BUG: np.ma.median returns wrong values for even number of elements

[mhvk]

Tests in astropy (astropy/astropy#5629) uncovered a bug in np.ma.median for the case where the median should be the mean of the two entries nearest the median:

np.ma.median(np.ma.MaskedArray([1., 1.]))
# 2.0
np.ma.median(np.ma.MaskedArray([1., 1., 1.]))
# 1.0
np.ma.median(np.ma.MaskedArray([0., 1., 1., 5.]))
# 2.0

The bug is not present in 1.11.2.

I have not yet investigated where the problem occurs, as it may be that this immediately rings a bell for @juliantaylor, who last seems to have committed to np.ma.extras

[charris]

Looks like a missing division by 2.

[charris]

Yep, the division at extras.py line 730 drops the result, https://github.com/numpy/numpy/blob/master/numpy/ma/extras.py#L730

@juliantaylor Heads up.

[mhvk]

Indeed, np.ma.median(np.ma.MaskedArray([2., 2.])) returns 4.

I think the culprit is a missing out=s here

...

Ah, I see you found the same. One thing not immediately clear to me is why here s is used directly while further down it is s.data.

[charris]

Yes, I wondered the same. I suspect the s.data form is correct.

[mhvk]

See #8414 for a fix.