Date:         Wed, 14 Feb 2007 22:14:47 -0600
Reply-To:     Hu Yang <sheepsas@GMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Hu Yang <sheepsas@GMAIL.COM>
Subject:      Re: question about proc glimmix
Comments: To: David L Cassell <davidlcassell@msn.com>
In-Reply-To:  <BAY103-F3090CD09D4692F4FA6BDF5B0970@phx.gbl>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed

hi,

We'd like to check our strategy for bias correction in GLMM. So we first
want to generate bias estimations by simulation and then correct them. We
hope the estimations done by proc glimmix (using PQL) are biased without any
corrections.

The dataset we simulated involved K=100 clusters of size nk=7. Conditionally
independent binary observations Ykl were generated within each cluster with
conditional response probabilities given for k=1,...,100 by:
logitE(Ykl|bk)=a0+a1*tl+a2*xk+a3*xk*tl+bk0+bk1*tl
where a0-a4 are fixed effect and bk0-bk1 are random effect, xk=1 for half
the sample and 0 for the other half and tl=l-4 for l=1,...,7. .


I did the estimations by the following:

proc glimmix data = logitDataset SCORING=6 MAXOPT=80;
class k;
_variance_=_mu_*(1-_mu_);
model Y/kl = tl xk tl*xk / S  dist=binomial link=logit;
random INT tl / G TYPE=CHOL sub=k;
 parms / lowerb=1e-4,.,1e-4;
 ODS OUTPUT G  = GOUT;
 ODS OUTPUT ParameterEstimates = FixedParaOUT;
 ODS OUTPUT ConvergenceStatus = cStatus;
run;

I want to know:
1) does proc glimmix do bias correction when using PQL?
2) something wrong with the codes above to estimate the model?

Many thanks!

Hu

On 2/14/07, David L Cassell <davidlcassell@msn.com> wrote:
>
> sheepsas@GMAIL.COM wrote:
> >
> >hi there,
> >
> >I am now using proc glimmix to estimate the fixed effects and the
> variance
> >matrix in a mixed linear model by PQL method. I know PQL can cause bias
> for
> >the
> >estimations. But I want to know if this proc can do something like 'bias
> >correction' or not when doing estimations by PQL? If yes, how can I
> control
> >it to do bias
> >correction or not? I did not find useful information from its help file.
> >
> >Can somebody help me? Thanks!
> >
> >Hu
>
> Hmm.  The standard methods for PROC GLIMMIX are residual likelihood
> or maximum likelihood on an expansion locus, yielding a Taylor series
> linearization of the GLMM.
>
> So what *exactly* are you using?  What's your code, and why?
>
> HTH,
> David
> --
> David L. Cassell
> mathematical statistician
> Design Pathways
> 3115 NW Norwood Pl.
> Corvallis OR 97330
>
> _________________________________________________________________
> Search for grocery stores. Find gratitude. Turn a simple search into
> something more.
>
> http://click4thecause.live.com/search/charity/default.aspx?source=hmemtagline_gratitude&FORM=WLMTAG
>