Jump to content


Photo

Modelling with different constant CV

PD model NLME weighting constant CV

  • Please log in to reply
6 replies to this topic

#1 LLLi

LLLi

    Advanced Member

  • Members
  • PipPipPip
  • 92 posts

Posted 08 September 2016 - 09:20 PM

Hi, 

 

I am studying the PD2-Inhibitory Imax model form Dr. Dan Weiner's textbook "PK&PD Data Analysis: concept and application". In dataset, the column0 is concentration and the column1 is blood pressure.

 

In this application the author wants to simulate differences in experimental error (constant CV) after fitting a dataset to PD 108 model (weighting=uniform, the values for Imax, IC50, n and E0 are 34.66, 139.6, 2.031 and 171.4, respectively). So the same dataset was fit with 0.5, 2, 3 or 5% constant CV. 

 

First I tried the PD 108 model with different constant CV but the results were the same for different constant CV. 

 

Then I tried Phoenix NLME. Type=Emax and check baseline, inhibitor and sigmoid. I chose residual error as mutiplicative and froze the SD as 0.005, 0.02, 0.03, and 0.05. But I still got similar results. Furthermore, there is no CV% for each parameter. (The results in the book show different mean estimates and precision CV%). 

 

Could someone help me to figure it out? Please see the attachment. Thank you!

 

LLLi

Attached Files



#2 serge guzy

serge guzy

    Advanced Member

  • Members
  • PipPipPip
  • 485 posts

Posted 09 September 2016 - 04:49 AM

Dear LLLi

Please find the attached project I made and did not see any problem.

First I added an id column and checked the population mode, naïve pool. I know that this sometime helps.

Second I created different models with larger cv up to the one you tested.

I then fit all those and saw that the parameters change a lot when starting with large cv but then start to converge to the same value which makes sense because the cv is so small that the parameters are not affected anymore. However the -2LL is different.

Now in the book, the weighted least square is used as optimizer and not the maximum likelihood as far as I know.

This would explain the difference  because you use a multiplicative error.

I got standard errors for all cases when I am using the population check box and naïve pool.

 

Then to me all seems OK.

 

Simon will try to attach the project. I cannot for some reason

Hope it helps

Best

Serge

 



#3 LLLi

LLLi

    Advanced Member

  • Members
  • PipPipPip
  • 92 posts

Posted 09 September 2016 - 02:43 PM

Dear LLLi

Please find the attached project I made and did not see any problem.

First I added an id column and checked the population mode, naïve pool. I know that this sometime helps.

Second I created different models with larger cv up to the one you tested.

I then fit all those and saw that the parameters change a lot when starting with large cv but then start to converge to the same value which makes sense because the cv is so small that the parameters are not affected anymore. However the -2LL is different.

Now in the book, the weighted least square is used as optimizer and not the maximum likelihood as far as I know.

This would explain the difference  because you use a multiplicative error.

I got standard errors for all cases when I am using the population check box and naïve pool.

 

Then to me all seems OK.

 

Simon will try to attach the project. I cannot for some reason

Hope it helps

Best

Serge

Dear Serge,

 

Thank you for your reply. I have received the project from Simon.

 

I tried to re-run your model but failed and get the following warning message:

Error in Population: 

Variable 'nGam' undefined
Variable 'nE0' undefined
'".\NLME7.exe"' is not recognized as an internal or external command,
operable program or batch file.
Model execution failed or cancelled
 
I found that Naive-pooled was the only method in run options and there were some words "other methods require" below it. I guess I have no licence for the population PK? This is the reason that the modelling failed?
 
According to your information, the reason why I got the similar results from different CV is that WinNonlin uses a different optimizer from the textbook. But I can not understand "This would explain the difference  because you use a multiplicative error." Would you please explain it with more detail?
 
Furthermore, can we use PD library model with different CV? Could you help to check if I set the weighting right in my PD models? In weighting options I chose "user defined". The weight column=column1*constant CV.
 
Thanks,
LLLi


#4 serge guzy

serge guzy

    Advanced Member

  • Members
  • PipPipPip
  • 485 posts

Posted 09 September 2016 - 02:57 PM

Dear LLLi, you should get the population license but I thought that naïve pool should work if you have only the winnonlion license.

Now multiplicative error is one of the option you have with Phoenix.

I am not using old winnonlion. I think you should just go with Phoenix.

Now go to Parameters/Structural and uncheck the check box under ran.

This will remove completely the random effects and rerun one of the models to see if it works.

My version does not require to remove the random effects if you use the naïve Pool. May be the fact you do not have the NLME license creates that problem.

 

About the optimizer, when you use the weighted least square and additive error, my experience is that you get same or very similar results than with the maximum likelihood. This is not the case if the error is multiplicative.

Best

Serge



#5 serge guzy

serge guzy

    Advanced Member

  • Members
  • PipPipPip
  • 485 posts

Posted 09 September 2016 - 03:08 PM

Dear LLLi, you should get the population license but I thought that naïve pool should work if you have only the winnonlion license.

Now multiplicative error is one of the option you have with Phoenix.

I am not using old winnonlion. I think you should just go with Phoenix.

Now go to Parameters/Structural and uncheck the check box under ran.

This will remove completely the random effects and rerun one of the models to see if it works.

My version does not require to remove the random effects if you use the naïve Pool. May be the fact you do not have the NLME license creates that problem.

 

About the optimizer, when you use the weighted least square and additive error, my experience is that you get same or very similar results than with the maximum likelihood. This is not the case if the error is multiplicative.

Best

Serge

ok, About the PD models, in weighing option, use last option which is 1/(Yhat*Yhat). This is multiplicative error.

best

Serge



#6 LLLi

LLLi

    Advanced Member

  • Members
  • PipPipPip
  • 92 posts

Posted 09 September 2016 - 03:40 PM

ok, About the PD models, in weighing option, use last option which is 1/(Yhat*Yhat). This is multiplicative error.

best

Serge

Dear Serge,

 

I remove the random effect and rerun the model. It works! 

 

In NLME, we can freeze the error stdev to a certain value. Can we also freeze stdev in PD library model?

 

Thank you!

LLLi



#7 serge guzy

serge guzy

    Advanced Member

  • Members
  • PipPipPip
  • 485 posts

Posted 09 September 2016 - 03:49 PM

I do not see in the old Winnonlin an option to freeze the standard deviation. You should shift all in Phoenix.

Best

Serge







Also tagged with one or more of these keywords: PD model, NLME, weighting, constant CV

0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users