Abhishek Chakraborty
Last answered:

01 Sept 2020

 Posted on:

28 Aug 2020
0

variance_inflation_factor for constants

I am using the function variance_inflation_factor from statsmodels.stats.outliers_influence for calculating VIF. My questions are,  
  1. Should we consider the constant (added with add_constant, as statsmodels doesn't add constant by default) while calculating VIF?
  2. What does a very high value of a VIF of constant suggest . My constant (or intercept) has a VIF of 70-100.
  3. Should we drop the constant (or intercept)  if the VIF is high but it is significant (i.e. p-value is low) ?
Note : All my variables are min maxed scaled except the categorical variables and the dependent variable.
1 answers ( 0 marked as helpful)
365 Team
Instructor
Posted on:

01 Sept 2020
0
Hi Abhishek, Thanks for reaching out.
  1. Conceptually it doesn't matter if we add or remove the constant from the VIF calculation. Overall it doesn't make a difference for the model, because the constant is not a predictor.
  2. A constant has irrational VIF values as it is not supposed to be calculated/factored.
  3. No, you should not drop the constant. You should simply disregard the VIF result you get for the constant.
Best, The 365 Team

Submit an answer