From: BAE, KYUN-SEOP <*KYUN-SEOP.BAE*>

Date: Thu, 13 Nov 2008 12:09:02 -0800

Hi, All,

This is just brief my understanding.

If common variable IRSET (a kind of reset indicator variable) is set to

1, this error message shows.

This variable is set with various conditions but without minimization

failure.

The condition I found is that

IRSETP==1 or IRSET > 0.1*ITN or (IRESET==1 and IER==0)

Here ITN is iteration count and IER (presumably integer/indicator/index

of) error return code.

One reason (IRESET==1) is that 'RESET' event on Hessian matrix.

You may have seen the message like 'RESET HESSIAN, TYPE I' during the

minimization.

You may not see this message if you do not use PRINT=1 option.

Anyway, this message means NONNEM experienced some difficulties during

minimization, but "recovered" without any significant error message.

So, user is requested to be careful and make decision with "COVARIANCE"

together.

Because, "COVARIANCE STEP" is done in different subroutines from

"MIMINIZATION STEP" subroutines, successful and reasonable covariance

step supports that "MINIMIZATION STEP" was right.

As others already mentioned, I don't care much about this message if

standard errors are reasonable.

Thanks,

=====================

Kyun-Seop Bae MD PhD

Email: kyun-seop.bae

-----Original Message-----

From: owner-nmusers

On Behalf Of Ken Kowalski

Sent: Thursday, November 13, 2008 07:36

To: 'Higashimori, Mitsuo'; nmusers

Subject: RE: [NMusers] NONMEM message

Mitsuo,

This is a new message specific to NONMEM VI. I must confess I don't

know what to make of this message myself. It would be informative if

someone could tell us what internals in NONMEM trigger this message

(i.e., "PROBLEMS OCCURRED WITH THE MINIMIZATION").

With respect to your two model runs note that they are really two

different parameterizations. In the first parameterization, where

TVCL = THETA(1) * THETA(2) ** SEX

note that THETA(1) represents the true value of CL for males and

THETA(2) represents the ratio of CL between females to males. In the

second parameterization, where

TVCL = THETA(1)

IF (SEX.EQ.1) TVCL = THETA(2)

note that THETA(1) and THETA(2) represent the true values of CL for

males and females, respectively. Thus, THETA(2) has a different

interpretation between these two parameterizations.

A third parameterization that you could consider is

TVCL = THETA(1) *(1 + THETA(2))**SEX or equivalently, TVCL = =

THETA(1) *

(1 +

THETA(2)*SEX)

where THETA(1) is again the true value of CL for males and THETA(2) is

the fractional change in CL for females relative to males.

Each of these parameterizations should result in the same model fit

(i.e., minimum value of the OFV) but one parameterization may be more

stable than another...it is similar to the issue with continuous

covariates where we center or scale the covariate based on the mean or

median value (i.e., centering or scaling will reduce the correlation in

the estimates between the intercept term and the covariate effect which

should lead to a more stable model and faster convergence to the minimum

OFV).

I would look at the COV step output and in particular, look at the

correlation of the estimates between THETA(1) and THETA(2) for these

different parameterizations. My guess is that the correlation is higher

for the first parameterization (given that you indicate that it gives

this warning message and that the second parameterization does not).

You can also use the PRINT=E option on the COV statement and look at =

the

ratio of the largest to the smallest eigenvalues to more globally assess

the stability of your model. In the end, if they all converge to the

same final OFV and if you really want to estimate the ratio of CLs

between males and females then so be it even if the model is less stable

and NONMEM gives you this warning message. On the other hand, if the

different parameterizations don't converge to the same OFV then you need

to look more closely at how you parameterize the covariate effect. If

you get a lower OFV with the second parameterization because it is more

stable and NONMEM has an easier time iterating to the minimum OFV then I

would go with that parameterization and if you want to estimate the

ratio of the CLs you can always estimate it as

THETA(2)/THETA(1) (i.e., which is equivalent to THETA(2) in the first

parameterization.

I hope this helps.

Ken

Kenneth G. Kowalski

President & CEO

A2PG - Ann Arbor Pharmacometrics Group, Inc.

110 E. Miller Ave., Garden Suite

Ann Arbor, MI 48104

Work: 734-274-8255

Cell: 248-207-5082

Fax: 734-913-0230

ken.kowalski

-----Original Message-----

From: owner-nmusers

On Behalf Of Higashimori, Mitsuo

Sent: Wednesday, November 12, 2008 10:06 PM

To: nmusers

Subject: [NMusers] NONMEM message

Dear all,

I have a following error(?) massage on a poplation analysis using NONMEM

VI.

0MINIMIZATION SUCCESSFUL

HOWEVER, PROBLEMS OCCURRED WITH THE MINIMIZATION.

REGARD THE RESULTS OF THE ESTIMATION STEP CAREFULLY, AND ACCEPT THEM

ONLY AFTER CHECKING THAT THE COVARIANCE STEP PRODUCES REASONABLE

OUTPUT.

I found it when I described a control stream to assess sex difference on

oral clearance as shown in below,

TVCL = THETA(1) * THETA(2) ** SEX

where SEX=0 for male and SEX=1 for female.

This message was displayed without any error message. It was not

dissapeared even though I changed the initial parameters. However, it

was solved when I changed the model definition. For example,

TVCL = THETA(1)

IF (SEX.EQ.1) TVCL = THETA(2)

Could you please let me know some details regarding the message.

Especially, I'd like to know

1. What impact does this error message give the analysis result?

2. Why does it depend on the model definition?

Thanks,

_/ _/ _/ Mitsuo Higashimori, Ph.D.

_/ _/ _/ Pharmacokinetic Group, Early Phase Development Department _/ _/

_/ Clinical Division, Research & Development _/ _/ _/ AstraZeneca K.K.

_/ _/ _/ E-mail: Mitsuo.Higashimori

Received on Thu Nov 13 2008 - 15:09:02 EST

Date: Thu, 13 Nov 2008 12:09:02 -0800

Hi, All,

This is just brief my understanding.

If common variable IRSET (a kind of reset indicator variable) is set to

1, this error message shows.

This variable is set with various conditions but without minimization

failure.

The condition I found is that

IRSETP==1 or IRSET > 0.1*ITN or (IRESET==1 and IER==0)

Here ITN is iteration count and IER (presumably integer/indicator/index

of) error return code.

One reason (IRESET==1) is that 'RESET' event on Hessian matrix.

You may have seen the message like 'RESET HESSIAN, TYPE I' during the

minimization.

You may not see this message if you do not use PRINT=1 option.

Anyway, this message means NONNEM experienced some difficulties during

minimization, but "recovered" without any significant error message.

So, user is requested to be careful and make decision with "COVARIANCE"

together.

Because, "COVARIANCE STEP" is done in different subroutines from

"MIMINIZATION STEP" subroutines, successful and reasonable covariance

step supports that "MINIMIZATION STEP" was right.

As others already mentioned, I don't care much about this message if

standard errors are reasonable.

Thanks,

=====================

Kyun-Seop Bae MD PhD

Email: kyun-seop.bae

-----Original Message-----

From: owner-nmusers

On Behalf Of Ken Kowalski

Sent: Thursday, November 13, 2008 07:36

To: 'Higashimori, Mitsuo'; nmusers

Subject: RE: [NMusers] NONMEM message

Mitsuo,

This is a new message specific to NONMEM VI. I must confess I don't

know what to make of this message myself. It would be informative if

someone could tell us what internals in NONMEM trigger this message

(i.e., "PROBLEMS OCCURRED WITH THE MINIMIZATION").

With respect to your two model runs note that they are really two

different parameterizations. In the first parameterization, where

TVCL = THETA(1) * THETA(2) ** SEX

note that THETA(1) represents the true value of CL for males and

THETA(2) represents the ratio of CL between females to males. In the

second parameterization, where

TVCL = THETA(1)

IF (SEX.EQ.1) TVCL = THETA(2)

note that THETA(1) and THETA(2) represent the true values of CL for

males and females, respectively. Thus, THETA(2) has a different

interpretation between these two parameterizations.

A third parameterization that you could consider is

TVCL = THETA(1) *(1 + THETA(2))**SEX or equivalently, TVCL = =

THETA(1) *

(1 +

THETA(2)*SEX)

where THETA(1) is again the true value of CL for males and THETA(2) is

the fractional change in CL for females relative to males.

Each of these parameterizations should result in the same model fit

(i.e., minimum value of the OFV) but one parameterization may be more

stable than another...it is similar to the issue with continuous

covariates where we center or scale the covariate based on the mean or

median value (i.e., centering or scaling will reduce the correlation in

the estimates between the intercept term and the covariate effect which

should lead to a more stable model and faster convergence to the minimum

OFV).

I would look at the COV step output and in particular, look at the

correlation of the estimates between THETA(1) and THETA(2) for these

different parameterizations. My guess is that the correlation is higher

for the first parameterization (given that you indicate that it gives

this warning message and that the second parameterization does not).

You can also use the PRINT=E option on the COV statement and look at =

the

ratio of the largest to the smallest eigenvalues to more globally assess

the stability of your model. In the end, if they all converge to the

same final OFV and if you really want to estimate the ratio of CLs

between males and females then so be it even if the model is less stable

and NONMEM gives you this warning message. On the other hand, if the

different parameterizations don't converge to the same OFV then you need

to look more closely at how you parameterize the covariate effect. If

you get a lower OFV with the second parameterization because it is more

stable and NONMEM has an easier time iterating to the minimum OFV then I

would go with that parameterization and if you want to estimate the

ratio of the CLs you can always estimate it as

THETA(2)/THETA(1) (i.e., which is equivalent to THETA(2) in the first

parameterization.

I hope this helps.

Ken

Kenneth G. Kowalski

President & CEO

A2PG - Ann Arbor Pharmacometrics Group, Inc.

110 E. Miller Ave., Garden Suite

Ann Arbor, MI 48104

Work: 734-274-8255

Cell: 248-207-5082

Fax: 734-913-0230

ken.kowalski

-----Original Message-----

From: owner-nmusers

On Behalf Of Higashimori, Mitsuo

Sent: Wednesday, November 12, 2008 10:06 PM

To: nmusers

Subject: [NMusers] NONMEM message

Dear all,

I have a following error(?) massage on a poplation analysis using NONMEM

VI.

0MINIMIZATION SUCCESSFUL

HOWEVER, PROBLEMS OCCURRED WITH THE MINIMIZATION.

REGARD THE RESULTS OF THE ESTIMATION STEP CAREFULLY, AND ACCEPT THEM

ONLY AFTER CHECKING THAT THE COVARIANCE STEP PRODUCES REASONABLE

OUTPUT.

I found it when I described a control stream to assess sex difference on

oral clearance as shown in below,

TVCL = THETA(1) * THETA(2) ** SEX

where SEX=0 for male and SEX=1 for female.

This message was displayed without any error message. It was not

dissapeared even though I changed the initial parameters. However, it

was solved when I changed the model definition. For example,

TVCL = THETA(1)

IF (SEX.EQ.1) TVCL = THETA(2)

Could you please let me know some details regarding the message.

Especially, I'd like to know

1. What impact does this error message give the analysis result?

2. Why does it depend on the model definition?

Thanks,

_/ _/ _/ Mitsuo Higashimori, Ph.D.

_/ _/ _/ Pharmacokinetic Group, Early Phase Development Department _/ _/

_/ Clinical Division, Research & Development _/ _/ _/ AstraZeneca K.K.

_/ _/ _/ E-mail: Mitsuo.Higashimori

Received on Thu Nov 13 2008 - 15:09:02 EST