Date:  Wed, 21 Sep 2005 14:31:20 0700 
ReplyTo:  anne olean <annekolean@YAHOO.COM> 
Sender:  "SAS(r) Discussion" <SASL@LISTSERV.UGA.EDU> 
From:  anne olean <annekolean@YAHOO.COM> 
Subject:  Genmod  Criteria For Assessing Fit 
ContentType:  text/plain; charset=iso88591 
Hi, I am fitting the following model using GEE:
proc genmod data=mydata;
class id tx;
model y = txweek blY /dist=normal type3;
repeated subject=id(tx)/type=ar(1) covb corrw;
output out=genout pred=preds reschi=resid;
run;
Y is a continuous outcome between 0 and 30 (upper
bound could be as large as 55 points, but the above is
the range of my sample) and mean is around 15 and the
data looks pretty normally distributed. Tx is
treatment assignment (A or B), and week ranges from 0
to 10. blY is a baseline measure of the outcome Y. The
data is collected over 10 weeks, and it is collected
biweekly (ie. week 0, 2, 4, 6, 8, 10). There are 100
subjects, two arms (placebo vs. treated group.) I am
interested in whether there is a treatment effect (eg.
does treatment improve outcome, i.e. is there a
reduction in Y over time). Since the data are repeated
measurements from the same subjects, I selected AR(1)
as the covariance structure since observations farther
apart are likely to be less correlated (However, I
also specified other correlation structures such as
exchangeable and independent).
The issue I have is this: When I fit the above model,
I get a huge Deviance and Value/DF, see below:
Criterion DF Value Value/DF
Deviance 705 29253.6097 41.4945
Scaled Deviance 705 709.0000 1.0057
Pearson ChiSquare 705 29253.6097 41.4945
Scaled Pearson X2 705 709.0000 1.0057
Log Likelihood 2324.7329
Why is that, and what can I do about it? Did I
misspecify the model? IS this a problem? Or can I
ignore it? I also plotted the residuals (histogram and
normal probability plot) and those graphs look
beautiful (ie, residuals are normally distributed, and
range between 17 and +17).
When I look at the parameter estimates, the
interaction was not significant (p = .75), but all
three main effects were significant(p <.03). If there
is a significant main effect but not an interaction
effect, how is that interpreted?
Any insight is as always immensely appreciated.
Anne K.
__________________________________
Yahoo! Mail  PC Magazine Editors' Choice 2005
http://mail.yahoo.com
