Dropping 'nan' with Pearson's r in scipy/pandas

Question

Quick question: Is there a way to use 'dropna' with the Pearson's r function in scipy? I'm using it in conjunction with pandas, and some of my data has holes in it. I know you used to be able suppress 'nan' with Spearman's r in older versions of scipy, but that functionality is now missing.

To my mind, this seems like a disimprovement, so I wonder if I'm missing something obvious.

My code:

for i in range(len(frame3.columns)):    
    correlation.append(sp.pearsonr(frame3.iloc[ :,i], control['CONTROL']))

Solution 1

You can use np.isnan like this:

for i in range(len(frame3.columns)):    
    x, y = frame3.iloc[ :,i].values, control['CONTROL'].values
    nas = np.logical_or(x.isnan(), y.isnan())
    corr = sp.pearsonr(x[~nas], y[~nas])
    correlation.append(corr)

Solution 2

You can also try creating temporary dataframe, and used pandas built-in method for computing pearson correlation, or use the .dropna method in the temporary dataframe to drup null values before using sp.pearsonr

for col in frame3.columns:    
     correlation.append(frame3[col].to_frame(name='3').join(control['CONTROL']).corr()['3']['CONTROL'])

Comments

[Daniel Gibson]

This is making some assumptions about joining, eg: the indices are compatible

[Steve Scott]

I got the error AttributeError: 'numpy.ndarray' object has no attribute 'isnan'

[ramesh]

@SteveScott: instead of x.isnan(), try np.isnan(x)