python - Pandas Correlation Groupby

Question

Assuming I have a dataframe similar to the below, how would I get the correlation between 2 specific columns and then group by the 'ID' column? I believe the Pandas 'corr' method finds the correlation between all columns. If possible I would also like to know how I could find the 'groupby' correlation using the .agg function (i.e. np.correlate).

What I have:

ID  Val1    Val2    OtherData   OtherData
A   5       4       x           x
A   4       5       x           x
A   6       6       x           x
B   4       1       x           x
B   8       2       x           x
B   7       9       x           x
C   4       8       x           x
C   5       5       x           x
C   2       1       x           x

What I need:

ID  Correlation_Val1_Val2
A   0.12
B   0.22
C   0.05

Thanks!

Answer 1

You pretty much figured out all the pieces, just need to combine them:

>>> df.groupby('ID')[['Val1','Val2']].corr()

             Val1      Val2
ID                         
A  Val1  1.000000  0.500000
   Val2  0.500000  1.000000
B  Val1  1.000000  0.385727
   Val2  0.385727  1.000000

In your case, printing out a 2x2 for each ID is excessively verbose. I don't see an option to print a scalar correlation instead of the whole matrix, but you can do something simple like this if you only have two variables:

>>> df.groupby('ID')[['Val1','Val2']].corr().iloc[0::2,-1]

ID       
A   Val1    0.500000
B   Val1    0.385727

For the more general case of 3+ variables

For 3 or more variables, it is not straightforward to create concise output but you could do something like this:

groups = list('Val1', 'Val2', 'Val3', 'Val4')
df2 = pd.DataFrame()
for i in range( len(groups)-1): 
    df2 = df2.append( df.groupby('ID')[groups].corr().stack()
                        .loc[:,groups[i],groups[i+1]:].reset_index() )

df2.columns = ['ID', 'v1', 'v2', 'corr']
df2.set_index(['ID','v1','v2']).sort_index()

Note that if we didn't have the groupby element, it would be straightforward to use an upper or lower triangle function from numpy. But since that element is present, it is not so easy to produce concise output in a more elegant manner as far as I can tell.