Passing a level argument to max/min prevents non-numeric columns from being dropped

[ajcr]

When passing a level argument and numeric_only=True to max and min, non-numeric columns are not dropped in the result:

>>> df = pd.DataFrame({'a': [1,1,2,2], 'b': [1,4,2,6], 
                       'c': [6,3,2,3], 'd': ['w','x','y','z']})
>>> df.set_index(['a', 'b'])
>>> df
     c  d
a b      
1 1  6  w
  4  3  x
2 2  2  y
  6  3  z

>>> df.max(level=0, numeric_only=True)
   c  d
a      
1  6  x
2  3  z

I would have expected that column 'd' would not be included here, as is the case if the level argument is omitted:

>>> df.max(numeric_only=True)
c    6
dtype: int64

[jorisvandenbossche]

Yes, that seems like a bug. PR always welcome!

[jreback]

The level arg dispatches to groupby, and because .max/.min handled for strings this 'works'.

So this may require explict filtering before dispatching to handle in a more uniform way. So _make_stat_function might want to do this (or maybe a simpler way)

[mroeschke]

The level keyword in aggregation functions was recently deprecated in #40869 so I don't this this bug will be relevant in the future once this keyword is removed. Closing but happy to reopen if I misunderstood anything