Jump to content


Photo

Failed to obtain PK parameter CV percents


  • Please log in to reply
11 replies to this topic

#1 sinyinlim

sinyinlim

    Member

  • Members
  • PipPip
  • 20 posts

Posted 24 July 2017 - 05:59 PM

Hello,

 

Can anyone help explain why would I fail to obtain PK parameter CV percents and omega standard error by solely changing the residual variability methods (CEps) from additive to multiplicative or combined?

 

Everything else seemed normal.

 

Thanks!!

 

 
 


#2 Ana Henry

Ana Henry

    Advanced Member

  • Members
  • PipPipPip
  • 220 posts

Posted 24 July 2017 - 06:17 PM

Hello,

 

Can anyone help explain why would I fail to obtain PK parameter CV percents and omega standard error by solely changing the residual variability methods (CEps) from additive to multiplicative or combined?

 

Everything else seemed normal.

 

Thanks!!

 

Hello,

The standard errors are calculated as a second step after fitting the model.  Standard errors for the parameters are computed using the Hessian matrix.  Changing the model residual error model will change the hessian matrix because you have different parameters (sigma is a parameter to be estimated) and in some cases the Hessian cannot be resolved.  Try going to the run options tab and changing the error estimation method to another method (e.g., Sandwich or Fisher).  It should not affect your fit.

 

Ana



#3 sinyinlim

sinyinlim

    Member

  • Members
  • PipPip
  • 20 posts

Posted 24 July 2017 - 10:52 PM

Hello,

The standard errors are calculated as a second step after fitting the model.  Standard errors for the parameters are computed using the Hessian matrix.  Changing the model residual error model will change the hessian matrix because you have different parameters (sigma is a parameter to be estimated) and in some cases the Hessian cannot be resolved.  Try going to the run options tab and changing the error estimation method to another method (e.g., Sandwich or Fisher).  It should not affect your fit.

 

Ana

Thanks for your help, Ana. Changing the error estimation method could not fix the problem. Had to reduce the number of parameter estimations.. 



#4 Simon Davis

Simon Davis

    Advanced Member

  • Administrators
  • 1,119 posts

Posted 31 July 2017 - 10:47 AM

Sinyinlim - did you try restarting your model using the final estimates from the previous run; quite often that helps me get CV% calculated sucessfully.  Swapping to QRPEM may also help - but sometimes with more complicated models you still won't get them unfortunately as you've seen.

 

 Simon.



#5 sinyinlim

sinyinlim

    Member

  • Members
  • PipPip
  • 20 posts

Posted 03 August 2017 - 06:08 PM

Hi Simon,

 

I tried swapping to QRPEM, it did calculate the CV% successfully, but like you said, not always. Thanks for the helpful comment!!



#6 Charvi

Charvi

    Member

  • Members
  • PipPip
  • 19 posts

Posted 23 August 2018 - 06:32 PM

Hello,

 

I am facing similar issues - such that no CV% are obtained in FOCE-ELS method, but they are obtained in QPREM.

I even used the estimates obtained from QPREM as my initial estimates for parameters while applying the FOCE-ELS and tried changing the method as well - Sandwich vs. Hessian vs. Fisher

Can you please share insights into what could be the issue?

Is there a particular error model that does not seem to work for FOCE ELS?

Can you please share some experiences and solutions?

 

Many Thanks,

Charvi



#7 bwendt@certara.com

bwendt@certara.com

    Advanced Member

  • Administrators
  • 188 posts

Posted 24 August 2018 - 09:09 AM

Hi Charvi,

 

as Ana noted in her first reply, standard errors are computed using the Hessian matrix and this computation might fail.  This is common in any regression software when you need to invert a matrix. The hessian might not be obtainable and then you cannot compute standard error. More often than not the model is overparametrized/unstable.
You can also try to change the error estimation method on the Run Options tab by using forward rather than central where forward is faster but less precise than central. Apart from that you typically start with hessian, then move to sandwich and last try fisher. If none of this works, then use the Bootstrap method.
 
Bernd


#8 Charvi

Charvi

    Member

  • Members
  • PipPip
  • 19 posts

Posted 24 August 2018 - 04:28 PM

Hi Bernd,

 

Thank you for your response, this helps.

 

Regards.

Charvi



#9 Charvi

Charvi

    Member

  • Members
  • PipPip
  • 19 posts

Posted 10 September 2018 - 06:45 PM

Hi All.

 

On a related note - as I understand QPREM does not consider bounds on fixed effects. In cases where bounds are necessary for certain parameters such as bioavailability, fraction etc. - is it not recommended to use it? Or is there any other way a fixed effect can be constrained while using QPREM?

 

Thanks,
Charvi
 



#10 bwendt@certara.com

bwendt@certara.com

    Advanced Member

  • Administrators
  • 188 posts

Posted 12 September 2018 - 09:51 AM

Hi Charvi,

 

you can use the ilogit transform to constrain anything between 0 and 1. Here is the PML to constrain bioavailability, F:

 

dosepoint(Aa, bioavail=F)

F=ilogit(Flogit)

stparm(Flogit=tvFlogit * exp(nFlogit))

fixef(tvFlogit = c(,0,))

ranef(diag(nFlogit)=c(0.1))



#11 Charvi

Charvi

    Member

  • Members
  • PipPip
  • 19 posts

Posted 12 September 2018 - 04:30 PM

Thanks Bernd for your response.

This helps.

 

Regards,

Charvi



#12 1711197384@qq.com

1711197384@qq.com

    Newbie

  • Members
  • Pip
  • 2 posts

Posted 20 August 2020 - 07:35 AM

 

Hi Charvi,

 

as Ana noted in her first reply, standard errors are computed using the Hessian matrix and this computation might fail.  This is common in any regression software when you need to invert a matrix. The hessian might not be obtainable and then you cannot compute standard error. More often than not the model is overparametrized/unstable.
You can also try to change the error estimation method on the Run Options tab by using forward rather than central where forward is faster but less precise than central. Apart from that you typically start with hessian, then move to sandwich and last try fisher. If none of this works, then use the Bootstrap method.
 
Bernd

 hi, Bernd

I have some questions, if you can help me solve it, I would be very grateful.

1.because fisher's  accuracy is not high, the CV% will become high, so i don't want to use it. so when I am writing an article, can I use cv% in bootstrap as the CV% in the final model.

2.after stepwise program, i got the covariates and the got final model with covariates, i found i can't obtain CV% and  95CI in the final model, so what can i do?

3. If I change the algorithm (such as QRPEM) now in the final model,is it correct? 

4.in the final model, if i can "accpet all fixed and random", and the run again until obtained the CV%.

Many Thanks,

Leeking.

 

 






0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users