log-log regression

[Snorre Haabesland]

Hi, i am trying to find out if it is possible to use the natural log of an interaction variable in a log-log regression. I have created an interaction of a continuous variable and a dummy-variable, but i get multicollinearity if i do so. Or is it ok to use a linear variable in a log-log regression?

[Clyde Schechter]

If you have two continuous predictors log_x1 and log_x2, and you want to include an interaction, you just do that as c.log_x1#c.log_x2: you do not apply logarithmic transformation to that.

But you should not be log transforming indicator ("dummy") variables at all. If they are coded 1/0, it is impossible because you cannot take the logarithm of 0. If they are coded 1/2 or something like that, all that taking the logarithm does is change the scale of the coefficient from a natural and easily understandable one to something incomprehensible. So don't log-transform that variable.

If you want the interaction between continuous log-transformed variable log_x1 and an indicator variable D, that is just i.D#c.log_x1.