Date: Mon, 12 Feb 2007 23:53:33 -0500
Reply-To: SUBSCRIBE SAS-L Cherish <cherish@GLOBAL-ANALYTICS.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: SUBSCRIBE SAS-L Cherish <cherish@GLOBAL-ANALYTICS.COM>
Subject: Negative Binomial Distribution- Variable Selection
I am new to this Archive. I found so many posts very useful. I have a
problem selecting best subset variables using Proc Genmod. I want to build a
negative binomial model which predicts how many cycles does it take before a
customer defaults for the first time. DV is a count variable(0,1,2...12). I
have around 800 variables to choose from. I want to use the variables which
are predictive at 99%. Is there some way in SAS I can do this. Basically i
am looking at any selection procedures for proc genmod, where we can allow a
variable only if it is significant at 99%(similar to forward or stepwise).
Currently am using the following method. I would like to know if it's correct.
I am building a linear model using the same DV and IDV's using forward
selection(proc reg) at 99% significance. I know that the variables which
enters the proc reg need not be predictive when we use proc genmod. But as I
didn't find any other way, am currently using this. Both are increasing
functions. Almost all the vars which are predictive at 99% in linear model
are coming out to be predictive at 99% in negative binomial regression also.
Does this method sound logically correct?