Date: Tue, 15 Oct 2002 14:25:18 -0700
Reply-To: Dale McLerran <stringplayer_2@YAHOO.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Dale McLerran <stringplayer_2@YAHOO.COM>
Subject: Re: repeated cov structure question...
In-Reply-To: <20021015185711.81035.qmail@web40707.mail.yahoo.com>
Content-Type: text/plain; charset=us-ascii
--- Kimberly Austin <austinkimberly@YAHOO.COM> wrote:
> Dale and other interested SAS-Lers,
> Thanks (again) for your input on my cov structure question. Of
> course, I have a few follow-up questions that I am currently mulling
> over:
> 1. How should I approach selection of a spatial cov structure?
> Initially I thought a linear structure was most intuitive given that
> the spatial (time) effect I am using is Julian date. However, after
> doing some snooping, I found that SP(LIN) would not run because the
> convergence criteria for REML wasn't met. I have tried a few other
> structures and found the exponential SP(EXP) structure does run.
> Before I do any further snooping I need to get a better grasp of how
> to "properly" approach selection of a cov structure. Should this be
> a concern, and if so, how should I approach it?
Did you try the SP(POW) covariance structure. If you had equally
spaced observations (though with some observations possibly missing),
the SP(POW) structure would be identical to an AR(1) covariance
structure. AR(1) is a common structure employed for time series
data. That is why I employed SP(POW) in the template which I
suggested yesterday. The structure SP(LIN) = V(1-pDij)1(pDij<=1)
is actually a nonlinear covariance structure. The covariance
decreases linearly with increased distance as long as pDij<=1.
When pDij>1, then the covariance is 0. So, this structure is
linear in two segments, but that makes the overall structure
nonlinear. I would think that you might run into estimation
problems with such a model.
> 2. I am still wrestling with the idea that I need to include my
> experimental unit as a random effect to account for between-subject
> variation with in my treatments (repeated observations were taken on
> experimental units within treatments only, i.e., units belonged to
> only one treatment group through the course of the study). However,
> when including a random statement [random unit(treatment)] in
> addition to a repeated statement, the number of subjects listed under
> "model information" in the output is given as 1 (where as without the
> random statement the correct number of subjects is given).
You probably had your code structured something like
proc mixed data=mydata;
class ID ...;
model response = ...;
random ID;
repeated / subject=ID type=sp(exp)(time);
run;
as opposed to
proc mixed data=mydata;
class ID ...;
model response = ...;
random intercept / subject=ID;
repeated / subject=ID type=sp(exp)(time);
run;
>Also,
> the maximum number of observations on a subject is given as my total
> sample size as opposed to the actual maximum number of observations
> made of a subject. So I guess my question is, when inluding subjects
> (within treatment) as a random effect to account for between-subject
> variation, in addition to using a repeated statement to account for
> correlation among repeated observations on a subject, does the
> repeated statement still "work" proplerly given these differences?
The repeated statement will still work. Note, though, that a
simple random effects model such as
proc mixed data=mydata;
class ID ...;
model response = ...;
random intercept / subject=ID;
run;
is identical in covariance structure to a repeated measures design
with compound symmetric error structure modeled by
proc mixed data=mydata;
class ID ...;
model response = ...;
repeated / subject=ID type=cs;
run;
The spatial covariance structures model a decay in the covariance
with increased distance. If there is little decay over time, then
the random and repeated effects may not be separable. However, to
the extent that you do have a person effect, and within the person
the covariance between observation close in time is larger than
the covariance between observations distant in time, then you
should be able to estimate both random and repeated effects.
> I hope that made sense.
> Any input, or relevant references to dig into that you know of, are
> welcomed.
> Thanks, again, for your time,
> Kimberly
Dale
=====
---------------------------------------
Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra@fhcrc.org
Ph: (206) 667-2926
Fax: (206) 667-5977
---------------------------------------
__________________________________________________
Do you Yahoo!?
Faith Hill - Exclusive Performances, Videos & More
http://faith.yahoo.com
