MODEL COMPARISON - TRANSFORMED TARGET VS ORIGINAL TARGET

[Francesco89]

Hello.

I have some regression models working with the transformed target (Log10 transformation on interval variable) and some decision trees working with the original (non transformed) interval target.

Can I use the Model Comparison node to assess regressions vs decision trees despite the different target distribution? Are the ASE resulting from the regression (using log10 target) comparable to the ASE resulting from decision trees (using original target)?

Thanks a lot

[ewv]

The ASE's shouldn't be comparable, since they are effectively in different units (ie - if your response were dollars, your ASE for your decision tree would have units of dollars^2 while your regression models ASE would be in units of (log(dollars))^2 )

That being said, I'm guessing SAS would un-transform your regression models response when doing Model Comparison so that it could do and apples to apples comparison for evaluating the best model. It would be worth giving this a shot to see how the results look - it should be pretty obvious if SAS actually reversed the transformation during the Model Comparison node after you run it