Date:         Tue, 12 Aug 2003 12:13:30 -0400
Reply-To:     Jay Weedon <jweedon@EARTHLINK.NET>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Jay Weedon <jweedon@EARTHLINK.NET>
Organization: http://extra.newsguy.com
Subject:      Re: estimation problem in PROC GENMOD
Content-Type: text/plain; charset=us-ascii

On 12 Aug 03 15:39:23 GMT, austinkimberly@YAHOO.COM (Kimberly Austin)
wrote:

>I am trying to model the probability of occurrence of a behavior using PROC GENMOD.  My primary interest is in the effect of a contaminant treatment which is included as a 2 level indicator variable TREAT with the control group as the reference level.  The covariates include the continuous variables temperature, and daylength, as well as a 3 level categorical variable for wind speed.
>When I fit the model including all these variables using the code:
>
>PROC GENMOD data=siez DESCENDING;
>        CLASS TREAT WIND;
>        MODEL siezure=temp dayl TREAT WIND / DIST=b LINK=logit LRCI;
>        ODS OUTPUT ModelFit;
>        ODS OUTPUT ParameterEstimates;
>RUN;
>
>the estimate for the effect of TREAT has 0 standard error and the confidence levels are equal to the estimate, which is a large value. When I fit the same model minus TREAT all estimates are fine and there is a better indication of model fit based on deviance/d.f (0.9 vs 0.3 when TREAT is included).
>
>I expected the variable for TREAT to have a strong positive effect on the probability of siezures, and it is the variable of interest in the study. Why the strange esimate, SE and CI for the TREAT parameter, and how should I address the problem?

I saw this the first time you posted it. It's hard to say what the
problem is without looking at your data or output, but you should look
at the possibility that there's multicollinearity or some other kind
of degeneracy in the data set. Is TREAT predictable from WIND or from
the covariates?

JW