Proc mixed, Convergence criteria met but final Hessian is not positive definite

[shoh]

I ran the code below in Proc mixed, where TRT is an integer type such as 21, 22, 23,..., etc. However, when this result and TRT are converted to text(character) and put into the model, the Covariance Parameter Estimates and Fit statistics are the same, but the results are different from the Type 3 estimate. Looking at the Model iteration statement, when I put in an integer, “Convergence criteria met” appears, but when I put in a character type, “Convergence criteria met but final Hessian is not positive definite” comes up, and when I look at the covariance matrix, the two are different. Can you tell me why the two are different even though they are specified as CLASS?

PROC MIXED DATA=Lipid_Chg;
by test;
where CHG ^=. and visit eq "PT2D28" and test eq "VLDL Cholesterol Cal";
CLASS subnum trt visit;
MODEL CHG= basen trt visit trt*visit /DDFM=KR;
RANDOM subnum;
LSMEANS trt*visit/CL DIFF ALPHA=0.05;
RUN;

[SteveDenham]

Could you share the first part of the output for both attempts?  The part I would like to see is the material up to and including the iteration history.

Thanks,

SteveDenham

[PaigeMiller]

A search of the SAS Communities for "Hessian not positive definite" returns several posts with answers marked correct.

[StatsMan]

Your WHERE statement restricts the data to a single level of VISIT and you have a TRT*VISIT interaction in the model. It would be unusual to fit a CLASS effect with a single level. It would be even more unusual to fit an interaction of that single level CLASS effect. 

Seeing the CLASS level table results might shed some light on what else could be causing problems here.