From: Yaning Wang 
Subject:[NMusers] fundamental questions about NONMEM
Date: Mon, 10 Mar 2003 20:07:20 -0500

Dear NONMEM users:
In order to compare PKBUGs/WinBUGs with NONMEM, I read NONMEM Users
guide part VII to have a better understanding about NONMEM. Here are
some questions I don?t understand after I read the manual.

1. On page 4, it seems to me that the approximation equation for
-2*log(Li) was obtained via first-order Laplacian approximation without
taking etahat as the posterior mode of eta, otherwise there should be
only the first four terms. Could someone recommend me a reference where
I can learn how to derive this version of Laplacian approximation at any
estimate of eta?

2. The last term in the approximation equation mentioned above should be
zero for the conditional estimation methods, such as Laplacian method
and FOCE method, but not zero for non-conditional estimation method (FO
method) and constrained conditional estimation method (Hybrid method),

3. In FOCE method (A.2.), the replacement of the hessian matrix with a
function of the gradient matrix (1/2*Gamma?*Gamma) was stated to be ?a
type of first-order approximation?, which is where and how ?first-order
approximation? happened, a second level of approximation on top of
Laplacian approximation (the first level of approximation) during
parameter estimation. Could someone recommend me a reference on this
type of first-order approximation? What is the relationship between this
type of first-order approximation and the first-order model
approximation (first-order Taylor expansion) introduced in A.6. and
NONMEM Users guide I?

4. In FO method (A.3.), the third level of approximation, all eta?s
being zero, was applied during parameter estimation in addition to the
two levels of approximation mentioned above.  This is a new definition
of FO method, compared with that described in NONMEM Users guide I. And
under A.6.(the centering FOCE method with the first-order model), it was
stated that the new FO method is equivalent to applying non-centering
FOCE to the first-order model. Based on this statement and the rationale
behind the hierarchical methods, Laplacian>FOCE>Hybrid>FO, can I say
that the centering FOCE method with the first-order model is somewhere
between FOCE and FO? Has anyone applied this method?

5. I thought all the error terms in NONMEM were modeled by normal
distribution until I read this manual. It is emphasized on page 1 that
normal assumption is only made to keep the discussion simple. If I want
to model the oral biavailability (F) with a beta(a, b) distribution, how
do I give the initial estimates for a and b and specify this
distribution in NONMEM?

Thanks a lot for any input.

Yaning Wang, Graduate student
Department of Pharmaceutics
College of Pharmacy
University of Florida


From:Yaning Wang 
Subject:[NMusers] fundamental questions (ignore 1 and 2) about NONMEM
Date:Wed, 12 Mar 2003 10:04:33 -0500

Ignore question 1 and 2. I figured them out. Any suggestion on 3, 4 and 5?