Dear all,
I have a question regarding the setting of ISAMPLE under SAEM and IMP
method in Nonmem 7.2.
I have noticed that in the NONMEM manual, the default setting of ISAMPLE
for IMP is 300, and 2 for SAEM. Well it was well explained about the
general guidance of ISAMPLE for IMP, i.e. might need to float to 1000-3000
in case of sparse samples. However, the setting for ISAMPLE under SAEM is
not explained that much.
Additionally, in the examples provided in the manual most of ISAMPLE
setting is in its hundreds or more for IMP, this option under SAEM is
mostly 2. I understand that with a lower value the computation is fast, but
why is there's such a big difference in terms of setting sample number per
subject?
The reason I am asking this is because of a diffitulty to achieve
convergence in the burn-in step using below estimation method:
*$EST MET=SAEM INTER NBURN=2000 NITER=1000 PRINT=10 NOABORT NOPRIOR=0
ISAMPLE=2 SIGL=6 CTYPE=0 SEED=150159 FILE=Saem.ext $EST MET=IMP INTER
NITER=2000 ISAMPLE=1000 PRINT=10 SEED=150987 SIGL=6, CTYPE=3 FILE=Imp.ext *
Thanks very much for your explanation.
Best Regards
--
Xinting
ISAMPLE under SAEM and IMP
Dear Xinting,
While I’m no expert in depth I will try to explain my understanding. ISAMPLE
means different things in IMP and in SAEM. In IMP it is the number of random
samples per individual, that are used to adjust the area of the sampling
density. This has a much smaller impact than raising the ISAMPLE in SAEM.
In SAEM ISAMPLE is instead the number of parameter vectors (Phis) that are
generated in mode 1 of MCMC expectation part of SAEM. Each Phi is generated for
an individual from the proposal density. The proposal and backwards densities
are evaluated for each Phi and the joint density is evaluated for the current
and the proposed Phi. A test statistic is generated and if the test agrees the
new vector of parameters is accepted.
I think you can simplify it to: A higher ISAMPLE increases the number of
parameter sets evaluated per individual in each step.
You can control the aspects of mode 1, mode 1A and so on more precisely using
the ISAMPLE_M1 etcetera options. All in all it is hard to find a set of options
that suits your specific problem, but perhaps someone on the list can help you
if you can share some more characteristics of your problem. If you have access
to NM73 I can recommend the AUTO=1 setting which has performed well for me. For
further reading I think the Technical Guide is what you want to look at.
I hope that helps.
Best regards
Henrik B. Nyberg
Quoted reply history
From: [email protected] [mailto:[email protected]] On
Behalf Of Xinting Wang
Sent: 08 April 2015 03:42
To: [email protected]
Subject: [NMusers] ISAMPLE under SAEM and IMP
Dear all,
I have a question regarding the setting of ISAMPLE under SAEM and IMP method in
Nonmem 7.2.
I have noticed that in the NONMEM manual, the default setting of ISAMPLE for
IMP is 300, and 2 for SAEM. Well it was well explained about the general
guidance of ISAMPLE for IMP, i.e. might need to float to 1000-3000 in case of
sparse samples. However, the setting for ISAMPLE under SAEM is not explained
that much.
Additionally, in the examples provided in the manual most of ISAMPLE setting is
in its hundreds or more for IMP, this option under SAEM is mostly 2. I
understand that with a lower value the computation is fast, but why is there's
such a big difference in terms of setting sample number per subject?
The reason I am asking this is because of a diffitulty to achieve convergence
in the burn-in step using below estimation method:
$EST MET=SAEM INTER NBURN=2000 NITER=1000 PRINT=10 NOABORT NOPRIOR=0 ISAMPLE=2
SIGL=6 CTYPE=0
SEED=150159 FILE=Saem.ext
$EST MET=IMP INTER NITER=2000 ISAMPLE=1000 PRINT=10 SEED=150987 SIGL=6, CTYPE=3
FILE=Imp.ext
Thanks very much for your explanation.
Best Regards
--
Xinting
--
LEGAL NOTICE
This message is intended for the use of the named recipient(s) only and may
contain
confidential and / or privileged information. If you are not the intended
recipient, please
contact the sender and delete this message. Any unauthorised use of the
information
contained in this message is prohibited.
Mango Business Solutions Limited is registered in England under No. 4560258
with its
registered office at Suite 3, Middlesex House, Rutherford Close, Stevenage,
Herts, SG1 2EF,
UK.
PLEASE CONSIDER THE ENVIRONMENT BEFORE PRINTING THIS EMAIL
Xinting Wang:
Generating each sample in SAEM is computationally more expensive in SAEM, and
so ISAMPLE of 2-10 is generally done. You may try increasing ISAMPLE=3 or 5 in
SAEM. Furthermore, convergence testing is only a guide, Review the .ext file
and be satisfied by visual inspection that the parameters are not directionally
changing, and are just randomly fluctuating.
Robert J. Bauer, Ph.D.
Vice President, Pharmacometrics R&D
ICON Early Phase
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com
Quoted reply history
From: [email protected] [mailto:[email protected]] On
Behalf Of Xinting Wang
Sent: Tuesday, April 07, 2015 7:42 PM
To: [email protected]
Subject: [NMusers] ISAMPLE under SAEM and IMP
Dear all,
I have a question regarding the setting of ISAMPLE under SAEM and IMP method in
Nonmem 7.2.
I have noticed that in the NONMEM manual, the default setting of ISAMPLE for
IMP is 300, and 2 for SAEM. Well it was well explained about the general
guidance of ISAMPLE for IMP, i.e. might need to float to 1000-3000 in case of
sparse samples. However, the setting for ISAMPLE under SAEM is not explained
that much.
Additionally, in the examples provided in the manual most of ISAMPLE setting is
in its hundreds or more for IMP, this option under SAEM is mostly 2. I
understand that with a lower value the computation is fast, but why is there's
such a big difference in terms of setting sample number per subject?
The reason I am asking this is because of a diffitulty to achieve convergence
in the burn-in step using below estimation method:
$EST MET=SAEM INTER NBURN=2000 NITER=1000 PRINT=10 NOABORT NOPRIOR=0 ISAMPLE=2
SIGL=6 CTYPE=0
SEED=150159 FILE=Saem.ext
$EST MET=IMP INTER NITER=2000 ISAMPLE=1000 PRINT=10 SEED=150987 SIGL=6, CTYPE=3
FILE=Imp.ext
Thanks very much for your explanation.
Best Regards
--
Xinting
Dear all,
Thanks very much for your explanation.
I tried to manipulate different ISAMPLE options after your suggestion, and
found that raising ISAMPLE from 2 to 5 would not produce a convergence
[CTYPE=3], but if the number is 300, then a convergence is rapidly
achieved. Additionally, this does not have an untolerable prolongation of
computation time. While I understand that this is only a guidance, and
judgement should be made based on the outcome from .ext file produced, it
seems to me that raising the number could actually help with convergence.
A follow-up question is, does the convergence criteria in IMP meaniningful,
or similar to SAEM, a reference to .ext file is also the preferred method
to evaluate if the method is appropriate? Many thanks!
Best Regards
Xinting
Quoted reply history
On 9 April 2015 at 01:02, Bauer, Robert <[email protected]> wrote:
> Xinting Wang:
>
> Generating each sample in SAEM is computationally more expensive in SAEM,
> and so ISAMPLE of 2-10 is generally done. You may try increasing ISAMPLE=3
> or 5 in SAEM. Furthermore, convergence testing is only a guide, Review
> the .ext file and be satisfied by visual inspection that the parameters are
> not directionally changing, and are just randomly fluctuating.
>
>
>
> Robert J. Bauer, Ph.D.
>
> Vice President, Pharmacometrics R&D
>
> ICON Early Phase
>
> Office: (215) 616-6428
>
> Mobile: (925) 286-0769
>
> [email protected]
>
> www.iconplc.com
>
>
>
> *From:* [email protected] [mailto:[email protected]]
> *On Behalf Of *Xinting Wang
> *Sent:* Tuesday, April 07, 2015 7:42 PM
> *To:* [email protected]
> *Subject:* [NMusers] ISAMPLE under SAEM and IMP
>
>
>
> Dear all,
>
> I have a question regarding the setting of ISAMPLE under SAEM and IMP
> method in Nonmem 7.2.
>
> I have noticed that in the NONMEM manual, the default setting of ISAMPLE
> for IMP is 300, and 2 for SAEM. Well it was well explained about the
> general guidance of ISAMPLE for IMP, i.e. might need to float to 1000-3000
> in case of sparse samples. However, the setting for ISAMPLE under SAEM is
> not explained that much.
>
> Additionally, in the examples provided in the manual most of ISAMPLE
> setting is in its hundreds or more for IMP, this option under SAEM is
> mostly 2. I understand that with a lower value the computation is fast, but
> why is there's such a big difference in terms of setting sample number per
> subject?
>
> The reason I am asking this is because of a diffitulty to achieve
> convergence in the burn-in step using below estimation method:
>
>
>
>
> *$EST MET=SAEM INTER NBURN=2000 NITER=1000 PRINT=10 NOABORT NOPRIOR=0
> ISAMPLE=2 SIGL=6 CTYPE=0 SEED=150159 FILE=Saem.ext $EST MET=IMP INTER
> NITER=2000 ISAMPLE=1000 PRINT=10 SEED=150987 SIGL=6, CTYPE=3 FILE=Imp.ext *
>
> Thanks very much for your explanation.
>
> Best Regards
>
>
> --
>
> Xinting
>
>
>
Dear Xinting:
Well, in NONMEM 7.3, because ISAMPLE is used for both SAEM and IMP, it is
filtered such that if the user sets ISAMPLE for SAEM higher than ISAMPLEMAX, it
is not changed from the default of 2 for SAEM. Your NONMEM report file will
tell you the actual ISAMPLE used for the SAEM step. Please refer to that. Of
course, $SIZES ISAMPLEMAX=??? will allow you to increase ISAMPLEMAX, should you
actually needed higher than 10. Typically, however, ISAMPLE should not be
greater than 10 for SAEM.
Your convergence success may be serendipitous, as it sometimes is with Monte
Carlo matters.
Meanwhile, I am afraid I did not understand your last sentence.
Robert J. Bauer, Ph.D.
Vice President, Pharmacometrics R&D
ICON Early Phase
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com
Quoted reply history
From: Xinting Wang [mailto:[email protected]]
Sent: Sunday, April 12, 2015 9:57 PM
To: Bauer, Robert; Leonid Gibiansky; [email protected]
Cc: [email protected]
Subject: Re: [NMusers] ISAMPLE under SAEM and IMP
Dear all,
Thanks very much for your explanation.
I tried to manipulate different ISAMPLE options after your suggestion, and
found that raising ISAMPLE from 2 to 5 would not produce a convergence
[CTYPE=3], but if the number is 300, then a convergence is rapidly achieved.
Additionally, this does not have an untolerable prolongation of computation
time. While I understand that this is only a guidance, and judgement should be
made based on the outcome from .ext file produced, it seems to me that raising
the number could actually help with convergence.
A follow-up question is, does the convergence criteria in IMP meaniningful, or
similar to SAEM, a reference to .ext file is also the preferred method to
evaluate if the method is appropriate? Many thanks!
Best Regards
Xinting
On 9 April 2015 at 01:02, Bauer, Robert
<[email protected]<mailto:[email protected]>> wrote:
Xinting Wang:
Generating each sample in SAEM is computationally more expensive in SAEM, and
so ISAMPLE of 2-10 is generally done. You may try increasing ISAMPLE=3 or 5 in
SAEM. Furthermore, convergence testing is only a guide, Review the .ext file
and be satisfied by visual inspection that the parameters are not directionally
changing, and are just randomly fluctuating.
Robert J. Bauer, Ph.D.
Vice President, Pharmacometrics R&D
ICON Early Phase
Office: (215) 616-6428<tel:%28215%29%20616-6428>
Mobile: (925) 286-0769<tel:%28925%29%20286-0769>
[email protected]<mailto:[email protected]>
http://www.iconplc.com
From: [email protected]<mailto:[email protected]>
[mailto:[email protected]<mailto:[email protected]>] On
Behalf Of Xinting Wang
Sent: Tuesday, April 07, 2015 7:42 PM
To: [email protected]<mailto:[email protected]>
Subject: [NMusers] ISAMPLE under SAEM and IMP
Dear all,
I have a question regarding the setting of ISAMPLE under SAEM and IMP method in
Nonmem 7.2.
I have noticed that in the NONMEM manual, the default setting of ISAMPLE for
IMP is 300, and 2 for SAEM. Well it was well explained about the general
guidance of ISAMPLE for IMP, i.e. might need to float to 1000-3000 in case of
sparse samples. However, the setting for ISAMPLE under SAEM is not explained
that much.
Additionally, in the examples provided in the manual most of ISAMPLE setting is
in its hundreds or more for IMP, this option under SAEM is mostly 2. I
understand that with a lower value the computation is fast, but why is there's
such a big difference in terms of setting sample number per subject?
The reason I am asking this is because of a diffitulty to achieve convergence
in the burn-in step using below estimation method:
$EST MET=SAEM INTER NBURN=2000 NITER=1000 PRINT=10 NOABORT NOPRIOR=0 ISAMPLE=2
SIGL=6 CTYPE=0
SEED=150159 FILE=Saem.ext
$EST MET=IMP INTER NITER=2000 ISAMPLE=1000 PRINT=10 SEED=150987 SIGL=6, CTYPE=3
FILE=Imp.ext
Thanks very much for your explanation.
Best Regards
--
Xinting