and expected values from the posterior probability thanks to @cbrummitt !!! :1st_place_medal:

• all distributions have a sample(mu) method

Bug fixes

• can now raise to negative power
• confidence and prediction intervals use correct degrees of freedom
• all public methods that accept data check for finite data

Improvements

• fixes to documentation

New Features:

• GAMs accept weights in fitting, gridsearch, likelihood, statistics...
• PoissonGAM accepts exposure

Changes

• better handling of PIRLS weights
• check for isfinite(...).all() in check_X, check_y

Bug Fixes

• constant covariates won't break SVD