Convergence issue with multilevel models.


No cake for spunky
As recommended I am running a chi square test to determine if random (and fixed) effects are statistically significant for a multilevel model (the Wald test reported with the normal software are known to be doubtful for a variety of reasons). I am having trouble even when using a few variables in the model (I have about 75 groups and several thousand cases if that matters). It is failing to converge or have a Hessian (matrix I believe) that is not 'positive definitive'.

The link has the macro I am using, its at the end in the appendix. Because it is so long I did not post it directly here. I think the problem with convergence is with the macro not running the standard multilevel model which has not shown this problem.

Any suggestions how to deal with the problems of not converging and a Hessian that is incorrect, if this matters, would be greatly appreciated.

This has the macro.


Less is more. Stay pure. Stay poor.
I am not in front of a computer, but are you sure they haven't incorporated that macro into the program. What were the three cov options that are available in the program? I seem to remember something like cov, cov., cov.. as options, but I'm out of the office


No cake for spunky

I am running the macro -so I think I have incorporated the macro into the program. This is what I run specifically.

proc mixed data=work.test4 method=reml covtest empirical
noclprint ;
class unitid_pri;
model dv=/ ddfm=contain s;
random intercept /subject=unitid_pri ;
ods output FitStatistics=fm SolutionF=SFfm ;
proc mixed data=work.test4 method=reml noclprint;
class unitid_pri;
model dv= /ddfm=contain s;
ods output FitStatistics=rm SolutionF=SFrm ;