Date: Wed, 14 Aug 1996 21:06:41 GMT
Sender: "SAS(r) Discussion" <SAS-L@UGA.CC.UGA.EDU>
From: Netnews Server <NETNEWS@AMERICAN.EDU>
Organization: Fred Hutchinson Cancer Research Center
Subject: Re: proc mixed w/ nested random factors
In <320F7E1D.4A5F@neuron.cpmc.columbia.edu>, "Dr. Steven P Ellis"
>I'd like to do a variance components analysis using proc mixed and a
>model in which the fixed part contains 2 fixed factors, a few
>covariates, and assorted interactions. In addition, there are 2 random
>factors. The first random factor is subjects (repeated measures). This
>is nested in the second random factor, blocks. Different blocks are
>independent. Different subjects are conditionally independent given
>blocks. I may also want to include an interaction of block and subject.
>I've discussed this with colleagues more learned than I concerning these
>matters. They have some ideas but none of them are sure of how to
>If someone out there knows how to handle this problem, I'd like to hear
>Thanks in advance.
> -- Steve Ellis
An outline of the statements which would fit the model you describe above
class blocks subjects time;
model response = <fixed effects, covariates> / s;
repeated time / subject=subjects(blocks) type=cs;
This model assumes that each subject, regardless of block, has a common
correlated error structure. The type=cs on the repeated statement will
induce estimation of a compound symmetric error structure. Of course,
there are quite a number of other error structures which you might wish
to model instead of compound symmetry, but I just put something in there
mostly as a reminder that you need to specify a type parameter or else
you end up with a diagonal R matrix, certainly not what you want.
As far as a model with block by subject interaction, doesn't a random
response for each subject within each block actually represent such a
model? Is there any need to explicitly fit such a model? If you were
to assume that each subject*block combination had a mean 0 drawn from
a normal distribution with variance sigma, then you are saying that
your blocks have no effect. At least that would be my understanding.
Doesn't sound to me like a model of interest. Or is there something
which I am missing?
Fred Hutchinson Cancer Research Center
1124 Columbia Street
Seattle, WA 98104