09-29-2015 09:51 AM
When estimating a multilevel model using proc mixed, is there a way to fix some of the covariances in the G matrix to 0 while freely estimating the other components? Here is the matrix I want to estimate:
cov cov var
0 0 0 var
0 0 0 0 var
0 0 0 0 0 var
09-30-2015 02:54 PM
Assuming that you have an unstructured matrix, and you can make reasonable assumptions regarding the variances and covariances to be estimated, then the PARMS statement should be what you need:
PARMS (var1) (cov12) (var2) (cov13) (cov23) (var3) (0) (0) (0) (var4) (0) (0) (0) (0) (var5) (0) (0) (0) (0) (0) (var6)/hold=7,8,9,11,12,13,14,16,17,18,19,20;
Insert appropriate values for var1-var6 and the covariances. OR set the variances to 1 and covariances to 0, and let the optimizer work.
10-09-2015 04:17 PM - edited 10-09-2015 04:17 PM
Thank you so much for helping me understand the PARMS statement. What would be the criteria for determining appropriate starting values for the variances and covariances if I don't use 1s and 0s?
10-12-2015 08:37 AM
Variances and covariances are easy to come by using PROC CORR with the COV option. These would be reasonable starting values.