Sum of grouped bool column has inconsistent type

[jkleint]

Summing a bool column after a groupby gives a bool result until there are two or more True values, when it becomes a float64. Seems like it should always be an (unsigned?) integer. Straight sum without a groupby always gives an int64. This is with 0.13.1.

pd.DataFrame([True]).groupby(lambda x: 0).sum()
      0
0  True

pd.DataFrame([True,True]).groupby(lambda x: 0).sum()
   0
0  2

pd.DataFrame([False]).groupby(lambda x: 0).sum()
       0
0  False

pd.DataFrame([False,False]).groupby(lambda x: 0).sum()
       0
0  False

pd.DataFrame([False,False,True]).groupby(lambda x: 0).sum()
      0
0  True

pd.DataFrame([False,False,True,True]).groupby(lambda x: 0).sum()
   0
0  2

pd.DataFrame([False,False]).sum()
0    0
dtype: int64

[jreback]

this is a dupe of #3752, but I like your examples better, so will keep this issue!

Its possible to fix, but hasn't been high on the list of priorities

[xflr6]

As for getting float64 instead of int64 as result, a possible workaround is to use count_nonzero from numpy instead of sum to aggregate:

>>> pd.DataFrame([True,True]).groupby(lambda x: 0).agg(pd.np.count_nonzero)[0]
0    2
Name: 0, dtype: int64

[ediphy-dwild]

for some additional context - sometimes the user may not know they are dealing with a bool type. this may occur when performing a groupby on the result of pd.get_dummies, which may return columns of type uint8, but not always. if get_dummies returns a uint16, the issue above is not triggered, and dummies_result.groupby(...).sum() returns int types. if any of the counts in dummies is small enough, the groupby result will be float.

[aflugge]

This is really very confusing as it means some code might work well as expected on some data while running into an error on other data. I would much appreciate if this could be fixed.