Date:         Thu, 12 Dec 2002 23:10:39 -0500
Reply-To:     Dali Fu <viciavicia@YAHOO.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Dali Fu <viciavicia@YAHOO.COM>
Subject:      GENMOD....scale parameter....QAIC

I am so happy to find this list group and hope to get help here.

I am doing survival analysis using PROC GENMOD with complementary log-log
link function.The best model was selected according to the smallest QAIC or
QAICC, calculated based on the log likelihood.The number of candidate
models is up to 480. The problems I have:

1. pscale or dscale??
It is stated in SAS online help V8 that scaled deviance or pearson chi-
square indicates the goodness-of-fit; however, in some of my models, the
two indicators go different directions (one >1 and the other <1).
Occordingly, the selected models are quite different if I specify pscale
vs. dscale in the model statement. Any comments for this behavior?

2. log-likelihood, need corrected or not when calculate QAIC?
SAS corrected the log-likelihood (llh) based on the scale parameter. When
calculate QAIC=-2*llh/c +2*k, is it right to back-caculate the "correct"
loglikelyhood value, which is the SAS llh times the square of scale
parameter?

Thanks,

Dali Fu