I had to transform several continuous variables before my analysis. I used Ladder ( Stata) to find the best transformation for each variable. Some were log transformed , while some were squared, etc.
Does that complicate interpretation?
The interpretation is not complicated. But my idea is to use the real Measure of the variables. Get the best of data using the best multi-regression model with Minitab.
If you transform predictors, only the interpretation of the respectice coefficients change. The coefficients give the expected change in the response per unit change of the (possibly transformed) predictor while holding all other predictors constant. If the predictor is log-transformed, one unit chenge on the log scale is a proportional (multiplicative) change on the linear scale. For instance, if the base of the log is 2, the coeffieint gives you the expected change in the reponse per doubeling of the respective predictor. There is not such a nice interpretation for other transformations. If the transformation is the square, the interpretation of the coefficient is its expected change per unit change of the squared predictor while holding all other predictors constant.
A model can be used for description or for explanation. If it is used for description, it does not much matter what transformations you use, since all you want is the model fit. It it is used for explanation, transformations should be selected to make sense (in the respective context).
Box-Cox transformation systen includes both your transformations. Better investigate if you REALLY do need transformations. Actuslly BIGGER problem may be CORRELATIONS among your Predictor variables [multicollinearity]. Because of that Ridge Regression & PCA regression are Handy. Also look at attached reprint.