I'm really scratching my head on this one.
Apparently, if you have a data frame with duplicate indices, subtracting columns consisting of datetimes breaks alignment, whereas subtracting regular int columns does not. Does "subtraction" mean something different in this context?
In [371]: import numpy as np
In [372]: import pandas as pd
In [373]: sec = np.datetime64(1, 's')
In [374]: df = pd.DataFrame({'a' : [sec, sec], 'b' : [0,0]}, index = [0,0])
In [375]: df['b'] - df['b']
Out[375]:
0 0
0 0
Name: b, dtype: int64
All good so far, that's what I'd expect.
In [376]: df['a'] - df['a']
Out[376]:
0 0 days
0 0 days
0 0 days
0 0 days
Name: a, dtype: timedelta64[ns]
What???
Same thing happens even if the col values themselves are not duplicates.
In [377]: df['a'].iloc[0] = np.datetime64(50, 's')
In [378]: df['a'] - df['a']
Out[378]:
0 00:00:00
0 -00:00:00.1000000
0 00:00:00.1000000
0 00:00:00
Name: a, dtype: timedelta64[ns]