Happy New Year,
I hope everybody is ready for a great 2025 !
I'll start my message/question by defining 2 different ways of coding a simple
power relationship between body weigh on clearance.
*
Coding 1
MU_1 = THETA(1) + THETA(2) * LOG(WGT/70)
CL = EXP( MU_1 + ETA(1) )
*
Coding 2
MU_1 = THETA(1)
CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
The reference and training materials for NONMEM clearly indicate that MU
variables should be time invariant within occasions and recommend using coding
2 when body weight is time varying. Nevertheless, it is possible for an analyst
to use coding 1. As far as I can tell from some limited testing, this is not a
"fatal" error. Either with FOCE(I) or SAEM/IMP, NONMEM reports a warning but
performs the model optimization. The table outputs also report CL as a time
varying variable changing as body weight changes.
So my questions are the following: when coding 1 is used and body weight is
time varying, what is NONMEM actually doing during model optimization? Does
NONMEM internally create occasions to break the records by interval of constant
body weight and constant MU1? Alternatively, does NONMEM internally calculate
an average of MU1? Something entirely different? What's the risk taken by an
analyst when using coding 1 versus coding 2?
Thank you in advance for you input
Sébastien Bihorel
Director, Quantitative Pharmacology
+1 914-648-9581
[email protected]
Regeneron - Internal
********************************************************************
This e-mail and any attachment hereto, is intended only for use by the
addressee(s) named above and may contain legally privileged and/or confidential
information. If you are not the intended recipient of this e-mail, any
dissemination, distribution or copying of this email, or any attachment hereto,
is strictly prohibited. If you receive this email in error please immediately
notify me by return electronic mail and permanently delete this email and any
attachment hereto, any copy of this e-mail and of any such attachment, and any
printout thereof. Finally, please note that only authorized representatives of
Regeneron Pharmaceuticals, Inc. have the power and authority to enter into
business dealings with any third party.
********************************************************************
MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any differences in the fit, parameter estimates, precision, convergence speed (number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does not have this restriction of time-independence even if you use Mu referencing, so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is that some kind of averaging is used to get time independent value of WT that is then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the differences. A more stringent test would be to use time-dependent and strongly influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
Quoted reply history
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70)
> CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that MU variables should be time invariant within occasions and recommend using coding 2 when body weight is time varying. Nevertheless, it is possible for an analyst to use coding 1. As far as I can tell from some limited testing, this is not a "fatal" error. Either with FOCE(I) or SAEM/IMP, NONMEM reports a warning but performs the model optimization. The table outputs also report CL as a time varying variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body weight is time varying, what is NONMEM actually doing during model optimization? Does NONMEM internally create occasions to break the records by interval of constant body weight and constant MU1? Alternatively, does NONMEM internally calculate an average of MU1? Something entirely different? What's the risk taken by an analyst when using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]
>
> Regeneron - Internal
>
> ********************************************************************
>
> This e-mail and any attachment hereto, is intended only for use by the addressee(s) named above and may contain legally privileged and/or confidential information. If you are not the intended recipient of this e-mail, any dissemination, distribution or copying of this email, or any attachment hereto, is strictly prohibited. If you receive this email in error please immediately notify me by return electronic mail and permanently delete this email and any attachment hereto, any copy of this e-mail and of any such attachment, and any printout thereof. Finally, please note that only authorized representatives of Regeneron Pharmaceuticals, Inc. have the power and authority to enter into business dealings with any third party.
>
> ********************************************************************
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]
www.iconplc.com
Quoted reply history
-----Original Message-----
From: [email protected] <[email protected]> On Behalf Of
Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel <[email protected]>; [email protected]
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]
>
>
> Regeneron - Internal
>
> ********************************************************************
> This e-mail and any attachment hereto, is intended only for use by the
> addressee(s) named above and may contain legally privileged and/or
> confidential information. If you are not the intended recipient of
> this e-mail, any dissemination, distribution or copying of this email,
> or any attachment hereto, is strictly prohibited. If you receive this
> email in error please immediately notify me by return electronic mail
> and permanently delete this email and any attachment hereto, any copy
> of this e-mail and of any such attachment, and any printout thereof.
> Finally, please note that only authorized representatives of Regeneron
> Pharmaceuticals, Inc. have the power and authority to enter into
> business dealings with any third party.
> ********************************************************************
Thank you for the clarification. Much appreciated.
Regeneron - Internal
Quoted reply history
________________________________
From: Bauer, Robert <[email protected]>
Sent: Saturday, January 11, 2025 2:31 AM
To: 'Leonid Gibiansky' <[email protected]>; Sébastien Bihorel
<[email protected]>; [email protected]
<[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that
subject. Robert J. Bauer, Ph. D. Senior Director Pharmacometrics R&D ICON Early
Phase 731
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
https://urldefense.com/v3/__ http://www.iconplc.com__;!!ODpDvJZr5w!GNRSQC0vmVcClv1zMJgsNsCcNeoMkxqOaEMlRn6_-jHqy6jald8QxKYV1I6AUPX3dritCEex3_fXtP0HHRqN6YOFPjqy8iI$
-----Original Message-----
From: [email protected]<mailto:[email protected]>
<[email protected]> On Behalf Of Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel <[email protected]>;
[email protected]<mailto:[email protected]>
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]<mailto:[email protected]>
>
>
> Regeneron - Internal
>
> ********************************************************************
> This e-mail and any attachment hereto, is intended only for use by the
> addressee(s) named above and may contain legally privileged and/or
> confidential information. If you are not the intended recipient of
> this e-mail, any dissemination, distribution or copying of this email,
> or any attachment hereto, is strictly prohibited. If you receive this
> email in error please immediately notify me by return electronic mail
> and permanently delete this email and any attachment hereto, any copy
> of this e-mail and of any such attachment, and any printout thereof.
> Finally, please note that only authorized representatives of Regeneron
> Pharmaceuticals, Inc. have the power and authority to enter into
> business dealings with any third party.
> ********************************************************************
Is this also true for the FOCEI when MU-referencing is present, or only for SAEM/IMP/etc "new" methods?
Thank you
Leonid
Quoted reply history
On 1/11/2025 2:31 AM, Bauer, Robert wrote:
> If a covariate varies across records within a subject, NONMEM obtains a simple average among the records and uses this as the covariate value for that subject.
>
> Robert J. Bauer, Ph.D.
> Senior Director
> Pharmacometrics R&D
> ICON Early Phase
> 731 Arbor way, suite 100
> Blue Bell, PA 19422
> Office: (215) 616-6428
> Mobile: (925) 286-0769
> [email protected] <mailto:[email protected]>
> www.iconplc.com http://www.iconplc.com
>
> -----Original Message-----
>
> From: [email protected] < mailto: [email protected] > < [email protected] > On Behalf Of Leonid Gibiansky
>
> Sent: Friday, January 10, 2025 5:21 PM
>
> To: Sébastien Bihorel < [email protected] >; [email protected] < mailto: [email protected] >
>
> Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
>
> Hi Sébastien,
>
> As you did these experiments, can you share the results: have you seen any differences in the fit, parameter estimates, precision, convergence speed (number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does not have this restriction of time-independence even if you use Mu referencing, so results should be identical or very close).
>
> As code is encrypted, only Bob can answer the question but my understanding is that some kind of averaging is used to get time independent value of WT that is then used by the SAEM/IMP algorithm for parameter update procedure.
>
> As WT changes slowly and not very significantly, it could be hard to see the differences. A more stringent test would be to use time-dependent and strongly influential ADA (0/1): how bad is the incorrect version 1 in this case?
>
> Thank you
> Leonid
>
> On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
> >
> > Happy New Year,
> >
> > I hope everybody is ready for a great 2025 !
> >
> > I'll start my message/question by defining 2 different ways of coding
> > a simple power relationship between body weigh on clearance.
> >
> > *
> > Coding 1
> >
> > MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
> >
> > *
> > Coding 2
> >
> > MU_1 = THETA(1)
> > CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
> >
> > The reference and training materials for NONMEM clearly indicate that
> > MU variables should be time invariant within occasions and recommend
> > using coding 2 when body weight is time varying. Nevertheless, it is
> > possible for an analyst to use coding 1. As far as I can tell from
> > some limited testing, this is not a "fatal" error. Either with FOCE(I)
> > or SAEM/IMP, NONMEM reports a warning but performs the model
> > optimization. The table outputs also report CL as a time varying
> > variable changing as body weight changes.
> >
> > So my questions are the following: when coding 1 is used and body
> > weight is time varying, what is NONMEM actually doing during model
> > optimization? Does NONMEM internally create occasions to break the
> > records by interval of constant body weight and constant MU1?
> > Alternatively, does NONMEM internally calculate an average of MU1?
> > Something entirely different? What's the risk taken by an analyst when
> > using coding 1 versus coding 2?
> >
> > Thank you in advance for you input
> >
> >
> > __
> > Sébastien Bihorel
> > Director, Quantitative Pharmacology
> > +1 914-648-9581
> > [email protected] <mailto:[email protected]>
> >
> >
> > Regeneron - Internal
> >
> > ********************************************************************
> > This e-mail and any attachment hereto, is intended only for use by the
> > addressee(s) named above and may contain legally privileged and/or
> > confidential information. If you are not the intended recipient of
> > this e-mail, any dissemination, distribution or copying of this email,
> > or any attachment hereto, is strictly prohibited. If you receive this
> > email in error please immediately notify me by return electronic mail
> > and permanently delete this email and any attachment hereto, any copy
> > of this e-mail and of any such attachment, and any printout thereof.
> > Finally, please note that only authorized representatives of Regeneron
> > Pharmaceuticals, Inc. have the power and authority to enter into
> > business dealings with any third party.
> > ********************************************************************
>
>
Hi Bob,
I am really puzzled by this statement. I would expect NONMEM to recognize time
varying covariates provide information about the fixed effects and used the
time specific value of the covariate to make a prediction.
Averaging the covariate across all the records for a subject seems like a poor
use of information.
Is your statement saying something special associated with the mu-referenced
transformation? If so would you please clarify your statement about averaging?
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: [email protected]<mailto:[email protected]>
web: http://holford.fmhs.auckland.ac.nz/
Quoted reply history
From: [email protected] <[email protected]> On Behalf Of
Bauer, Robert
Sent: Saturday, 11 January 2025 8:32 pm
To: 'Leonid Gibiansky' <[email protected]>; Sébastien Bihorel
<[email protected]>; [email protected]
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com
-----Original Message-----
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]<mailto:[email protected]>
>
>
> Regeneron - Internal
>
> ********************************************************************
> This e-mail and any attachment hereto, is intended only for use by the
> addressee(s) named above and may contain legally privileged and/or
> confidential information. If you are not the intended recipient of
> this e-mail, any dissemination, distribution or copying of this email,
> or any attachment hereto, is strictly prohibited. If you receive this
> email in error please immediately notify me by return electronic mail
> and permanently delete this email and any attachment hereto, any copy
> of this e-mail and of any such attachment, and any printout thereof.
> Finally, please note that only authorized representatives of Regeneron
> Pharmaceuticals, Inc. have the power and authority to enter into
> business dealings with any third party.
> ********************************************************************
Hello Nick:
The statement I made pertains only to EM algorithms (ITS, SAEM, IMP), and
BAYES. The classic methods (FOCEI, Laplace), do not engage in averaging the
covariates across records even when thetas are MU referenced, as the classic
algorithms do not use EM update methods to advance the theta estimates.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com/
Quoted reply history
From: Nick Holford <[email protected]>
Sent: Saturday, January 11, 2025 9:29 PM
To: Bauer, Robert <[email protected]>; 'Leonid Gibiansky'
<[email protected]>; Sébastien Bihorel
<[email protected]>; [email protected]
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
I am really puzzled by this statement. I would expect NONMEM to recognize time
varying covariates provide information about the fixed effects and used the
time specific value of the covariate to make a prediction.
Averaging the covariate across all the records for a subject seems like a poor
use of information.
Is your statement saying something special associated with the mu-referenced
transformation? If so would you please clarify your statement about averaging?
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: [email protected]<mailto:[email protected]>
web: http://holford.fmhs.auckland.ac.nz/
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Bauer, Robert
Sent: Saturday, 11 January 2025 8:32 pm
To: 'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com
-----Original Message-----
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]<mailto:[email protected]>
>
>
> Regeneron - Internal
>
> ********************************************************************
Hi Bob,
Thanks for explaining that EM and BAYES methods are a form of naïve pooled data
analysis for the individual.
I will make sure I stick to the classic methods when dealing with clinical data
with time varying covariates such as body mass, post-menstrual age and serum
creatinine.
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: [email protected]<mailto:[email protected]>
web: http://holford.fmhs.auckland.ac.nz/
Quoted reply history
From: Bauer, Robert <[email protected]>
Sent: Tuesday, 14 January 2025 5:34 am
To: Nick Holford <[email protected]>; 'Leonid Gibiansky'
<[email protected]>; Sébastien Bihorel
<[email protected]>; [email protected]
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hello Nick:
The statement I made pertains only to EM algorithms (ITS, SAEM, IMP), and
BAYES. The classic methods (FOCEI, Laplace), do not engage in averaging the
covariates across records even when thetas are MU referenced, as the classic
algorithms do not use EM update methods to advance the theta estimates.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com/
From: Nick Holford <[email protected]<mailto:[email protected]>>
Sent: Saturday, January 11, 2025 9:29 PM
To: Bauer, Robert <[email protected]<mailto:[email protected]>>;
'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
I am really puzzled by this statement. I would expect NONMEM to recognize time
varying covariates provide information about the fixed effects and used the
time specific value of the covariate to make a prediction.
Averaging the covariate across all the records for a subject seems like a poor
use of information.
Is your statement saying something special associated with the mu-referenced
transformation? If so would you please clarify your statement about averaging?
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: [email protected]<mailto:[email protected]>
web: http://holford.fmhs.auckland.ac.nz/
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Bauer, Robert
Sent: Saturday, 11 January 2025 8:32 pm
To: 'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com
-----Original Message-----
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]<mailto:[email protected]>
>
>
> Regeneron - Internal
>
> ********************************************************************
Hi Bob,
I agree with the sentiments of Nick’s email below that if EM and Bayes’ methods
are simply averaging the time-varying covariates and essentially treating them
as time-invariant values set to the average covariate values, then this would
be a MAJOR deficiency in how EM and Bayes methods in NONMEM handle time-varying
covariates. Afterall, the main reason we investigate time-varying covariates
is to evaluate whether certain parameters can change over time within a subject
where the time-varying covariates may help explain some of the within-subject
variation (e.g., IOV). If the EM and Bayes’ methods as implemented in NONMEM
treat the time-varying covariates as time-invariant at the arithmetic mean
value, then the predictions will not be properly considering these time-varying
covariates. However, I suspect that this is not the case, and the EM and
Bayes’ methods as implemented in NONMEM are actually considering the
time-varying nature of these covariates and the confusion comes from an
ambiguous explanation of what NONMEM is doing. Let me see if I can explain
what NONMEM is doing, and you can correct me if I’m wrong or further elaborate
on Nick’s and my concerns.
I assume that EM and Bayes methods are actually using the time-varying
covariates, however, the EM and Bayes’ methods perform centering and/or scaling
based on the subject-specific arithmetic mean values of the covariates when
mu-referencing is implemented. This is to enhance numerical stability when
estimating the fixed effects associated with time-varying covariates. Thus,
the EM and Bayes’ methods are actually using the time-varying values of the
covariates in the prediction of the responses and the averaging of the
covariates within a subject is only implemented for MU referencing. Is my
understanding correct?
Thanks,
Ken
Quoted reply history
From: [email protected] <[email protected]> On Behalf Of
Nick Holford
Sent: Monday, January 13, 2025 5:49 PM
To: Bauer, Robert <[email protected]>; 'Leonid Gibiansky'
<[email protected]>; Sébastien Bihorel
<[email protected]>; [email protected]
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
Thanks for explaining that EM and BAYES methods are a form of naïve pooled data
analysis for the individual.
I will make sure I stick to the classic methods when dealing with clinical data
with time varying covariates such as body mass, post-menstrual age and serum
creatinine.
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: <mailto:[email protected]> [email protected]
web: http://holford.fmhs.auckland.ac.nz/ http://holford.fmhs.auckland.ac.nz/
From: Bauer, Robert <[email protected] <mailto:[email protected]>
>
Sent: Tuesday, 14 January 2025 5:34 am
To: Nick Holford <[email protected] <mailto:[email protected]> >;
'Leonid Gibiansky' <[email protected]
<mailto:[email protected]> >; Sébastien Bihorel
<[email protected] <mailto:[email protected]> >;
[email protected] <mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hello Nick:
The statement I made pertains only to EM algorithms (ITS, SAEM, IMP), and
BAYES. The classic methods (FOCEI, Laplace), do not engage in averaging the
covariates across records even when thetas are MU referenced, as the classic
algorithms do not use EM update methods to advance the theta estimates.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
<mailto:[email protected]> [email protected]
http://www.iconplc.com/ www.iconplc.com
From: Nick Holford <[email protected] <mailto:[email protected]>
>
Sent: Saturday, January 11, 2025 9:29 PM
To: Bauer, Robert <[email protected] <mailto:[email protected]>
>; 'Leonid Gibiansky' <[email protected]
<mailto:[email protected]> >; Sébastien Bihorel
<[email protected] <mailto:[email protected]> >;
[email protected] <mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
I am really puzzled by this statement. I would expect NONMEM to recognize time
varying covariates provide information about the fixed effects and used the
time specific value of the covariate to make a prediction.
Averaging the covariate across all the records for a subject seems like a poor
use of information.
Is your statement saying something special associated with the mu-referenced
transformation? If so would you please clarify your statement about averaging?
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: <mailto:[email protected]> [email protected]
web: http://holford.fmhs.auckland.ac.nz/ http://holford.fmhs.auckland.ac.nz/
From: [email protected] <mailto:[email protected]>
<[email protected] <mailto:[email protected]> > On Behalf
Of Bauer, Robert
Sent: Saturday, 11 January 2025 8:32 pm
To: 'Leonid Gibiansky' <[email protected]
<mailto:[email protected]> >; Sébastien Bihorel
<[email protected] <mailto:[email protected]> >;
[email protected] <mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected] <mailto:[email protected]>
www.iconplc.com http://www.iconplc.com
-----Original Message-----
From: [email protected] <mailto:[email protected]>
<[email protected] <mailto:[email protected]> > On Behalf
Of Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel <[email protected]
<mailto:[email protected]> >; [email protected]
<mailto:[email protected]>
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected] <mailto:[email protected]>
>
>
> Regeneron - Internal
>
> ********************************************************************
Hello Ken:
Yes, the record (time)-averaging of the covariate values is done only in the EM
update process of estimating the thetas, but otherwise, the time-variant
covariate is used as coded in the model, for individual predictions for example.
If you want to use time-varying covariates in EM/Bayes methods, simply do not
mu reference those thetas involved with time-varying covariates (or turn off
that mu reference equation with the MUM=N() option). NONMEM will then use a
gradient technique that will update the thetas, without averaging the
time-varying covariate, but this process requires more computation time.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com/
Quoted reply history
From: [email protected] <[email protected]>
Sent: Monday, January 13, 2025 4:02 PM
To: 'Nick Holford' <[email protected]>; Bauer, Robert
<[email protected]>; 'Leonid Gibiansky' <[email protected]>;
'Sébastien Bihorel' <[email protected]>; [email protected]
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
I agree with the sentiments of Nick's email below that if EM and Bayes' methods
are simply averaging the time-varying covariates and essentially treating them
as time-invariant values set to the average covariate values, then this would
be a MAJOR deficiency in how EM and Bayes methods in NONMEM handle time-varying
covariates. Afterall, the main reason we investigate time-varying covariates
is to evaluate whether certain parameters can change over time within a subject
where the time-varying covariates may help explain some of the within-subject
variation (e.g., IOV). If the EM and Bayes' methods as implemented in NONMEM
treat the time-varying covariates as time-invariant at the arithmetic mean
value, then the predictions will not be properly considering these time-varying
covariates. However, I suspect that this is not the case, and the EM and
Bayes' methods as implemented in NONMEM are actually considering the
time-varying nature of these covariates and the confusion comes from an
ambiguous explanation of what NONMEM is doing. Let me see if I can explain
what NONMEM is doing, and you can correct me if I'm wrong or further elaborate
on Nick's and my concerns.
I assume that EM and Bayes methods are actually using the time-varying
covariates, however, the EM and Bayes' methods perform centering and/or scaling
based on the subject-specific arithmetic mean values of the covariates when
mu-referencing is implemented. This is to enhance numerical stability when
estimating the fixed effects associated with time-varying covariates. Thus,
the EM and Bayes' methods are actually using the time-varying values of the
covariates in the prediction of the responses and the averaging of the
covariates within a subject is only implemented for MU referencing. Is my
understanding correct?
Thanks,
Ken
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Nick Holford
Sent: Monday, January 13, 2025 5:49 PM
To: Bauer, Robert <[email protected]<mailto:[email protected]>>;
'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
Thanks for explaining that EM and BAYES methods are a form of naïve pooled data
analysis for the individual.
I will make sure I stick to the classic methods when dealing with clinical data
with time varying covariates such as body mass, post-menstrual age and serum
creatinine.
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: [email protected]<mailto:[email protected]>
web: http://holford.fmhs.auckland.ac.nz/
From: Bauer, Robert <[email protected]<mailto:[email protected]>>
Sent: Tuesday, 14 January 2025 5:34 am
To: Nick Holford <[email protected]<mailto:[email protected]>>;
'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hello Nick:
The statement I made pertains only to EM algorithms (ITS, SAEM, IMP), and
BAYES. The classic methods (FOCEI, Laplace), do not engage in averaging the
covariates across records even when thetas are MU referenced, as the classic
algorithms do not use EM update methods to advance the theta estimates.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com/
From: Nick Holford <[email protected]<mailto:[email protected]>>
Sent: Saturday, January 11, 2025 9:29 PM
To: Bauer, Robert <[email protected]<mailto:[email protected]>>;
'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Bob,
I am really puzzled by this statement. I would expect NONMEM to recognize time
varying covariates provide information about the fixed effects and used the
time specific value of the covariate to make a prediction.
Averaging the covariate across all the records for a subject seems like a poor
use of information.
Is your statement saying something special associated with the mu-referenced
transformation? If so would you please clarify your statement about averaging?
Best wishes,
Nick
--
Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
email: [email protected]<mailto:[email protected]>
web: http://holford.fmhs.auckland.ac.nz/
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Bauer, Robert
Sent: Saturday, 11 January 2025 8:32 pm
To: 'Leonid Gibiansky'
<[email protected]<mailto:[email protected]>>; Sébastien
Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
If a covariate varies across records within a subject, NONMEM obtains a simple
average among the records and uses this as the covariate value for that subject.
Robert J. Bauer, Ph.D.
Senior Director
Pharmacometrics R&D
ICON Early Phase
731 Arbor way, suite 100
Blue Bell, PA 19422
Office: (215) 616-6428
Mobile: (925) 286-0769
[email protected]<mailto:[email protected]>
http://www.iconplc.com
-----Original Message-----
From: [email protected]<mailto:[email protected]>
<[email protected]<mailto:[email protected]>> On Behalf
Of Leonid Gibiansky
Sent: Friday, January 10, 2025 5:21 PM
To: Sébastien Bihorel
<[email protected]<mailto:[email protected]>>;
[email protected]<mailto:[email protected]>
Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying covariates
Hi Sébastien,
As you did these experiments, can you share the results: have you seen any
differences in the fit, parameter estimates, precision, convergence speed
(number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI does
not have this restriction of time-independence even if you use Mu referencing,
so results should be identical or very close).
As code is encrypted, only Bob can answer the question but my understanding is
that some kind of averaging is used to get time independent value of WT that is
then used by the SAEM/IMP algorithm for parameter update procedure.
As WT changes slowly and not very significantly, it could be hard to see the
differences. A more stringent test would be to use time-dependent and strongly
influential ADA (0/1): how bad is the incorrect version 1 in this case?
Thank you
Leonid
On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
>
> Happy New Year,
>
> I hope everybody is ready for a great 2025 !
>
> I'll start my message/question by defining 2 different ways of coding
> a simple power relationship between body weigh on clearance.
>
> *
> Coding 1
>
> MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
>
> *
> Coding 2
>
> MU_1 = THETA(1)
> CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
>
> The reference and training materials for NONMEM clearly indicate that
> MU variables should be time invariant within occasions and recommend
> using coding 2 when body weight is time varying. Nevertheless, it is
> possible for an analyst to use coding 1. As far as I can tell from
> some limited testing, this is not a "fatal" error. Either with FOCE(I)
> or SAEM/IMP, NONMEM reports a warning but performs the model
> optimization. The table outputs also report CL as a time varying
> variable changing as body weight changes.
>
> So my questions are the following: when coding 1 is used and body
> weight is time varying, what is NONMEM actually doing during model
> optimization? Does NONMEM internally create occasions to break the
> records by interval of constant body weight and constant MU1?
> Alternatively, does NONMEM internally calculate an average of MU1?
> Something entirely different? What's the risk taken by an analyst when
> using coding 1 versus coding 2?
>
> Thank you in advance for you input
>
>
> __
> Sébastien Bihorel
> Director, Quantitative Pharmacology
> +1 914-648-9581
> [email protected]<mailto:[email protected]>
>
>
> Regeneron - Internal
>
> ********************************************************************
Dear NMusers,
Documentation relevant to this discussion (Intro to NONMEM 7.5.1. p
185-186):
Time dependent covariates, or covariates changing with each record within
an individual, cannot
be part of the MU_ equation. For example
MU_3=THETA(1)*TIME+THETA(2)
should not be done. Or, consider
MU_3=THETA(2)*WT
Where WT is not constant within an individual, but varies with observation
record (time). This
would also not be suitable. However, we could phrase as
MU_3=THETA(2)
CL=WT*(MU_3+ETA(3))
where MU_3 represents a population mean clearance per unit weight, which is
constant with
time (observation record), and is more universal among subjects. The MU
variables may vary
with inter-occasion, but not with time.
Suppose we have a situation where WT has an unknown power term associated
with it modeled
as THETA(3) in this example:
CL=THETA(2)*WT**THETA(3)*EXP(ETA(1))
Normally, we could efficiently linear model this as follows:
MU_1=THETA(2)+THETA(3)*LOG(WT)
CL=EXP(MU_1+ETA(1))
with THETA(2) transformed into the log of clearance domain. However, if WT
changes record
by record within the individual, then LOG(WT) may not be in the Mu
modeling. We would then
remove the THETA(3)*LOG(WT) term from MU_1:
MU_1=LOG(THETA(2))
CL=WT**THETA(3)*EXP(MU_1+ETA(1))
And THETA(3) itself would not be MU modeled.
Best wishes Pyry
Quoted reply history
On Tue, 14 Jan 2025 at 23:00, <[email protected]> wrote:
> I got an email indicating my message below may not have gone through so,
> I’m sending it again. Apologies if you receive this message twice. -Ken
>
>
>
>
>
> Hi Bob,
>
>
>
> I agree with the sentiments of Nick’s email below that if EM and Bayes’
> methods are simply averaging the time-varying covariates and essentially
> treating them as time-invariant values set to the average covariate values,
> then this would be a MAJOR deficiency in how EM and Bayes methods in NONMEM
> handle time-varying covariates. Afterall, the main reason we investigate
> time-varying covariates is to evaluate whether certain parameters can
> change over time within a subject where the time-varying covariates may
> help explain some of the within-subject variation (e.g., IOV). If the EM
> and Bayes’ methods as implemented in NONMEM treat the time-varying
> covariates as time-invariant at the arithmetic mean value, then the
> predictions will not be properly considering these time-varying
> covariates. However, I suspect that this is not the case, and the EM and
> Bayes’ methods as implemented in NONMEM are actually considering the
> time-varying nature of these covariates and the confusion comes from an
> ambiguous explanation of what NONMEM is doing. Let me see if I can explain
> what NONMEM is doing, and you can correct me if I’m wrong or further
> elaborate on Nick’s and my concerns.
>
>
>
> I assume that EM and Bayes methods are actually using the time-varying
> covariates, however, the EM and Bayes’ methods perform centering and/or
> scaling based on the subject-specific arithmetic mean values of the
> covariates when mu-referencing is implemented. This is to enhance
> numerical stability when estimating the fixed effects associated with
> time-varying covariates. Thus, the EM and Bayes’ methods are actually
> using the time-varying values of the covariates in the prediction of the
> responses and the averaging of the covariates within a subject is only
> implemented for MU referencing. Is my understanding correct?
>
>
>
> Thanks,
>
>
>
> Ken
>
>
>
>
>
>
>
> *From:* [email protected] <[email protected]> *On
> Behalf Of *Nick Holford
> *Sent:* Monday, January 13, 2025 5:49 PM
> *To:* Bauer, Robert <[email protected]>; 'Leonid Gibiansky' <
> [email protected]>; Sébastien Bihorel <
> [email protected]>; [email protected]
> *Subject:* RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying
> covariates
>
>
>
> Hi Bob,
>
>
>
> Thanks for explaining that EM and BAYES methods are a form of naïve pooled
> data analysis for the individual.
>
>
>
> I will make sure I stick to the classic methods when dealing with clinical
> data with time varying covariates such as body mass, post-menstrual age and
> serum creatinine.
>
>
>
> Best wishes,
>
> Nick
>
>
>
> --
>
> Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
>
> mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
>
> email: [email protected]
>
> web: http://holford.fmhs.auckland.ac.nz/
>
>
>
> *From:* Bauer, Robert <[email protected]>
> *Sent:* Tuesday, 14 January 2025 5:34 am
> *To:* Nick Holford <[email protected]>; 'Leonid Gibiansky' <
> [email protected]>; Sébastien Bihorel <
> [email protected]>; [email protected]
> *Subject:* RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying
> covariates
>
>
>
> Hello Nick:
>
> The statement I made pertains only to EM algorithms (ITS, SAEM, IMP), and
> BAYES. The classic methods (FOCEI, Laplace), do not engage in averaging
> the covariates across records even when thetas are MU referenced, as the
> classic algorithms do not use EM update methods to advance the theta
> estimates.
>
>
>
> Robert J. Bauer, Ph.D.
>
> Senior Director
>
> Pharmacometrics R&D
>
> ICON Early Phase
>
> 731 Arbor way, suite 100
>
> Blue Bell, PA 19422
>
> Office: (215) 616-6428
>
> Mobile: (925) 286-0769
>
> [email protected]
>
> www.iconplc.com
>
>
>
> *From:* Nick Holford <[email protected]>
> *Sent:* Saturday, January 11, 2025 9:29 PM
> *To:* Bauer, Robert <[email protected]>; 'Leonid Gibiansky' <
> [email protected]>; Sébastien Bihorel <
> [email protected]>; [email protected]
> *Subject:* RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying
> covariates
>
>
>
> Hi Bob,
>
>
>
> I am really puzzled by this statement. I would expect NONMEM to recognize
> time varying covariates provide information about the fixed effects and
> used the time specific value of the covariate to make a prediction.
>
> Averaging the covariate across all the records for a subject seems like a
> poor use of information.
>
> Is your statement saying something special associated with the
> mu-referenced transformation? If so would you please clarify your statement
> about averaging?
>
>
>
> Best wishes,
>
> Nick
>
>
>
> --
>
> Nick Holford, Professor Emeritus Clinical Pharmacology, MBChB, FRACP
>
> mobile: NZ+64(21) 46 23 53 ; FR+33(6) 62 32 46 72
>
> email: [email protected]
>
> web: http://holford.fmhs.auckland.ac.nz/
>
>
>
> *From:* [email protected] <[email protected]> *On
> Behalf Of *Bauer, Robert
> *Sent:* Saturday, 11 January 2025 8:32 pm
> *To:* 'Leonid Gibiansky' <[email protected]>; Sébastien Bihorel <
> [email protected]>; [email protected]
> *Subject:* RE: [EXTERNAL] Re: [NMusers] MU referencing and time-varying
> covariates
>
>
>
> If a covariate varies across records within a subject, NONMEM obtains a
> simple average among the records and uses this as the covariate value for
> that subject.
>
> Robert J. Bauer, Ph.D.
> Senior Director
> Pharmacometrics R&D
> ICON Early Phase
> 731 Arbor way, suite 100
> Blue Bell, PA 19422
> Office: (215) 616-6428
> Mobile: (925) 286-0769
> [email protected]
> www.iconplc.com
>
> -----Original Message-----
> From: [email protected] <[email protected]> On
> Behalf Of Leonid Gibiansky
> Sent: Friday, January 10, 2025 5:21 PM
> To: Sébastien Bihorel <[email protected]>;
> [email protected]
> Subject: [EXTERNAL] Re: [NMusers] MU referencing and time-varying
> covariates
>
> Hi Sébastien,
>
> As you did these experiments, can you share the results: have you seen any
> differences in the fit, parameter estimates, precision, convergence speed
> (number of iteration), and evaluation time for SAEM/IMP (I think, FOCEI
> does not have this restriction of time-independence even if you use Mu
> referencing, so results should be identical or very close).
>
> As code is encrypted, only Bob can answer the question but my
> understanding is that some kind of averaging is used to get time
> independent value of WT that is then used by the SAEM/IMP algorithm for
> parameter update procedure.
>
> As WT changes slowly and not very significantly, it could be hard to see
> the differences. A more stringent test would be to use time-dependent and
> strongly influential ADA (0/1): how bad is the incorrect version 1 in this
> case?
>
> Thank you
> Leonid
>
> On 1/10/2025 2:56 PM, Sébastien Bihorel wrote:
> >
> > Happy New Year,
> >
> > I hope everybody is ready for a great 2025 !
> >
> > I'll start my message/question by defining 2 different ways of coding
> > a simple power relationship between body weigh on clearance.
> >
> > *
> > Coding 1
> >
> > MU_1 = THETA(1) + THETA(2) * LOG(WGT/70) CL = EXP( MU_1 + ETA(1) )
> >
> > *
> > Coding 2
> >
> > MU_1 = THETA(1)
> > CL = EXP( MU_1 + ETA(1) ) * ( WGT/70 )**THETA(2)
> >
> > The reference and training materials for NONMEM clearly indicate that
> > MU variables should be time invariant within occasions and recommend
> > using coding 2 when body weight is time varying. Nevertheless, it is
> > possible for an analyst to use coding 1. As far as I can tell from
> > some limited testing, this is not a "fatal" error. Either with FOCE(I)
> > or SAEM/IMP, NONMEM reports a warning but performs the model
> > optimization. The table outputs also report CL as a time varying
> > variable changing as body weight changes.
> >
> > So my questions are the following: when coding 1 is used and body
> > weight is time varying, what is NONMEM actually doing during model
> > optimization? Does NONMEM internally create occasions to break the
> > records by interval of constant body weight and constant MU1?
> > Alternatively, does NONMEM internally calculate an average of MU1?
> > Something entirely different? What's the risk taken by an analyst when
> > using coding 1 versus coding 2?
> >
> > Thank you in advance for you input
> >
> >
> > __
> > Sébastien Bihorel
> > Director, Quantitative Pharmacology
> > +1 914-648-9581
> > [email protected]
> >
> >
> > Regeneron - Internal
> >
> > ********************************************************************
>
>
>
>