What’s new#

v0.5.5 (September 2024)#

This is a minor release with several bugfixes, and major updates to the internal structure and sphinx documentation.

v0.5.4 (January 2024)#

This is a minor release with several bugfixes and no new features. The new version is tested for Python 3.8-3.11 (but should also work with Python 3.12). See the full changelog for 0.5.4.

This release requires pandas≥1.5. We recommend scipy≥1.11.0.

v0.5.3 (December 2022)#

Bugfixes

Fixed a bug where the boolean value returned by pingouin.anderson() was inverted. It returned True when the data was NOT coming from the tested distribution, and vice versa. PR 308.
Fixed misleading documentation and input_type in the pingouin.convert_effsize() function. When converting from a Cohen’s d effect size to a correlation coefficient, the resulting correlation is not a Pearson correlation but instead a point-biserial correlation. To avoid any confusion, input_type='r' has been deprecated and replaced with input_type='pointbiserialr'. For more details, see issue 302.

New function

We have added the pingouin.ptests() function to calculate a T-test (T- and p-values) between all pairs of columns in a given dataframe. This is the T-test equivalent of pingouin.rcorr(). It can only be used as a pandas.DataFrame method, not as a standalone function. The output is a square dataframe with the T-values on the lower triangle and the p-values on the upper triangle.

>>> import pingouin as pg
>>> df = pg.read_dataset('pairwise_corr').iloc[:30, 1:]
>>> df.columns = ["N", "E", "O", "A", "C"]
>>> df.ptests()
        N       E      O      A    C
N       -     ***    ***    ***  ***
E  -8.397       -                ***
O  -8.585  -0.483      -         ***
A  -9.026   0.278  0.786      -  ***
C  -4.759   3.753  4.128  3.802    -

Improvements

Effect sizes are now calculated using an exact method instead of an approximation based on T-values in pingouin.pairwise_tukey() and pingouin.pairwise_gameshowell(). PR 328.
pingouin.normality() does not raise an AssertionError anymore if one of the groups in group has ≤ 3 samples. PR 324.
Added customization options to pingouin.plot_rm_corr(), which now takes optional keyword arguments to pass through to seaborn.regplot() and seaborn.scatterplot(). PR 312.
Changed some plotting functions to increase compatibility with seaborn.FacetGrid. As explained in issue 306, the major change is to generate matplotlib.axes using default parameters instead of accepting fig and dpi keyword arguments. This change applies to pingouin.plot_blandaltman(), pingouin.plot_paired(), pingouin.plot_circmean(), and pingouin.qqplot(). In the future, open a matplotlib.axes and pass it through using the ax parameter to use custom figure settings with these functions. Other minor changes include the addition of the square keyword argument to pingouin.plot_circmean() and pingouin.qqplot() to ensure equal aspect ratios, and the removal of scatter_kws as a keyword argument in pingouin.plot_blandaltmann() (now alter the scatter parameters using general **kwargs). PR 314.

v0.5.2 (June 2022)#

Bugfixes

The eta-squared (n2) effect size was not properly calculated in one-way and two-way repeated measures ANOVAs. Specifically, Pingouin followed the same behavior as JASP, i.e. the eta-squared was the same as the partial eta-squared. However, as explained in issue 251, this behavior is not valid. In one-way ANOVA design, the eta-squared should be equal to the generalized eta-squared. Note that, as of March 2022, this bug is also present in JASP. We have therefore updated the unit tests to use JAMOVI instead.

Warning

Please double check any effect sizes previously obtained with the pingouin.rm_anova() function.

Fixed invalid resampling behavior for bivariate functions in pingouin.compute_bootci() when x and y were not paired. PR 281.
Fixed bug where confidence (previously ci) was ignored when calculating the bootstrapped confidence intervals in pingouin.plot_shift(). PR 282.

Enhancements

The pingouin.pairwise_ttests() has been renamed to pingouin.pairwise_tests(). Non-parametric tests are also supported in this function with the parametric=False argument, and thus the name “ttests” was misleading (see issue 209).
Allow pingouin.bayesfactor_binom() to take Beta alternative model. PR 252.
Allow keyword arguments for logistic regression in pingouin.mediation_analysis(). PR 245.
Speed improvements for the Holm and FDR correction in pingouin.multicomp(). PR 271.
Speed improvements univariate functions in pingouin.compute_bootci() (e.g. func="mean" is now vectorized).
Rename eta to eta_squared in pingouin.power_anova() and pingouin.power_rm_anova() to avoid any confusion. PR 280.
Use black code formatting.
Add support for DataMatrix objects. PR 286.

Dependencies

Force scikit-learn<1.1.0 to avoid bug in pingouin.logistic_regression(). PR 272.

v0.5.1 (February 2022)#

This is a minor release, with several bugfixes and improvements. This release is compatible with SciPy 1.8 and Pandas 1.4.

Bugfixes

Added support for SciPy 1.8 and Pandas 1.4. PR 234.
Fixed bug where pingouin.rm_anova() and pingouin.mixed_anova() changed the dtypes of categorical columns in-place (issue 224).

Enhancements

Faster implementation of pingouin.gzscore(), adding all options available in zscore: axis, ddof and nan_policy. Warning: this functions is deprecated and will be removed in pingouin 0.7.0 (use scipy.stats.gzscore() instead). PR 210.
Replace use of statsmodels’ studentized range distribution functions with more SciPy’s more accurate scipy.stats.studentized_range(). PR 229.
Add support for optional keywords argument in the pingouin.homoscedasticity() function (issue 218).
Add support for the Jarque-Bera test in pingouin.normality() (issue 216).

Lastly, we have also deprecated the Gitter forum in favor of GitHub Discussions. Please use Discussions to ask questions, share ideas / tips and engage with the Pingouin community!

v0.5.0 (October 2021)#

This is a MAJOR RELEASE with several important bugfixes. We recommend all users to upgrade to this new version.

BUGFIX - Repeated measurements

This release fixes several critical issues related to how Pingouin handles missing values in repeated measurements. The following functions have been corrected:

pingouin.rm_anova()
pingouin.mixed_anova()
pingouin.pairwise_ttests(), only for mixed design or two-way repeated measures design.

A full description of the issue, with code and example, can be found at: raphaelvallat/pingouin#206. In short, in Pingouin <0.5.0, listwise deletion of subjects (or rows) with missing values was not strictly enforced in repeated measures or mixed ANOVA, depending on the input data format (if missing values were explicit or implicit). Pingouin 0.5.0 now uses a stricter complete-case analysis regardless of the input data format, which is the same behavior as JASP.

Furthermore, the pingouin.remove_rm_na() has been deprecated. Instead, listwise deletion of rows with missing values in repeated measurements is now performed using:

>>> data_piv = data.pivot_table(index=subject, columns=within, values=dv)
>>> data_piv = data_piv.dropna()  # Listwise deletion
>>> data = data_piv.melt(ignore_index=False, value_name=dv).reset_index()

BUGFIX - Strict listwise deletion in pairwise_ttests when repeated measures are present

This is related to the previous issue. In mixed design, listwise deletion (complete-case analysis) was not strictly enforced in pingouin.pairwise_ttests() for the between-subject and interaction T-tests. In other words, the between-subject and interaction T-tests were calculated using a pairwise-deletion approach, even with nan_policy="pairwise". The same issue occured in two-way repeated measures design, in which no strict listwise deletion was performed prior to calculating the T-tests, even with nan_policy="pairwise".

This has now been fixed such that Pingouin will always perform a strict listwise deletion whenever repeated measurements are present when nan_policy="listwise" (default). This complete-case analysis behavior can be disabled with nan_policy="pairwise", in which case missing values will be removed separately for each contrast. This may not be appropriate for post-hoc analysis following a repeated measures or mixed ANOVA, which is always conducted on complete-case data.

BUGFIX - Homoscedasticity

The pingouin.homoscedasticity() gave WRONG results for wide-format dataframes because the test was incorrectly calculated on the transposed data. See issue 204.

Enhancements

Partial correlation functions (pingouin.pcorr() and pingouin.partial_corr()) now use numpy.linalg.pinv() with hermitian=True, which improves numerical stability. See issue 198.
Added support for integer column names in most functions. Previously, Pingouin raised an error if the column names were integers. See issue 201.
pingouin.pairwise_corr() now works when the column names of the dataframe are integer, and better support numpy.arrays in the columns argument.
Added support for wide-format dataframe in pingouin.friedman() and pingouin.cochran()

v0.4.0 (August 2021)#

Major upgrade of the dependencies. This release requires Python 3.7+, SciPy 1.7+, NumPy 1.19+ and Pandas 1.0+. Pingouin uses the alternative argument that has been added to several statistical functions of Scipy 1.7+ (see below). However, SciPy 1.7+ requires Python 3.7+. We recommend all users to upgrade to the latest version of Pingouin.

Major enhancements#

Directional testing

The tail argument has been renamed to alternative in all Pingouin functions to be consistent with SciPy and R (#185). Furthermore, "alternative='one-sided'" has now been deprecated. Instead, alternative must be one of “two-sided” (default), “greater” or “less”. Again, this is the same behavior as SciPy and R.

Added support for directional testing with "alternative='greater'" and "alternative='less'" in pingouin.corr() (#176). As a result, the p-value, confidence intervals and power of the correlation will change depending on the directionality of the test. Support for directional testing has also been added to pingouin.power_corr() and pingouin.compute_esci().

Finally, the tail argument has been removed from pingouin.rm_corr(), pingouin.circ_corrcc() and pingouin.circ_corrcl() to be consistent with the original R / Matlab implementations.

Partial correlation

Major refactoring of pingouin.partial_corr(), which now uses the same method as the R ppcor package, i.e. based on the inverse covariance matrix rather than the residuals of a linear regression. This new approach is faster and works better in some cases (such as Spearman partial correlation with binary variables, see issue 147). One caveat is that only the Pearson and Spearman correlation methods are now supported in partial/semi-partial correlation.

Box M test

Added the pingouin.box_m() function to calculate Box’s M test for equality of covariance matrices (#175).

Minor enhancements#

pingouin.wilcoxon() now supports a pre-computed array of differences, similar to scipy.stats.wilcoxon() (issue 186).
pingouin.mwu() and pingouin.wilcoxon() now support keywords arguments that are passed to the lower-level scipy functions.
Added warning in pingouin.partial_corr() with method="skipped": the MCD algorithm does not give the same output in Python (scikit-learn) than in the original Matlab library (LIBRA), and this can lead to skipped correlations that are different in Pingouin than in the Matlab robust correlation toolbox (see issue 164).
pingouin.ancova() always uses statsmodels, regardless of the number of covariates. This fixes LinAlg errors in pingouin.ancova() and pingouin.rm_corr() (see issue 184).
Avoid RuntimeWarning when calculating CI and power of a perfect correlation in pingouin.corr() (see issue 183).
Use scipy.linalg.lstsq() instead of numpy.linalg.lstsq() whenever possible to better check for NaN and Inf in input (see issue 184).
flake8 requirements for max line length has been changed from 80 to 100 characters.

v0.3.12 (May 2021)#

Bugfixes

This release fixes a critical error in pingouin.partial_corr(): the number of covariates was not taken into account when calculating the degrees of freedom of the partial correlation, thus leading to incorrect results (except for the correlation coefficient which remained unaffected). For more details, please see issue 171.

In addition to fixing the p-values and 95% confidence intervals, the statistical power and Bayes Factor have been removed from the output of pingouin.partial_corr(), at least temporary until we can make sure that these give exact results.

We have also fixed a minor bug in the robust skipped and shepherd correlation (see pingouin.corr()), for which the calculation of the confidence intervals and statistical power did not take into account the number of outliers. These are now calculated only on the cleaned data.

Warning

We therefore strongly recommend that all users UPDATE Pingouin (pip install -U pingouin) and CHECK ANY RESULTS obtained with the pingouin.partial_corr() function.

Enhancements

Major refactoring of pingouin.plot_blandaltman(), which now has many additional parameters. It also uses a T distribution instead of a normal distribution to estimate the 95% confidence intervals of the mean difference and agreement limits. See issue 167.
For clarity, the z, r2 and adj_r2 have been removed from the output of pingouin.corr() and pingouin.pairwise_corr(), as these can be readily calculated from the correlation coefficient.
Better testing against R for pingouin.partial_corr() and pingouin.corr().

v0.3.11 (April 2021)#

Bugfixes

Fix invalid computation of the robust skipped correlation in pingouin.corr() (see issue 164).
Passing a wrong tail argument to pingouin.corr() now always raises an error (see PR 160). In previous versions of pingouin, using any method other than "pearson" and a wrong tail argument such as "two-tailed" or "both" (instead of the correct "two-sided") may have resulted in silently returning a one-sided p-value.
Reverted changes made in pingouin.pairwise_corr() which led to Pingouin calculating the correlations between the DV columns and the covariates, thus artificially increasing the number of pairwise comparisons (see issue 162).

v0.3.10 (February 2021)#

Bugfix

This release fixes an error in the calculation of the p-values in the pingouin.pairwise_tukey() and pingouin.pairwise_gameshowell() functions (see PR156). Old versions of Pingouin used an incorrect algorithm for the studentized range approximation, which resulted in (slightly) incorrect p-values. In most cases, the error did not seem to affect the significance of the p-values. The new version of Pingouin now uses statsmodels internal implementation of the Gleason (1999) algorithm to estimate the p-values.

Please note that the Pingouin p-values may be slightly different than R (and JASP), because it uses a different algorithm. However, this does not seem to affect the significance levels of the p-values (i.e. a p-value below 0.05 in JASP is likely to be below 0.05 in Pingouin, and vice versa).

We therefore recommend that all users UPDATE Pingouin (pip install -U pingouin) and CHECK ANY RESULTS obtained with the pingouin.pairwise_tukey() and pingouin.pairwise_gameshowell() functions.

v0.3.9 (January 2021)#

Bugfix

This release fixes a CRITICAL ERROR in the pingouin.pairwise_ttests() function (see issue 151). The bug concerns one-way and two-way repeated measures pairwise T-tests. Until now, Pingouin implicitly assumed that the dataframe was sorted such that the ordering of the subject was the same across all repeated measurements (e.g. the third values in the repeated measurements always belonged to the same subject). This led to incorrect results when the dataframe was not sorted in such a way.

We therefore strongly recommend that all users UPDATE Pingouin (pip install -U pingouin) and CHECK ANY RESULTS obtained with the pingouin.pairwise_ttests() function. Note that the bug does not concern non-repeated measures pairwise T-test, since the ordering of the values does not matter in this case.

Furthermore, and to prevent a similar issue, we have now disabled marginal=False in two-way repeated measure design. As of this release, marginal=False will therefore only have an impact on the between-factor T-test(s) of a mixed design.

Deprecation

a. Removed the Glass delta effect size. Until now, Pingouin invalidly assumed that the control group was always the one with the lowest standard deviation. Since this cannot be verified, and to avoid any confusion, the Glass delta effect size has been completely removed from Pingouin. See issue 139.

Enhancements

pingouin.plot_paired() now supports an arbitrary number of within-levels as well as horizontal plotting. See PR 133.
pingouin.linear_regression() now handles a rank deficient design matrix X by producing a warning and trying to calculate the sum of squared residuals without relying on np.linalg.lstsq(). See issue 130.
pingouin.friedman() now has an option to choose between Chi square test or F test method.
Several minor improvements to the documentation and GitHub Actions. See PR150.
Added support for kwargs in pingouin.corr() (see issue 138).
Added confidence argument in pingouin.ttest() to allow for custom CI (see issue 152).

v0.3.8 (September 2020)#

Bugfixes

Fix a bug in in pingouin.ttest() in which the confidence intervals for one-sample T-test with y != 0 were invalid (e.g. pg.ttest(x=[4, 6, 7, 4], y=4)). See issue 119.

New features

Added a pingouin.options module which can be used to set default options. For example, one can set the default decimal rounding of the output dataframe, either for the entire dataframe, per column, per row, or per cell. See PR120. For more details, please refer to notebooks/06_others.ipynb.
```
import pingouin as pg
pg.options['round'] = None  # Default: no rounding
pg.options['round'] = 4
pg.options['round.column.CI95%'] = 2
pg.options['round.row.T-test'] = 2
pg.options['round.cell.[T-test]x[CI95%]'] = 2
```

Enhancements

pingouin.linear_regression() now returns the processed X and y variables (Xw and yw for WLS) and the predicted values if as_dataframe=False. See issue 112.
The Common Language Effect Size (CLES) in pingouin.mwu() is now calculated using the formula given by Vargha and Delaney 2000, which works better when ties are present in data. This is consistent with the pingouin.wilcoxon() and pingouin.compute_effsize() functions. See issue 114.
Better handling of kwargs arguments in pingouin.plot_paired() (see PR 116).
Added boxplot_in_front argument to the pingouin.plot_paired(). When set to True, the boxplot is displayed in front of the lines with a slight transparency. This can make the overall plot more readable when plotting data from a large number of subjects. (see PR 117).
Better handling of Categorical columns in several functions (e.g. ANOVA). See issue 122.
multivariate_normality() now also returns the test statistic. This function also comes with better unit testing against the MVN R package.
pingouin.pairwise_corr() can now control for all covariates by excluding each specific set of column-combinations from the covariates to use for this combination, similar to pingouin.pcorr(). See PR 124.
Bayes factor formatting is now handled via the options module. The default behaviour is unchanged (return as formatted string), but can easily be disabled by setting pingouin.options[“round.column.BF10”] = None. See PR 126.

v0.3.7 (July 2020)#

Bugfixes

This hotfix release brings important changes to the pingouin.pairwise_tukey() and pingouin.pairwise_gameshowell() functions. These two functions had been implemented soon after Pingouin’s first release and were not as tested as more recent and widely-used functions. These two functions are now validated against JASP.

We strongly recommend that all users upgrade their version of Pingouin (pip install -U pingouin).

Fixed a bug in pingouin.pairwise_tukey() and pingouin.pairwise_gameshowell() in which the group labels (columns A and B) were incorrect when the between column was encoded as a pandas.Categorical with non-alphabetical categories order. This was caused by a discrepancy in how Numpy and Pandas sorted the categories in the between column. For more details, please refer to issue 111.
Fixed a bug in pingouin.pairwise_gameshowell() in which the reported standard errors were slightly incorrect because of a typo in the code. However, the T-values and p-values were fortunately calculated using the correct standard errors, so this bug only impacted the values in the se column.
Removed the tail and alpha argument from the in pingouin.pairwise_tukey() and pingouin.pairwise_gameshowell() functions to be consistent with JASP. Note that the alpha parameter did not have any impact. One-sided p-values were obtained by halving the two-sided p-values.

Error

Please check all previous code and results that called the pingouin.pairwise_tukey() or pingouin.pairwise_gameshowell() functions, especially if the between column was encoded as a pandas.Categorical.

Deprecation

We have now removed the pingouin.plot_skipped_corr() function, as we felt that it may not be useful or relevant to many users (see issue 105).

v0.3.6 (July 2020)#

Bugfixes

Changed the default scikit-learn solver in pingouin.logistic_regression() from ‘lbfgs’ to ‘newton-cg’ in order to get results that are always consistent with R or statsmodels. Previous version of Pingouin were based on the ‘lbfgs’ solver which internally applied a regularization of the intercept that may have led to different coefficients and p-values for the predictors of interest based on the scaling of these predictors (e.g very small or very large values). The new ‘newton-cg’ solver is scaling-independent, i.e. no regularization is applied to the intercept and p-values are therefore unchanged with different scaling of the data. If you prefer to keep the old behavior, just use: pingouin.logistic_regression(..., solver='lbfgs').
Fixed invalid results in pingouin.logistic_regression() when fit_intercept=False was passed as a keyword argument to scikit-learn. The standard errors and p-values were still calculated by taking into account an intercept in the model.

Warning

We highly recommend double-checking all previous code and results that called the pingouin.logistic_regression() function, especially if it involved non-standardized predictors and/or custom keywords arguments passed to scikit-learn.

Enhancements

Added within_first boolean argument to pingouin.pairwise_ttests(). This is useful in mixed design when one want to change the order of the interaction. The default behavior of Pingouin is to return the within * between pairwise tests for the interaction. Using within_first=False, one can now return the between * within pairwise tests. For more details, see issue 102 on GitHub.
pingouin.list_dataset() now returns a dataframe instead of simply printing the output.
Added the Palmer Station LTER Penguin dataset, which describes the flipper length and body mass for different species of penguins. It can be loaded with pingouin.read_dataset('penguins').
Added the Tips dataset. It can be loaded with pingouin.read_dataset('tips').

v0.3.5 (June 2020)#

Enhancements

Added support for weighted linear regression in pingouin.linear_regression(). Users can now pass sample weights using the weights argument (similar to lm(..., weights) in R and LinearRegression.fit(X, y, sample_weight) in scikit-learn).
The \(R^2\) in pingouin.linear_regression() is now calculated in a similar manner as statsmodels and R, which give different results as sklearn.metrics.r2_score() when, and only when, no constant term (= intercept) is present in the predictor matrix. In that case, scikit-learn (and previous versions of Pingouin) uses the standard \(R^2\) formula, which assumes a reference model that only includes an intercept:

\[R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2}\]

However, statsmodels, R, and newer versions of Pingouin use a modified formula, which uses a reference model corresponding to noise only (i.e. no intercept, as explained in this post):

\[R_0^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i y_i^2}\]

Note that this only affects the (rare) cases when no intercept is present in the predictor matrix. Remember that Pingouin automatically add a constant term in pingouin.linear_regression(), a behavior that can be disabled using add_intercept=False.
Added support for robust biweight midcorrelation ('bicor') in pingouin.corr() and pingouin.pairwise_corr().
The Common Language Effect Size (CLES) is now calculated using the formula given by Vargha and Delaney 2000, which works better when ties are present in data.

\[\text{CL} = P(X > Y) + .5 \times P(X = Y)\]

This applies to the pingouin.wilcoxon() and pingouin.compute_effsize() functions. Furthermore, the CLES is now tail-sensitive in the former, but not in the latter since tail is not a valid argument. In pingouin.compute_effsize(), the CLES thus always corresponds to the proportion of pairs where x is higher than y. For more details, please refer to PR #94.
Confidence intervals around a Cohen d effect size are now calculated using a central T distribution instead of a standard normal distribution in the pingouin.compute_esci() function. This is consistent with the effsize R package.

Code

Added support for unsigned integers in dtypes safety checks (see issue #93).

v0.3.4 (May 2020)#

Bugfixes

The Cohen \(d_{avg}\) for paired samples was previously calculated using eq. 10 in Lakens 2013. However, this equation was slightly different from the original proposed by Cumming 2012, and Lakens has since updated the equation in his effect size conversion spreadsheet. Pingouin now uses the correct formula, which is \(d_{avg} = \frac{\overline{X} - \overline{Y}}{\sqrt{\frac{(\sigma_1^2 + \sigma_2^2)}{2}}}\).
Fixed minor bug in internal function pingouin.utils._flatten_list that could lead to TypeError in pingouin.pairwise_ttests() with within/between factors encoded as integers (see issue #91).

New functions

Added pingouin.convert_angles() function to convert circular data in arbitrary units to radians (\([-\pi, \pi)\) range).

Enhancements

Better documentation and testing for descriptive circular statistics functions.
Added safety checks that angles is expressed in radians in circular statistics function.
pingouin.circ_mean() and pingouin.circ_r() now perform calculations omitting missing values.
Pingouin no longer changes the default matplotlib style to a Seaborn-default (see issue #85).
Disabled rounding of float in most Pingouin functions in order to reduce numerical imprecision. For more details, please refer to issue #87. Users can still round the output using the pandas.DataFrame.round() method, or changing the default precision of Pandas DataFrame with pandas.set_option.
Disabled filling of missing values by '-' in some ANOVAs functions, which may have lead to dtypes issues.
Added partial eta-squared (np2 column) to the output of pingouin.ancova() and pingouin.welch_anova().
Added the effsize option to pingouin.anova() and pingouin.ancova() to return different effect sizes. Must be one of 'np2' (partial eta-squared, default) or 'n2' (eta-squared).
Added the effsize option to pingouin.rm_anova() and pingouin.mixed_anova() to return different effect sizes. Must be one of 'np2' (partial eta-squared, default), 'n2' (eta-squared) or ng2 (generalized eta-squared).

Code and dependencies

Compatibility with Python 3.9 (see PR by tirkarthi).
To avoid any confusion, the alpha argument has been renamed to angles in all circular statistics functions.
Updated flake8 guidelines and added continuous integration for Python 3.8.
Added the tabulate package as dependency. The tabulate package is used by the pingouin.print_table() function as well as the pandas.DataFrame.to_markdown() function.

v0.3.3 (February 2020)#

Bugfixes

Fixed a bug in pingouin.pairwise_corr() caused by the deprecation of pandas.core.index in the new version of Pandas (1.0). For now, both Pandas 0.25 and Pandas 1.0 are supported.
The standard deviation in pingouin.pairwise_ttests() when using return_desc=True is now calculated with np.nanstd(ddof=1) to be consistent with Pingouin/Pandas default unbiased standard deviation.

New functions

Added pingouin.plot_circmean() function to plot the circular mean and circular vector length of a set of angles (in radians) on the unit circle.

v0.3.2 (January 2020)#

Hotfix release to fix a critical issue with pingouin.pairwise_ttests() (see below). We strongly recommend that you update to the newest version of Pingouin and double-check your previous results if you’ve ever used the pairwise T-tests with more than one factor (e.g. mixed, factorial or 2-way repeated measures design).

Bugfixes

MAJOR: Fixed a bug in pingouin.pairwise_ttests() when using mixed or two-way repeated measures design. Specifically, the T-tests were performed without averaging over repeated measurements first (i.e. without calculating the marginal means). Note that for mixed design, this only impacts the between-subject T-test(s). Practically speaking, this led to higher degrees of freedom (because they were conflated with the number of repeated measurements) and ultimately incorrect T and p-values because the assumption of independence was violated. Pingouin now averages over repeated measurements in mixed and two-way repeated measures design, which is the same behavior as JASP or JAMOVI. As a consequence, and when the data has only two groups, the between-subject p-value of the pairwise T-test should be (almost) equal to the p-value of the same factor in the pingouin.mixed_anova() function. The old behavior of Pingouin can still be obtained using the marginal=False argument.
Minor: Added a check in pingouin.mixed_anova() to ensure that the subject variable has a unique set of values for each between-subject group defined in the between variable. For instance, the subject IDs for group1 are [1, 2, 3, 4, 5] and for group2 [6, 7, 8, 9, 10]. The function will throw an error if there are one or more overlapping subject IDs between groups (e.g. the subject IDs for group1 AND group2 are both [1, 2, 3, 4, 5]).
Minor: Fixed a bug which caused the pingouin.plot_rm_corr() and pingouin.ancova() (with >1 covariates) to throw an error if any of the input variables started with a number (because of statsmodels / Patsy formula formatting).

Enhancements

Upon loading, Pingouin will now use the outdated package to check and warn the user if a newer stable version is available.
Globally removed the export_filename parameter, which allowed to export the output table to a .csv file. This helps simplify the API and testing. As an alternative, one can simply use pandas.to_csv() to export the output dataframe generated by Pingouin.
Added the correction argument to pingouin.pairwise_ttests() to enable or disable Welch’s correction for independent T-tests.

v0.3.1 (December 2019)#

Bugfixes

Fixed a bug in which missing values were removed from all columns in the dataframe in pingouin.kruskal(), even columns that were unrelated. See raphaelvallat/pingouin#74.
The pingouin.power_corr() function now throws a warning and return a np.nan when the sample size is too low (and not an error like in previous version). This is to improve compatibility with the pingouin.pairwise_corr() function.
Fixed quantile direction in the pingouin.plot_shift() function. In v0.3.0, the quantile subplot was incorrectly labelled as Y - X, but it was in fact calculating X - Y. See raphaelvallat/pingouin#73

v0.3.0 (November 2019)#

New functions

Added pingouin.plot_rm_corr() to plot a repeated measures correlation

Enhancements

Added the relimp argument to pingouin.linear_regression() to return the relative importance (= contribution) of each individual predictor to the \(R^2\) of the full model.
Complete refactoring of pingouin.intraclass_corr() to closely match the R implementation in the psych package. Pingouin now returns the 6 types of ICC, together with F values, p-values, degrees of freedom and confidence intervals.
The pingouin.plot_shift() now 1) uses the Harrel-Davis robust quantile estimator in conjunction with a bias-corrected bootstrap confidence intervals, and 2) support paired samples.
Added the axis argument to pingouin.harrelldavis() to support 2D arrays.

Older versions#

v0.2.9 (September 2019)

Bugfixes

Disabled default l2 regularization of coefficients in pingouin.logistic_regression(). As pointed out by Eshin Jolly in PR54, scikit-learn automatically applies a penalization of coefficients, which in turn makes the estimation of standard errors and p-values not totally correct/interpretable. This regularization behavior is now disabled, resulting in the same behavior as R glm(..., family=binomial).

Code and dependencies

Pandas methods are now internally defined using the pandas_flavor package package.
Internal code refactoring of the pingouin.pairwise_ttests() (to slightly speed up computation and improve memory usage).
The first argument of the pingouin.anova(), pingouin.ancova(), pingouin.welch_anova(), pingouin.pairwise_ttests(), pingouin.pairwise_tukey(), pingouin.pairwise_gameshowell(), pingouin.welch_anova(), pingouin.kruskal(), pingouin.friedman(), pingouin.cochran(), pingouin.remove_rm_na() functions is now data instead of dv (to be consistent with other Pingouin functions). This will cause error if the user runs previous Pingouin code with positional-only arguments. As a general rule, you should always pass keywords arguments (read more here).
For clarity, pingouin.fdr(), pingouin.bonf(), pingouin.holm() have been deprecated from the API and must be called via pingouin.multicomp().
pingouin.pairwise_ttests() output does not include the CLES column by default anymore. Users must explicitly pass effsize='CLES'.
The remove_na argument of pingouin.cronbach_alpha() has been replaced with nan_policy (‘pairwise’, or ‘listwise’).
Disabled Travis / AppVeyor testing for Python 3.5 While most functions should work just fine, please note that only Python >3.6 is supported now.

New functions

Added pingouin.harrelldavis(), a robust quantile estimation method (to be used in a future version of the pingouin.plot_shift() function). See PR63 by Nicolas Legrand.
The pingouin.ancova() can now directly be used a Pandas method, e.g. data.ancova(...).
The pingouin.pairwise_tukey() can now directly be used a Pandas method, e.g. data.pairwise_tukey(...).
Added Sidak one-step correction to pingouin.multicomp() (method='sidak').

Enhancements

Added support for pairwise deletion in pingouin.pairwise_ttests() (default is listwise deletion), using the nan_policy argument.
Added support for listwise deletion in pingouin.pairwise_corr() (default is pairwise deletion), using the nan_policy argument.
Added the interaction boolean argument to pingouin.pairwise_ttests(), useful if one is only interested in the main effects.
Added correction_uniform boolean argument to pingouin.circ_corrcc(). See PR64 by Dominik Straub.

Contributors

Raphael Vallat
Eshin Jolly
Nicolas Legrand
Dominik Straub

v0.2.5 (May 2019)

MAJOR BUG FIXES

Fixed error in p-values for one-sample one-sided T-test (pingouin.ttest()), the two-sided p-value was divided by 4 and not by 2, resulting in inaccurate (smaller) one-sided p-values.
Fixed global error for unbalanced two-way ANOVA (pingouin.anova()), the sums of squares were wrong, and as a consequence so were the F and p-values. In case of unbalanced design, Pingouin now computes a type II sums of squares via a call to the statsmodels package.
The epsilon factor for the interaction term in two-way repeated measures ANOVA (pingouin.rm_anova()) is now computed using the lower bound approach. This is more conservative than the Greenhouse-Geisser approach and therefore give (slightly) higher p-values. The reason for choosing this is that the Greenhouse-Geisser values for the interaction term differ than the ones returned by R and JASP. This will be hopefully fixed in future releases.

New functions

Added pingouin.multivariate_ttest() (Hotelling T-squared) test.
Added pingouin.cronbach_alpha() function.
Added pingouin.plot_shift() function.
Several functions of pandas can now be directly used as pandas.DataFrame methods.
Added pingouin.pcorr() method to compute the partial Pearson correlation matrix of a pandas.DataFrame (similar to the pcor function in the ppcor package).
The pingouin.partial_corr() now supports semi-partial correlation.

Enhancements

The pingouin.rm_corr() function now returns a pandas.DataFrame with the r-value, degrees of freedom, p-value, confidence intervals and power.
pingouin.compute_esci() now works for paired and one-sample Cohen d.
pingouin.bayesfactor_ttest() and pingouin.bayesfactor_pearson() now return a formatted str and not a float.
pingouin.pairwise_ttests() now returns the degrees of freedom (dof).
Better rounding of float in pingouin.pairwise_ttests().
Support for wide-format data in pingouin.rm_anova()
pingouin.ttest() now returns the confidence intervals around the difference in means.

Missing values

pingouin.remove_na() and pingouin.remove_rm_na() are now external function documented in the API.
pingouin.remove_rm_na() now works with multiple within-factors.
pingouin.remove_na() now works with 2D arrays.
Removed the remove_na argument in pingouin.rm_anova() and pingouin.mixed_anova(), an automatic listwise deletion of missing values is applied (same behavior as JASP). Note that this was also the default behavior of Pingouin, but the user could also specify not to remove the missing values, which most likely returned inaccurate results.
The pingouin.ancova() function now applies an automatic listwise deletion of missing values.
Added remove_na argument (default = False) in pingouin.linear_regression() and pingouin.logistic_regression() functions
Missing values are automatically removed in the pingouin.anova() function.

Contributors

Raphael Vallat
Nicolas Legrand