1
$\begingroup$

I want to regress $x_1, x_2, x_3$ and $x_4$ on $y$.

Question: When fitting a regression model in the regular fashion without any transformations the coefficient for $x_4$ is positive. However when I log transform the data (all the variables) the sign of the coefficient for $x_4$ becomes negative.

Any idea why a log transform would change the sign?

$\endgroup$
1
  • $\begingroup$ Two questions. Can you show a snippet of the data ? Can you show the code how you did this transformation? $\endgroup$ – Alvaro Joao Feb 21 '16 at 14:27
2
$\begingroup$

At first I thought this must have to do with multicollinearity. But then I tried this out with even a single predictor and you can observe this as well.

The reason is quite simple: it's the noise in the data and the fact that we can't estimate the function (and its derivative) perfectly with finite samples. Moreover, if you believe there is an underlying true (log-)linear relationship than either the log-log model or the original variable model is not linear.

set.seed(1)
XX <- matrix(exp(rnorm(20)), ncol = 1)
yy <- exp(rnorm(nrow(XX)) + 0.1 * XX)

mod.orig <- lm(yy~XX)
mod.log.log <- lm(log(yy)~log(XX))
mod.log.orig <- lm(log(yy)~XX)

layout(matrix(1:3, ncol = 3))
plot(XX, yy)
abline(mod.orig, col = 2)
abline(h = 1, lty = 2, col = 4)

plot(log(XX), log(yy))
abline(mod.log.log, col = 2)
abline(h = 0, lty = 2, col = 4)

plot(XX, log(yy))
abline(mod.log.orig, col = 2)
abline(h = 0, lty = 2, col = 4)

enter image description here

Now in your 4-variable case, I am sure multicollinearity plays a role as well (and correlation of X vs log(X) variables is also very much affected by noise).

Update: I forgot to add the third option before which is the true model in linear terms: $\log y = \alpha + \beta \cdot x$. I added this option as a third option as well.

$\endgroup$
2
  • $\begingroup$ thanks ! Do you know any remedy for this problem? $\endgroup$ – Parash Dejmar Feb 21 '16 at 19:52
  • $\begingroup$ I can think of two options for you: 1) see if you have any context-dependent theory that tells you if you should model it as a linear or log-linear model; 2) if you don't have that, maybe use variable selection methods (LASSO, stepwise) and hope that sign-changing variable drops out of the model. $\endgroup$ – Georg M. Goerg Feb 22 '16 at 1:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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