Dear all,
I have some population PK data which are in general very sparse (95% have only
1 blood sample between 2 successive doses). I developed a population PK model
with the one-compartment model with 1st order absorption. The progress is
generally okay except that whenever a random effect, i.e. *(1+ETA(1)), is used
to describe distribution of Vd, OMEGA would be estimated to be very large
(around 45% in terms of CV, with 80% Shrinkage), despite statistical
significance (dOF approx. -5.5). So I dropped the random effect and expressed
Vd in terms of a single fixed effect. When the final model has come out, I
performed bootstrap and found that most estimates are accurate except Vd, which
has a very large standard error and bias (mean 232, bias 49, SE 156), while the
estimates for CL and other parameters look normal. I then constructed the
predictive plots for the developed model using both the original estimates
(i.e. estimates using my original dataset) (#1) and estimates from one of the
bootstrap runs which has an extreme estimate of Vd (9xx) (#2), and found out
that the two plots of plasma profiles are quite different in terms of the shape
(#1 is "taller", #2 is much flatter) but have similar average Cp.
These seem to be suggesting that given my sparse data, it is impossible to
require accurate estimations of both CL and Vd. Apart from fixing Vd to a fixed
value, is there any other possible solutions? Or is there anything that I might
have overlooked?
Thanks and regards,
Matthew
Large errors in the estimation of volume of distribution (Vd) for sparse data
Hi Mathew,
Have you tried using an exponential model for vd ? like this: Vd =
TEHTA(1)*EXP(ETA(1))
Ahmad.
Quoted reply history
From: [email protected] [mailto:[email protected]] On
Behalf Of HUI, Ka Ho
Sent: Tuesday, 10 November 2015 1:13 AM
To: [email protected]
Subject: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Dear all,
I have some population PK data which are in general very sparse (95% have only
1 blood sample between 2 successive doses). I developed a population PK model
with the one-compartment model with 1st order absorption. The progress is
generally okay except that whenever a random effect, i.e. *(1+ETA(1)), is used
to describe distribution of Vd, OMEGA would be estimated to be very large
(around 45% in terms of CV, with 80% Shrinkage), despite statistical
significance (dOF approx. -5.5). So I dropped the random effect and expressed
Vd in terms of a single fixed effect. When the final model has come out, I
performed bootstrap and found that most estimates are accurate except Vd, which
has a very large standard error and bias (mean 232, bias 49, SE 156), while the
estimates for CL and other parameters look normal. I then constructed the
predictive plots for the developed model using both the original estimates
(i.e. estimates using my original dataset) (#1) and estimates from one of the
bootstrap runs which has an extreme estimate of Vd (9xx) (#2), and found out
that the two plots of plasma profiles are quite different in terms of the shape
(#1 is "taller", #2 is much flatter) but have similar average Cp.
These seem to be suggesting that given my sparse data, it is impossible to
require accurate estimations of both CL and Vd. Apart from fixing Vd to a fixed
value, is there any other possible solutions? Or is there anything that I might
have overlooked?
Thanks and regards,
Matthew
Thanks for your responses!
Nitin, I encountered an error when generating VPC by PsN. It says "No DV values
found after filtering original data. At lib/tool/npc.subs.pm line 2215." What
does it mean?
Felix, Past published data suggested similar parameter estimates and models
compared to my final model. This is PO and I fixed Ka at a pre-estimated value
(So no estimation of fixed or random effect).
Ahmad, Yes. The CV is even larger.
Matthew
Quoted reply history
From: Abu Helwa, Ahmad Yousef Mohammad - abuay010
[mailto:[email protected]]
Sent: Tuesday, November 10, 2015 5:34 AM
To: HUI, Ka Ho <[email protected]>; [email protected]
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Hi Mathew,
Have you tried using an exponential model for vd ? like this: Vd =
TEHTA(1)*EXP(ETA(1))
Ahmad.
From: felix boakye-agyeman [mailto:[email protected]]
Sent: Tuesday, November 10, 2015 12:41 AM
To: HUI, Ka Ho <[email protected]>
Subject: Re: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Hello,
Do you have historical data to compare you data to? (Do you know if you are
hitting a local minimum)
Is this iv or po, if its po how is your Ka?
You may also be over-parameterized due to your data
From: Kaila, Nitin [mailto:[email protected]]
Sent: Tuesday, November 10, 2015 12:14 AM
To: HUI, Ka Ho <[email protected]>
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Matthew.
Construct visual predictive check (VPC) plots, using all the estimates of the
bootstrap runs, as that will be a more true estimate of overall variability in
the Cp predictions.
Use the -rawres option in PsN to perform the VPC, and then compare your
original final model VPC plot with the VPC plot with all estimates of the
bootstrap.
Nitin
From: [email protected]<mailto:[email protected]>
[mailto:[email protected]] On Behalf Of HUI, Ka Ho
Sent: Monday, November 9, 2015 9:43 AM
To: [email protected]<mailto:[email protected]>
Subject: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Dear all,
I have some population PK data which are in general very sparse (95% have only
1 blood sample between 2 successive doses). I developed a population PK model
with the one-compartment model with 1st order absorption. The progress is
generally okay except that whenever a random effect, i.e. *(1+ETA(1)), is used
to describe distribution of Vd, OMEGA would be estimated to be very large
(around 45% in terms of CV, with 80% Shrinkage), despite statistical
significance (dOF approx. -5.5). So I dropped the random effect and expressed
Vd in terms of a single fixed effect. When the final model has come out, I
performed bootstrap and found that most estimates are accurate except Vd, which
has a very large standard error and bias (mean 232, bias 49, SE 156), while the
estimates for CL and other parameters look normal. I then constructed the
predictive plots for the developed model using both the original estimates
(i.e. estimates using my original dataset) (#1) and estimates from one of the
bootstrap runs which has an extreme estimate of Vd (9xx) (#2), and found out
that the two plots of plasma profiles are quite different in terms of the shape
(#1 is "taller", #2 is much flatter) but have similar average Cp.
These seem to be suggesting that given my sparse data, it is impossible to
require accurate estimations of both CL and Vd. Apart from fixing Vd to a fixed
value, is there any other possible solutions? Or is there anything that I might
have overlooked?
Thanks and regards,
Matthew
Hi Matthew,
Very large standard error and bias of Vd suggest that Vd is not well
identified. Or in other word, your data didn't contain sufficient information
to fit Vd. Loosely speaking it is a problem of over-parameterization, because
you have only one measurement point, but you try to fit 2 parameters (Vd and
clearance).
Zheng
Quoted reply history
________________________________
From: [email protected] <[email protected]> on behalf of
HUI, Ka Ho <[email protected]>
Sent: Tuesday, 10 November 2015 3:12 PM
To: Kaila, Nitin; Abu Helwa, Ahmad Yousef Mohammad - abuay010; felix
boakye-agyeman; [email protected]
Subject: [NMusers] RE: Large errors in the estimation of volume of distribution
(Vd) for sparse data
Thanks for your responses!
Nitin, I encountered an error when generating VPC by PsN. It says “No DV values
found after filtering original data. At lib/tool/npc.subs.pm line 2215.” What
does it mean?
Felix, Past published data suggested similar parameter estimates and models
compared to my final model. This is PO and I fixed Ka at a pre-estimated value
(So no estimation of fixed or random effect).
Ahmad, Yes. The CV is even larger.
Matthew
From: Abu Helwa, Ahmad Yousef Mohammad - abuay010
[mailto:[email protected]]
Sent: Tuesday, November 10, 2015 5:34 AM
To: HUI, Ka Ho <[email protected]>; [email protected]
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Hi Mathew,
Have you tried using an exponential model for vd ? like this: Vd =
TEHTA(1)*EXP(ETA(1))
Ahmad.
From: felix boakye-agyeman [mailto:[email protected]]
Sent: Tuesday, November 10, 2015 12:41 AM
To: HUI, Ka Ho <[email protected]>
Subject: Re: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Hello,
Do you have historical data to compare you data to? (Do you know if you are
hitting a local minimum)
Is this iv or po, if its po how is your Ka?
You may also be over-parameterized due to your data
From: Kaila, Nitin [mailto:[email protected]]
Sent: Tuesday, November 10, 2015 12:14 AM
To: HUI, Ka Ho <[email protected]>
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Matthew.
Construct visual predictive check (VPC) plots, using all the estimates of the
bootstrap runs, as that will be a more true estimate of overall variability in
the Cp predictions.
Use the –rawres option in PsN to perform the VPC, and then compare your
original final model VPC plot with the VPC plot with all estimates of the
bootstrap.
Nitin
From: [email protected]<mailto:[email protected]>
[mailto:[email protected]] On Behalf Of HUI, Ka Ho
Sent: Monday, November 9, 2015 9:43 AM
To: [email protected]<mailto:[email protected]>
Subject: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Dear all,
I have some population PK data which are in general very sparse (95% have only
1 blood sample between 2 successive doses). I developed a population PK model
with the one-compartment model with 1st order absorption. The progress is
generally okay except that whenever a random effect, i.e. *(1+ETA(1)), is used
to describe distribution of Vd, OMEGA would be estimated to be very large
(around 45% in terms of CV, with 80% Shrinkage), despite statistical
significance (dOF approx. -5.5). So I dropped the random effect and expressed
Vd in terms of a single fixed effect. When the final model has come out, I
performed bootstrap and found that most estimates are accurate except Vd, which
has a very large standard error and bias (mean 232, bias 49, SE 156), while the
estimates for CL and other parameters look normal. I then constructed the
predictive plots for the developed model using both the original estimates
(i.e. estimates using my original dataset) (#1) and estimates from one of the
bootstrap runs which has an extreme estimate of Vd (9xx) (#2), and found out
that the two plots of plasma profiles are quite different in terms of the shape
(#1 is “taller”, #2 is much flatter) but have similar average Cp.
These seem to be suggesting that given my sparse data, it is impossible to
require accurate estimations of both CL and Vd. Apart from fixing Vd to a fixed
value, is there any other possible solutions? Or is there anything that I might
have overlooked?
Thanks and regards,
Matthew
Regarding the error message from PsN vpc: I can see from the message
that you are using a *very* old version of PsN. I suggest that you
install the latest version and try again.
Best regards,
Kajsa Harling
On 11/10/2015 05:12 AM, HUI, Ka Ho
wrote:
Thanks for your responses!
Nitin, I encountered an
error when generating VPC by PsN. It says “No DV values
found after filtering original data.
At lib/tool/npc.subs.pm line 2215.” What does
it mean?
Felix, Past published data suggested similar
parameter estimates and models compared to my final model.
This is PO and I fixed Ka at a pre-estimated value (So no
estimation of fixed or random effect).
Ahmad, Yes. The CV is even larger.
Matthew
Quoted reply history
From: Abu Helwa, Ahmad
Yousef Mohammad - abuay010
[ mailto: [email protected] ]
Sent: Tuesday, November 10, 2015 5:34 AM
To: HUI, Ka Ho < [email protected] > ;
[email protected]
Subject: RE: Large errors in the estimation of volume
of distribution (Vd) for sparse data
Hi Mathew,
Have you tried using an exponential model for
vd ? like this: Vd = TEHTA(1)*EXP(ETA(1))
Ahmad.
From: felix boakye-agyeman
[ mailto: [email protected] ]
Sent: Tuesday, November 10, 2015 12:41 AM
To: HUI, Ka Ho < [email protected] >
Subject: Re: [NMusers] Large errors in the estimation
of volume of distribution (Vd) for sparse data
Hello,
Do you have
historical data to compare you data to? (Do you know if you
are hitting a local minimum)
Is this iv or po, if
its po how is your Ka?
You may also be
over-parameterized due to your data
From: Kaila, Nitin
[ mailto: [email protected] ]
Sent: Tuesday, November 10, 2015 12:14 AM
To: HUI, Ka Ho
< [email protected] >
Subject: RE: Large errors in the estimation of
volume of distribution (Vd) for sparse data
Matthew.
Construct
visual predictive check (VPC) plots, using all the estimates
of the bootstrap runs, as that will be a more true estimate
of overall variability in the Cp predictions.
Use the
–rawres option in PsN to perform the VPC, and then compare
your original final model VPC plot with the VPC plot with
all estimates of the bootstrap.
Nitin
From:
[email protected]
[ mailto: [email protected] ]
On Behalf Of HUI, Ka Ho
Sent: Monday, November 9, 2015 9:43 AM
To: [email protected]
Subject: [NMusers] Large errors in the estimation
of volume of distribution (Vd) for sparse data
Dear all,
I have some population
PK data which are in general very sparse (95% have only 1
blood sample between 2 successive doses). I developed a
population PK model with the one-compartment model with 1st
order absorption. The progress is generally okay except that
whenever a random effect, i.e. *(1+ETA(1)), is used to
describe distribution of Vd, OMEGA would be estimated to be
very large (around 45% in terms of CV, with 80% Shrinkage),
despite statistical significance (dOF approx. -5.5). So I
dropped the random effect and expressed Vd in terms of a
single fixed effect. When the final model has come out, I
performed bootstrap and found that most estimates are
accurate except Vd, which has a very large standard error
and bias (mean 232, bias 49, SE 156), while the estimates
for CL and other parameters look normal. I then constructed
the predictive plots for the developed model using both the
original estimates (i.e. estimates using my original
dataset) (#1) and estimates from one of the bootstrap runs
which has an extreme estimate of Vd (9xx) (#2), and found
out that the two plots of plasma profiles are quite
different in terms of the shape (#1 is “taller”, #2 is much
flatter) but have similar average Cp.
These seem to be
suggesting that given my sparse data, it is impossible to
require accurate estimations of both CL and Vd. Apart from
fixing Vd to a fixed value, is there any other possible
solutions? Or is there anything that I might have
overlooked?
Thanks and regards,
Matthew
Thank you Zheng. Yes most subjects have only 1 point between doses. I was just
wondering if there are techniques that may improve the results.
And thank you Kajsa. Yes I updated my PsN and tried again. It still did not
work at first but later I spotted some coding mistakes in my control file.
Under $INPUT block I used CP=DV instead DV, which caused the error.
Thanks and regards,
Matthew
Quoted reply history
From: Kajsa Harling [mailto:[email protected]]
Sent: Monday, November 16, 2015 9:41 PM
To: HUI, Ka Ho <[email protected]>; [email protected]
Cc: Kaila, Nitin <[email protected]>; Abu Helwa, Ahmad Yousef Mohammad -
abuay010 <[email protected]>; felix boakye-agyeman
<[email protected]>
Subject: Re: [NMusers] RE: Large errors in the estimation of volume of
distribution (Vd) for sparse data
Regarding the error message from PsN vpc: I can see from the message that you
are using a *very* old version of PsN. I suggest that you install the latest
version and try again.
Best regards,
Kajsa Harling
From: Zheng Liu [mailto:[email protected]]
Sent: Monday, November 16, 2015 1:08 PM
To: HUI, Ka Ho <[email protected]>; Kaila, Nitin
<[email protected]>; Abu Helwa, Ahmad Yousef Mohammad - abuay010
<[email protected]>; felix boakye-agyeman
<[email protected]>; [email protected]
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Hi Matthew,
Very large standard error and bias of Vd suggest that Vd is not well
identified. Or in other word, your data didn't contain sufficient information
to fit Vd. Loosely speaking it is a problem of over-parameterization, because
you have only one measurement point, but you try to fit 2 parameters (Vd and
clearance).
Zheng
On 11/10/2015 05:12 AM, HUI, Ka Ho wrote:
Thanks for your responses!
Nitin, I encountered an error when generating VPC by PsN. It says "No DV values
found after filtering original data. At lib/tool/npc.subs.pm line 2215." What
does it mean?
Felix, Past published data suggested similar parameter estimates and models
compared to my final model. This is PO and I fixed Ka at a pre-estimated value
(So no estimation of fixed or random effect).
Ahmad, Yes. The CV is even larger.
Matthew
From: Abu Helwa, Ahmad Yousef Mohammad - abuay010
[mailto:[email protected]]
Sent: Tuesday, November 10, 2015 5:34 AM
To: HUI, Ka Ho
<[email protected]><mailto:[email protected]>;
[email protected]<mailto:[email protected]>
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Hi Mathew,
Have you tried using an exponential model for vd ? like this: Vd =
TEHTA(1)*EXP(ETA(1))
Ahmad.
From: felix boakye-agyeman [mailto:[email protected]]
Sent: Tuesday, November 10, 2015 12:41 AM
To: HUI, Ka Ho
<[email protected]><mailto:[email protected]>
Subject: Re: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Hello,
Do you have historical data to compare you data to? (Do you know if you are
hitting a local minimum)
Is this iv or po, if its po how is your Ka?
You may also be over-parameterized due to your data
From: Kaila, Nitin [mailto:[email protected]]
Sent: Tuesday, November 10, 2015 12:14 AM
To: HUI, Ka Ho
<[email protected]><mailto:[email protected]>
Subject: RE: Large errors in the estimation of volume of distribution (Vd) for
sparse data
Matthew.
Construct visual predictive check (VPC) plots, using all the estimates of the
bootstrap runs, as that will be a more true estimate of overall variability in
the Cp predictions.
Use the -rawres option in PsN to perform the VPC, and then compare your
original final model VPC plot with the VPC plot with all estimates of the
bootstrap.
Nitin
From: [email protected]<mailto:[email protected]>
[mailto:[email protected]] On Behalf Of HUI, Ka Ho
Sent: Monday, November 9, 2015 9:43 AM
To: [email protected]<mailto:[email protected]>
Subject: [NMusers] Large errors in the estimation of volume of distribution
(Vd) for sparse data
Dear all,
I have some population PK data which are in general very sparse (95% have only
1 blood sample between 2 successive doses). I developed a population PK model
with the one-compartment model with 1st order absorption. The progress is
generally okay except that whenever a random effect, i.e. *(1+ETA(1)), is used
to describe distribution of Vd, OMEGA would be estimated to be very large
(around 45% in terms of CV, with 80% Shrinkage), despite statistical
significance (dOF approx. -5.5). So I dropped the random effect and expressed
Vd in terms of a single fixed effect. When the final model has come out, I
performed bootstrap and found that most estimates are accurate except Vd, which
has a very large standard error and bias (mean 232, bias 49, SE 156), while the
estimates for CL and other parameters look normal. I then constructed the
predictive plots for the developed model using both the original estimates
(i.e. estimates using my original dataset) (#1) and estimates from one of the
bootstrap runs which has an extreme estimate of Vd (9xx) (#2), and found out
that the two plots of plasma profiles are quite different in terms of the shape
(#1 is "taller", #2 is much flatter) but have similar average Cp.
These seem to be suggesting that given my sparse data, it is impossible to
require accurate estimations of both CL and Vd. Apart from fixing Vd to a fixed
value, is there any other possible solutions? Or is there anything that I might
have overlooked?
Thanks and regards,
Matthew