From: Ã…strand, Magnus <*Magnus.Astrand*>

Date: Thu, 12 Feb 2015 12:12:52 +0000

Dear all, an alternative which I try to use on strictly positive parameters=

is to estimate on log-scale. Then I think often the assymptitic approximat=

ion is more true and the resulting measures of parameter uncertainty are mo=

re reliable.

BW

Magnus

________________________________________

Från: owner-nmusers

haouch Aziz <Aziz.Chaouch

Skickat: den 12 februari 2015 12:01:51

Till: Eleveld, DJ; pascal.girard

Kopia: nmusers

Ämne: [NMusers] RE: Standard errors of estimates for strictly positive pa=

rameters

Dear Douglas, dear Pascal,

Thanks a lot for your answers. I guess the main point here is constrained v=

s unconstrained optimization as the asymptotic covariance matrix of estimat=

es (as returned by $COV) is "well defined" only in the latter case. When fi=

tting model 1, one would normally constrain THETA(1) to be positive by usin=

g something like:

$THETA

(0, 15, 50) ; TVCL

In this situation I wonder whether it makes sense at all to consider the ou=

tput of $COV. It seems model 2 would be here preferable (unconstrained opti=

mization). If model 1 is fitted without boundary constraints on THETA(1), t=

he covariance matrix of estimates may have "some" meaning but the optimizat=

ion in NONMEM is then likely to crash if it encounters a negative value at =

some point, which again speaks somehow in favor of model 2 (unless one is n=

ot interested in the output of $COV).

Now what about $OMEGA? Here NONMEM knows that these are variances and there=

fore we do not need to explicitly (i.e. manually) place boundary constraint=

s on the diagonal elements of the omega matrix. However something must acco=

unt for it internally. The covariance matrix of estimates returned by $COV =

also contains elements that refer to omega so I'm unsure how these are trea=

ted. For diagonal elements of the omega matrix, does NONMEM optimize log(om=

ega) or omega? Or does it uses a Cholesky decomposition of the Omega matrix=

and optimize elements on that scale? Again, unless the optimization on ome=

ga is unconstrained, can we really trust the output of $COV? Basically the =

question here is how would you construct an asymptotic 95% confidence inter=

val for a diagonal element of Omega (i.e. a variance) based on the informat=

ion from the covariance matrix of estimates?

The covariance matrix of estimate is of importance to me because I'm consid=

ering published studies and I do not have access to the data so I cannot re=

fit the model with an alternative parametrization. Results from $COV (in ls=

t file when available from the authors) is then the only available piece of=

information about the uncertainty of the estimation process.

Kind regards,

Aziz Chaouch

________________________________________

De : Eleveld, DJ [mailto:d.j.eleveld

Envoyé : mercredi, 11. février 2015 22:26

À : Chaouch Aziz; nmusers

Objet : RE: Standard errors of estimates for strictly positive parameters

Hi Aziz,

Just some comments off the top of my head in a quite informal way: I'm not =

really sure that these are the same problem because they dont start with th=

e same information in the form of parameter constraints. In model 1 you are=

asking the optimizer for the unconstrained maximum likelihood solution for=

TVCL. OK, this is reasonable in a lot of situations, but not necessairily =

in all situations.

In model 2 you add information by forcing TVCL and CL to be positive. If yo=

u think of the optimal solution as some point in N-dimensional space which =

has to be searched for, in model 2 you are saying "dont even look in the sp=

ace where TVCL or CL is negative". Even stronger, in model 2 you are also s=

aying "dont even get close to zero" because the log-normal distribution van=

ishes towards zero.

Which solution of these is best for some particular application depends on =

a lot of things. One of the things I would think about in this situation is=

whether or not my a priori beliefs match with the structual constraints of=

the model. Do I really think that the "true" CL could be zero? If yes, the=

n model 2 is hard to defend in that case.

You description of your situation regarding standard errors is a part of th=

e same thing. When you extrapolate standard errors into low-probability are=

as you are checking the boundaries of the probability area. It should not b=

e suprising that model 1 might tell you that CL is negative since this was =

part of the solution space which you allowed. With model 2 your model struc=

ture says "dont even look there"

In short, although these two models might look similar, I think they are re=

ally quite different. This becomes most clear when you consider the low-pro=

bability space.

Sorry for the vauge language.

Warm regards,

Douglas

________________________________________

De : pascal.girard

Envoyé : mercredi, 11. février 2015 18:30

À : Chaouch Aziz; nmusers

Objet : RE: Standard errors of estimates for strictly positive parameters

Dear Aziz,

NM does not return the asymptotic SE of THETA(1) in model 1 on the log-scal=

e. So I would use model 2.

With best regards / Mit freundlichen Grüßen / Cordialement

Pascal

________________________________________

From: owner-nmusers

of Chaouch Aziz [Aziz.Chaouch

Sent: Wednesday, February 11, 2015 5:21 PM

To: nmusers

Subject: [NMusers] Standard errors of estimates for strictly positive param=

eters

Hi,

I'm interested in generating samples from the asymptotic sampling distribut=

ion of population parameter estimates from a published PKPOP model fitted w=

ith NONMEM. By definition, parameter estimates are asymptotically (multivar=

iate) normally distributed (unconstrained optimization) with mean M and cov=

ariance C, where M is the vector of parameter estimates and C is the covari=

ance matrix of estimates (returned by $COV and available in the lst file).

Consider the 2 models below:

Model 1:

TVCL = THETA(1)

CL = TVCL*EXP(ETA(1))

Model 2:

TVCL = EXP(THETA(1))

CL = TVCL*EXP(ETA(1))

It is clear that model 1 and model 2 will provide exactly the same fit. How=

ever, although in both cases the standard error of estimates (SE) will refe=

r to THETA(1), the asymptotic sampling distribution of TVCL will be normal =

in model 1 while it will be lognormal in model 2. Therefore if one is inter=

ested in generating random samples from the asymptotic distribution of TVCL=

, some of these samples might be negative in model 1 while they'll remain n=

icely positive in model 2. The same would happen with bounds of (asymptotic=

) confidence intervals: in model 1 the lower bound of a 95% confidence inte=

rval for TVCL might be negative (unrealistic) while it would remain positiv=

e in model 2.

This has obviously no impact for point estimates or even confidence interva=

ls constructed via non-parametric bootstrap since boundary constraints can =

be placed on parameters in NONMEM. But what if one is interested in the asy=

mptotic covariance matrix of estimates returned by $COV? The asymptotic sam=

pling distribution of parameter estimates is (multivariate) normal only if =

the optimization is unconstrained! Doesn't this then speak in favour of mod=

el 2 over model 1? Or does NONMEM take care of it and returns the asymptoti=

c SE of THETA(1) in model 1 on the log-scale (when boundary constraints are=

placed on the parameter)?

Thanks,

Aziz Chaouch

________________________________

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

esseerde(n). Anderen dan de geadresseerde(n) mogen geen gebruik maken van d=

it bericht, het niet openbaar maken of op enige wijze verspreiden of vermen=

igvuldigen. Het UMCG kan niet aansprakelijk gesteld worden voor een incompl=

ete 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.

________________________________

Confidentiality Notice: This message is private and may contain confidentia=

l and proprietary information. If you have received this message in error, =

please notify us and remove it from your system and note that you must not =

copy, distribute or take any action in reliance on it. Any unauthorized use=

or disclosure of the contents of this message is not permitted and may be =

unlawful.

Received on Thu Feb 12 2015 - 07:12:52 EST

Date: Thu, 12 Feb 2015 12:12:52 +0000

Dear all, an alternative which I try to use on strictly positive parameters=

is to estimate on log-scale. Then I think often the assymptitic approximat=

ion is more true and the resulting measures of parameter uncertainty are mo=

re reliable.

BW

Magnus

________________________________________

Från: owner-nmusers

haouch Aziz <Aziz.Chaouch

Skickat: den 12 februari 2015 12:01:51

Till: Eleveld, DJ; pascal.girard

Kopia: nmusers

Ämne: [NMusers] RE: Standard errors of estimates for strictly positive pa=

rameters

Dear Douglas, dear Pascal,

Thanks a lot for your answers. I guess the main point here is constrained v=

s unconstrained optimization as the asymptotic covariance matrix of estimat=

es (as returned by $COV) is "well defined" only in the latter case. When fi=

tting model 1, one would normally constrain THETA(1) to be positive by usin=

g something like:

$THETA

(0, 15, 50) ; TVCL

In this situation I wonder whether it makes sense at all to consider the ou=

tput of $COV. It seems model 2 would be here preferable (unconstrained opti=

mization). If model 1 is fitted without boundary constraints on THETA(1), t=

he covariance matrix of estimates may have "some" meaning but the optimizat=

ion in NONMEM is then likely to crash if it encounters a negative value at =

some point, which again speaks somehow in favor of model 2 (unless one is n=

ot interested in the output of $COV).

Now what about $OMEGA? Here NONMEM knows that these are variances and there=

fore we do not need to explicitly (i.e. manually) place boundary constraint=

s on the diagonal elements of the omega matrix. However something must acco=

unt for it internally. The covariance matrix of estimates returned by $COV =

also contains elements that refer to omega so I'm unsure how these are trea=

ted. For diagonal elements of the omega matrix, does NONMEM optimize log(om=

ega) or omega? Or does it uses a Cholesky decomposition of the Omega matrix=

and optimize elements on that scale? Again, unless the optimization on ome=

ga is unconstrained, can we really trust the output of $COV? Basically the =

question here is how would you construct an asymptotic 95% confidence inter=

val for a diagonal element of Omega (i.e. a variance) based on the informat=

ion from the covariance matrix of estimates?

The covariance matrix of estimate is of importance to me because I'm consid=

ering published studies and I do not have access to the data so I cannot re=

fit the model with an alternative parametrization. Results from $COV (in ls=

t file when available from the authors) is then the only available piece of=

information about the uncertainty of the estimation process.

Kind regards,

Aziz Chaouch

________________________________________

De : Eleveld, DJ [mailto:d.j.eleveld

Envoyé : mercredi, 11. février 2015 22:26

À : Chaouch Aziz; nmusers

Objet : RE: Standard errors of estimates for strictly positive parameters

Hi Aziz,

Just some comments off the top of my head in a quite informal way: I'm not =

really sure that these are the same problem because they dont start with th=

e same information in the form of parameter constraints. In model 1 you are=

asking the optimizer for the unconstrained maximum likelihood solution for=

TVCL. OK, this is reasonable in a lot of situations, but not necessairily =

in all situations.

In model 2 you add information by forcing TVCL and CL to be positive. If yo=

u think of the optimal solution as some point in N-dimensional space which =

has to be searched for, in model 2 you are saying "dont even look in the sp=

ace where TVCL or CL is negative". Even stronger, in model 2 you are also s=

aying "dont even get close to zero" because the log-normal distribution van=

ishes towards zero.

Which solution of these is best for some particular application depends on =

a lot of things. One of the things I would think about in this situation is=

whether or not my a priori beliefs match with the structual constraints of=

the model. Do I really think that the "true" CL could be zero? If yes, the=

n model 2 is hard to defend in that case.

You description of your situation regarding standard errors is a part of th=

e same thing. When you extrapolate standard errors into low-probability are=

as you are checking the boundaries of the probability area. It should not b=

e suprising that model 1 might tell you that CL is negative since this was =

part of the solution space which you allowed. With model 2 your model struc=

ture says "dont even look there"

In short, although these two models might look similar, I think they are re=

ally quite different. This becomes most clear when you consider the low-pro=

bability space.

Sorry for the vauge language.

Warm regards,

Douglas

________________________________________

De : pascal.girard

Envoyé : mercredi, 11. février 2015 18:30

À : Chaouch Aziz; nmusers

Objet : RE: Standard errors of estimates for strictly positive parameters

Dear Aziz,

NM does not return the asymptotic SE of THETA(1) in model 1 on the log-scal=

e. So I would use model 2.

With best regards / Mit freundlichen Grüßen / Cordialement

Pascal

________________________________________

From: owner-nmusers

of Chaouch Aziz [Aziz.Chaouch

Sent: Wednesday, February 11, 2015 5:21 PM

To: nmusers

Subject: [NMusers] Standard errors of estimates for strictly positive param=

eters

Hi,

I'm interested in generating samples from the asymptotic sampling distribut=

ion of population parameter estimates from a published PKPOP model fitted w=

ith NONMEM. By definition, parameter estimates are asymptotically (multivar=

iate) normally distributed (unconstrained optimization) with mean M and cov=

ariance C, where M is the vector of parameter estimates and C is the covari=

ance matrix of estimates (returned by $COV and available in the lst file).

Consider the 2 models below:

Model 1:

TVCL = THETA(1)

CL = TVCL*EXP(ETA(1))

Model 2:

TVCL = EXP(THETA(1))

CL = TVCL*EXP(ETA(1))

It is clear that model 1 and model 2 will provide exactly the same fit. How=

ever, although in both cases the standard error of estimates (SE) will refe=

r to THETA(1), the asymptotic sampling distribution of TVCL will be normal =

in model 1 while it will be lognormal in model 2. Therefore if one is inter=

ested in generating random samples from the asymptotic distribution of TVCL=

, some of these samples might be negative in model 1 while they'll remain n=

icely positive in model 2. The same would happen with bounds of (asymptotic=

) confidence intervals: in model 1 the lower bound of a 95% confidence inte=

rval for TVCL might be negative (unrealistic) while it would remain positiv=

e in model 2.

This has obviously no impact for point estimates or even confidence interva=

ls constructed via non-parametric bootstrap since boundary constraints can =

be placed on parameters in NONMEM. But what if one is interested in the asy=

mptotic covariance matrix of estimates returned by $COV? The asymptotic sam=

pling distribution of parameter estimates is (multivariate) normal only if =

the optimization is unconstrained! Doesn't this then speak in favour of mod=

el 2 over model 1? Or does NONMEM take care of it and returns the asymptoti=

c SE of THETA(1) in model 1 on the log-scale (when boundary constraints are=

placed on the parameter)?

Thanks,

Aziz Chaouch

________________________________

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

esseerde(n). Anderen dan de geadresseerde(n) mogen geen gebruik maken van d=

it bericht, het niet openbaar maken of op enige wijze verspreiden of vermen=

igvuldigen. Het UMCG kan niet aansprakelijk gesteld worden voor een incompl=

ete 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.

________________________________

Confidentiality Notice: This message is private and may contain confidentia=

l and proprietary information. If you have received this message in error, =

please notify us and remove it from your system and note that you must not =

copy, distribute or take any action in reliance on it. Any unauthorized use=

or disclosure of the contents of this message is not permitted and may be =

unlawful.

Received on Thu Feb 12 2015 - 07:12:52 EST