Date: Wed, 25 Feb 1998 10:19:44 -600
Reply-To: Louise Lawson <llawson@MRG.SOPH.UAB.EDU>
Sender: "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From: Louise Lawson <llawson@MRG.SOPH.UAB.EDU>
Organization: University of Alabama at Birmingham
Subject: GENMOD v GLM, redux
Thanks to everyone who set me straight on GLM. I have to admit that
I think in terms of repeated measures analysis, and thus evaluated
GLM from that perspective. I apologize if I confused any of my
fellow students.
Here's a second try, using all the helpful information y'all sent.
Again, I would appreciate your corrections if I still have
it wrong.
GLM uses least squares with an identity link to fit a generalized
linear model. The assumptions are that the dependent variable is
normally distibuted (the jury is divided on how and when this
assumption can be violated) and that residuals are i.i.d. random
variables and sum to zero. GENMOD is also a generalized linear
model, which uses maximum likelihood estimation and allows you to
specify the distribution of your dv and the link function you want
to use. Jon uses it for any type of count data, and I tend to use it
if I need to actually calculate a prevalence rate ratio (as opposed
to estimating it with proc LOGISTIC).
Repeated measures analysis with GLM is an ANOVA model that expects
the dv to be continuous and normally distributed, the predictors to
be categorical and requires that the cluster sizes be equal. It is
computationally the least expensive, and what you want to use when
the above requirements are met (which has never happened for me,
but I can dream). GLM has a random statement, but you really
shouldn't use it if you have important random effects (see below).
In GENMOD, the repeated statement invokes a marginal generalized
linear model (aka GEE, generalized estimating equations) that treats
within subject effects as nuisance parameters. You are not required
to have a balanced design, and this is theoretically the best model
to use if you have a few repeated measures with a large number of
subjects . GENMOD cannot model random effects.
If you have random effects, you need to use proc MIXED, which also
requires a continuous, normal dv. If this isn't true in your case,
but you do have random effects, you can model them with the GLIMMIX
macro, which invokes proc MIXED. You also want to use proc MIXED
or GLIMMIX if you have repeated measures with a small sample size or
a large number of observations relative to the number of
experimental units (subjects), or if you're specifically interested
in testing within subject (repeated statment) or between subject
(random statement) effects.
M. Louise Lawson, MPH
Progammer/Analyst II
University of Alabama at Birmingham
Department of Epidemiology
Mortimer Jordan Hall, Room 108
Birmingham, AL 35294-2010
(205)975-5716
| 
