233335_en-US

liliconcom

VIF value < 1 when using Generalized Regression within Fit Model

Hi, 

 

I noticed from time to time, I get VIF < 1 when fitting data to best subset, lasso or double lasso under Generalized Regression (with normal distribution).  For validation method, I chose BIC for all different models.  When I got the result, I would right click in the "Active Parameter Estimates" -> Column -> VIF.   when I extract the active estimates to fit in a OLS model, the coefficients for the predictors are different and the VIFs were all > 1.  Based on the VIF description in JMP, I built model using the predictor with VIF < 1 as the response and the rest of the variables as predictors, the OLS model showed R2 ~ 0.2 and adjusted R2 between 0.079 - 0.2.  So I am not sure why VIF can be < 1.

 

 

 

 

Thank you so much for your time and help.

Re: VIF value < 1 when using Generalized Regression within Fit ModelThat does seem strange. It implies a negative R^2 for the regression of this predictor against the other predictors.

I would recommend contacting technical support with this one (email: support@jmp.com).Re: VIF value < 1 when using Generalized Regression within Fit Model

Here is a explanation of how VIF can be less than 1. It has to due with the shrinkage (biasing) of the parameter estimates.

Re: VIF value < 1 when using Generalized Regression within Fit ModelThanks, @markbailey. I did wonder if it was because of shrinkage.Re: VIF value < 1 when using Generalized Regression within Fit Model

Thank you so much for the link Mark.  I guess I am still wondering if we can still think of them as the absolute VIF where we set a limit at anything lower than 4 still indicates minimum collinearity.

 

I also got this answer from José Quaresma (SAS Global Technical Support) which I want to share with everyone too.