Python Pandas: how to add a totally new column to a data frame inside of a groupby/transform operation

Question

I want to mark some quantiles in my data, and for each row of the DataFrame, I would like the entry in a new column called e.g. "xtile" to hold this value.

For example, suppose I create a data frame like this:

import pandas, numpy as np
dfrm = pandas.DataFrame({'A':np.random.rand(100), 
                         'B':(50+np.random.randn(100)), 
                         'C':np.random.randint(low=0, high=3, size=(100,))})

And let's say I write my own function to compute the quintile of each element in an array. I have my own function for this, but for example just refer to scipy.stats.mstats.mquantile.

import scipy.stats as st
def mark_quintiles(x, breakpoints):
    # Assume this is filled in, using st.mstats.mquantiles.
    # This returns an array the same shape as x, with an integer for which
    # breakpoint-bucket that entry of x falls into.

Now, the real question is how to use transform to add a new column to the data. Something like this:

def transformXtiles(dataFrame, inputColumnName, newColumnName, breaks):
    dataFrame[newColumnName] = mark_quintiles(dataFrame[inputColumnName].values, 
                                              breaks)
    return dataFrame

And then:

dfrm.groupby("C").transform(lambda x: transformXtiles(x, "A", "A_xtile", [0.2, 0.4, 0.6, 0.8, 1.0]))

The problem is that the above code will not add the new column "A_xtile". It just returns my data frame unchanged. If I first add a column full of dummy values, like NaN, called "A_xtile", then it does successfully over-write this column to include the correct quintile markings.

But it is extremely inconvenient to have to first write in the column for anything like this that I may want to add on the fly.

Note that a simple apply will not work here, since it won't know how to make sense of the possibly differently-sized result arrays for each group.

Answer 1

What problems are you running into with apply? It works for this toy example here and the group lengths are different:

In [82]: df
Out[82]: 
   X         Y
0  0 -0.631214
1  0  0.783142
2  0  0.526045
3  1 -1.750058
4  1  1.163868
5  1  1.625538
6  1  0.076105
7  2  0.183492
8  2  0.541400
9  2 -0.672809

In [83]: def func(x):
   ....:     x['NewCol'] = np.nan
   ....:     return x
   ....: 

In [84]: df.groupby('X').apply(func)
Out[84]: 
   X         Y  NewCol
0  0 -0.631214     NaN
1  0  0.783142     NaN
2  0  0.526045     NaN
3  1 -1.750058     NaN
4  1  1.163868     NaN
5  1  1.625538     NaN
6  1  0.076105     NaN
7  2  0.183492     NaN
8  2  0.541400     NaN
9  2 -0.672809     NaN