Dear nmusers,
In one of my PK analysis, if I chose Y=F+F*ERR(1)+ERR(2), I got lower
objective value=1130.907, but $COV step failed (S matrix is singular);
when I chose Y=F+ERR(1), I got $COV successful but higher objective
value=1351.735. I prefer to choose the addivtive error model even if with
higher objective value because it seems comfortable to see $COV
successful. Can anyone give me other suggestions?
Thanks
Yuanyue (Paul) Gao
School of Pharmacy
University of Pittsburgh
716 Salk Hall
3501 Terrace Street
Pittsburgh, PA 15261
Phone: 412-648-8546
E-mail: [email protected]
CCV or Additive error models?
Yuanyue,
It is has been well established and discussed at length on nmusers that a successful $COV step says nothing about the suitability of the model to describe the data.
Quoted reply history
On the other hand the OFV is a very reliable statistic for model selection. You should use the combined error model.
Nick
On 1/07/2011 5:17 p.m., [email protected] wrote:
> Dear nmusers,
>
> In one of my PK analysis, if I chose Y=F+F*ERR(1)+ERR(2), I got lower
> objective value=1130.907, but $COV step failed (S matrix is singular);
> when I chose Y=F+ERR(1), I got $COV successful but higher objective
> value=1351.735. I prefer to choose the addivtive error model even if with
> higher objective value because it seems comfortable to see $COV
> successful. Can anyone give me other suggestions?
>
> Thanks
>
> Yuanyue (Paul) Gao
>
> School of Pharmacy
> University of Pittsburgh
> 716 Salk Hall
> 3501 Terrace Street
> Pittsburgh, PA 15261
> Phone: 412-648-8546
> E-mail: [email protected]
--
Nick Holford, Professor Clinical Pharmacology
Dept Pharmacology& Clinical Pharmacology
University of Auckland,85 Park Rd,Private Bag 92019,Auckland,New Zealand
tel:+64(9)923-6730 fax:+64(9)373-7090 mobile:+64(21)46 23 53
email: [email protected]
http://www.fmhs.auckland.ac.nz/sms/pharmacology/holford
Dear Yuanyue,
If this is a standard PK analysis I would first like to investigate what a
simple CCV model does: Y=F+F*ERR(1) or Y=F*EXP(ERR(1)), as this would seem
to me the default for any model derived on concentration data using standard
assay technology.
If you know you have low epsilon shrinkage (look for the line saying
EPSshrink(%) in your output, and trust Mats and Rada* that you need to stay
well below 20%) you could make a diagnostic plot of the absolute values of
your individual weighted residuals against PRED or IPRED. If you have what
could be judged as a horizontal smooth through the data, your model is fine.
If you have pronounced higher values at the low end when you use a constant
CV model, you add on the additive component. My guess would be that with
your additive-only model you would see an increase in your IWRES values with
increasing PRED/IPRED.
No matter what Nick says, I still prefer a succesful covariance step, as
lack of one can be an indication of an over-parameterised model. I do know
that almost all the biologically relevant models struggle with
over-parameterisation, but given a choice I would rather not
(unnecessarily?) have this flaw in my residual error model.
Cheers,
Rik
*MO Karlsson and RM Savic. Diagnosing model diagnostics. CPT 2007
82:1:17-20.
Quoted reply history
-----Original Message-----
From: [email protected] [mailto:[email protected]] On
Behalf Of Nick Holford
Sent: 01 July 2011 6:11 PM
To: nmusers
Subject: Re: [NMusers] CCV or Additive error models?
Yuanyue,
It is has been well established and discussed at length on nmusers that a
successful $COV step says nothing about the suitability of the model to
describe the data.
On the other hand the OFV is a very reliable statistic for model selection.
You should use the combined error model.
Nick
On 1/07/2011 5:17 p.m., [email protected] wrote:
> Dear nmusers,
>
> In one of my PK analysis, if I chose Y=F+F*ERR(1)+ERR(2), I got lower
> objective value=1130.907, but $COV step failed (S matrix is singular);
> when I chose Y=F+ERR(1), I got $COV successful but higher objective
> value=1351.735. I prefer to choose the addivtive error model even if
> with higher objective value because it seems comfortable to see $COV
> successful. Can anyone give me other suggestions?
>
> Thanks
>
>
> Yuanyue (Paul) Gao
>
> School of Pharmacy
> University of Pittsburgh
> 716 Salk Hall
> 3501 Terrace Street
> Pittsburgh, PA 15261
> Phone: 412-648-8546
> E-mail: [email protected]
>
--
Nick Holford, Professor Clinical Pharmacology Dept Pharmacology& Clinical
Pharmacology University of Auckland,85 Park Rd,Private Bag
92019,Auckland,New Zealand
tel:+64(9)923-6730 fax:+64(9)373-7090 mobile:+64(21)46 23 53
email: [email protected]
http://www.fmhs.auckland.ac.nz/sms/pharmacology/holford
Rik,
What standard concentration assay technology does not have background noise?
If you are going to decide on a residual error model based on mechanism then you should at least start by default with an additive term.
You don't have to look at shrinkage, residual plots, or standard errors to test different residual error models. The OBJ is all you need.
Nick
Quoted reply history
On 4/07/2011 8:55 a.m., Rik Schoemaker wrote:
> Dear Yuanyue,
>
> If this is a standard PK analysis I would first like to investigate what a
> simple CCV model does: Y=F+F*ERR(1) or Y=F*EXP(ERR(1)), as this would seem
> to me the default for any model derived on concentration data using standard
> assay technology.
>
> If you know you have low epsilon shrinkage (look for the line saying
> EPSshrink(%) in your output, and trust Mats and Rada* that you need to stay
> well below 20%) you could make a diagnostic plot of the absolute values of
> your individual weighted residuals against PRED or IPRED. If you have what
> could be judged as a horizontal smooth through the data, your model is fine.
> If you have pronounced higher values at the low end when you use a constant
> CV model, you add on the additive component. My guess would be that with
> your additive-only model you would see an increase in your IWRES values with
> increasing PRED/IPRED.
>
> No matter what Nick says, I still prefer a succesful covariance step, as
> lack of one can be an indication of an over-parameterised model. I do know
> that almost all the biologically relevant models struggle with
> over-parameterisation, but given a choice I would rather not
> (unnecessarily?) have this flaw in my residual error model.
>
> Cheers,
>
> Rik
>
> *MO Karlsson and RM Savic. Diagnosing model diagnostics. CPT 2007
> 82:1:17-20.
>
> -----Original Message-----
> From: [email protected] [mailto:[email protected]] On
> Behalf Of Nick Holford
> Sent: 01 July 2011 6:11 PM
> To: nmusers
> Subject: Re: [NMusers] CCV or Additive error models?
>
> Yuanyue,
>
> It is has been well established and discussed at length on nmusers that a
> successful $COV step says nothing about the suitability of the model to
> describe the data.
>
> On the other hand the OFV is a very reliable statistic for model selection.
> You should use the combined error model.
>
> Nick
>
> On 1/07/2011 5:17 p.m., [email protected] wrote:
>
> > Dear nmusers,
> >
> > In one of my PK analysis, if I chose Y=F+F*ERR(1)+ERR(2), I got lower
> > objective value=1130.907, but $COV step failed (S matrix is singular);
> > when I chose Y=F+ERR(1), I got $COV successful but higher objective
> > value=1351.735. I prefer to choose the addivtive error model even if
> > with higher objective value because it seems comfortable to see $COV
> > successful. Can anyone give me other suggestions?
> >
> > Thanks
> >
> > Yuanyue (Paul) Gao
> >
> > School of Pharmacy
> > University of Pittsburgh
> > 716 Salk Hall
> > 3501 Terrace Street
> > Pittsburgh, PA 15261
> > Phone: 412-648-8546
> > E-mail: [email protected]
>
> --
> Nick Holford, Professor Clinical Pharmacology Dept Pharmacology& Clinical
> Pharmacology University of Auckland,85 Park Rd,Private Bag
> 92019,Auckland,New Zealand
> tel:+64(9)923-6730 fax:+64(9)373-7090 mobile:+64(21)46 23 53
> email: [email protected]
> http://www.fmhs.auckland.ac.nz/sms/pharmacology/holford
--
Nick Holford, Professor Clinical Pharmacology
Dept Pharmacology& Clinical Pharmacology
University of Auckland,85 Park Rd,Private Bag 92019,Auckland,New Zealand
tel:+64(9)923-6730 fax:+64(9)373-7090 mobile:+64(21)46 23 53
email: [email protected]
http://www.fmhs.auckland.ac.nz/sms/pharmacology/holford