Re: FW: OMEGA HAS A NONZERO BLOCK

From:Nick Holford Subject:Re: FW: [NMusers] OMEGA HAS A NONZERO BLOCK Date:Fri, 04 Oct 2002 16:25:10 +1200 Ken, Thanks for bringing together these comments. 1. I don't agree that the minimum objective function value (MOF) is the ultimate criterion when it means the model parameters are intrinsically unreasonable. NONMEM can have problems estimating covariances. I do not recall ever seeing NONMEM estimate a covariance of zero but have frequently seen correlations close to 1 (or -1). I repeat my assertion that a correlation of 1 (or -1) is never reasonable because it means that one parameter differs only by a scale factor from another and thus only one parameter needs to be estimated. If the deterministic structure of the model is compatible with one parameter then change this part of the model rather than introducing it via a trick via the random effects part of the model. 2. When NONMEM is unable to estimate a reasonable value then I suggest that a reasonable heuristic is to fix the value e.g. a correlation of 0.5 between CL and V, rather than assume correlation is zero. The method proposed by Mats can be used to do this. I was not suggesting that Steve try this reparameterization and still try to estimate the THETA defining the covariance. I agree that this is unlikely to solve the problem although it should be noted that some reparameterizations can have unexpected benefits for estimation. 3. Bayesian estimation of the covariance can be approximated using NONMEM with the undocumented, unsupported and widely discussed PRIOR method in NONMEM V 1.1. This method would work quite nicely with the way Mats has suggested using a THETA to represent the covariance. I agree that a strong prior would probably be needed to avoid the difficulties NONMEM has in estimating covariances when the information in the data is low. The difference between a fixed parameter and one estimated with a strong prior is not likely to be of any practical importance so I would pragmatically tend to use the fixed parameter method. As Steve is such a keen WinBugger I wonder if he can tell us how his problem behaves with a truely Bayesian estimation method? 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.health.auckland.ac.nz/pharmacology/staff/nholford/