additional keys in groupby indices when NAs are present

[josepm]

In [386]: h = pd.DataFrame({'a':[1,2,1,np.nan,1], 'b':[1,2,3,3,2], 'c':[2,3,1,4,2]})

In [387]: gh=h.groupby(['a', 'b'])

In [388]: gh.groups.keys()
Out[388]: [(1.0, 2), (nan, 3), (1.0, 3), (1.0, 1), (2.0, 2)]

In [389]: gh.indices.keys()
Out[389]: [(1.0, 2), (1.0, 3), (2.0, 3), (1.0, 1), (2.0, 2)]  # Incorrect

The tuple (2.0, 3) should not be here.
The problem goes away when there are no NAs

[ledmonster]

pandas.core.groupby:_get_indices_dict seems causing the problem.

In [61]: label_list = [ping.labels for ping in gh.grouper.groupings]
In [76]: keys = [_values_from_object(ping.group_index) for ping in gh.grouper.groupings]
In [77]: keys
Out[77]: [array([ 1.,  2.]), array([1, 2, 3])]
In [78]: label_list
Out[78]: [array([ 0,  1,  0, -1,  0]), array([0, 1, 2, 2, 1])]
In [79]: _get_indices_dict(label_list, keys)
Out[79]:
{(1.0, 2): array([4]),
 (1.0, 3): array([2]),
 (2.0, 3): array([3]),
 (1.0, 1): array([0]),
 (2.0, 2): array([1])}

[jreback]

groupbying with a nan group (or in a level) is not really supported. In general these cause weird types of issues. I'll mark it as a bug though.