RE : Simulating different populations

From: "GIRARD PASCAL" PASCAL.GIRARD@adm.univ-lyon1.fr Subject: RE : [NMusers] Simulating different populations Date: Thu, 11 Aug 2005 12:43:13 +0200 Hi Marc and Max, My 20 cents... In order to do what Marc suggests, you make the assumptions: (1) that NONMEM fixed effects posterior distribution is multivariate normal (MVN) (2) parameterized with mean equal to the THETA estimates and variance to $COV var-cov matrix from which you exclude random effects (co)variance lines and columns, and (3) that you have identified all covariates and/or multimodal distributions using mixture models. Those assumptions are probably not true for many parameters. For example clearance is more likely to be log-normally distributed. F bioavailability is classically bounded by 0 and 1. So in order to use your MVN assumption, you will need first to re-parameterize your model with exponential, logistic functions, ... (for more details, see http://www.page-meeting.org/page/page2002/PascalGirardPage2002.pdf ) Assuming that you have a function or subroutine to resample from MVN (R and Splus do it very easily), second difficulty is to pass var-cov matrix from NONMEM output to your MVN resampling function. Depending on the complexity of your model, this matrix can be very large and keying all those numbers a real pain in ... and source of errors! So for doing this, I have developed an INFN subroutine (to use with PREDPP or an equivalent for $PRED) that outputs THETAs and var-cov matrix in an ASCII file. Probably, "Marc and others" have developed similar tools. Anyway, I am still dreaming of an integrated tool doing all this without extra-coding... Hope it helps, Dr Pascal Girard EA 3738, Ciblage Thrapeutique en Oncologie Fac Mdecine Lyon-Sud, BP12 69921 OULLINS Cedex France Tel +33 (0)4 26 23 59 54 / Fax +33 (0)4 26 23 59 49