Re: OMEGA HAS A NONZERO BLOCK

From:Mats Karlsson Subject:Re: [NMusers] OMEGA HAS A NONZERO BLOCK Date:Mon, 07 Oct 2002 21:16:57 +0200 Dear all, Just two things that are on the boundary of this discussion. First, the use of OFV/MOF to discriminate between models incorporating a covariance terms or not has been suggested. We know that the likelihood ratio test doesn't work well in many situations. In the most recent issue of JPKPD, we describe simulations that indicate that covariance terms are even worse than variance terms (which are worse than fixed effects), when it comes to providing appropriate tail areas for hypothesis testing using the LR test. The inclusion of covariance terms often give large drops in OFV, even if the covariance truly is zero. Second, the solution everyone appears to gravitate towards is that additional knowledge to be gained from the "domain experts" is useful. As I see it there are two types of domain experts. Those of the particular drug in question and those who are experienced in assessing interindividual variability components in PK models in general (i.e. us). If the former can't say much about a particular parameter, I think that we could still provide a better assessment regarding a parameter than just fix it to a boundary (I agree with Ken that assuming a variance of zero in a parameter like Ka or V is as unrealistic as assuming a correlation of one). Providing some knowledge based on previously analysed drugs is considerably easier for variance components that for covariances, but why not turn to the literature or experience for guidance. This could be done very simplistically (e.g. find an estimate for a similar drug and fix it to this) or quite sophisticated (taking into account protein binding, route of elimination, include as a prior, etc), depending on the importance of the particular component for the purpose of the modelling. Mats