ENH: pivot/groupby index with nan

[jreback]

ENH: maybe for now just provide a warning if dropping the nan rows when pivotting...

rom ml

http://stackoverflow.com/questions/16860172/python-pandas-pivot-table-silently-drops-indices-with-nans

This is effectivly trying to groupby on a NaN, currently not allowed

In [13]: a = [['a', 'b', 12, 12, 12], ['a', nan, 12.3, 233., 12], ['b', 'a', 123.23, 123, 1], ['a', 'b', 1, 1, 1.]]

In [14]: df = DataFrame(a, columns=['a', 'b', 'c', 'd', 'e'])

In [15]: df.groupby(['a','b']).sum()
Out[15]: 
          c    d   e
a b                 
a b   13.00   13  13
b a  123.23  123   1

Workaround to fill the index with a dummy, pivot, and replace

    In [31]: df2 = df.copy()

    In [32]: df2['dummy'] = np.nan

    In [33]: df2['b'] = df2['b'].fillna('dummy')

    In [34]: df2
    Out[34]: 
       a      b       c    d   e  dummy
    0  a      b   12.00   12  12    NaN
    1  a  dummy   12.30  233  12    NaN
    2  b      a  123.23  123   1    NaN
    3  a      b    1.00    1   1    NaN

    In [35]: df2.pivot_table(rows=['a', 'b'], values=['c', 'd', 'e'], aggfunc=sum)
    Out[35]: 
       a      b       c    d   e
    0  a      b   13.00   13  13
    1  a  dummy   12.30  233  12
    2  b      a  123.23  123   1

    In [36]: df2.pivot_table(rows=['a', 'b'], values=['c', 'd', 'e'], aggfunc=sum).replace('dummy',np.nan)
    Out[36]: 
       a    b       c    d   e
    0  a    b   13.00   13  13
    1  a  NaN   12.30  233  12
    2  b    a  123.23  123   1

[jreback]

@wesm so this brings up the issue of groupby with nan in the index. I suppose could include the nan group via a groupby option?

[hayd]

http://stackoverflow.com/questions/18429491/grouping-getting-rows-with-nan-values-divided-from-all-other-rows-in-pandas/18431417#18431417

[hayd]

adding dropna=True to groupby seems reasonable.

Behaviour currently in docs as "this is how R works", but doesn't really say why...

[mattvivier]

@jreback - I apologize, I'm a bit new to github, but is your last action indicating that this (adding dropna=True) should've been implemented in 0.16.0? If so, should I file another bug as I'm experiencing similar behavior in 0.16.2?

[hayd]

modified the milestone: 0.16.0, Next Major Release on Mar 5

Github's a bit unclear, but implementing this was moved from 0.16.0 to next major release (0.17).

[mattvivier]

@hayd thanks for the clarification

[Zenadix]

+1 for adding dropna=True to groupby

[peterpanmj]

@hayd I am using 0.17, but it seems there is still no dropna= True option for groupby

[jreback]

this is still an open issue
pull requests to implement are welcome

[multiloc]

+1 for optional dropna in groupby. The automatic drops can lead to strange behavior:

In [1]:
import pandas as pd
import numpy as np

In [2]:
df = pd.DataFrame({'foo': [1, 2, 3], 'bar': [4, 5, np.nan], 'qux': [None, None, None]})

In [3]:
df

Out[3]:
bar foo qux
0   4   1   None
1   5   2   None
2   NaN 3   None

In [4]:
grouped = df.groupby(by=['foo', 'bar'])

In [5]:
keys = grouped.groups.keys()
keys

Out[5]:
[(3, nan), (2, 5.0), (1, 4.0)]

In [6]:
grouped.get_group(keys[0])
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-6-8c0d3ec7c89a> in <module>()
----> 1 grouped.get_group(keys[0])

.../env/local/lib/python2.7/site-packages/pandas/core/groupby.py in get_group(self, name, obj)
    646         inds = self._get_index(name)
    647         if not len(inds):
--> 648             raise KeyError(name)
    649 
    650         return obj.take(inds, axis=self.axis, convert=False)

KeyError: (3, nan)

[hmgaudecker]

+1 for the dropna option. Just got bitten by this again...

[stefanw]

+1 for the dropna option as well, got bitten by it with some NaT values. This is most unexpected, especially with multiple column groupbys. Explicitly excluding NaN rows before grouping seems more pythonic to me, but a backwards-compatible option with dropna=False would work as well.

[nbonnotte]

I'm working on a solution.

Problem: there is already a dropna parameter in DataFrame.pivot_table, which is used for something slightly different:

dropna : boolean, default True
    Do not include columns whose entries are all NaN

I'll gladly accept any input. The best solution would probably to just call that option with another name, both for .groupby and .pivot_table. Is there any other similar parameter somewhere in pandas whose name could be used again?

[madeleinemf]

I would also really like an option that allows dropna=False to allow a group keyed by Nan values.

[TomAugspurger]

In #29716, @sorenwolfers noted that this affects .sum(level=...) as well

import pandas as pd
import numpy as np
df = pd.DataFrame({'ind':[np.nan,np.nan,'a','a'],'col':[0,1,2,3]}).set_index('ind')
df.sum(level='ind')

[franz101]

Any update?
Getting this error especially with using groupie and transform

[franz101]

Any update?
Getting this error especially with using groupie and transform

[jreback]

@franz101 happy to have a PR for this
there are several attempts so lots of source material

[charlesdong1991]

take

[json2d]

seems really easy to get caught by this one

[OFFICE365USER]

Got bitten by this. Again. Can't believe this has been open since 2013. Consistency with R is not a reason to do something!

[stevenlis]

Could you guys pls consider adding a warning for the users at least?

Btw, I'm using 1.0.3 and set dropna=Fasle will make pivot_table extremely slow.

[andrewdalecramer]

People are aware that it's an issue. It's not a difficult issue to solve, I've done it before and it worked fine, the main impediment there is just getting the option through all the layers of indirection down to the actual work code. The reason it hasn't been solved by me is all the testing and work that goes around it, which I was unable to complete. If someone is familiar with the testing apparatus I'd be glad to talk them through my old PR (which would need to be updated as it is quite old).

On my part, I've rolled off the project using pandas a long time ago and am primarily using R now (which handles this just fine) so I don't have a strong reason to spend the days required to learn how testing works in pandas.

[jreback]

if you look at #30584 this issue is completed for groupby, wasn't hard actually. will be in 1.1 (soon).

but just like anything else in pandas, its all volunteer, if you want an issue fixed then pushing a PR is the best way.

[OFFICE365USER]

People are aware that it's an issue. It's not a difficult issue to solve, I've done it before and it worked fine, the main impediment there is just getting the option through all the layers of indirection down to the actual work code. The reason it hasn't been solved by me is all the testing and work that goes around it, which I was unable to complete. If someone is familiar with the testing apparatus I'd be glad to talk them through my old PR (which would need to be updated as it is quite old).

On my part, I've rolled off the project using pandas a long time ago and am primarily using R now (which handles this just fine) so I don't have a strong reason to spend the days required to learn how testing works in pandas.

People are not aware of this issue. An interim solution would be groupby() to throw an error for an unsupported case with a flag to allow the broken exclude-nan "feature".

[andrewdalecramer]

Sorry, contributers are aware it's an issue. You're welcome to become one and submit a PR.

[zxymath]

There are some problem when groupby category column

categories = ['x%s' % str(i) for i in range(2)]
d = pd.Categorical(np.random.choice(categories, size=100), categories=categories)
d[80:] = np.nan
data = pd.DataFrame(np.random.rand(len(d), 1))
print(data.groupby(d, dropna=False).count())

Out:
     0
x0  35
x1  45

[LaveryRM]

As a business analyst I was trained to apply two QC rules:

never let the number of rows, or the var of interest total, change w/o knowing why.
Check every path through nested IFs and through conditional transforms.

This was basic Business Analyst /data scientist QC and should happens on EVERY merge/filter step in a in a project. Because of the need to do this frequently, I suggest EVERY summation should default to preserving the count and the total.

An option to remove NaNs in an index is appropriate but not as the default.

[colomb8]

why is this issue closed? the problem persist on pandas 1.4.0.

[soerenwolfers]

On 1.5.2 (and possibly lower) you can pass dropna=False, and it will work with both regular indices and multi-indices.