From: musor000@optonline.net
Subject: [NMusers] bias and good fit vs centering
Date: Thu, 18 Aug 2005 20:38:13 -0400
Hello NONMEM Users,
We are trying to fit a model which combines 2 studies (IV injection - 1 compartment
and SC injection - 2 compartments). We use METHOD=1, POSTHOC, LAPLACIAN). Plots
all subjects look very good, but there is a problem. Populations estimates are quite
different from the means of individual estimates. For example, population estimate
of absorbtion constant (Ka) is 5 times as big as mean of Ka for individuals. It is
larger than the largest individual Ka!!! I tried CENTERING option. In this case
there is no bias, but fit is much worse (objective function) and plots do not look good.
We do not have any confidence in the results. What do we have to use? Population
estimates or means of individual estimates? How do experienced and sophisticated
NONMEM users deal with issue like that?
Thank you,
Pavel Kovalenko
bias and good fit vs centering
From: "Wang, Yaning" WangYA@cder.fda.gov
Subject: RE: [NMusers] bias and good fit vs centering
Date: Fri, 19 Aug 2005 09:55:43 -0400
Pavel:
If your residual error model is not additive (Y=F+ERR(1)), I would rather
use FOCE with interaction (METHOD=1 INT) instead of LAPLACIAN. In the
current version of NONMEM (version V), LAPLACIAN ignore the interaction
between ETA and EPSILON when it does exit. If you have a model like
Y=F(eta)+F(eta)*ERR(1), LAPLACIAN will first simplify the model to
Y=F(eta)+F(0)*ERR(1) and then move on. In my opinion, for a linear mixed
model with a proportional error model, LAPLACIAN is equal to FO.
Despite the better approximation of LAPLACIAN method for the marginal
likelihood in general, lack of interaction option for LAPLACIAN may
ultimately make LAPLACIAN worse than FOCE with interaction when the residual
error model includes an interaction.
Yaning Wang, Ph.D.
Pharmacometrician
Office of Clinical Pharmacology and Biopharmaceutics
Center of Drug Research and Evaluation
Food and Drug Administration
Office: 301-827-9763
From: "Mouksassi Mohamad-Samer" mohamad-samer.mouksassi@umontreal.ca
Subject: [NMusers] bias and goodfit vs centering
Date: Fri, 19 Aug 2005 11:01:15 -0400
Pavel,
Dr Stuart Beal described a way to use the Laplacian aproximation with
interaction in NONMEM V. This can be found at http://www.cognigencorp.com/nonmem/nm/99jul162002.html.
I have never tried it so far but maybe some other NONMEM users may have some
experience in its application and advantages.
MOUKSASSI Mohamad-Samer Pharm.D.
tudiant Ph.D. Sciences Parmaceutiques
Universit de Montral
Facult de Pharmacie
Pavillon Jean-Coutu
Tel : 343-6111 ext 0388
From: "Wang, Yaning" WangYA@cder.fda.gov
Subject: RE: [NMusers] bias and good fit vs centering
Date: Fri, 19 Aug 2005 12:14:38 -0400
Pavel:
If "It is larger than the largest individual Ka!!!", that means your
histogram of ETA for Ka will be totally on the left side of zero. This is a
clear violation of the between-subject variability assumption. Did you try
to fit the SC data alone?
Yaning
From: musor000@optonline.net [mailto:musor000@optonline.net]
Subject: Re: RE: [NMusers] bias and good fit vs centering
Sent: Friday, August 19, 2005 8:57 PM
Hello Yaning,
You are right. I had to use the interaction option. Also, I used the same
"constant CV" error for both studies and different additive erors. It fixed
the problem.
Thank you,
Pavel
From: Tom
Subject: RE: [NMusers] bias and good fit vs centering
Date: Monday, August 22, 2005 2:29 pm
Hi,Pavel:
I saw your question on NONMEM user list. Did Dr. Yaning Wang's
suggestion (FOCE interaction) fix your problem?
I would really appreciate it if you can share the results.
Thanks
Tom
From: musor000@optonline.net [mailto:musor000@optonline.net]
Subject: Re: RE: [NMusers] bias and good fit vs centering
Sent: Tuesday, August 23, 2005 9:06 PM
Hello Tom,
Yes, it fixed my problem. I used the interaction option and FOCE. Also, I
used the same "constant CV" error for both studies but different additive
erors, which helped a lot. It did not provide a better fit for individual
subjects (the plots looked very similar to the Laplacial plots), but
population parameters looked much better. They were centered.
Kind regards,
Pavel
_______________________________________________________