Dear users,
I am modeling a PK data set with sparse data (2-3 samples/ind,~250
individuals). When I plotted TPST vs Conc, I do have a reasonable spread of
concentrations across the PK profile. I started modeling (2-comp better model)
with FOCEI and was getting numerical difficulties and wanted to use the EM
based methods in NM7.2 I used ITS followed by SAEM and IMP as below with an
intention to use quick ITS run for passing better (?) initial estimates to SAEM
and then IMP for standard errors and objective function evaluation:
$EST METHOD=ITS INTERACTION FILE=EXAMPLE1.EXT NITER=500 PRINT=5 NOABORT SIGL=4
CTYPE=3 CITER=10 CALPHA=0.05 NOPRIOR=1 NSIG=2
$EST METHOD=SAEM INTERACTION NBURN=5000 NITER=500 PRINT=100 SEED=1556678
ISAMPLE=2
$EST METHOD=IMP MAPITER=0 INTERACTION EONLY=1 NITER=5 ISAMPLE=3000 PRINT=1
SIGL=8 NOPRIOR=1
I got the following for each estimation method for termination message:
ITS method- OPTIMIZATION NOT COMPLETED
SAEM- Stochastic mode: Convergence Achieved
STOCHASTIC PORTION COMPLETED
REDUCED STOCHASTIC PORTION COMPLETED
IMP- OPTIMIZATION NOT COMPLETED
My parameter estimates from SAEM make good sense to the previous reports and
SEs from IMP are reasonable.
Questions:
1. How important for me to increase ITS iterations to bring optimization to be
successful for passing initial estimates to SAEM? What to do if ITS never gets
successful optimization after increasing iterations, it is likely given sparse
data?
2. Which method gives a better estimate of correlation matrix and condition
number (calculated from Eigenvalues)SAEM or IMP?
3. Are SEs valid from IMP as the termination message says "optimization not
complete". Please suggest tips for getting IMP optimization to be complete.
4. If I continue to make changes to model, can I use IMP OBJ value similar to
FOCEI OBJFUN value?
I appreciate your guidance and referring me to any literature regarding these
questions.
Regards,
Ayyappa
Regards,
Ayyappa
SAEM/IMP best practices
Hi Ayyappa;
It seems to me that using ITS, CTYPE=3 and CALPHA=0.05, you will never achieve
convergence as tested for by NONMEM, because your criteria are very strict.
But, that doesn't mean that ITS has not converged. You could plot the chain out
and see where it stabilizes, or you could use less stringent termination
criteria (e.g., CTYPE=1, CALPHA=0.01). Looking at your SAEM code, I would
increase your NITER to at least 2000-3000 if ISAMPLE=2.
Michael Fossler, Pharm.D., Ph.D., F.C.P.
Senior Director, CPMS
CPMS - US
RD Projects Clinical Platforms & Sciences
GSK
Upper Merion West
King of Prussia, PA
Email [email protected]<mailto:[email protected]>
Tel +1 610 270 4797
Cell 443-350-1194
http://www.gsk.com/ | http://twitter.com/GSK |
http://www.youtube.com/user/gskvision |
http://www.facebook.com/glaxosmithkline |
http://www.flickr.com/photos/glaxosmithkline
[cid:[email protected]]
Quoted reply history
From: [email protected] [mailto:[email protected]] On
Behalf Of Ayyappa Chaturvedula
Sent: Saturday, November 09, 2013 7:56 PM
To: [email protected]
Subject: [NMusers] SAEM/IMP best practices
Dear users,
I am modeling a PK data set with sparse data (2-3 samples/ind,~250
individuals). When I plotted TPST vs Conc, I do have a reasonable spread of
concentrations across the PK profile. I started modeling (2-comp better model)
with FOCEI and was getting numerical difficulties and wanted to use the EM
based methods in NM7.2 I used ITS followed by SAEM and IMP as below with an
intention to use quick ITS run for passing better (?) initial estimates to SAEM
and then IMP for standard errors and objective function evaluation:
$EST METHOD=ITS INTERACTION FILE=EXAMPLE1.EXT NITER=500 PRINT=5 NOABORT SIGL=4
CTYPE=3 CITER=10 CALPHA=0.05 NOPRIOR=1 NSIG=2
$EST METHOD=SAEM INTERACTION NBURN=5000 NITER=500 PRINT=100 SEED=1556678
ISAMPLE=2
$EST METHOD=IMP MAPITER=0 INTERACTION EONLY=1 NITER=5 ISAMPLE=3000 PRINT=1
SIGL=8 NOPRIOR=1
I got the following for each estimation method for termination message:
ITS method- OPTIMIZATION NOT COMPLETED
SAEM- Stochastic mode: Convergence Achieved
STOCHASTIC PORTION COMPLETED
REDUCED STOCHASTIC PORTION COMPLETED
IMP- OPTIMIZATION NOT COMPLETED
My parameter estimates from SAEM make good sense to the previous reports and
SEs from IMP are reasonable.
Questions:
1. How important for me to increase ITS iterations to bring optimization to be
successful for passing initial estimates to SAEM? What to do if ITS never gets
successful optimization after increasing iterations, it is likely given sparse
data?
2. Which method gives a better estimate of correlation matrix and condition
number (calculated from Eigenvalues)SAEM or IMP?
3. Are SEs valid from IMP as the termination message says "optimization not
complete". Please suggest tips for getting IMP optimization to be complete.
4. If I continue to make changes to model, can I use IMP OBJ value similar to
FOCEI OBJFUN value?
I appreciate your guidance and referring me to any literature regarding these
questions.
Regards,
Ayyappa
Regards,
Ayyappa
<<inline: image003.png>>