Re: Help: Non-positive semi-definite message

From: Alan Xiao <Alan.Xiao@cognigencorp.com> Subject: Re: Help: Non-positive semi-definite message Date: Thu, 16 Aug 2001 09:35:50 -0400 Dear Vladimir. Yep. I found that the absolute magnitude of SEM (= STD ERR/MEAN) are systematically smaller with the option than not, although the relative magnitude is consistent (meaning if a parameter has a higher SEM without the option, it also has higher SEM with the option.) About the over-parameterization, it's the first problem I tried to fight with. But first I focused on covariates. About number of etas, I tried if I could increase it, rather than decrease it, because the number of etas and their positions were tested during the construction of the structural model and it always did well during the addition of the first 23 covariates to the model. But I did find that in the structural model, eta5 and eta6 did not show correlation but after addition of covariates to their parameters, they are correlated. Then I tried $OMEGA BLOCK. Now, one question, if we drop an eta from a parameter with covariates, (the addition of covariates significantly decreases the eta value, from 0.4** down to 0.04**), does this mean we push the interindividual variability be basically from the covariates? because dropping an eta is equivalent to push it as zero, although we can claim that data might be not enough to support it since eta is not estimated, which is different from being estimated as zero. Does any one have experiences that after addition of covariates to a parameter, the estimate of eta on that parameter goes to zero? - I know, theoretically it's a little bizzarr but bizzarr things sometimes happen. Thanks, Alan. -- ***** Alan Xiao, Ph.D *************** ***** PK/PD Scientist *************** ***** Cognigen Corporation ********** ***** Tel: 716-633-3463 ext 265 ******