iSixSigma

VIF Question

Six Sigma – iSixSigma Forums Old Forums General VIF Question

Viewing 3 posts - 1 through 3 (of 3 total)
  • Author
    Posts
  • #50871

    newbie
    Participant

    What would be the procedure if running a regression analysis and multicollinearity is exposed as a result of a high (>10) VIF value but:

    All predictors are strongly correlated (from correlation matrix)
    All coefficients “make sense” in their effects
    Now what?  Thanks!
     

    0
    #175429

    Robert Butler
    Participant

    I’d recommend the following:
    1. Normalize all of your variables (scale and center so that all go from -1 to 1).
    2. Re-run the analysis and look at the VIF’s for each one of the predictor variables.
    3. Take the variable with the highest VIF (i.e. VIF > 10) drop it from the model statement and re-run.
    4. Repeat 3 until you have a subset of the X’s that meet the VIF criteria – run backward elimination and stepwise (forward selection with replacement) and see if the two methods reduce to the same model.
       a. If they do – you can go on with the rest of the regression analysis using the final terms in the model.
      b. If they don’t you will have to decide which model makes more sense physically.  If they both make sense then you will have to look at the results of the rest of the regression analysis – lof, residual patterns, etc. to decide which of the two does a better job of fitting your data.
    5. Don’t bother with the simple correlation matrix – it is only looking at one variable vs. another and it can be misleading. The VIF looks at one variable as it is regressed on all of the other variables in the matrix. The best bet, of course is to use VIF’s in conjunction with condition indices but as I’ve said before, as far as I know most packages don’t have this capability.

    0
    #175430

    newbie
    Participant

    Awesome…thanks Robert!

    0
Viewing 3 posts - 1 through 3 (of 3 total)

The forum ‘General’ is closed to new topics and replies.