msg from atul

From: bvatul <bvatul@ufl.edu> Subject: msg from atul Date: Tue, 14 Nov 2000 15:02:11 -0500 Hello All Could someone please clarify my doubts 1. I am using ADVAN3 TRANS4 to fit the observed concentrations of a drug. I have two data sets from two studies; rich (phaseI) as well as sparse (phaseIII). The rich data set could be explained by ADVAN3 using proportional error model. However when I am analysing the sparse data set, in which the patients are administered concurrently many drugs, the model is unable to explain the variability correctly. There are five outliers out of 84 patients where I am observing high weighted residuals value of 15. The dose is given as a short term infusion. Could anyone please share views about how to go with the outliers. I have tried FO as well as FOCE and various error models (Proportional and additive). One more message which I observed was "OCCURS DURING SEARCH FOR ETA AT NONZEROVALUE OF ETA. A ROOT OF THE CHARACTERISTIC EQUATION IS 0 BECAUSE K*K21 IS MUCH SMALLER THAT (K+K12+K21)**2. PERHAPS K OR K21 IS VERY SMALL OR K12 IS VERY LARGE". 2. When I am analysing sparse data set seperately (the model parameterised in terms of Q, CL and V) the values of intercompartmental clearance (Q=6) differs from rich data set (Q=70). I do not have enough points in the distribution phase. Will such a situation lead to this discrepancy. or is it due to any model misspecification. Can I fix the value of Q and its omega which I observed in rich data set to sparse data set and analyse. Will it be meaningful? What is the criterion of fixing omega? Does any one forsee any problem if I analyse sparse and rich data set of the same drug together and later on check for potential covariates. Or should I analyse sparse data set separately. Thanks in advance Atul

From: "Bachman, William" <bachmanw@globomax.com> Subject: RE: msg from atul Date: Tue, 14 Nov 2000 15:51:39 -0500 Atul: Just a few quick comments off the top of my head: It is possible that either the pharmacokinetics or the residual error structure is different between your phase I and III subjects (or both). This can be tested in the former case by including STUDY as a covariate and testing for inclusion as a covariate on model parameters (and significance) and in the latter case by coding a different residual error model for phase I and phase III data. It can often be helpful to model sparse data with rich data if the sparse data is lacking in information. However, you need to be aware that the populations from which the data arise may not be the same and to account for the differences if possible. I don't foresee a problem looking for covariates in modeling the combined data as long as one of the potential covariates is STUDY. Bill William J. Bachman, Ph.D. Senior Scientist GloboMax LLC 7250 Parkway Drive, Suite 430 Hanover, MD 21076 Telephone: (410) 782-2212 FAX: (410) 712-0737

From: Nick Holford <n.holford@auckland.ac.nz> Subject: Re: msg from atul Date: Wed, 15 Nov 2000 10:01:37 +1300 Atul, I suggest you fit the dense and sparse data simultaneously using FOCE INTERACTION (INTERACTION is essential with proportional error models). Try using different error model parameters for the dense and sparse data. Consider removing outlier concentrations if they do not look reasonable. Ignore the error message about root of the characteristic equation. Good luck! Nick Nick Holford, Divn Pharmacology & Clinical Pharmacology University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556 http://www.phm.auckland.ac.nz/Staff/NHolford/nholford.htm

From: LSheiner <lewis@c255.ucsf.edu> Subject: Re: msg from atul Date: Tue, 14 Nov 2000 14:11:38 -0800 All - Picking up on the importance of "interaction", I recently had occasion to reconsider this, and came up with the following idea for the case that FOCE with interaction appeared to be indicated but was prohibitively time-consuming. What about taking a GLS-type approach? For concreteness, asume $ERROR fragment is: $ERROR IPRED = F Y = F + F*EPS(1) + EPS(2) Then: 1. Fit using FO with POSTHOC, and table IPRED (these will be good predictions, even if there is some bias in pop'n model, especially if data are fairly dense per individual, as data often are in precisely those cases that tend to give us biased estimates without INTERACTION ...) 2. Fit again, using previously tabled IPRED as an input data item called W, and use $ERROR IPRED = F Y = F + W*EPS(1) + EPS(2) One could, of course, cycle through this again an step 3, inputting the IPRED from step 2 as a new W ... Any thoughts? LBS. -- _/ _/ _/_/ _/_/_/ _/_/_/ Lewis B Sheiner, MD (lewis@c255.ucsf.edu) _/ _/ _/ _/_ _/_/ Professor: Lab. Med., Bioph. Sci., Med. _/ _/ _/ _/ _/ Box 0626, UCSF, SF, CA, 94143-0626 _/_/ _/_/ _/_/_/ _/ 415-476-1965 (v), 415-476-2796 (fax)

From: "Piotrovskij, Vladimir [JanBe]" <VPIOTROV@janbe.jnj.com> Subject: RE: msg from atul Date: Wed, 15 Nov 2000 08:53:15 +0100 Atul, Concerning outliers: if your drug is metabolized and CYP2D6-dependent you might assume a subpopulation of 'poor metabolizers'. In this case, implementing a mixture of two (log-)normals for CL may help. If polymorphic metabolism in not an issue these 'outliers' ('outliers', not real outliers) may be the result of an obvious deviation of concentration distribution from normality. If you apply a so-called 'transform-both-side' approach and fit the model to the logarithm of concentrations (model prediction, F, should be transformed accordingly) you will most probably get rid of the 'outliers'. The residual error for log-transformed concentrations will be a simple additive one, and no INTERACTION will be needed, too. Best regards, Vladimir ---------------------------------------------------------------------- Vladimir Piotrovsky, Ph.D. Janssen Research Foundation Clinical Pharmacokinetics (ext. 5463) B-2340 Beerse Belgium Email: vpiotrov@janbe.jnj.com From: Mats Karlsson <Mats.Karlsson@biof.uu.se> Subject: Re: msg from atul Date: Wed, 15 Nov 2000 19:06:50 +0100 Dear Lewis, As part of Ulrika Wählby's work on actual versus nominal significance levels in NONMEM it became clear (we think) that whilst FO provides upwards biased significance levels FOCE+INTERACTION did not (roughly speaking). As you pointed out FOCE may be very time-consuming so we did investigate the GLS approach with FO as an quick alternative. Unfortunately, although it did improve things compared to FO, it was not as good as FOCE+INTER, and not good enough to say that nominal and actual significance levels agreed. However, that's not to say that for other purposes the idea may not be good (after all it did better than standard FO). Just one note, if you use IPRED as W and have sparse data, it may be almost the same as weighting against the observed concentration. Therefore I would not use it unless I had a decent number of observations per subject (which is the circumstance we studied). You may with sparse data actually want to use PRED from a previous run rather than IPRED. Best regards, Mats -- Mats Karlsson, PhD Professor of Biopharmaceutics and Pharmacokinetics Div. of Biopharmaceutics and Pharmacokinetics Dept of Pharmacy Faculty of Pharmacy Uppsala University Box 580 SE-751 23 Uppsala Sweden phone +46 18 471 4105 fax +46 18 471 4003 mats.karlsson@biof.uu.se