PROC HPSPLIT understanding performance

[jawon]

When I add more independent vars, without changing any options, I'm seeing that the resulting model can be worse than with fewer independent vars. The fit statistics like misclassifications and AUC are worse. Why is that? 

I would think if the new vars don't contribute, it would just get ignored, not make the model worse. This makes me wonder what my strategy should be for adding more vars. My understanding is that an advantage of decision trees is that I can essentially throw in the kitchen sink and HPSPLIT will figure out what to use and what not to use. Naive of me?

[PGStats]

Your assumptions are correct, in theory, at least. But if you are adding more noise to what was only noise, then anything can happen.

In other words, did you have a significant model to start with?

If you haven't tried it, I would suggest using CRITERION ENTROPY (the default) or CRITERION GINI.

[jawon]

Significant model? I have 1,000 records to work with and the model is misclassifying 35%, so maybe not.

Not sure I get the comment about adding more noise. Even if I had mostly noise, why would adding more variables result in a worse decision tree?