mixed model: NOTE: Convergence criteria met but final hessian is not positive definite.

[fengyuwuzu]

When I ran a mixed model, I got a note of

NOTE: Convergence criteria met but final hessian is not positive definite.

I added "parms /old" and still got he same note. Is this because I used "type=cs" in the second random statement?


How to take care of this note message, or having this note does mean my model is not valid? Thanks

proc mixed data=addm1 method=reml;
   class TRTPN(ref='1') VISITNUM(ref='1') SUBJID;
   model CHG=TRTPN VISITNUM TRTPN*VISITNUM/solution;
   random intercept/subject= SUBJID type=un;
   random intercept/subject=pknum(SUBJID) type=cs;
   lsmeans TRTPN/pdiff adjust=SCHEFFE;
   parms /ols;
run;

[Rick_SAS · Accepted solution]

There are several possible reasons, which are discussed in the excellent paper "Tips and Strategies for Mixed Modeling with
SAS/STAT Procedures" (Kiernan, Tao, and Gibbs, 2012).  In my experience, a common reason is that the model does not fit the data, either because it is misspecified or because the sample is too small.

[fengyuwuzu]

Thank you Rick. I read this paper this morning, and the idea to include 

parms /ols;

is from this paper. I will read it more carefully 

Probably due to the small sample size, and we have 5 arms.