Estimation method selection

Dear nmusers I am new to the field of pop pk. I am developing a Pop PK model of a compound by combining sparse PK data from 3 different mouse PK studies. I have following doubts/queries: 1. When we tried diff. estimation methods in nm 7.1 such as FO, FOCE or laplace. We got lowest OFV with laplace. Can we select the estimation method based on only OFV. We got the estimates in both FO and laplace but variability was reaching the boundary in case of FO. -Is it correct to make decision based on OFV across different models (one comp vs two comp). The diagnostic plots seem to be similar however variability was less with laplacian method as compared to FO. Which method of estimation would be most appropriate. - Is it correct to compare OFV across difeerent estimation methods (viz FO; FOCE; Laplace) - Is it correct to compare OFV across difeerent PK models (viz 1 comp; 2 comp etc) Inputs from experts are welcomed. Tausif Ahmed PhD Piramal Life Sciences Ltd, India

On 10/9/2011 2:01 PM, tausif Ahmed wrote: > 1. When we tried diff. estimation methods in nm 7.1 such as FO, FOCE > or laplace. We got lowest OFV with laplace. Can we select the > estimation method based on only OFV. We got the estimates in both FO > and laplace but variability was reaching the boundary in case of FO. When I went to a training course in NONMEM 7, a prominent developer told us that "FO can F.O.". I found that quite amusing, and I believe that I have never used FO since... -- Paul Matthias Diderichsen, PhD Quantitative Solutions B.V. +31 624 330 706

Dear Tausif Ahmed, Witty comments can be fun but here is a serious attempt to help you. 1. OFV values are not comparable between methods like FO, FOCE and LAPLACE. You can apply an Monte Carlo Importance Sampling step after each of these estimation methods to generate comparable OFV values (see example below and read more in ftp://nonmem.iconplc.com/Public/nonmem712/intro712.pdf). However I am not sure that I think that the estimation method needs to be chosen in that way. For your average continuous data problem FOCE with interaction should be a sufficient method. FO are well known to be biased and should only be used if practical problems such as runtimes etc. prohibits you from using other methods. ; After the SAEM method, obtain good estimates of the marginal density (objective function), along with good estimates of the standard errors. $EST METHOD=IMP INTERACTION EONLY=1 NITER=5 ISAMPLE=3000 PRINT=1 SIGL=8 SEED=123334 CTYPE=3 CITER=10 CALPHA=0.05 2. OFV values are comparable across different hierarchical structure models. There are a two parameter (i.e. approximately a 2 degree of freedom) difference between a 1 compartment and a 2 compartment model. However, look out because there are several possible caveats in these comparisons especially with regards to censored observations [I,II,III]. I. Byon W, Fletcher CV, Brundage RC. Impact of censoring data below an arbitrary quantification limit on structural model misspecification. J Pharmacokinet Pharmacodyn. 2008 Feb;35(1):101-16. II. Ahn JE, Karlsson MO, Dunne A, Ludden TM. Likelihood based approaches to handling data below the quantification limit using NONMEM VI. J Pharmacokinet Pharmacodyn. 2008 Aug;35(4):401-21. III. Bergstrand M, Karlsson MO. Handling data below the limit of quantification in mixed effect models. AAPS J. 2009 Jun;11(2):371-80. Kind regards, Martin Bergstrand, PhD Pharmacometrics Research Group Dept of Pharmaceutical Biosciences Uppsala University Sweden [email protected] Visiting scientist: Mahidol-Oxford Tropical Medicine Research Unit, Bangkok, Thailand Phone: +66 8 9796 7611