Help: Non-positive semi-definite message

From: Alan Xiao <Alan.Xiao@cognigencorp.com> Subject: Help: Non-positive semi-definite message Date: Wed, 15 Aug 2001 09:54:13 -0400 Dear NMUSERs, Can any one systematically address the possible causes of and possible solutions to the message of "Nonpositive semidefinite but nonsingular" ? I have a 5-compartment model to simultaneously fit one parent drug (2 compartments) and two SERIAL metabolites ( 1 and 2 compartments) over a database of 600 patients and >7000 conc records. The conversion between the two metabolites are reversible. With a final model including 29 covariate terms on 5 different clearances (intact elimination and conversion) (40 THETAs + 5 ETAs + 3 EPSs), I always got the above message. I tried changing initial guesses, merging different covariates (if applicable), using omega block, etc. the message was always there just like your shadow under the sun. Note that, with 23 covariates in the model (38 THETAs) during forward selection, there is no problem to go through the $COV step. After that, I dropped the $COV step from the model for further forward selection and backward elimination (of covariates) just because it cost too much time (>30 hours for each run with $COV on a high speed UNIX system in average). Also note that, no matter how diverse the initial guesses are, the final esitmates are the same once it is successfully minimized (still with the above message). (Of course I just tried a few sets of initial guesses, out of numberous possible combinations. - This is based on one of explanations that this message might mean minimization is not global but local). In addition, there are some patients in the dataset with missing covariate values and they were imputated with population MEANs. Also there are some obvious outliers which can not be appropriately expressed by any covariate funcitons. I appreciate if any one can share his/her experience of how to handle this kind of problem. (I've already searched discussions on this topic previously posted in this news group). Thanks, Alan. -- ***** Alan Xiao, Ph.D *************** ***** PK/PD Scientist *************** ***** Cognigen Corporation ********** ***** Tel: 716-633-3463 ext 265 ******