From: "Joern Loetsch"
Subject: [NMusers] MIXTURE modeling
Date: Thu, 17 Nov 2005 11:29:13 +0100

Dear NONMEM users,
I am working on a mixture model and would be grateful for some advice.
I have a one-compartment model with first-order absorption, and get two
subpopulations with different CL, the same Ka (factor almost equal to one
between the Ka's of the two subpopulations, and A value of V that varies in
one subpopulation but is almost the same for all subjects in the second
subpopulation (very low ETA in that subpopulation). 
1. Can I join the Ka's to have only one THETA(Ka) for both subpopulation? 
2. How do I identify co-variates? Same procedure as without MIX, only
separately for the two subpopulations? 
3. The objective function went down by 55.14 from the model without
covariates and without MIX, but the subpopulations do not appear meaningful.
This applies also to the quite high (as compared to the estimate without
MIX) volume of distribution that does not vary interindividually in one
subpopulation. How do I deal with this result.  
Thank you in advance for your advice.
J. Lötsch

Prof. Dr. med. Jörn Lötsch
pharmazentrum frankfurt/ZAFES
Institut für Klinische Pharmakologie 
Johann Wolfgang Goethe-Universität 
Theodor-Stern-Kai 7 
D-60590 Frankfurt am Main


From: "Gobburu, Jogarao V"
Subject: RE: [NMusers] MIXTURE modeling
Date: Thu, 17 Nov 2005 06:17:25 -0500

Hello, please see my responses below:
0. First of all, I am not sure what prompted you to apply a mixture model
and how rich the data are.
1. You do not have to estimate different modes for the two populations for
all the parameters. You should limit it to the ones that you think are
reasonable, ie., you need some prior expectation (eg: poor vs fast
metabolizers) or graphical evidence.
2. Covariate selection should be similar with or without MIX. However, it
would be important to explore if one or several of the covariates can
explain the mixture.
3. Bruce Green and Nick Holford presented some work on using log-likelihood
ratios to select mixture models at the Annual AAPS meeting last week
(asymptotic vs. empirical via re-randomization tests). They show that the
estimation method (FO vs FOCE) makes a big difference. They may have more to
add on this. Regarding volume for one of the sub-pops, do you have adequate
data to estimate this parameter?  



From: "Bruce Green"
Subject: RE: [NMusers] MIXTURE modeling
Date: Fri, 18 Nov 2005 09:52:22 +1000

Hi Joern,

Mixture models can be useful if you can clearly identify two subpopulations
by just looking at the data. However, relying on the dOBJ as a diagnostic
for the presence of a mixture is probably not a good idea, especially if you
are using the FO method. We identified that the delta OBJ might be really
large (in the magnitude of thousands) to reject the null. Which method in
NONMEM are you using, what sort of error structure do you have?