Hi everyone,
Have run into a problem that has me absolutely stumped and would appreciate some assistance.

I'm attempting to compute the ICC from a dataset using mixed models (Level 1 = Individual, Level 2 = Agency). I'm principally running SPSS (syntax below), but tried it with SAS and ran into the same problem. I continually get the "Hessian matrix is not positive definite" warning with the covariance estimate coming out zero. Yet there seems to be significant variability both within and across agencies when I look at the data. The models I'm running are empty except for the random effect and I've played with the step-halvings, iterations, etc. endlessly without any change. The inclusion of additional fixed factors actually DOES lead to a positive definite matrix in some cases (the opposite of what I'd expect based on readings suggesting this means you need to simplify the model), but they obviously can't be included for when I'm calculating the ICC.

Any ideas? Can this only mean that the ICC is effectively zero, even if eyeballing the data indicates variance across and within organizations? I don't actually need



MIXED dv
/CRITERIA=CIN(95) MXITER(1000) MXSTEP(1000) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0.1,
RELATIVE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE)
/FIXED=| SSTYPE(3)
/METHOD=REML
/PRINT=HISTORY(100)
/RANDOM=INTERCEPT | SUBJECT(Agency) COVTYPE(VC).