Date: Wed, 23 May 2007 11:16:15 -0500
Reply-To: "Swank, Paul R" <Paul.R.Swank@UTH.TMC.EDU>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: "Swank, Paul R" <Paul.R.Swank@UTH.TMC.EDU>
Subject: Re: GLIMMIX and 3 random effects (nested?)
Content-Type: text/plain; charset="us-ascii"
The least squares means you get from GLIMMIX need to be
back-transformed. Given that, they will likely be smaller than the raw
means because of the skew. They will be more like geometric means.
Second issue. Using an AR(1) for of the variance covariance matrix is
not helpful with two time points since it assumes that as measures get
firther apart in time, the correaltion between measures will be reduced.
But with two time points, there is only one time difference. The
question is whether the variances at the two time points are the same.
If so, then compund symmetry should work. If not then UN (unstructured)
might be better.
Paul R. Swank, Ph.D. Professor
Director of Reseach
Children's Learning Institute
University of Texas Health Science Center-Houston
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of
Sent: Tuesday, May 22, 2007 11:18 PM
Subject: GLIMMIX and 3 random effects (nested?)
I am quite new to SAS and modeling and appreciate any help.
The goal of my analysis is to detected difference between two treatments
on 500 patients (equally split between treatment A and treatment B). The
outcome is a number of certain occasions per month (many has 0 but some
has very high number of occasions). I assume that the outcome has
Poisson distribution and use following model
PROC GLIMMIX data=workingdata ORDER=INTERNAL; CLASS TREATMENT TIME
SITE_id COHORT PATIENT_id; MODEL Outcome = TREATMENT
TREATMENT*TIME / dist=poisson solution;
However, I do not know how to model random effect.
The patients were recruited at 12 SITES (S1, S2, ...S12), and were split
for the treatment into cohorts. In each site, there were 9 to 11 cohort
and each cohort about 5 patients. Each patient was observed at 2 time
Thus that means that I have 3 random factors: SITE, COHORT and PATIENT.
Obviously, the COHORT is nested within SITE. Since the subject was
observed twice in time, I thought about modeling it as AR(1).
SITE_id ... are unique ids of sites
PATIENT_id ... are unique ids of patients COHORT ... are number of
cohort, but not unique. COHORT=1 is first cohort but is nested within
Here is my attempt:
RANDOM SITE_id COHORT(SITE_id);
RANDOM PATIENT_id / type = AR(1);
I tried to check my model by comparing the observed means by TIME and
TREATMENT with predicted means using statement
LSMEANS TREATMENT*TIME / pdiff;
but the results I am getting are very much lower than the observed
I am also unsure about a statement
as it seems needed in the model.
I really appreciate any comments.
Thank you very much,