RE: $OMEGA blocks and log-likelihood profiling

From: "Gastonguay, Marc" marc.gastonguay@snet.net Subject: RE: [NMusers] $OMEGA blocks and log-likelihood profiling Date: Wed, June 2, 2004 10:27 am Ken, Nick, Mark, NMusers: In addition to problems with over-parameterization, it's possible that a large fraction of bootstrap runs can fail (e.g. minimization unsuccessful) even when the model is not over-parameterized: particularly when one of the conditional estimation methods is used (e.g. methods other than FO). I've learned from Tom Ludden that NONMEM V can sometimes run into local minima when searching for individual ETA estimates - resulting in convergence difficulties for the population model. I suspect that this may be happening with a fraction of the bootstrap failures under conditional estimation. I've found that placing constraints on the individual ETA estimates through PRED-exit codes can be helpful in dramatically reducing the fraction of failed bootstrap runs. For example (assuming an exponential inter-individual variance model and 3 variance terms in OMEGA): IF(ABS(ETA(1)).GT.10) EXIT 1 102 IF(ABS(ETA(2)).GT.10) EXIT 1 103 IF(ABS(ETA(3)).GT.10) EXIT 1 104 ... (also use NOABORT option in $EST) I hope this is helpful. Marc Gastonguay