RE: different results from two computers

From: Mark Sale - Next Level Solutions com Subject: RE: [NMusers] different results from two computers Date: Wed, 16 Aug 2006 04:57:22 -0700 Nick et al Update on the stability of NONMEM/importance of successful minimization. After fixing a problem with the summation of the likelihood, we have managed to make NONMEM converge successfully indenpendent of sequence of subjects. You may recall the discussion that, for one problem at least, whether NONMEM converged was dependent on the sequence of subjects in the data set. Nick (appropriately) cited this as some evidence of the lack of importance of successful minimization, since for this this problem at least it seemed to be a random event (since the sequence of subjects was randomly defined). We seemed to have solved that source of instability in NONMEM with some fairly simple changes in the code, and now have a version of NONMEM VI where convergence is independent of sequence of subjects in the data set, and is essentially the best case from the 100 native NONMEM runs. Perhaps eventually, other sources of numerical instability will be identified and we can all agree that minimization really does mean something about the quality of the model/data set. I'll post the results of the native NONMEM and the modified NONMEM on my web site (www.NextLevelSolns.com/downloads) probably tomorrow. I'm looking for cases to try where this approach will make models that refused to do a covariance step now will do one. Anyone with interesting, relatively fast problems (< 10 minutes) that fail covariance, please contact me. Mark Sale MD Next Level Solutions, LLC www.NextLevelSolns.com