# calculating variance inflation factor for logistic regression using statsmodels (or python)?

I am making a logistic regression model using Statsmodels while following the book "Discovering statistics using R" by Andy Field, Jeremy Miles, and Zoë Field . While following along the example I went on to calculate the VIF to check multicollinearity between variables in logistic regression model using following code:

import pandas as pd
import statsmodels.api as sm
from statsmodels.stats.outliers_influence import variance_inflation_factor

pen_df.drop(['Unnamed: 4'], inplace=True, axis=1)
pen_df['Scoredx']  = pen_df['Scored'].replace({'Scored':1, 'Missed':0})

p02 = sm.Logit(pen_df['Scoredx'], pen_df[['const', 'PSWQ', 'Previous', 'Anxious']]).fit()

copy_df = pen_df.copy()
copy_df.drop(['Scored','Scoredx'], inplace=True, axis=1)
from statsmodels.stats.outliers_influence import variance_inflation_factor
vif = pd.Series([variance_inflation_factor(copy_df.values, i)
for i in range(1, copy_df.shape[1])],
index=copy_df.columns[1:])

print(vif)


Which gave the output

However , the output in the book comes as follows

Upon going through the answer by Alexander in this post and this_documentation, I come to understand that VIF in statsmodels use OLS and due to that there may be this discrepancy in my answer. I want to know that how to calculate VIF in this case(logit model) using statsmodels or more generally python to match the answer given in the book.

I have added datafile just in the case it may be useful for reproducibility.

• The statsmodels function uses a design matrix and does not add a constant or demean. use df = p02.model.exog in the vif function. – Josef Jul 1 '20 at 14:07
• For nonlinear models like Logit or GLM, we can have two vif versions, one for the original design matrix, and one for the weighted IRLS design matrix. Using exog, then the vif for the original data is computed. – Josef Jul 1 '20 at 14:10
• @Josef I appreciate your explanation, however while Using exog without adding an extra constant , as you suggested I get the following output imgur.com/a/Kpuuo5D , also I don't know much about IRLS , so I would appreciate if I could know why the results vary when compared to results obtained from R. – p47hf1nd3r Jul 1 '20 at 14:57

Reverse engineering what R package car does for vif of GLM. The computation is based on the covariance of the parameter estimates. It also uses Generalized VIF which is defined for terms instead of single columns of the design matrix. In the example, every term is one column, so this does not make a difference.

A corresponding Python code for the vif for columns based on the estimated model using statsmodels is:

cov = p02.cov_params()
corr = cov / p02.bse / p02.bse[:, None]
np.diag(np.linalg.inv(corr.values[1:, 1:]))[[1, 0, 2]]
​
array([35.22707635,  1.08976625, 35.58192988])


statsmodels currently only has vif based on the original design matrix.

(I have not yet seen a reference for GLM vif that provides background for this computation using covariance of parameter estimates.)