Date: Thu, 7 Nov 2002 08:31:20 +1000
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Hockey Richard <rhockey@MATER.ORG.AU>
Subject: Re: proc MIXED with binary dependent variable
Content-Type: text/plain; charset="iso-8859-1"
Alternative methods for estimating the ICC for binary outcomes is given in:
Ridout MS, Demetrio CG, Firth D.
Estimating intraclass correlation for binary data.
Biometrics. 1999 Mar;55(1):137-48.
Also the MIXOR software estimates the ICC but I'm a bit doubtful of how?
> -----Original Message-----
> From: Dale McLerran [mailto:stringplayer_2@YAHOO.COM]
> Sent: Thursday, 7 November 2002 6:34 AM
> To: SAS-L@LISTSERV.UGA.EDU
> Subject: Re: proc MIXED with binary dependent variable
> --- Candy Kane <candykane@CANDYLAND.ORG> wrote:
> > On 5 Nov 02 19:37:58 GMT, stringplayer_2@YAHOO.COM (Dale McLerran)
> > wrote:
> > >CK,
> > >
> > >Note, too, that for a binary response, the intraclass correlation
> > >cannot be gotten directly from the GLIMMIX output. An estimate
> > >of the variance of the random effect can be gotten from the
> > >output, but the intraclass correlation cannot be gotten. Just
> > >out of curiosity, how did you compute the intraclass correlation?
> > Thanks for the interest DAle. I calculated it as
> > variance due to the clustering variable / (variance due to the
> > clustering variable + variance remaining)
> > from the final proc mixed run. Is this naive?
> > CK
> YES, it is naive!!! For a binomial response fitted using the
> GLIMMIX procedure, the reported "pure error" variance is a
> scale parameter which indicates whether your binomial response
> is underdispersed or overdispersed. If you have estimated a
> reasonable model, the scale parameter will typically be about 1.0.
> The scale parameter indicates whether there is more or less
> variance in the response than a binomial response should have,
> nothing more and nothing less. Note, too, that the between subject
> variance is measured for the intercept in eta. Variance of the
> intercept does produce additional variance in the binomial
> response, but the amount of variance induced on the response
> depends on the value of eta. If eta reflects a probability
> near 0.5, then the variance in the response will be greater than
> if eta reflects a probability near 0 or 1. The takehome is that
> computing the intraclass correlation from the variance components
> reported by GLIMMIX is fraught with difficulty.
> One simple, reasonable, approach for estimating the intraclass
> correlation is to pass the binomial response variable directly
> through the procedure MIXED. Treat the response as though it
> was gaussian. We do not want to employ this approach for
> estimation of the probability of a response, conditional on
> the covariates in the model. However, just as a method of
> obtaining the approximate variance of the response which is
> due to the random intercept as well as the binomial response
> variance, this can work reasonably well.
> There are other methods which have been suggested in the literature.
> However, the above approach will generally serve quite well.
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: firstname.lastname@example.org
> Ph: (206) 667-2926
> Fax: (206) 667-5977
> Do you Yahoo!?
> HotJobs - Search new jobs daily now