Changelog

0.X.X

New features

Maintenance and fixes

Documentation

Deprecation

0.15.0

New features

• Add default priors for binomial and bernoulli families with logit link (#830)
• Add horseshoe prior (#836)
• Handle multivariate responses with HSGP (#856)

Maintenance and fixes

• Change the JAX random number generator key for 32 bit systems (#833)
• Change rename to replace in pre-render.py (#843)
• Fix out of sample prediction for multivariate families. It would not work for tables where the number of rows was different from the one used to fit the model (#847)
• Check variables before trying to access them in posterior predictive sampling (#851)
• Pass kwargs to nutpie + create env.yml file (#855)

Documentation

• Fix typos and incomplete doc strings (#765)
• Clarify elpd differences interepretation (#825)
• Fix the contributing readme link (#837)
• Add example using offset (#842)
• Fix model formula in negative binomial notebook (#859)
• Fix formatting in t-test examples (#861)
• Fix issue 812 Broken link (#862)
• Update repository documentation files (#865)

Deprecation

0.14.0

New features

• Add configuration facilities to Bambi (#745)
• Interpet submodule now outputs informative messages when computing default values (#745)
• Bambi supports weighted responses (#761)
• Bambi supports constrained responses (#764)
• Implement compute_log_likelihood() method to compute the log likelihood on a model (#769)
• Add a class InferenceMethods that allows users to access the available inference methods and kwargs (#795)

Maintenance and fixes

• Fix bug in predictions with models using HSGP (#780)
• Fix get_model_covariates() utility function (#801)
• Upgrade PyMC dependency to >= 5.13 (#803)
• Use pm.compute_deterministics() to compute deterministics when bayeux based samplers are used (#803)
• Wrap all the parameters of the response distribution (the likelihood) with a pm.Deterministic (#804)
• Keep bayeux-ml as the single direct JAX-related dependency (#804)
• The response component only holds response information about the response, not about predictors of the parent parameter (#804)
• Resolve import error associated with bayeux (#822)

Documentation

• Our Code of Conduct now includes how to send a report (#783)
• Add polynomial regression example (#809)
• Add Contact form to our webpage (#816)

Deprecation

• f"{response_name}_obs" has been replaced by "__obs__" as the dimension name for the observation index (#804)
• f"{response_name}_{parameter_name}" is no longer the name for the name of parameters of the likelihood. Now Bambi uses "{parameter_name}" (#804)
• kind in Model.predict() now use "response_params" and "response" instead of "mean" and "pps" (#804)
• include_mean has been replaced by include_response_params in Model.fit() (#804)

0.13.0

This is the first version of Bambi that is released with a Governance structure. Added in #709.

New features

• Bambi now supports censored responses (#697)
• Implement "exponential" and "weibull" families (#697)
• Add "kidney" dataset (#697)
• Add interpret submodule (#684, #695, #699, #701, #732, #736)
  • Implements comparisons, predictions, slopes, plot_comparisons, plot_predictions, and plot_slopes
• Support censored families

Maintenance and fixes

• Bump quartodoc version to 0.6.1 (#720)
• Replace univariate_ordered with ordered (#724)
• Add missing docstring for center_predictors (#726)
• Fix bugs in plot_comparison (#731)

Documentation

• Add docstrings to utility functions (#696)
• Migrate documentation to Quarto (#712)
• Add case study for MRP (#716)
• Add example about ordinal regression (#719)
• Add example about zero inflated models (#725)
• Add example about predictions for new groups (#734)

Deprecation

• Drop official suport for Python 3.8 (#720)
• Change plots submodule name to interpret (#705)

0.12.0

New features

• Implement new families "ordinal" and "sratio" for modeling of ordinal responses (#678)
• Allow families to implement a custom create_extra_pps_coord() (#688)
• Allow predictions on new groups (#693)

Maintenance and fixes

• Robustify how Bambi handles dims (#682)
• Fix links in FAQ (#686)
• Update additional dependencies install command (#689)
• Update predict pps docstring (#690)
• Add warning for aliases athat aren’t used (#691)

Documentation

• Add families to the Getting Started guide (#683)

Deprecation

0.11.0

New features

• Add support for Gaussian Processes via the HSGP approximation (#632)
• Add new families: "zero_inflated_poisson", "zero_inflated_binomial", and "zero_inflated_negativebinomial" (#654)
• Add new families: "beta_binomial" and "dirichlet_multinomial" (#659)
• Allow plot_cap() to show predictions at the observation level (#668)
• Add new families: "hurdle_gamma", "hurdle_lognormal", "hurdle_negativebinomial", and "hurdle_poisson" (#676)

Maintenance and fixes

• Modify how HSGP is built in PyMC when there are groups (#661)
• Modify how Bambi is imported in the tests (#662)
• Prevent underscores from being removed in dim names (#664)
• Bump sphix dependency to a version greater than 7 (#672)

Documentation

• Document how to use custom priors (#656)
• Fix name of arviz traceplot function in the docs (#666)
• Add example that shows how plot_cap() works (#670)

Deprecation

0.10.0

New features

• Refactored the codebase to support distributional models (#607)
• Added a default method to handle posterior predictive sampling for custom families (#625)
• plot_cap() gains a new argument target that allows to plot different parameters of the response distribution (#627)

Maintenance and fixes

• Moved the tests directory to the root of the repository (#607)
• Don’t pass dims to the response of the likelihood distribution anymore (#629)
• Remove requirements.txt and replace with pyproject.toml config file to distribute the package (#631)

Documentation

• Update examples to work with the new internals (#607)
• Fixed figure in the Sleepstudy example (#607)
• Add example using distributional models (#641)

Deprecation

• Removed versioned documentation webpage (#616)
• Removed correlated priors for group-specific terms (#607)
• Dictionary with tuple keys are not allowed for priors anymore (#607)

0.9.3

Maintenance and fixes

• Update to PyMC >= 5, which means we use PyTensor instead of Aesara now (#613, #614)

0.9.2

New features

• Implement censored() (#581)
• Add Formula class (#585)
• Add common numpy transforms to extra_namespace (#589)
• Add AsymmetricLaplace family for Quantile Regression (#591)
• Add ‘transforms’ argument to plot_cap() (#594)
• Add panel covariates to plot_cap() and make it more flexible (#596)

Maintenance and fixes

• Reimplemented predictions to make better usage of xarray data structures (#573)
• Keep 0 dimensional parameters as 0 dimensional instead of 1 dimensional (#575)
• Refactor terms for modularity and extensibility (#582)
• Remove seed argument from model.initial_point() (#592)
• Add build check function on prior predictive and plot prior (#605)

Documentation

• Add quantile regression example (#608)

Deprecation

• Remove automatic_priors argument from Model (#603)
• Remove string option for data input in Model (#604)

0.9.1

New features

• Add support for jax sampling via numpyro and blackjax samplers (#526)
• Add Laplace family (#524)
• Improve Laplace computation and integration (#555 and #563)

Maintenance and fixes

• Ensure order variable is preserved when ploting priors (#529)
• Treat offset accordingly (#534)
• Refactor tests to share data generation code (#531)

Documentation

• Update documentation following good inferencedata practices (#537)
• Add logos to repo and docs (#542)

Deprecation

• Deprecate method argument in favor of inference_method (#554)

0.9.0

New features

• Bambi now uses PyMC 4.0 as it’s backend. Most if not all your previous model should run the same, without the need of any change.
• Add Plot Conditional Adjusted Predictions plot_cap (#517)

Maintenance and fixes

• Group specific terms now work with numeric of multiple columns (#516)

0.8.0

New features

• Add VonMises ("vonmises") built-in family (#453)
• Model.predict() gains a new argument include_group_specific to determine if group-specific effects are considered when making predictions (#470)
• Add Multinomial ("multinomial") built-in family (#490)

Maintenance and fixes

• Add posterior predictive sampling method to “categorical” family (#458)
• Require Python >= 3.7.2 to fix NoReturn type bug in Python (#463)
• Fixed the wrong builtin link given by link="inverse" was wrong. It returned the same result as link="cloglog" (#472)
• Replaced plain dictionaries with namedtuples when same dictionary structure was repeated many times (#472)
• The function check_full_rank() in utils.py now checks the array is 2 dimensional (#472)
• Removed _extract_family_prior() from bambi/families as it was unnecesary (#472)
• Removed bambi/families/utils.py as it was unnecesary (#472)
• Removed external links and unused datasets (#483)
• Replaced "_coord_group_factor" with "__factor_dim" and "_coord_group_expr" with "__expr_dim" in dimension/coord names (#499)
• Fixed a bug related to modifying the types of the columns in the original data frame (#502)

Documentation

• Add circular regression example (#465)
• Add Categorical regression example (#457)
• Add Beta regression example (#442)
• Add Radon Example (#440)
• Fix typos and clear up writing in some docs (#462)
• Documented the module bambi/defaults (#472)
• Improved documentation and made it more consistent (#472)
• Cleaned Strack RRR example (#479)

Deprecation

• Removed old default priors (#474)
• Removed draws parameter from Model.predict() method (#504)

0.7.1

Maintenance and fixes

• Fixed bug related to the shape of 2 level categorical group-specific effects (#441)

0.7.0

New features

• Add “categorical” built-in family (#426)
• Add include_mean argument to the method Model.fit() (#434)
• Add .set_alias() method to Model (#435)

Maintenance and fixes

• Codebase for the PyMC backend has been refactored (#408)
• Fix examples that averaged posterior values across chains (#429)
• Fix issue #427 with automatic priors for the intercept term (#430)

Documentation

• Add StudentT regression example, thanks to @tjburch (#414)
• Add B-Spline regression example with cherry blossoms dataset (#416)
• Add hirarchical linear regression example with sleepstudy dataset (#424)

0.6.3

• Use formulae 0.2.0 (#411)

0.6.2

Maintenance and fixes

• Change default priors for the ‘t’ family (#403)

Documentation

• Add installation instructions with conda (#406)
• Corrected a typo: pary_id -> party_id (#402)
• Add donation information (#409)

0.6.1

Maintenance and fixes

• Documentation for all versions is built from scratch when there’s a release. This ensures older versions link to the current stable release. (#396)
• Add new axis to prior predictive samples to represent 1 chain in the InferenceData object we return (#397)
• Move Family, Likelihood and Link to the families submodule and improved some docstrings (#399)

Documentation

• Add example with hierarchical binomial model (#398)

0.6.0

New features

• Add alternative default priors (#360)
• Add StudentT family (#367)
• Add Beta family (#368)
• Implement both in-sample and out-of-sample model predictions (#372)
• Add function to load datasets (#375)
• Add option to specify potentials (#379)
• Add Binomial family (#386)

Maintenance and fixes

• Automatic switch initialization method from “jitter-adapt_diag” to “adapt_diag” when sampling fails (#383)
• Predictors are internally centered when there is an intercept. This generally results in improved sampling efficiency (#385)
• Improve documentation and error message in Model.graph() (#390)

Documentation

• Add Negative Binomial family example notebook (#346)
• Fixed typos and improved many notebooks (#374, #377, #382)

0.5.0

New features

• It is possible to specify priors for parameters in the response distribution (#335)
• Add probit and cloglog link functions (#340)

Maintenance and fixes

• Informative message when default priors fail because of perfect separation. Model can be fit with custom priors (#330)
• Breaking changes to the API. All the information related to the model goes in Model() instantiation now (#333)
• Fix gamma family (#337)
• Non-default links are properly passed to statsmodels (#337)
• Fix Wald family (#340)
• Fix Negative binomial family (#340)
• Add informative message when link function is not available for a given family (#340)
• Update formulae version to 0.0.10 (#348)

Documentation

• Notebooks are updated to the new API (#336)
• Add badges, update introduction and minor style changes in webpage (#344)
• Add example using Gamma and Wald families (#345)
• Webpage theme has been updated to PyData theme (#347)
• Add model evaluation to logistic regression example (#350)

0.4.1

New features

• Add option to save a figure from model.graph() by passing the name of a file. Figure format and resolution can also be set (#317)
• Objects of class Prior, Family and Model have nicer print methods (#326)

Maintenance and fixes

• Add negative binomial family to config file, which was missing (#324)
• Add test to check model compilation with families available (#327)
• Update formulae to version 0.0.9 (#329)

Documentation

• Fix gamma docstring (#328)

0.4.0

New features

• Use formulae to parse model formulas (#299)
• Add model representation (#300)

Maintenance and fixes

• Remove deprecation warning related to pm.sample returning idata (#295)

Documentation

• Add citation to Bambi preprint (#290)
• Remove reference to pystan (#292)

0.3.0

New features

• Add posterior predictive sampling (#250)
• Add prior predictive sampling (#244)
• Add gamma, negativebinomial and wald families (#207)

Maintenance and fixes

• Use pm.sample_prior_predictive function to sample and plot from prior (#238)
• Fix FutureWarning: Support for multi-dimensional indexing (#242)
• Use last version of black (#245)
• fix broken link increase Python version (#227)
• Add black style check on lint (#220)
• Some linting while re-reading library (#219)
• Remove future warning when converting the trace to InferenceData (#213)
• Include missing files for sdist (#204)
• Fixed if-else comparison that prevented HalfTStudent prior from being used (#205)
• Sidestep plotting flat priors in plot_priors() (#258)
• GLM.fit_constrained in automatic priors now uses start_params = None (#265)
• Categorical Term within Model now have Term.categorical equal to True(#269)
• Use logging instead of warnings (#270)
• Omits ploting group-level effects and offset variables (#276)
• Logistic regression works with no explicit index (#277)
• Add argument to optionally keep offsets in InferenceData (#288)
• Add argument to optionally keep group level effects and offsets variables in plot_prior (#288)

Documentation

• Update example notebooks (#232)
• add missing notebooks (#229)
• Fix notebooks (#222)
• Clean docs (#200)
• Added notebook using Bambi and ArviZ for model comparison (#267)
• Use same color palette in all notebooks (#282)
• Fix divergences in examples (two divergences remaining in Strack RRR example) (#282)

Deprecation

• Drop support python 3.6 (#218)
• Remove stan backend and replace sd with sigma (#205)
• Deprecate samples argument in favor of draws (#247)

0.2.0 The First Python 3 (and ArviZ) Bambino

New features

• Add laplace approximation (#184) (only for educational use, do not use for real problems)
• Use arviz (#182, #178, #166, #159)

Maintenance and fixes

• Update requirements (#191)
• Change default sd prior and update docs (#189)
• Add f-strings and support python 3.6+ (#188)
• Fix parallel sampling (#186)
• Lint code (#175, #173, #171, #167)
• Move coverage configuration to setup.cfg (#168)
• Add long description to setup.py; light linting on setup.py (#162)
• Black list external/ and tests/from pylint

Documentation

• Add missing example (#194)
• Update docs and fix typos (#185, #181)
• Add missing items to readme and code of conduct (#180)
• Simplify readme (#179)
• Unify docstring style and remove not used code (#169)

Deprecation

• Deprecate Stan backend (#183)

0.1.5 (The last legacy Python Bambino)

New features

• Use a callable as link function (#147)

Maintenance and fixes

• Update to Python 3, black and some pylint (#158)
• Fix test warnings (#144)
• Reorder requirements; Add matplotlib to requirements.txt (#143)
• Reorder imports; Only import necessary submodules from statsmodels (#142)
• Update travis config (#135)

Documentation

• Add contributing guide (#146)
• Update notebooks (#140)

Deprecation

• Last version to support Python 2.7

0.1.1 (2017 December 11)

• Minor release for bugfixes and minor improvements. Changes include:
• Bug that was causing an incorrect link function to be used in the PyMC3 backend when fitting logistic models.
• Fixed handling of missing values in categorical variables.
• Fixed bug in set_priors() when passing numerical values for scale.
• Improved internal handling of custom priors.
• Preliminary Sphinx docs (WIP; thanks to @ejolly).

0.1.0 (2017 March 31)

This is a major release that introduces several new features, significant API changes, and a large number of bug fixes and minor improvements. Notable changes include:

• Support for Stan as the sampling back-end (in addition to PyMC3), via the PyStan package.
• Dropped support for the add_term API; all model specification is now done via formulas.
• Expanded support for arbitrary random effects specifications; any formula now supported by patsy can be passed in as the left-hand side of a random effects specification (e.g., previously, ‘(a*b)|c’ would not have worked).
• Completely refactored Results classes that no longer depend on PyMC3, providing a completely generic representation of sampler results, independent of any back-end.
• Refactored plotting and summary methods implemented on the abstract MCMCResults classes rather than at the back-end level.
• Much better compilation and sampling performance for models that include random effects with many levels. In many cases, performance should now be comparable to the most efficient native implementations of the models in the respective back-ends.
• All random effects priors now use the “non-centered” parameterization by default, significantly reducing bias for some models.
• Improved naming conventions that are more consistent with other packages (e.g., random effects now include the ‘|’ operator in term names).
• Refactored Term class, including a separate subclass for RandomTerms, and a number of other associated changes to the internal object model.
• Updated documentation and notebooks, including two new notebooks featuring well-developed examples (datasets included).
• Improved handling of NA values in continuous columns.
• Support for flat priors everywhere (by setting auto_scale=False).
• Numerous bug fixes and minor improvements

0.0.5 (2017 January 17)

• Weakly informative default priors now work the same for all response families & link functions
• Minor bug fixes/tweaks

0.0.4 (2016 October 11)

• Fixes referencing of Theano ops after PyMC3 namespace clean-up
• Added example Jupyter notebooks
• Improved handling of priors
• Improved prior plots and result summaries
• Improved access to MCMC trace results
• Add handling for datasets with NaN values
• Added travis-ci and coveralls support
• Minor bug fixes/tweaks

0.0.3 (2016 September 4)

First official release.