MAINT: stats.wilcoxon: fix failure with multidimensional x with NaN and slice length > 50

[mdhaber]

Reference issue

Closes gh-20591

What does this implement/fix?

This fixes a bug in which under specific conditions (x.ndim > 1, x.shape[axis] > 50, np.isnan(x).any(), nan_policy='propagate', and method='auto', for instance) wilcoxon could fail and raise an unexpected error.

[mdhaber]

@tirthasheshpatel This is a pretty short fix; do you have a moment to take a look before we branch 1.13.1?

[tirthasheshpatel]

LGTM! Thanks for the fix!

Just to confirm: Passing each slice individually would return z depending on which method it branches to but the behavior is unchanged for vectorized (multi-dimensional) inputs, right?

[mdhaber]

Passing each slice individually would return z depending on which method it branches to...

Er... no, or that's what happened in 1.13. This fixes that.

import numpy as np
import scipy
from scipy import stats
x = np.arange(1, 100)
scipy.__version__  # 1.12.0
# branches to `approx`, but does not return `zstatistic`
hasattr(stats.wilcoxon(x, method='auto'), 'zstatistic')  # False

In SciPy 1.13.0/main:

import numpy as np
import scipy
from scipy import stats
x = np.arange(1, 100)
scipy.__version__  # 1.13.0
# branches to `approx` and returns `zstatistic` - this is the problem
hasattr(stats.wilcoxon(x, method='auto'), 'zstatistic')  # **True**

Because method='auto' is allowed to add zstatistic to the result object, NaNs cause problems. For example, with nan_policy='omit' in 1.13:

rng = np.random.default_rng(25893459825282452)
x = rng.random((2, 55))
x[1, :10] = np.nan
# Slice is length 55 > 50, so method='approx' is used, and `zstatistic` is added
hasattr(stats.wilcoxon(x[0], method='auto', nan_policy='omit'), 'zstatistic')  # True
# Slice after removing NaNs is length 45 < 50, so method='exact' is used, and `zstatistic` is not added
hasattr(stats.wilcoxon(x[1], method='auto', nan_policy='omit'), 'zstatistic')  # False
# Error because one slice has three return values and the other has only two
stats.wilcoxon(x, axis=1, method='auto', nan_policy='omit')  # error

By ensuring that zstatistic is only added when method='approx' is specified explicitly, we ensure that either every slice has zstatistic or every slice does not have zstatistic, avoiding this problem.

[tirthasheshpatel]

Oh, OK. That makes sense. Thanks for the clarification!

Tests look good too, so merging! Thanks!