`Nan` in the multi index get cast into string after a subtraction

[alphaho]

related #6322

Hi,

I've found this accidentally that after the subtraction between df1 and df2, the nan in the multi index has turned to a string rather than the original float type. And it's also weird that this issue only happens with the parse_dates set to False.

The following is the code to reproduce:

import StringIO
from pandas import DataFrame

index = ['ID', 'Index']

input1 = '''"value","ID","Index"
"0.0","B9",""
"0.0","B",""
'''

input2 = '''"value","ID","Index"
"1.5375969","M","Blah"
'''


df1 = DataFrame.from_csv(StringIO.StringIO(input1), index_col=index, parse_dates=False)
df2 = DataFrame.from_csv(StringIO.StringIO(input2), index_col=index, parse_dates=False)
## The following two lines can work as expected
#df1 = DataFrame.from_csv(file1, index_col=index, )
#df2 = DataFrame.from_csv(file2, index_col=index, )

diff = df2['value'] - df1['value']

print diff.index[0], diff.index[1]  # => ('B', 'nan') ('B9', 'nan')
print df1.index[0], df1.index[1]  # => ('B9', nan) ('B', nan)

versions of pandas and it's dependencies:

INSTALLED VERSIONS
------------------
commit: None
python: 2.7.5.final.0
python-bits: 64
OS: Darwin
OS-release: 13.1.0
machine: x86_64
processor: i386
byteorder: little
LC_ALL: None
LANG: zh_HK.UTF-8

pandas: 0.13.1
Cython: None
numpy: 1.8.1
scipy: None
statsmodels: None
IPython: 1.1.0
sphinx: None
patsy: None
scikits.timeseries: None
dateutil: 2.2
pytz: 2014.2
bottleneck: None
tables: None
numexpr: None
matplotlib: None
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
sqlalchemy: None
lxml: None
bs4: None
html5lib: None
bq: None
apiclient: None

[jreback]

I see the result you are getting
but in general you SHOULD not have nan's in the index (or multi-index) at all, except possibly as an intermediate operation. It really doesn't make sense for most operations. In fact you can't even reset_index on df1 as its unclear what to do with the 2nd level.

Bottom line, this is a 'feature' that maybe can be implemented at some point, but its quite tricky.

So avoid for now; I'll leave the issue open (and link to another).

In [28]: df2['value'] - df1['value']
Out[28]: 
ID  Index
B   nan     NaN
B9  nan     NaN
M   Blah    NaN
Name: value, dtype: float64

In [29]: df2
Out[29]: 
             value
ID Index          
M  Blah   1.537597

[1 rows x 1 columns]

In [30]: df1
Out[30]: 
          value
ID Index       
B9 NaN        0
B  NaN        0

[2 rows x 1 columns]

[alphaho]

Thanks, I'll avoid using the nan as index for now.

[dsm054]

As of 0.20.2, this no longer seems to be an issue:

In [19]: diff.index
Out[19]: 
MultiIndex(levels=[['B', 'B9', 'M'], ['Blah']],
           labels=[[0, 1, 2], [-1, -1, 0]],
           names=['ID', 'Index'],
           sortorder=0)

In [20]: diff.index[0][1] is np.nan
Out[20]: True

In [21]: diff.index[1][1] is np.nan
Out[21]: True

So I think this can be closed.

[jreback]

can u do a pr with this test?

[dsm054]

Sure, will do.