From: Eleveld-Ufkes, DJ <*d.j.eleveld*>

Date: Wed, 20 Nov 2019 13:36:38 +0000

Hi Sumeet,

The most common with power covariate relationships is indeed covariate=0 =

and an exponent that is estimated negative (or traverses into that area dur=

ing estimation). Given your problem is power function and BMI and cannot co=

rrectly be 0 then I agree it should work. Double check that you have no BMI=

=0 for any individuals on any records. No missing values? Maybe NONMEM is=

treating missing values as 0?

On possibility is outliers are causing numerical instability in the power m=

odel. In my experience exponential models are less prone to numerical stabi=

lity than power models. Maybe examine the exponential fit to see the genera=

l tendencies between CL and BMI. Exponential models don't blowup like power=

models do. For not-all-too-extreme cases the results should not be drastic=

ally different from a power model. It might give you a hint as to why the p=

ower model is misbehaving. A single outlier can cause considerable problems=

when N is small.

As an aside about automated SCM, is there not anywhere a correction to the =

dOFV threshold for multiple testing? You are testing 5 models on CL and sti=

ll keeping the threshold at 3.84, so won't you be getting more than expecte=

d false-positive detection of a significant covariate?

Warm regards,

Douglas Eleveld

From: owner-nmusers

Behalf Of Singla, Sumeet K

Sent: woensdag 20 november 2019 05:43

To: nmusers

Subject: [NMusers] Error in Power-model for covariate selection

Hello Everyone,

I am using PsN enabled SCM option in Pirana to analyze selection of covaria=

tes in the PK 2- Comp model. As long as I am using just linear and exponent=

ial covariate model, everything runs fine. However, as soon as I add power-=

model in the mix, "some" power models involving continuous covariates on pa=

rameters fail to run and it gives me the following error:

"HESSIAN OF POSTERIOR DENSITY IS NON-POSITIVE-DEFINITE DURING SEARCH"

I understand that individual PK parameter search might be moving into 0 or =

negative territory. I am trying to fix it but didn't have much success. I h=

ave set lower bounds, removed lower bounds, changed order of model in valid=

states option in scm configuration file, dataset doesn't contain any 0 or =

negative value, but nothing is working. FYI: I don't need to test hockey-st=

ick relation as literature and data doesn't support it, power model can onl=

y be used on continuous covariates and I have turned on parallel states opt=

ion.

This is how part of my scm results, followed by scm configuration file, fol=

lowed by base model for PK 2-Comp looks like:

MODEL TEST BASE OFV NEW OFV TEST OFV (DROP) G=

OAL dDF SIGNIFICANT PVAL

CLAGE-2 PVAL 2618.02603 2616.67228 1.35375 > 3=

.84150 1 0.244620

CLAGE-5 PVAL 2618.02603 2616.36273 1.66330 > 3=

.84150 1 0.197160

CLAGE-4 PVAL 2618.02603 2616.65232 1.37371 > 3=

.84150 1 0.241180

CLBMI-2 PVAL 2618.02603 2612.96657 5.05946 > 3=

.84150 1 YES! 0.024492

CLBMI-5 PVAL 2618.02603 FAILED FAILED > 3=

.84150 1 999

SCM FILE:

"

search_direction=both

p_forward=0.05

p_backward=0.01

continuous_covariates=BMI,AGE

categorical_covariates=USER,SEX

parallel_states=1

retries=2

threads=6

tweak_inits=1

;;1-NotIncluded, 2-LINEAR, 3-Hockey Stick Relation, 4-Exponential, 5-Power

[test_relations]

CL=AGE,SEX,BMI,USER

V1=AGE,SEX,BMI,USER

V2=AGE,SEX,BMI,USER

Q=AGE,SEX,BMI,USER

[valid_states]

continuous = 1,2,5,4

categorical = 1,2

"

NONMEM Control Stream:

$SUBROUTINE ADVAN3 TRANS4

$PK

TVV1 = THETA(1) ;Central Volume of distributi=

on in L

V1 = TVV1*EXP(ETA(1))

TVCL = THETA(2)

CL = TVCL*EXP(ETA(2)) ; Clearance L/h

TVQ = THETA(3)

Q = TVQ*EXP(ETA(3)) ;Intercompartment Clearance

TVV2 = THETA(4)

V2 = TVV2*EXP(ETA(4)) ;Peripheral volume in L

S1=V1

$ERROR

IPRED=F

Y= F + F*ERR(1); Proportional Error

$THETA

(0, 16); [V1] based on PK 2 Comp

(0, 255); [CL] based on PK 2 Comp

(0, 33.5); [Q] based on PK 2 Comp

(0, 29.7); [V2] based on PK 2 Comp

$OMEGA

(0, 0.08); [P] omega(1,1)

(0, 0.159); [P] omega(2,2)

(0, 0.140); [P] omega(3,3)

(0, 0.19); [P] omega(4,4)

$SIGMA

(0, 0.06) ;sigma1

$EST METHOD=1 PRINT=5 MAXEVAL=9999 SIG=3 NOABORT

Regards,

Sumeet K. Singla

Ph.D. Candidate

Division of Pharmaceutics and Translational Therapeutics

College of Pharmacy | University of Iowa

Iowa City, Iowa

sumeet-singla

518.577.5881

________________________________

De inhoud van dit bericht is vertrouwelijk en alleen bestemd voor de geadre=

sseerde(n). Anderen dan de geadresseerde(n) mogen geen gebruik maken van di=

t bericht, het niet openbaar maken of op enige wijze verspreiden of vermeni=

gvuldigen. Het UMCG kan niet aansprakelijk gesteld worden voor een incomple=

te aankomst of vertraging van dit verzonden bericht.

The contents of this message are confidential and only intended for the eye=

s of the addressee(s). Others than the addressee(s) are not allowed to use =

this message, to make it public or to distribute or multiply this message i=

n any way. The UMCG cannot be held responsible for incomplete reception or =

delay of this transferred message.

Received on Wed Nov 20 2019 - 08:36:38 EST

Date: Wed, 20 Nov 2019 13:36:38 +0000

Hi Sumeet,

The most common with power covariate relationships is indeed covariate=0 =

and an exponent that is estimated negative (or traverses into that area dur=

ing estimation). Given your problem is power function and BMI and cannot co=

rrectly be 0 then I agree it should work. Double check that you have no BMI=

=0 for any individuals on any records. No missing values? Maybe NONMEM is=

treating missing values as 0?

On possibility is outliers are causing numerical instability in the power m=

odel. In my experience exponential models are less prone to numerical stabi=

lity than power models. Maybe examine the exponential fit to see the genera=

l tendencies between CL and BMI. Exponential models don't blowup like power=

models do. For not-all-too-extreme cases the results should not be drastic=

ally different from a power model. It might give you a hint as to why the p=

ower model is misbehaving. A single outlier can cause considerable problems=

when N is small.

As an aside about automated SCM, is there not anywhere a correction to the =

dOFV threshold for multiple testing? You are testing 5 models on CL and sti=

ll keeping the threshold at 3.84, so won't you be getting more than expecte=

d false-positive detection of a significant covariate?

Warm regards,

Douglas Eleveld

From: owner-nmusers

Behalf Of Singla, Sumeet K

Sent: woensdag 20 november 2019 05:43

To: nmusers

Subject: [NMusers] Error in Power-model for covariate selection

Hello Everyone,

I am using PsN enabled SCM option in Pirana to analyze selection of covaria=

tes in the PK 2- Comp model. As long as I am using just linear and exponent=

ial covariate model, everything runs fine. However, as soon as I add power-=

model in the mix, "some" power models involving continuous covariates on pa=

rameters fail to run and it gives me the following error:

"HESSIAN OF POSTERIOR DENSITY IS NON-POSITIVE-DEFINITE DURING SEARCH"

I understand that individual PK parameter search might be moving into 0 or =

negative territory. I am trying to fix it but didn't have much success. I h=

ave set lower bounds, removed lower bounds, changed order of model in valid=

states option in scm configuration file, dataset doesn't contain any 0 or =

negative value, but nothing is working. FYI: I don't need to test hockey-st=

ick relation as literature and data doesn't support it, power model can onl=

y be used on continuous covariates and I have turned on parallel states opt=

ion.

This is how part of my scm results, followed by scm configuration file, fol=

lowed by base model for PK 2-Comp looks like:

MODEL TEST BASE OFV NEW OFV TEST OFV (DROP) G=

OAL dDF SIGNIFICANT PVAL

CLAGE-2 PVAL 2618.02603 2616.67228 1.35375 > 3=

.84150 1 0.244620

CLAGE-5 PVAL 2618.02603 2616.36273 1.66330 > 3=

.84150 1 0.197160

CLAGE-4 PVAL 2618.02603 2616.65232 1.37371 > 3=

.84150 1 0.241180

CLBMI-2 PVAL 2618.02603 2612.96657 5.05946 > 3=

.84150 1 YES! 0.024492

CLBMI-5 PVAL 2618.02603 FAILED FAILED > 3=

.84150 1 999

SCM FILE:

"

search_direction=both

p_forward=0.05

p_backward=0.01

continuous_covariates=BMI,AGE

categorical_covariates=USER,SEX

parallel_states=1

retries=2

threads=6

tweak_inits=1

;;1-NotIncluded, 2-LINEAR, 3-Hockey Stick Relation, 4-Exponential, 5-Power

[test_relations]

CL=AGE,SEX,BMI,USER

V1=AGE,SEX,BMI,USER

V2=AGE,SEX,BMI,USER

Q=AGE,SEX,BMI,USER

[valid_states]

continuous = 1,2,5,4

categorical = 1,2

"

NONMEM Control Stream:

$SUBROUTINE ADVAN3 TRANS4

$PK

TVV1 = THETA(1) ;Central Volume of distributi=

on in L

V1 = TVV1*EXP(ETA(1))

TVCL = THETA(2)

CL = TVCL*EXP(ETA(2)) ; Clearance L/h

TVQ = THETA(3)

Q = TVQ*EXP(ETA(3)) ;Intercompartment Clearance

TVV2 = THETA(4)

V2 = TVV2*EXP(ETA(4)) ;Peripheral volume in L

S1=V1

$ERROR

IPRED=F

Y= F + F*ERR(1); Proportional Error

$THETA

(0, 16); [V1] based on PK 2 Comp

(0, 255); [CL] based on PK 2 Comp

(0, 33.5); [Q] based on PK 2 Comp

(0, 29.7); [V2] based on PK 2 Comp

$OMEGA

(0, 0.08); [P] omega(1,1)

(0, 0.159); [P] omega(2,2)

(0, 0.140); [P] omega(3,3)

(0, 0.19); [P] omega(4,4)

$SIGMA

(0, 0.06) ;sigma1

$EST METHOD=1 PRINT=5 MAXEVAL=9999 SIG=3 NOABORT

Regards,

Sumeet K. Singla

Ph.D. Candidate

Division of Pharmaceutics and Translational Therapeutics

College of Pharmacy | University of Iowa

Iowa City, Iowa

sumeet-singla

518.577.5881

________________________________

De inhoud van dit bericht is vertrouwelijk en alleen bestemd voor de geadre=

sseerde(n). Anderen dan de geadresseerde(n) mogen geen gebruik maken van di=

t bericht, het niet openbaar maken of op enige wijze verspreiden of vermeni=

gvuldigen. Het UMCG kan niet aansprakelijk gesteld worden voor een incomple=

te aankomst of vertraging van dit verzonden bericht.

The contents of this message are confidential and only intended for the eye=

s of the addressee(s). Others than the addressee(s) are not allowed to use =

this message, to make it public or to distribute or multiply this message i=

n any way. The UMCG cannot be held responsible for incomplete reception or =

delay of this transferred message.

Received on Wed Nov 20 2019 - 08:36:38 EST