Date:         Fri, 25 Mar 2011 13:49:44 -0400
Reply-To:     William Shakespeare <shakespeare_1040@HOTMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         William Shakespeare <shakespeare_1040@HOTMAIL.COM>
Subject:      Re: RES: Generalized linear mixed model

I have no data.  I'm imagining something like Pinhiero's and Bate's rat
pup data:

-----------Level2------------  ----Level 1------
Litter  Treatment Litter_Size  Pup_id Weight Sex

They used a mixed model with weight as an outcome.  I'm imagining the
outcome as binary-maybe survive/died or defect/no defect, etc. and
simplifying to a very simple model to start with, say, sex as the only
predictor and including a random intercept corresponding to Litter.  I was
thinking about using proc glimmix but am also wondering if nlmixed might
be a better choice.  The question about estimation stems from several
things but mainly from some who suggest in noraml mixed models to use REML
to test random effects and ML for fixed.  I don't know if there's a
corresponding school of thought regarding generalized mixed models but I
seem to remember reading somewhere that the estimation method in glimmix
uses a pseudo likelhood which cannot be used for a LR test.  It's been
suggested that I fit the most complex model and simply look at the
significance tests to simplify and I can see the appeal in that.  At the
same time I can imagine some investigator wanting to do something more
sophisticated.  The question is how.