I am currently running a mixed log-linear model which is in this form:

Log yit = Xit + X2it + (1|individu)

I suspect a multicollinearity ( cor (Xit , X2it) close to 1 ). Do you think it makes sense to scale the explanatory variables (Xit). Or is there another alternative? If yes, how may I interpret the result( 1 SD?).

I already tried to linearize (Log yit = log Xit + (log Xit )2 + (1|individu)) but the result is not better. Thanks

  • 1
    $\begingroup$ with such a high correlation you could either omit one of them or orthogonalize through PCA. $\endgroup$
    – utobi
    Commented Apr 13, 2023 at 15:51
  • $\begingroup$ Thank you @utobi, I and come back to you $\endgroup$
    – Ado
    Commented Apr 13, 2023 at 16:58
  • $\begingroup$ for orthogonalization it is important to use standardized covariates with the mean removed and variance scaled to 1. $\endgroup$
    – Ado
    Commented Apr 13, 2023 at 18:57
  • $\begingroup$ No, standardization it’s not relevant and it won’t solve the collinearity either. $\endgroup$
    – utobi
    Commented Apr 13, 2023 at 19:00
  • $\begingroup$ but it is required before the orthogonalization, isn't it? $\endgroup$
    – Ado
    Commented Apr 13, 2023 at 19:07


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.