BUG: partial nan levels missing when grouping

[jreback]

from #10468

In [40]: df = DataFrame({'A' : [1,2,3]},index=pd.MultiIndex.from_tuples([(1,np.nan),(1,1),(2,1)],names=['first','second']))

In [41]: df
Out[41]: 
              A
first second   
1     NaN     1
      1       2
2     1       3

In [42]: df.groupby(level='second').sum()
Out[42]: 
        A
second   
1       5

In [43]: df.groupby(level=['first','second']).sum()
Out[43]: 
              A
first second   
1     1       2
2     1       3

I suppose that [43] is missing the (1,nan) group here.

[fbnrst]

Maybe this example makes my point from #10468 a little clearer:

In [18]: df1 = pd.DataFrame({'A' : [1,2,3]},index=pd.MultiIndex.from_tuples([(1,np.nan),(1,1),(2,1)],names=['first','second']))

In [19]: df1
Out[19]: 
              A
first second   
1     NaN     1
      1       2
2     1       3

[3 rows x 1 columns]

In [20]: df1.groupby(level=['first','second']).groups
Out[20]: {(1, 1): [(1, nan), (1, 1.0)], (2, 1): [(2, 1.0)]}

In [21]: df2 = df1.reset_index()

In [22]: df2
Out[22]: 
   first  second  A
0      1     NaN  1
1      1       1  2
2      2       1  3

[3 rows x 3 columns]

In [23]: df2.groupby(['first','second']).groups
Out[23]: {(1, nan): [0], (1, 1.0): [1], (2, 1.0): [2]}

Out[20] skips (1, nan), but Out[23] doesn't. As the documentation says NaN in the grouping key will be skipped here, I would rather argue that Out[23] should skip (1, nan). Out[43] from jreback above would actually be correct.