Date:         Tue, 5 Nov 2002 08:04:57 -0500
Reply-To:     Candy Kane <candykane@CANDYLAND.ORG>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Candy Kane <candykane@CANDYLAND.ORG>
Organization: http://extra.newsguy.com
Subject:      Re: proc MIXED with binary dependent variable
Content-Type: text/plain; charset=us-ascii

On 2 Nov 2002 20:39:24 -0800, andrewyupeng@yahoo.com (Andrew Yu)
wrote:

>I have a need to analyze a dummy outcomes variable (say, mortality in
>hospitals) with nested design (many patients in about 15 hospitals).
>The purpose is to compare the quality of hospitals adjusting for
>patient as well as some hospital factors (say, number of beds for each
>hospital).
>
>I know this require a fixed-effect model. However, my outcomes is a
>dummary variable. Can I still use Proc MIXED? if not, what procedure
>should I use?

I've just completed almost exactly the mortality analysis you mention
(guess I'll beat you to publication, just kidding). I used GLIMMIX.
Dale mentions the size of intraclass correlation: In my analyses this
was estimated to be modest (0.25-0.30). I had tens of thousands of
patients in several dozen hospitals; the regression models contained
up to three dozen terms. Convergence was rarely an issue, jobs
generally took between 5 & 20min to complete on my 2.2GHz Xeon PC.

CK