topic Re: Change convergence status if Hessian has negative eigenvalue in Statistical Procedures

Change convergence status if Hessian has negative eigenvalue

Maya1 — Tue, 27 Sep 2016 09:00:43 GMT

Hello,

I want to discard results from nlmixed if the optimization did not converge or if there are any problems with the Hessian Matrix (either not p.d. or negative eigenvalue)

Is there any chance to change the options in the convergence status to non-convergence if there are any problems with the Hessian Matrix? Otherwise I would have to read out the Hessian and test with iml whether the eigenvalues are > 0, which is somewhat inconvenient.

Thanks in advance!

Re: Change convergence status if Hessian has negative eigenvalue

SteveDenham — Tue, 27 Sep 2016 11:11:50 GMT

I have never tried any of these, but there seems to be several options to PROC NLMIXED that may be of use here. The section on Covariance Matrix Tolerances and the section in Details about the covariance matrix offer some hope that you can get at what you want. It appears that use of the G4= and COVSING= options may get at this problem. If I had to search for more help on this, Lex Jansen's site might have some papers.

I hate to say "Try some of these and see what happens" but sometimes that might be what you have to do.

Steve Denham

Re: Change convergence status if Hessian has negative eigenvalue

Rick_SAS — Tue, 27 Sep 2016 11:39:01 GMT

Read the article "Monitor convergence during simulation studies in SAS." The article describes how to use the Convergence Status table to monitor the status of a nonlinear optimization.