Modeling of inter study variability (ISV) as a third random effect level

From: "Harnisch, Lutz, HMR/DE" <Lutz.Harnisch@hmrag.com> Subject: Modeling of inter study variability (ISV) as a third random effect level Date: Mon, 31 May 1999 16:21:28 +0200 I don't know how to implement into NONMEM the model of a third random effect level, like ISV. Sampling individual ETA's from an additional OMEGA-element and allocating those for individual studies didn't work. I don't know how to maintain the individual ETA from one individual to the next within a single study. And furthermore how to get a new ETA for the first record of the next study. Therefore I tried a different approach: to sample all study ETA's at NEWIND.EQ.0 (i.e. at the first record in the dataset) with $OMEGA BLOCK(1) 0.1 ; first sub study ETA(1) $OMEGA BLOCK(1) SAME ; sec study ETA(2) ... $OMEGA BLOCK(1) SAME ; last substudy ETA(n) and allocated them as factors for instance on CL by IF (NEWIND.EQ.0) THEN IS01 = EXP(ETA(1)) IS02 = EXP(ETA(2)) ... ISn = EXP(ETA(n)) ENDIF FCL = 1 IF (STUD.EQ.ST01) THEN FCL = IS01 ELSIF (STUD.EQ.ST02) THEN FCL = IS02 ... ENDIF CL = THETA(x) * FCL But the coding above causes all the gradients for the PK-model to be zero, even if I fix ETA(1) to zero, which means for my understanding that, all factors should become 1, and therefore, the gradients should be different from zero. I really like to know if coding of ISV with NONMEM is possible, any help would be highly appreciated regards Lutz Harnisch Hoechst Marion Roussel | Product Realization | Biodynamics ( http://popkin.fra.hmrag.com/) Building H840 | Room 449 | D-65926-Frankfurt-Main | Germany phone +49-69-305-16481 | fax +49-69-305-81990 | mailto:lutz.harnisch@hmrag.com

Date: Mon, 31 May 1999 17:46:11 +0200 From: Pascal Girard <pg@upcl.univ-lyon1.fr> Subject: Re: Modeling of inter study variability (ISV) as a third random effect level Dear Lutz, Silvy Laporte has presented an implementation of such an ISV model at the last PAGE meeting in Wuppertal and a poster at ASCPT 98 in New Orleans. We have recently submitted a paper on this topic. The way you implement it does not work since at NEWIND=0, NONMEM has no information to compute random study effect except for the study that appears on the very first record of your dataset. We implemented ISV just like Mats implemented IOV. The trick is that in order to have a common ETA for every patient in one study, but different ETAs from one individual to another you have to set study number as the ID (ID=STUD) but no longer the patient identifier, create <<as many ETAs as the maximum number>> of patients in the biggest studies, then constraint all these ETAs to have the same variance. cheers, Pascal Pascal Girard ------------------------------------------------------------------- Service Pharmacologie Clinique PG@upcl.univ-lyon1.fr BP 3041,162, avenue Lacassagne Tel : +33 (0)4 78 78 57 26 69394 LYON Cedex 03 FRANCE Fax : +33 (0)4 78 78 57 19

From: "Harnisch, Lutz, HMR/DE" <Lutz.Harnisch@hmrag.com> Subject: AW: Modeling of inter study variability (ISV) as a third random e ffec t level Date: Mon, 31 May 1999 18:36:13 +0200 Dear Pascal, Thank you very much for your ideas. We were puzzled about this topic already at PAGE in Wuppertal. In our pooled phase I analysis, the maximum number of individuals within a study was larger than 20 subjects/study, however, the number of studies were below 20. Therefore, the introduction of more than 20 IOV corresponding to the maximum number of individuals might extent the limits in NONMEM, taking also into account that 20 IOVs for each parameter in the PK model are required, probably. Therefore we decided to take only a single THETA for each study, and to summarize them in a second stage. But what do you think about taking a small random sample of individuals from each study (around 5 IOVs), and apply your approach? regards Lutz

From: "Piotrovskij, Vladimir [JanBe]" <VPIOTROV@janbe.jnj.com> Subject: RE: Modeling of inter study variability (ISV) as a third random effect level Date: Tue, 1 Jun 1999 14:11:16 +0200 Dear, NONMEM is not the best choice for modelling multiple nested random effects. The new S-PLUS library called nmle version 3.0 is more suitable. It is on beta-testing now and can be downloaded from the following sites: http://nlme.stat.wisc.edu/Beta http://cm.bell-labs.com/stat/NLME/Beta Since it is already the 8th release, it is quite reliable, I believe. Best regards, Vladimir Vladimir Piotrovsky, PhD Clinical Pharmacokinetics, ext 5463 Janssen Research Foundation 2340 Beerse, Belgium e-mail: vpiotrov@janbe.jnj.com