RE: $OMEGA blocks and log-likelihood profiling

From: Nick Holford n.holford@auckland.ac.nz Subject: RE:[NMusers] $OMEGA blocks and log-likelihood profiling Date: Thu, June 3, 2004 6:58 am Hi, Thanks to all of you for your opinions. Ken does not consider over-parameterization is a form of malformed model (but I think this is hairsplitting). However, it is my impression is that it is not the model per se that is causing the failure to converge. Random (non-parametric bootstrap) samples of the original data are able to use the same model to converge and in some cases complete the $COV step. So this means to me that the failure is not a systematic feature of the model. It might however be a systematic problem with the NONMEM code or the compiler/processor combination which may take different execution paths depending on the numbers it has to work with. I looked at one of the bootstrap runs that ran the $COV step. The highest CV% (SEE/estimate*100) was 85% for between occasion variability in V1 but all the rest were less than 28%). The highest estimation error correlation was 0.93 (between renal and non-renal clearance) but not enough to declare overparameterization (See Ken's contribution http://www.cognigencorp.com/nonmem/nm/99may012003.html). It also has eigenvalues ranging from 7.35E-03 to 5.26 suggesting ill-conditioning is not severe (once again see Ken's comments). Do these values invalidate the model's ability to describe and predict the influence of covariates in predicting individual clearance values (the main point of the model building study)? I also tried Marc's suggestion to use EXIT to avoid high etas. Exiting if ETA was >=5 did not let the model complete its initialization phase. Q and V2 caused exits if ETA was >= 5.5 and <10. The runs terminated in the same way as the unconstrained run on the original data "DUE TO PROXIMITY OF LAST ITERATION EST. TO A VALUE AT WHICH THE OBJ. FUNC. IS INFINITE (ERROR=136) AT THE LAST COMPUTED INFINITE VALUE OF THE OBJ. FUNCT.". I only tried this on the original dataset. It took about a month to do the unconstrained bootstrap runs and I have other things to do right now. With regard to Jeff's comment about overparameterized bootstrap runs tending to have many "sitting on its null value" -- this certainly wasn't the case for the parameters defining the covariate effects. But it's not clear to me what he means by the 'correct' bootstrap distribution. It seems to me that the performance of a model should be judged by it's purpose and not by criteria determined by numerical idiosyncrasies. NONMEM V is a bit of a dog when it comes to local minima as Marc reminds us and the conditional methods are clearly less robust (but nevertheless IMHO more believable). So I prefer to base decisions about model performance in how good the fit is by assessment of its merits in the intended application rather than the ability to jump over certain arbitrary hurdles. Nick -- Nick Holford, Dept Pharmacology & Clinical Pharmacology University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556 http://www.health.auckland.ac.nz/pharmacology/staff/nholford/