# Is adjusted R-squared appropriate to compare models with different response variables?

I heard that adjusted R-squared from two model are comparable only if two models use the same response variables. So if the response variable in one model is Y and the other one is log(Y), then I should not use adjusted R-squared; rather I should use AIC, BIC, etc instead. Is this true?

• You can't compare AIC or BIC of observations under different transformations either. They depend on the likelihood, which may be quite different on different transformation scales. In particular, in the case of regression models, AIC and BIC will be functions of the estimate of $\sigma^2$, which depends on the transformation. See here and here for some related questions. Oct 9, 2013 at 0:49
• I should note that neither of those questions have answers either. Oct 9, 2013 at 1:08
• Models can be used for varying purposes. Suppose that the goal is to have the narrowest (in the original, not-log scale) 95% prediction intervals. One way to argue that comparison using R-sq is a bad idea is to find a simple example of data such that the log-transformed version has a higher R-sq and yet produces wider confidence intervals. No such example comes to mind yet. Oct 9, 2013 at 3:50
• No, even without knowing the math behind the scene, if your proposed phenomenon is correct, then two models with the same set of predictors each predicting the original $y$ and $ln(y)$ will have the same AIC and BIC, which is not the case. Oct 9, 2013 at 14:07