What’s New

Historically pymer4 versioning was a bit all over the place but has settled down since 0.5.0. This page includes the most notable updates between versions but github is the best place to checkout more details and releases.

0.8.1

  • Compatibility Updates:

    • This version includes a noarch build that should be installable on arm-based macOS platforms (e.g. M1, M2, etc)

    • This version drops support for Python 3.7 and adds support for 3.9-3.11

  • Breaking changes:

    • This version also uses joblib for model saving and loading and drops supported hdf5 files previously handled with the deepdish library as it is no longer actively maintained. This means that 0.8.1 will not be able to load models saved with earlier versions of pymer4!

  • Fixes:

0.8.0

  • NOTE:

    • there was no 0.7.9 release as there were enough major changes to warrant a new minor release version

    • this version unpins the maximum versions of rpy2 and pandas

    • if there are install issues with the conda release accompanying this version you should be able to successfully install into a conda environment using pip with the following: conda install 'r-lmerTest' 'r-emmeans' rpy2 -c conda-forge followed by pip install pymer4

  • Bug fixes:

    • fixed as issue where Lmer with family='binomial' was not converting logits into probabilities correctly

    • fixes #79

    • fixes #88

    • fixes #113

    • fixes #114

    • generally more robust conversion of R types to pandas

  • New features:

    • Lm models now support family='binomial' and uses the LogisticRegression class from scikit-learn with no regularization for estimation. Estimates and errors have been verified against the glm implementation in R

    • new lrt function for estimating likelihood-ratio tests between Lmer models thanks to @dramanica. This replicates the functionality of anova() in R for lmer models.

    • new .confint() method for Lmer models thanks to @dramanica. This allows computing confidence intervals on 1 or more paramters of an already fit model including random effects which are not computed by default when calling .fit()

0.7.8

  • Maintenance release that pins rpy2 >= 3.4.5,< 3.5.1 due to R -> Python dataframe conversion issue on recent rpy2 versions that causes a recursion error.

  • Pending code changes to support rpy2 >= 3.5.1 are tracked on this development branch. Upcoming releases will drop support for rpy2 < 3.5.X

  • Clearer error message when making circular predictions using Lmer models

0.7.7

  • This version is identical to 0.7.6 but supports R >= 4.1

  • Installation is also more flexible and includes instructions for using conda-forge and optimized libraries (MKL) for Intel CPUs

0.7.6

  • Bug fixes:

    • fixes an issue in which a Lmer model fit using categorical predictors would be unable to use .predict or would return fitted values instead of predictions on new data. random effect and fixed effect index names were lost thanks to Mario Leaonardo Salinas for discovering this issue

0.7.5

  • This version is identical to 0.7.4 and simply exists because a naming conflict that resulted in a failed released to Anaconda cloud. See release notes for 0.7.4 below

0.7.4

  • Compatibility updates:

    • This version drops official support for Python 3.6 and adds support for Python 3.9. While 3.6 should still work for the most part, development support and testing against this version of Python will no longer continue moving forward.

  • New features:

    • utils.result_to_table function nicely formats the model.coefs output for a fitted model. The docstring also contains instructions on using this in conjunction with the gspread-pandas library for “exporting” model results to a google sheet

0.7.3

  • Bug fixes:

    • fix issue in which random effect and fixed effect index names were lost thanks to @jcheong0428 and @Shotgunosine for the quick PRs!

0.7.2

  • Bug fixes:

    • fix bug in which boot_func would fail iwth y=None and paired=False

  • Compatibility updates:

    • add support for rpy2>=3.4.3 which handles model matrices differently

    • pin maximum pandas<1.2. This is neccesary until our other dependency deepdish adds support. See this issue

0.7.1

  • Pymer4 will be on conda as of this release!

    • install with conda install -c ejolly -c defaults -c conda-forge pymer4

    • This should make installation much easier

    • Big thanks to Tom Urbach for assisting with this!

  • Bug fixes:

    • design matrix now handles rfx only models properly

    • compatibility with the latest version of pandas and rpy2 (as of 08/20)

    • Lmer.residuals now save as numpy array rather than R FloatVector

  • New features:

    • stats.tost_equivalence now takes a seed argument for reproducibility

  • Result Altering Change:

    • Custom contrasts in Lmer models are now expected to be specified in human readable format. This should be more intuitive for most users and is often what users expect from R itself, even though that’s not what it actually does! R expects custom contrasts passed to the contrasts() function to be the inverse of the desired contrasts. See this vignette for more info.

    • In Pymer4, specifying the following contrasts: model.fit(factors = {"Col1": {'A': 1, 'B': -.5, 'C': -.5}})) will estimate the difference between A and the mean of B and C as one would expect. Behind the scenes, Pymer4 is performing the inversion operation automatically for R.

  • Lots of other devops changes to make testing, bug-fixing, development, future releases and overall maintenance much easier. Much of this work has been off-loaded to automated testing and deployment via Travis CI.

0.7.0

  • dropped support for versions of rpy2 < 3.0

  • Result Altering Change: Lm standard errors are now computed using the square-root of the adjusted mean-squared-error (np.sqrt(res.T.dot(res) / (X.shape[0] - X.shape[1]))) rather than the standard deviation of the residuals with DOF adjustment (np.std(res, axis=0, ddof=X.shape[1])). While these produce the same results if an intercept is included in the model, they differ slightly when an intercept is not included. Formerly in the no-intercept case, results from pymer4 would differ slightly from R or statsmodels. This change ensures the results are always identical in all cases.

  • Result Altering Change: Lm rsquared and adjusted rsquared now take into account whether an intercept is included in the model estimation and adjust accordingly. This is consistent with the behavior of R and statsmodels

  • Result Altering Change: hc1 is the new default robust estimator for Lm models, changed from hc0

  • API change: all model residuals are now saved in the model.residuals attribute and were formerly saved in the model.resid attribute. This is to maintain consistency with model.data column names.

  • New feature: addition of pymer4.stats module for various parametric and non-parametric statistics functions (e.g. permutation testing and bootstrapping)

  • New feature: addition of pymer4.io module for saving and loading models to disk

  • New feature: addition of Lm2 models that can perform multi-level modeling by first estimating a separate regression for each group and then performing inference on those estimates. Can perform inference on first-level semi-partial and partial correlation coefficients instead of betas too.

  • New feature: All model classes now have the ability to rank transform data prior to estimation, see the rank argument of their respective .fit() methods.

  • New features for Lm models:

    • Lm models can transform coefficients to partial or semi-partial correlation coefficients

    • Lm models can also perform weight-least-squares (WLS) regression given the weights argument to .fit(), with optional dof correction via Satterthwaite approximation. This is useful for categorical (e.g. group) comparison where one does not want to assume equal variance between groups (e.g. Welch’s t-test). This remains an experimental feature

    • Lm models can compute hc1 and hc2 robust standard errors

  • New documentation look: the look and feel of the docs site has been completely changed which should make getting information much more accessible. Additionally, overview pages have now been turned into downloadable tutorial jupyter notebooks

  • All methods/functions capable of parallelization now have their default n_jobs set to 1 (i.e. no default parallelization)

  • Various bug fixes to all models

  • Automated testing on travis now pins specific r and r-package versions

  • Switched from lsmeans to emmeans for post-hoc tests because lsmeans is deprecated

  • Updated interactions with rpy2 api for compatibility with version 3 and higher

  • Refactored package layout for easier maintainability

0.6.0

  • Dropped support for Python 2

  • upgraded rpy2 dependency version

  • Added conda installation instructions

  • Accepted JOSS version

0.5.0

  • Lmer models now support all generalized linear model family types supported by lme4 (e.g. poisson, gamma, etc)

  • Lmer models now support ANOVA tables with support for auto-orthogonalizing factors using the .anova() method

  • Test statistic inference for Lmer models can now be performed via non-parametric permutation tests that shuffle observations within clusters

  • Lmer.fit(factors={}) arguments now support custom arbitrary contrasts

  • New forest plots for visualizing model estimates and confidence intervals via the Lmer.plot_summary() method

  • More comprehensive documentation with examples of new features

  • Submission to JOSS

0.4.0

  • Added .post_hoc() method to Lmer models

  • Added .simulate() method to Lmer models

  • Several bug fixes for Python 3 compatibility

0.3.2

  • addition of simulate module

0.2.2

  • Official pyipi release

0.2.1

  • Support for standard linear regression models

  • Models include support for robust standard errors, boot-strapped CIs, and permuted inference

0.2.0

  • Support for categorical predictors, model predictions, and model plots

0.1.0

  • Linear and Logit multi-level models