--- On Thu, 6/19/08, A.B. <alicia.bingo@GMAIL.COM> wrote:

> From: A.B. <alicia.bingo@GMAIL.COM> > Subject: The repeated statement in Proc Mixed > To: SAS-L@LISTSERV.UGA.EDU > Date: Thursday, June 19, 2008, 12:43 PM > In the following model, I have treatment group and time and > baseline score > as covariate. There are 4 time points. My question is: do I > need to put > "time" in the repeated statement? What's the > difference if I don't put it > in? > > The results do differ. As I read the SAS documents on proc > mixed, it seems > to be related to missing values. Could someone explain it > more? > > Thanks! > * > > Model 1: > > proc* *mixed* data=work method=ml noclprint; > > class id treatment time; > > model score = treatment time score_bl/s; > > repeated time/type=un subject=id; > > lsmeans time treatment ; > * > > run*; > > > > *Model 2: * > > proc *mixed* data=work method=ml noclprint; > > class id treatment time; > > model score = treatment time score_bl/s; > > repeated /type=un subject=id; > > lsmeans time treatment ; > * > > run*; > -- > A.B.

Alicia,

Yes, you are correct that specifying the TIME effect is important when there are subjects who have missing values at select time points. Suppose that you have data with 3 time points and that you assume an unstructured covariance for the residuals as shown below:

_ _ Cov = | 4.25 1.92 0.26 | | 1.92 3.76 1.43 | | 0.26 1.43 4.07 | - -

Now, if you have an individual who has observations only at times 1 and 3, then the covariance of the residuals for these two observations would be 0.26. That covariance structure is specified when there is explicit reference to a time effect.

Without specification of the time effect, there is no information to attribute the first observation to time 1 and the second observation to time 3. In that case, the procedure MIXED would attribute the first observation to time 1 and the second observation to time 2. When the empirical covariance of "time 1" and "time 2" observations is constructed (with quotes to indicate that these are not real time 1 and real time 2 but assumed time 1 and 2 observations), then the covariance between observations assumed to be from times 1 and 2 is diminished because we have attributed a real time 3 observation as a time 2 observation. When there are a large number of time misattributions, your estimated covariance structure can be way off the mark.

Does this help?

Dale

--------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------