Contributor
Posts: 51

# Proc Mixed random statement sub=option/repeated measures, sp(pow) cov

Hello all !
I've created mixed model versions to fit a continuous response from an observational study with (clustered) subjects (IDs) being located at different "centers" (location), The response has been measured 3 times, but time intervals have been unequally spaced. I tried to use the sp(pow) method as proposed in "SAS for Mixed Models 2nd edition, Chapter 5.4" and timeD being the averaged elapsed days from timepoint 1 (1, 66, 204) at all locations.

I face two problems with the first code below:
First, I get an in complete and questionable R Matrix and R Correlation Matrix, with exactly the same values for the variance (on the diagonal), and missing values for the covariance and correlation between time 1 and 3.

Does anyone know what I have done wrong in applying the spatial cov structure sp(pow) ?

Second, table of dimensions for X and Z matrices are correct, but the number for subjects and max. observations per subject doesn't seem to be correctly recognised by the code (subject=1, and number of max. observations per subject corresponds to the number of observations/datalines).

Proc mixed data=x;
class ID geno time location;
model haem = geno time geno*time;
random location;
repeated time / subject= ID(location) type=sp(pow) (timeD) r rcorr; run;

With an additional subject options for the random statement the dimensions for subjects and max. observations per subject seems fine (max. 3 observations per subject), but inference on treatment (geno) and time effects is different (!) now with the F-statistic based on fewer denominator degrees of freedom (two-thirds of the den df in the above model):

Proc mixed data=x;
class ID geno time location;
model haem = geno time geno*time;
random location / subject= ID(location);
repeated time / subject= ID(location) type=sp(pow) (timeD) r rcorr; run;

What is the meaning of the subject= in the random statement? Which of them is the correct syntax?
Including the random location in the code doesn't give a Null Model Likelihood Ratio Test in the output either for comparison with the independent error model. Is there another way to do this ?

I'd appreciate very much your explanation and experience.
Thanks very much.
Discussion stats
• 0 replies
• 260 views
• 0 likes
• 1 in conversation