In a regression, there is a independent variable x, say x is a positive number indicating number of months. Is it appropriate to include both the original value(X) and its Log transformation(logX) as the independent variables in a regression? Does it have multi-col-linearity issue even if the VIF test is passed?

If it is appropriate, what is the rational behind this, I understand X and X square can be both included to capture the non linear effects. But never see X and LogX before.

  • $\begingroup$ Maybe a dup: stats.stackexchange.com/questions/277316/… $\endgroup$ Commented Jan 28, 2020 at 21:48
  • $\begingroup$ Thanks it is a little different, as he claimed X^2 and X, My case is X and Log X, I guess one possible explanation is that col-linearity test only concern about linear relationship in IVs, while logX and X are non linear relationship $\endgroup$
    – Bruce
    Commented Jan 28, 2020 at 22:24
  • $\begingroup$ It's fine. The easiest way to think about it is that it is just a more flexible non-linear representation of the effect. They can't be perfectly collinear unless you have fewer than 3 distinct values of the exposure. $\endgroup$
    – AdamO
    Commented Jan 28, 2020 at 22:51

1 Answer 1


There is not a problem in principle, but you should do it for some reason (not just blindly throw it in.) You should have told us your application, with some context, to get a better answer. As for multicolinearity, as logs are locally linear, if the $x$ variable you are transforming varies over a small range, there would be a problem. But, in that case neither would there be any point in including both. But you say VIF does not indicate any problems, so that is not the case.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.