Statistics Review: kruskal (Trac #112)

[scipy-gitbot]

Original ticket http://projects.scipy.org/scipy/ticket/112 on 2006-04-02 by @rkern, assigned to @rkern.

The function kruskal in file source:trunk/Lib/stats/stats.py needs review.

Please look over the StatisticsReview guidelines and add your comments below.

[aeklant]

@rgommers ping

To Do:

• Add examples
• Perhaps add more references
• Deal with empty input. It currently returns (nan, nan) or ZeroDivisonError. What would be the proper result here?
• nan input gives inconsistent result, sometimes (nan, nan) and sometimes other results like:

In [31]: stats.kruskal([np.nan], [np.nan])
Out[31]: KruskalResult(statistic=1.0, pvalue=0.31731050786291404)

which looks like a job for the nan_check policy.

[rgommers]

For the docstring also:

• add a See Also, you could reference mannwhitneyu, f_oneway, friedmanchisquare.
• references: the original paper would be useful (https://dx.doi.org/10.1080%2F01621459.1952.10483441)

Verifying the results against R may be useful (and if it matches, add a test or note about that match to an existing test). http://www.r-tutor.com/elementary-statistics/non-parametric-methods/kruskal-wallis-test

Nan plan sounds good. Empty input gives KruskalResult(statistic=nan, pvalue=nan)which sounds right to me. The ZeroDivisionError needs fixing.

The code could do with a minor cleanup:

• some comments (all except the one about Compute sum^2/n) are too obvious and can be removed
• na is a terrible variable name, can change to num_groups.

[rgommers]

Some discussion on exact calculation of p-values for small sample sizes can be found in gh-87. No need to do anything with that discussion/code, but good to link these PRs together.