08-06-2013 04:48 AM
I am looking at building a model in EM on a large number of variables many of which are categorical . It is using survey response answers
to try and find the most important aspects of the service we provide to predict likelihood to recommend us.
I am looking at two ways to narrow down the choice of variables before running a regression model node.
Firstly I interactively use a decision tree node so that I can see all the variables ranked by logworth.
Secondly I run a separate path into a variable selection node. Both decision tree and variable selection run from the same
data partition and replacement nodes.
The results seem to differ massively though. For the top one or two variables the two methods agree but the variable selection node
shows me variables ranked near the top that the decision tree tells me is irrelevant and vice versa.
I don't really know much about the variable selection but I have selected it to use chi squared as a measurement and used a 95% confidence level.
(minimum chi squared of 3.84)
Is there a reason the two would give me such different answers? Which should I trust?