[Fwd: Re: Non-positive semi-definite message]

-------- Original Message -------- Subject: Re: Non-positive semi-definite message Date: Wed, 15 Aug 2001 16:52:12 -0400 From: Alan Xiao <Alan.Xiao@cognigencorp.com> To: bvatul <bvatul@ufl.edu> Oh, boy, I just tried MATRIX=S option in $COV, it just took 20 minutes (I used $MSFI), compared to about 15-20 hours without this option (also use $MSFI). And it past the $cov step and the results are consistent with those of models during the forward selection process (all of the base models except the last few used $cov but without this option). Now, any suggestions about the exact explanations to this phenomenon? Thanks a lot - Atul. (Hi, Atul, I'll buy you a beer next time when we meet - might be on the AAPS meeting?). Alan. Alan Xiao wrote: > Dear Atul, > > Thanks a lot. > > Your third choice (MATRIX=S) is really what I first time heard about > using it. Does this mean we go around Matrix R (or more > appropriately, just inverse R) ? The NONMEM manual just mentioned > that "MATRIX=S requests that the inverse S matrix be used. " At this > case, does the statistical inference change? (since R and S are two > different matrices - Hessian and Cross-Product Gradient). Has > anyone explored the differences - in parameter estimates, statistical > inferences, etc. - between options MATRIX=R and MATRIX=S over a good > model (which does not produce the message "non-positive ...")? Does > this message only happen to MATRIX R (or strictly, inverse R)? or > also possible to matrix S (inverse S)? > > I did try SIGDIG = 4 and the first round results showed "Rounding > Error". The second round of results (after changing initial guesses) > are not out yet (it takes more than 30 hours). I did not go so far as > to SIGDIG = 5. Here, I'm wondering if anyone tried even further > before, such as SIGDIG = 6 or 7 (if applicable with any reasons)? > > For the first choice - minimizing (the number of) covariates, it's > really helpful during the forward selection step (if using $cov) since > it could pick up redundant (or correlated) covariates depending on the > magnitude of p values. However, if it's after the backward > elimination, ... Couldn't the backward elimination process remove > the redundant (correlated) covariates from the model? Or simply the > magnitude of p values is not lower enough (to remove redundant or > correlated covariates)? If so, how lower is low enough (generally in > industry) - a silly question, huh? - I use p = 0.001 right now. If > not so, does this mean we can not (completely) rely on the backward > elimination step based on the magnitude of the change of the objective > function values (chi square distribution assumption)? > > Sorry for a lot of questions. > > Thanks, > > Alan. > -- ***** Alan Xiao, Ph.D *************** ***** PK/PD Scientist *************** ***** Cognigen Corporation ********** ***** Tel: 716-633-3463 ext 265 ******