Comaprison of two models

Dear NMusers I would like to compare two models. Let’s say the model M1 and the model M2. The model M1 is a simple one with just one observation compartment (Y = IPRE*(1+ERR)). The second one is a more complex one with three observation compartments (Y1 = IPRE1*(1+ERR1); Y2=IPRE2*(1+ERR2); Y3=IPRE3*(1+ERR3)). The data sets are identical with regards to the first observation compartment. Y form M1 is in fact Y1 from M2 and Y2 and Y3 from M2 are additional observations which should improve the model because of additional information or perhaps not because of additional noise. If I am interested in comparing the two models focusing on the first observation (i.e. Y form M1 and Y1 form M2, respectively), I cannot use the OFV, since OFV2 (OFV for M2) will be a global measure of the fit including Y2 and Y3 from M2. So, how can I perform an estimation of M2 including the three observations and then isolate the contribution of Y1 to the global OFV2? May I assume additional properties of OFV, i.e. OFTtotal = OFV1+OFV2+OFV3? Is it possible to code the model so that only OFV1 will be computed? Many thanks in advance. Let me know if you need additional information. Best regards Robert ___________________________________________ Robert M. Kalicki, MD Postdoctoral Fellow Department of Nephrology and Hypertension Inselspital University of Bern Switzerland Address: Klinik und Poliklinik für Nephrologie und Hypertonie KiKl G6 Freiburgstrasse 15 CH-3010 Inselspital Bern Tel +41(0)31 632 96 63 Fax +41(0)31 632 14 58

Hi Robert, Total objective function is the sum of the OF of all observations, so it could be possible to compare OF related to Y1 directly (although I do not know how to extract this info from Nonmem). However, objective function is not the right way to compare these models. For example, assume that you have parent-metabolite model, and have the model that describe both quantities (Y1 and Y2 in your case). When the model have to describe both parent and metabolite, it has less flexibility to describe the metabolite (because it needs also fit the parent data). Thus, the fit of the metabolite data will not be as good as if you remove the parent-drug observations. However, if you do remove parent-drug observations, you may not be able to define all the model parameters, or you may end up with the model that is not mechanistic but more empirical, and this model can be less predictive if you need to extrapolate to other dose levels, route of administration, etc. If you need to compare how well your models fit data, use diagnostic plots: VPC, distributions of residuals, random effects, etc. You may directly compare model predictions (PRED of model 1 vs PRED of model 2, same with IPREDs), or compare diagnostic plots side-by-side. RSE of parameter estiamtes would be another thing to compare: more data may result in more precise estimates. The choice of the model may depend on why do you need the model, how will you use it. If in the above example the parent drug is inactive, and you need PK only to describe the PK-PD model of the active metabolite, you could be more interested in the precise description of the metabolite PK data, and may want to omit the parent data, just to save time. If on the other hand you are interested in the covariates that may influence parent-to-metabolite transformation, you may want to spend time investigating the entire system rather than only active metabolite. Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566