OK, I was misconstruing the subject when I calculated 78. However, here is an approach using @sld's data that may work. Note that since the subjects within a population are unique that the random matrix (G matrix) is all zeroes for the off-diagonal entries. So here is what I called 'Try8';
title1 "Try8";
proc mixed data=test_long method=reml ;*asycov;
ods output g=g8;
class varname pop id;
model y= varname;
random pop/subject=pop*id type=un g;
repeated varname / subject=pop*id type=un r;
run;
The final hessian is not positive definite, so I then tried porting to PROC GLIMMIX and used the following:
title1 "Try9";
proc glimmix data=test_long ;*method=reml ;*asycov;
*ods output g=g9;
class varname pop id;
model y= varname/ddfm=kr2;
random pop/subject=pop*id type=un g;
random _residual_ / subject=pop*id type=cs g;
run;
Still reports a lot of off diagonal zeroes, and the covariances between varnames isn't modeled, so i tried the following which worked well:
proc glimmix data=test_long ;
*ods output g=g9;
nloptions tech=congra;
class varname pop id;
model y= varname/ddfm=kr2;
random varname/subject=pop*id type=un residual g;
random pop/ subject=pop*id type=un(1) g;
run;
It still reports the off-diagonal zeroes for the second random statement but I think it captures everything.
I am at a loss as to why the OP's data returns an infinite likelihood, and will be curious if this last approach solves that problem.
Steve Denham
