get first and last values in a groupby

Question

I have a dataframe df

df = pd.DataFrame(np.arange(20).reshape(10, -1),
                  [['a', 'a', 'a', 'a', 'b', 'b', 'b', 'c', 'c', 'd'],
                   ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j']],
                  ['X', 'Y'])

How do I get the first and last rows, grouped by the first level of the index?

I tried

df.groupby(level=0).agg(['first', 'last']).stack()

and got

          X   Y
a first   0   1
  last    6   7
b first   8   9
  last   12  13
c first  14  15
  last   16  17
d first  18  19
  last   18  19

This is so close to what I want. How can I preserve the level 1 index and get this instead:

      X   Y
a a   0   1
  d   6   7
b e   8   9
  g  12  13
c h  14  15
  i  16  17
d j  18  19
  j  18  19

Answer 1

Option 1

def first_last(df):
    return df.ix[[0, -1]]

df.groupby(level=0, group_keys=False).apply(first_last)

---

Option 2 - only works if index is unique

idx = df.index.to_series().groupby(level=0).agg(['first', 'last']).stack()
df.loc[idx]

---

Option 3 - per notes below, this only makes sense when there are no NAs

I also abused the agg function. The code below works, but is far uglier.

df.reset_index(1).groupby(level=0).agg(['first', 'last']).stack() \
    .set_index('level_1', append=True).reset_index(1, drop=True) \
    .rename_axis([None, None])

---

Note

per @unutbu: agg(['first', 'last']) take the firs non-na values.

I interpreted this as, it must then be necessary to run this column by column. Further, forcing index level=1 to align may not even make sense.

Let's include another test

df = pd.DataFrame(np.arange(20).reshape(10, -1),
                  [list('aaaabbbccd'),
                   list('abcdefghij')],
                  list('XY'))

df.loc[tuple('aa'), 'X'] = np.nan

---

def first_last(df):
    return df.ix[[0, -1]]

df.groupby(level=0, group_keys=False).apply(first_last)

df.reset_index(1).groupby(level=0).agg(['first', 'last']).stack() \
    .set_index('level_1', append=True).reset_index(1, drop=True) \
    .rename_axis([None, None])

Sure enough! This second solution is taking the first valid value in column X. It is now nonsensical to have forced that value to align with the index a.