# Nonlinear transformation vs Nonlinear regression

Apologies, if the question has some error since my knowledge of statistics is limited. I am trying to predict DV using 9 IV. Using curve fit option in SPSS, I am finding that 6 of my IV have higher $R^2$ when a logarithmic or cubic curve is used over linear one.

I am confused, in such scenario. should I be transforming my variables and then run linear regression or should I use non-linear regression option??

• $R^2$ is a very poor measure of out-of-sample performance. Transformations should have theoretical justification and not some measure of goodness of fit. – Frans Rodenburg Dec 3 '17 at 12:44
• this is a variation on the usual theme of stepwise regression, which is often unreliable for inference, but it can sometimes work well for prediction. However, as it is, the question doesn't give us enough details for a good response. Can you 1) describe the actual problem you're trying to solve, so that we can understand if there are physical reasons which can motivate the transformations, 2) show us your data and 3) repeat the same exercise you did, but using cross-validation or traning-validation-test set, and see if the approach with the transformed variables still comes out winning? – DeltaIV Dec 3 '17 at 13:27
• hello, thank you both for your feedback. Luckily, I was able to delegate work to RA working under my supervision who has better statistical knowledge. I apologies for poor question style due to confidentiality of data. – Martan Dec 6 '17 at 11:09