RE: NM7 Importance Sampling: Objective function goes to zero, but results look reasonable
Hi Bob,
NONMEM 7.2 behaves as expected in this regard!
I do have another question. I find a very large numerical difference
between the final SAEMOBJ and the IMP OBJ. For example, in SAEMOBJ
-21378, IMP OBJ=-1120. The latter seems very reasonable as it is
similar to the OBJ value of a terminated FOCE run.
Also, with the same model and data, changing the initial values results
in a rather different SAEMOBJ (-21103), but finally ending up with the
same IMP OBJ (and parameter estimates).
Is this large difference worrisome? In the few SAEM applications I have
seen in Monolix, the difference between these the SAEMOBJ and the IMP
OBJ is not that large.
Kind regards,
Filip De Ridder
Janssen Research & Development, Beerse, Belgium.
Quoted reply history
From: Bauer, Robert [mailto:[email protected]]
Sent: Monday, 10 October 2011 17:30
To: De Ridder, Filip [JRDBE]; [email protected]
Subject: RE: [NMusers] NM7 Importance Sampling: Objective function goes
to zero, but results look reasonable
Dear Filip:
You may wish to try using NONMEM 7.2, which has an improved transition
process between SAEM and IMP EONLY=1. Add the extra option MAPITER=0:
$EST METHOD=IMP LIKE NITER=20 CTYPE=3 ISAMPLE=300 NOABORT EONLY=1
MAPITER=0
Robert J. Bauer, Ph.D.
Vice President, Pharmacometrics
ICON Development Solutions
Tel: (215) 616-6428
Mob: (925) 286-0769
Email: [email protected]
Web: www.icondevsolutions.com <outbind://220/www.icondevsolutions.com>
________________________________
From: [email protected] [mailto:[email protected]]
On Behalf Of De Ridder, Filip [JRDBE]
Sent: Monday, October 10, 2011 4:41 AM
To: [email protected]
Subject: [NMusers] NM7 Importance Sampling: Objective function goes to
zero, but results look reasonable
Dear NMUsers,
When using SAEM/IMP in NM7.1., I get a peculiar results from IMP. After
a successful SAEM-run resulting in reasonable point estimates, the
objective function is first declared as 0.00000 and later as "NaN" at
the final iteration. I do get standard-errors and these look
reasonable, but I am not sure how trustworthy they are given the NaN
OBJ.
This is the $EST call:
$EST METHOD=SAEM LAPLACE LIKE NBURN=1000 NITER=2000 PRINT=10 CTYPE=3
CINTERVAL=10 NOABORT
$EST METHOD=IMP LIKE NITER=20 CTYPE=3 ISAMPLE=300 NOABORT EONLY=1
FOCE/Laplace terminates due to rounding errors, but gives reasonable
point estimates close to the SAEM result.
Below is a selection of the search.
Kind regards,
Filip De Ridder
Janssen Research & Development, Beerse, Belgium.
#METH: Stochastic Approximation Expectation-Maximization
EM/BAYES SETUP
THETAS THAT ARE MU MODELED:
1 2 3 4 5 6 7 8 9 11 12
THETAS THAT ARE SIGMA-LIKE:
MONITORING OF SEARCH:
Stochastic/Burn-in Mode
iteration -1000 SAEMOBJ= 26003.614680275896
iteration -990 SAEMOBJ= -1092.1038437297473
...
iteration -670 SAEMOBJ= -3257.6199048158469
iteration -660 SAEMOBJ= -3454.3544417057101
Convergence achieved
Reduced Stochastic/Accumulation Mode
iteration 0 SAEMOBJ= -3365.8558463970126
iteration 10 SAEMOBJ= -3739.6849187475736
...
iteration 1960 SAEMOBJ= -3823.2755123680026
iteration 1970 SAEMOBJ= -3823.2896277883360
iteration 1980 SAEMOBJ= -3823.4174894810230
iteration 1990 SAEMOBJ= -3823.3225200406459
iteration 2000 SAEMOBJ= -3823.2969976068907
Elapsed estimation time in seconds: 2862.05
#TERM:
STOCHASTIC PORTION COMPLETED
REDUCED STOCHASTIC PORTION COMPLETED
#METH: Objective Function Evaluation by Importance Sampling
EM/BAYES SETUP
THETAS THAT ARE MU MODELED:
1 2 3 4 5 6 7 8 9 11 12
THETAS THAT ARE SIGMA-LIKE:
MONITORING OF SEARCH:
iteration 0 OBJ= 0.0000000000000000
iteration 10 OBJ= NaN
Elapsed estimation time in seconds: 1472.26
iteration 20 OBJ= NaN
#TERM:
OPTIMIZATION NOT TESTED