[SciPy-User] scipy.stats.nanmedian

[Keith Goodman]

I noticed a couple of issues with nanmedian in scipy.stats:

>> from scipy.stats import nanmedian
>> nanmedian(1)
---------------------------------------------------------------------------
ValueError: axis must be less than arr.ndim; axis=0, rank=0.
>> nanmedian(True)
---------------------------------------------------------------------------
ValueError: axis must be less than arr.ndim; axis=0, rank=0.
>> nanmedian(np.array(1))
---------------------------------------------------------------------------
ValueError: axis must be less than arr.ndim; axis=0, rank=0.
>> nanmedian(np.array([1, 2, 3]))
   array(2.0)

Changing the function from the original:

def nanmedian(x, axis=0):
    x, axis = _chk_asarray(x,axis)
    x = x.copy()
    return np.apply_along_axis(_nanmedian,axis,x)

to this (I know, it is not pretty):

def nanmedian(x, axis=0):
    if np.isscalar(x):
        return float(x)
    x, axis = _chk_asarray(x, axis)
    if x.ndim == 0:
        return float(x.tolist())
    x = x.copy()
    x = np.apply_along_axis(_nanmedian, axis, x)
    if x.ndim == 0:
        x = float(x.tolist())
    return x

gives the expected results:

>> nanmedian(1)
   1.0
>> nanmedian(True)
   1.0
>> nanmedian(np.array(1))
   1.0
>> nanmedian(np.array([1, 2, 3]))
   2.0

which agree with numpy:

>> np.median(1)
   1.0
>> np.median(True)
   1.0
>> np.median(np.array(1))
   1.0
>> np.median(np.array([1, 2, 3]))
   2.0

I'm keeping a local copy of the changes I made for my own package. But
it would be nice (for me) if this was fixed upstream. Are the changes
above good enough for scipy?

(Another difference from np.median that I noticed is that the default
axis for np.median is None and for scipy.stats.nanmean it is 0. But
perhaps it is too late to change that.)