Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Multicollinearity: vif with or without uncentered

    I am using the vif command to test for multicollinearity, however I can't seem to underestand the difference between a regular -vif command and a -vif, uncentered? They give me this output:
    Click image for larger version

Name:	vif.PNG
Views:	1
Size:	14.6 KB
ID:	1542456


    Two very different results. Do I just roll with the regular vif command or do I have a problem because of the VIF value exceeding 10 at the uncentered one?

    Kind regards!

  • #2
    Click image for larger version

Name:	pwcorr.PNG
Views:	1
Size:	14.1 KB
ID:	1542462


    This is my correlation table, I suppose there is multicollinearity in the highlighted part. How can I solve those? My dataset does really contain few variables, so I'm not sure if it's appropriate to delete those?

    Or do I have to make 'mean centered variables' by creating a new variable which says: var - average(var) ?
    Last edited by Kletser Jok; 23 Mar 2020, 10:36.

    Comment


    • #3
      Kletser:
      it is not clear from your original post teh presence/absence of a constant in your linerar regression model.
      As per -regress postestimation- entry in Stata .pdf manual (page 2212), the -unncentered. option may be used when you do not have a constant in your linear regression model.
      Kind regards,
      Carlo
      (StataNow 18.5)

      Comment


      • #4
        Originally posted by Carlo Lazzaro View Post
        Kletser:
        it is not clear from your original post teh presence/absence of a constant in your linerar regression model.
        As per -regress postestimation- entry in Stata .pdf manual (page 2212), the -unncentered. option may be used when you do not have a constant in your linear regression model.
        Hey Carlo, thank you for responding so quickly.
        How do I know if there is a constant in my lineair regression model? Are you referring to the intercept?
        Because I have cross section data and thus I do have a constant or am I Interpreting this wrong?

        You are saying I "may" use the ,uncentered command, but without it my VIFs are perfect.

        Comment


        • #5
          Click image for larger version

Name:	reg.PNG
Views:	1
Size:	21.0 KB
ID:	1542478

          This is my current output. Everything is insignificant which is very bad. However I'll probably need to correct for a lot of OLS assumptions.
          This implies I do have a constant in my regression and I need to use the ,uncentered command?

          Comment


          • #6
            Kletser:
            thanks for clarifying (pedantis sidelight: for the future please use CODE delimiters instead of screenshots. Thanks).
            You do have a constant (or intercept) in your OLS: hence, do not use the -uncentered- option in -estat vif-.
            Obtaining significant results or not is not the issue: give a true and fair representation odf the data generating process instead.
            That said:
            - see -linktest- to see whether or not your model is ill-specified;
            - search for a possible non-linear relationship between -Age- and the regressand:
            just replace -Age- with:
            Code:
            c.Age##c.Age
            in the right-hand side of your regression equation.
            Kind regards,
            Carlo
            (StataNow 18.5)

            Comment


            • #7
              Carlo. Thank you very much! You've been a great help!

              I'm not familiar with the -linktest- command but my hatsq is insignificant which means that my variables are correctly specified (from a quick read on the internet).

              I do find it bizar/strange that none of the actual variables are significant.
              I'll check my assumptions first but definitely will report a fair representation!

              Thank you very much.



              Comment

              Working...
              X