groupby.nth() labelling conventions changed from 0.17 -> 0.18

[ssanderson]

In pandas 0.17.1, grouping a DatetimeIndex with monthly frequency produced labels keyed at the start of the month, and nth() produces a label keyed by the selected row. In 0.18, we unconditionally return groups labelled by the end of the month.

Old Pandas

In [25]: pd.__version__
Out[25]: u'0.17.1'

In [26]: dates = pd.date_range('2014', periods=90)

In [27]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='M')).nth(5)
Out[27]:
2014-01-06   NaN
2014-02-06   NaN
2014-03-06   NaN
dtype: float64

Latest Pandas

In [71]: pd.__version__
Out[71]: u'0.18.1'

In [72]: dates = pd.date_range('2014', periods=90)

In [73]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='M')).nth(5)
Out[73]:
2014-01-31   NaN
2014-02-28   NaN
2014-03-31   NaN
dtype: float64

Expected Output

I expect the behavior from 0.17.1.

[ssanderson]

It looks like there are similar issues for other resampling frequencies:

Weekly resample in 0.17:

In [28]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='W')).nth(5)
Out[28]:
2014-01-11   NaN
2014-01-18   NaN
2014-01-25   NaN
2014-02-01   NaN
2014-02-08   NaN
2014-02-15   NaN
2014-02-22   NaN
2014-03-01   NaN
2014-03-08   NaN
2014-03-15   NaN
2014-03-22   NaN
2014-03-29   NaN
dtype: float64

Weekly resample in 0.18:

In [4]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='W')).nth(5)
Out[4]:
2014-01-12   NaN
2014-01-19   NaN
2014-01-26   NaN
2014-02-02   NaN
2014-02-09   NaN
2014-02-16   NaN
2014-02-23   NaN
2014-03-02   NaN
2014-03-09   NaN
2014-03-16   NaN
2014-03-23   NaN
2014-03-30   NaN
dtype: float64

[jreback]

M is the offset for month end
so I think 0.18.1 is correct here

[jorisvandenbossche]

I don't think this has to do with the resampling frequency, rather with a change in nth.

nth has been changed to a 'reducer', with the consequence it takes the grouper as the index by default (instead of a 'filter' which keeps the original index). For example, for mean, which is also a reducer, nothing has changed (so also in 0.17, the end of the month is used):

In [1]: dates = pd.date_range('2014', periods=90)

In [2]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='M')).nth(5)
Out[2]:
2014-01-06   NaN
2014-02-06   NaN
2014-03-06   NaN
dtype: float64

In [3]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='M')).mean()
Out[3]:
2014-01-31   NaN
2014-02-28   NaN
2014-03-31   NaN
Freq: M, dtype: float64

In [4]: pd.__version__
Out[4]: u'0.17.1'

In [1]: dates = pd.date_range('2014', periods=90)

In [2]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='M')).nth(5)
Out[2]:
2014-01-31   NaN
2014-02-28   NaN
2014-03-31   NaN
dtype: float64

In [3]: pd.Series(index=dates).groupby(pd.TimeGrouper(freq='M')).mean()
Out[3]:
2014-01-31   NaN
2014-02-28   NaN
2014-03-31   NaN
Freq: M, dtype: float64

In [4]: pd.__version__
Out[4]: '0.18.1+199.gc9a27ed'

In 0.17, the index you see in the nth case is the original index from the series, because nth acted like a 'filter' (keeping the original index for the filtered rows).
To get the original behaviour of nth, you can get this with as_index=False:

In [9]: pd.__version__
Out[9]: '0.18.1+199.gc9a27ed'

In [10]: pd.DataFrame(index=dates, columns=['a']).groupby(pd.TimeGrouper(freq='
    ...: M'), as_index=False).nth(5)
Out[10]:
              a
2014-01-06  NaN
2014-02-06  NaN
2014-03-06  NaN

The problem with this is that this only seems to work with Series (if I use the original Series example, you get an error).
There was recently also somebody who reported this on the mailing list (https://groups.google.com/forum/#!topic/pydata/4ZjOP0Lfjdc), and I am actually wondering if we should change the behaviour back to the 0.17 one.

[ssanderson]

I could see use-cases for getting the group label (the current behavior) or for getting the label of the selected row (the old behavior), but I would expect that to be an argument to nth() rather than to TimeGrouper. I don't know a ton about the internals of groupby(), but is it possible that the choice of filter vs. reducer could be made dynamically based on arguments to nth()? I'm imagining an API like:

pd.DataFrame(...).groupby(...).nth(5, label='group')  # Current behavior

pd.DataFrame(...).groupby(...).nth(5, label='selection')  # pandas 0.17 behavior

[jorisvandenbossche]

@ssanderson Note that as_index is an argument to groupby, not to TimeGrouper

[rhshadrach]

The current behavior appears correct. I think the only thing left in this issue is @jorisvandenbossche's comment that as_index=False does not work on a Series. I think this issue would be resolved if as_index=False worked on a SeriesGroupBy, see #36507. Currently, one could do:

s.to_frame(s.name).groupby(pd.TimeGrouper(freq='M'), as_index=False).nth(5)

[rhshadrach]

Looking at this again, with the exception of .nth(0), .nth(-1) and the list/slice equivalents (e.g. .nth([0])), nth always acts like a filter. E.g.

df = pd.DataFrame(
    {'a': [1, 1, 1, 2, 2, 3], 'b': range(6)}
)
gb = df.groupby('a')
result = gb.nth(1)
print(result)
#    b
# a
# 1  1
# 2  4

This is not a reducer as group a=3 is dropped, but the output is in the shape of a reducer instead of a filter. In particular, there should be no different between groupby(...).head(5) and groupby(...).nth([0, 1, 2, 3, 4]).

I think nth should have the result of a filter here, namely

   a  b
0  1  1
4  2  4