Help with PROC MIXED for repeated measures analysis with time-dependent covariate

I am a non-statistician, and would appreciate help on this.

A measurement (Y) is repeated on each subject on each of eight time points.
A covariate (covar) is also measured at each of these time points. There is
prior knowledge that there is a linear relationship between Y and covar.

The eight time points reflect different experimental conditions, and I would
like to examine the differences between the time points, adjusting for the
effect of the covariate.

The data is organized as single line per measurement in time.

I found it difficult to find information about proc mixed in one place. I
have not been able to find relevant examples of how to analyse this, but
some information from lurking on this group and some SUGI papers have
prompted me to make two guesses:

proc mixed data = mydata;
class id time; *subjects have been assigned unique id numbers;
model y = time covar time*covar / solution;
repeated time / subject = id type = cs;

or maybe?

proc mixed data = mydata;
class id time;
model y = time /solution;
repeated time / subject = id type = cs;
random intercept covar / subject = id;

Am I close? I don't understand whether to include the interaction term in
the first model. Also I am unsure how to include the covariate as a random
effect in the model. I understand I'll also have to specify the covariance
structure and add the relevant contrast statements.

I would very much appreciate suggestions.

Keith Wong
keithw@med.usyd.edu.au

Thanks for the help!

Does the model specification in my first guess (copied below) make the
continuous time-dependent covariate COVAR a fixed effect? Does it matter?

proc mixed data = mydata;
class id time; *subjects have been assigned unique id numbers;
model y = time covar time*covar / solution;
repeated time / subject = id type = cs;



Keith Wong


----- Original Message -----
From: "Paul R Swank" <Paul.R.Swank@UTH.TMC.EDU>
Newsgroups: comp.soft-sys.sas
Sent: Thursday, August 07, 2003 3:05 AM
Subject: Re: Help with PROC MIXED for repeated measures analysis with
time-dependent covariate


> Go with the first one. However, you need to ask yourself if the relation
of
> the covariate to the dv will change over time because that's what the
> interaction is addressing. You could leave it in and then drop it if it is
> not significant. If it is significant it will certainly complicate the
> interpretation. You may also want to consider if compound symmetry is the
> appropriate form for the var-cov matrix. If the time points are equally
> spaced, autoregresive might be a good alternative, depending on how far
> apart the assessments are.
>
> Paul R. Swank, Ph.D.
> Professor, Developmental Pediatrics
> Medical School
> UT Health Science Center at Houston