Your multiple NN report is here
Generalized RSquare is calculated based on the difference in loglikelihood value between the model, L(Beta hat), and null, L(Null), where n is the # observations.
See page 10 of Chris Gotwalt’s paper. (https://community.jmp.com/kvoqx44227/attachments/kvoqx44227/discussions/19686/1/JMP%20Neural%20Network%20Methodology.pdf).
(Note that RSquare reported for individual NN uses the same formula).
An overall Generalized RSquare when there are multiple responses is calculated by combining the differences, L(Beta hat)-L(Null), from individual NNs, so the overall difference is larger than the difference in an individual model. This results in a higher Generalized RSquare value since no # of models run is taken into considearion.
I personally find the overall Generalized RSquare not very useful.
Thank you Jian,
I understand the concept and also your concerns. that was very helpful!
Bernd
Hello,
when you run NN with two or more target variables you get the individual statistics of fit plus a Generalized RSquare table (s. attached). That rsquare value is as far as I observed it much higher than all the individual rsquare values. I didn't find an explanation in JMP Help and hover help is not available. Could someone point me to an explanation and interpretation help, please.
