Data Transformation in Multiple regression

[Lantan]

Hi guys!
I am trying to run a multiple regression. Some of the independent varibles need tranformation to achieve normal distribution. Some
have to be log - transformed and others sqrt transformed.
I would really like to know, if it is okay to make a model with differently transformed independent variables in the SAME model!?!?

[trinker]

Yes this is ok.

May I ask why you have to transform the IVs. They don't have to be normal. Did the diagnostic graphs from the model without transformation suggest the need for a trasnformation?

[TheEcologist]

Well interpretations gets a little tricky with all those transformation (each transformation does its own thing: e.g log transforms converts absolute differences into relative [as in percentage] differences). Additionally you should look at the residuals of your model as they should be normal - not necessarily each of your variables separately.

Another option is maybe to use a GLM analysis as these types of regression analyses relax the assumptions of normality.
So first check if you models' residuals are normal, then move towards transformations if you must but its better to use a GLM if it is more suited. No way we can tell with the information you have given though!

[Lantan]

Yes this is ok.

May I ask why you have to transform the IVs. They don't have to be normal. Did the diagnostic graphs from the model without transformation suggest the need for a trasnformation?

Thank you very much for your reply!
The residuals look crapy with untransformed data. Thats why I decieded to tranform them, searching for the best transformation I could not remember if it was okay to do different transformations.
Thanks again

[Lantan]

Well interpretations gets a little tricky with all those transformation (each transformation does its own thing: e.g log transforms converts absolute differences into relative [as in percentage] differences). Additionally you should look at the residuals of your model as they should be normal - not necessarily each of your variables separately.

Another option is maybe to use a GLM analysis as these types of regression analyses relax the assumptions of normality.
So first check if you models' residuals are normal, then move towards transformations if you must but its better to use a GLM if it is more suited. No way we can tell with the information you have given though!

Thank you very much too for your anwser!!!
I tried glm as well. But the Akaike Information Criterion told me that the linear model was better. But now I ask myself, if it is possible to compare
lm and glm with AIC? Do you happen to know that?