|Date: ||Tue, 13 May 2003 13:08:45 -0700|
|Reply-To: ||Kimberly Austin <austinkimberly@YAHOO.COM>|
|Sender: ||"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>|
|From: ||Kimberly Austin <austinkimberly@YAHOO.COM>|
|Subject: ||Re: repeated measures MANOVA follow-up question|
|Content-Type: ||text/plain; charset=us-ascii|
And again thank you for your responses to my repeated measures MANOVA problem.
I have run into a few difficulties with fitting a repeated measures MANOVA model in proc mixed. First, when using the repeated statement in the following code:
class unit var treatment time;
model y=var var*treatment var*cov / noint;
repeated var*time / subject=unit type=UN@CS;
results in the error "two repteated effects are required with this covariance structure". I believe that both sources of correlation, the mutliple responses, var, and the repeated measurements on the experimental units, unit, need to be placed immediately after the "repeated":
repeated var*time unit/ type=UN@CS
but this model will not coverge on parameter estimates.
However, for the data set I am working with, measurements were taken over a period of many months. Treating "time" as a class variable results in a variable with over 200 levels, making the interaction var*time too complex.
My question remains, how to properly model both the correlation among the multiple responses, "var", and the correlation among repeated measurements on the same experimental units, "unit", which were taken at unequal intervals and are not balanced among the units.
It seems the further I venture into this problem the more I am prone to take a more simplistic approach, perhaps collasping repeated measurements on units into one mean value for each unit on which to base the analysis?
Any suggestions to this problem would be greatly appreciated.
"...it is silly to design a study for which the tools to analyze the resulting data properly are not available."
-Hamer and Simpson. 1998. SUGI:23
Dale McLerran <stringplayer_2@YAHOO.COM> wrote:
What happens when you remove the intercept and main effect of
treatment is just this: you end up with a model for each
response which has the form
Y(i) = b0(i) + b1(i)*treatment
where (i) indicates which response is being modeled. The terms
b0(i) are obtained from the indicator variable effect estimates
while the terms b1(i) are obtained from the indicator by
treatment interaction term estimates. Note that if you include
intercept and main effects in your model, then you end up
fitting the model
Y(i) = b0(k) +
That is, you obtain a model in which the intercept is the
intercept for the last response variable in your set and the
variable indicator effect is an intercept offset for the i-th
response from the intercept for the last response. Similarly,
the main treatment effect is just the treatment effect for the
last response category, while the response indicator by
treatment effects are offsets of treatment effects for the
i-th response from the last response variable. In this model,
the type III tests of effects become tests of whether there
is variability in the intercept term across responses and
whether there is variability in the treatment effect across
responses. Now, that may be useful in some situations, but
in general for a multivariate response one would not want to
assume that the intercept and treatment effects have common
values for all responses.
If you have completely balanced data, then PROC GLM is able to
produce the same results as when the UN@CS covariance structure
is employed in PROC MIXED. PROC GLM requires, however, that
the entire response vector Y1 Y2 Y3 Y4 be discarded if you are
missing a single value from the set. PROC MIXED does not
require that you throw away data if a response is missing.
Also, PROC MIXED allows you to model covariance structures
other than compound symmetry for the repeated measures. So,
yes, I would favor PROC MIXED over PROC GLM.
Do you Yahoo!?
The New Yahoo! Search - Faster. Easier. Bingo.