From: "Jing Li" jli29@jhmi.edu Subject: [NMusers] difficulty with FOCE Date: Thu, 13 Oct 2005 10:38:12 -0400 Hi, I am running a 3-compartmental PK model with a non-linear elimation (Michealis-Menten) from the central compartment. The data were from 3 trials with total patients of 224. The drug was given as 20-min iv infusion weekly or once every three weeks at different dose leves. When it was run with first order (FO) estimation, the model got converged successfully and produced very reasonable PK parameters (THETAs), inter-individual variabilities (ETAs), and relative standard errors of estimations. The fitting looks pretty good. However, when it runs with first order conditional estimation (FOCE), convergence becomes almost impossible. Iteration becomes extremely slow, and after several iterations, there is always an error message like "...Occurs during search for ETA at a nonzero value of ETA. Numerical difficulties with integration routine. Maximum No. of evaluaitons of differential equations exceeded 100000". Does anyone have some good ideas to solve this problem? I appreciate! Also, is it acceptable to report parameters produced by FO? I am wondering if there are some guidelines or references for the choice of FO and FOCE. Thanks you in advance for any help! Sincerely, Jing _______________________________________________________ From: Justin Wilkins justin.wilkins@farmbio.uu.se Subject: Re: [NMusers] difficulty with FOCE Date: Thu, 13 Oct 2005 17:48:30 +0200 Hi Jing, Without seeing your model it's hard to narrow down the specific cause of your problem, but a typical rule of thumb to follow when this sort of thing happens is to make sure that your model is not overparameterized. FOCE is much more sensitive to unneeded model parameters than FO. I would test each of your parameters - starting with ETAs, and the ones with the largest RSE - to make sure it belongs in the model. If the 95% CI of a parameter (calculated from the standard error) overlaps zero, it's probably not needed. Check your FOCE output file for zero gradients (or very low gradient values) in any of the parameters as another clue as to which to start investigating first. If any of the FOCE parameter estimates look odd or very low, remove them and try again. If you're using ADVAN6, you might try using ADVAN8 instead, especially if your system is stiff (if there are large differences in the quantities used in your differential equations). It would be very helpful if you could post your model code so we can get a better idea of where the fault lies! Use of FO is associated with inflated rates of type I error (see publications on this topic by Wählby et al). Despite FOCE's relatively poor stability and greatly-increased run times, if you can get you model to converge using it - preferably with the INTER option - I would be much more confident in your parameter estimates. FO is useful for 'quick and dirty' model-based analysis but it is very definitely not as good as FOCE. I would feel satisfied with FO only where the time needed for FOCE was excessive. Justin _______________________________________________________