Date:         Thu, 25 Oct 2007 10:17:54 -0500
Reply-To:     Mary <mlhoward@avalon.net>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Mary <mlhoward@AVALON.NET>
Subject:      Logistic regression when I have only a small number of people in
              the event category.
Content-Type: text/plain; charset="iso-8859-1"

I have a question about Logistic Regression when I have a small number of people in the event category.

I'm running a model looking at whether people have the disease or not based on the value of a Genetic SNP, which is a categorical variable coded 0, 1, 2, and the base reference is 0.   Thus, in a simple model, I'm doing this:


proc logistic data=set1;
   class snp(param=ref ref='0');
   model disease_flag(event='1')=snp;
run;

On one of the diseases that I'm looking at, I've got a very small n size for the number of people that have the disease (n=19), and I have 300 controls.  

My question is, if I get a significant Likelihood ratio for the overall model, but then when I look at the individual odds ratios (1 vs 0, 2 vs 0), if those individual odds ratios are not significant, can I still rely on the overall likelihood ratio or is that not valid as well?   On some of the odds ratios I do get SAS's message that the validity of the model is questionable, but on others I just get the odds ratios whose confidence intervals include 1 and thus are not significant, but the overall likelihood ratio is significant. 

If I can't use logistic regression in looking at a disease in which very few people have the disease, might there be another alternative?

Thanks.

-Mary