Date: Sat, 28 Jul 2001 17:21:03 -0400
Reply-To: Steve Rowe <steverowe@EMAIL.MSN.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Steve Rowe <steverowe@EMAIL.MSN.COM>
Subject: 'weight' variable problem?
Content-Type: text/plain; charset=ISO-8859-1
Some SAS/STAT procedures use a ‘weight’ variable. Specifically, proc
discrim allows a weight statement/variable, and says it is similar to the
freq statement/variable. The weights are described as ‘relative weights’,
but there is no suggestion that the weights must add to 1 ... yet I wonder?
I have a report that proc discrim in V8 can produce an incorrect
actual_X_predicted table when there are weights. Specifically, using client-
supplied weights that summed to about 50 times the unweighted base size,
the table showed that every observation was classified into "other" because
it did not meet the a specified minimum posterior probability threshold of
0.5. When the weights were divided by 50, the procedure worked fine. Does
anyone know anything more about this?