Partition continuous data std. dev tree

feelsgood

Hi to All,

I have a continuous data set and I do a partition analysis. Data has 2500 points. I do splitting until I have an over rsquare of 0.98.
I choose (best tree branch) standard deviation is lowest as best prediction tree branch. But is it a reasonable decision? I wonder you guys can help me about this issue.
Thanks and regards,
Hasan

Peter_Bartell

I suggest that finding a 'best' model is utimately NOT answered by some one size fits all analysis pathway. The 'best' model is the model that helps you solve the practical problem at hand. Reliance on model diagnostics statistics as the sole criteria (and which one to pick?) ignores the issues of domain expertise and practicality of the proposed model to the solution of the problem. For example, is the purpose of the exercise explanatory or predictive? Have you considered cross validation as part of your workflow? You used a tree based method. What about other modeling methods, say regression based from simple OLS to alternative methods like PLS or if you have JMP Pro, generalized regression methods, or Neural Networks? And there are many other considerations.

 

For a deeper dive into these ideas I suggest taking a look at the "Building Better Models" sequence of JMP on demand webinars:

 

https://www.jmp.com/en_us/events/ondemand/webcasts/by-series/building-better-models.html