Jump to content


Photo

Select residual error model and VPC plot with negative value

VPC error model

  • Please log in to reply
4 replies to this topic

#1 joybaker

joybaker

    Advanced Member

  • Members
  • PipPipPip
  • 45 posts

Posted 19 November 2022 - 03:32 AM

Hello all, I use additive\ multiplicative\add&mul error models to evaluate my base model. All error models can fit OK, but when I check the  VPC results, the 5%/95% bounds under/overestimate the observed data for additive and add&mul error models.

 

So my question is whether my VPC setting for these two error models is right.

Can I do something to improve it.

What can lead to the negative values in this study (but not other studies)?

 

 

The file only shows one study, and I fit the same model to the whole studies. Actually the VPC problem only happened for this one. And overall the add&mul error model can reduce AIC significantly (~100 compared with multiplicative error model, and even better than additive) overall and give reasonable VPC to all studies except this one, which makes it a dilema, how can I select the final error model, as mentioning a bad VPC for one study is not a strong reason to select multiplicative over the mixed error model. This study represents 20% of the data points, 10% of the subjects, and the only food effect study.  

 

Thanks

Attached Thumbnails

  • VPC.png

Attached Files



#2 Simon Davis

Simon Davis

    Advanced Member

  • Administrators
  • 1,316 posts

Posted 19 November 2022 - 05:13 AM

Hi Joy,

 

First think I did was plot your data and would also want to double check that you have the correct units for the DV column you used as a lot (96/300) of your data is below your BQL of 0.000868 with peaks only 10x that.  Is that correct ?

why do you think that food is affecting Ka, I'm not sure I see it in these data, maybe a tlag or F would be better.
 

I think there is a lot of variability in your data, and therefore the VPC is telling you to be cautious as you can't be that confident of the parameter estimates, what sort of CV% were you getting in your fitting of a 2com model?  are you sure 1 com would not be sufficient, maybe if you have data from other studies it would be more informative to fit them all in one model?

 

  Simon.

Attached Thumbnails

  • joy_abs.jpg
  • Joy_BQL.jpg

Attached Files



#3 joybaker

joybaker

    Advanced Member

  • Members
  • PipPipPip
  • 45 posts

Posted 19 November 2022 - 12:35 PM

Hi Simon,

 

Thank you for the reply. Yes indeed, the whole data made me chose the current model that it may not be a good model for this data set. The BQL is right, I have changed the dataset a little bit to ignore some of the BQL (see attached). I think this has dragged the attention for my question a little bit. 

What I don't understand is the VPC difference between multiplicative/additive/mixed error model. I understand your point that the variability of the parameter estimation makes the VPC like this. But why the multiplicative error model gives a less exaggerate prediction than additive and mixed error model. For example, the mul erro model dose not give very negative VPC value.

Hi Joy,

 

First think I did was plot your data and would also want to double check that you have the correct units for the DV column you used as a lot (96/300) of your data is below your BQL of 0.000868 with peaks only 10x that.  Is that correct ?

why do you think that food is affecting Ka, I'm not sure I see it in these data, maybe a tlag or F would be better.
 

I think there is a lot of variability in your data, and therefore the VPC is telling you to be cautious as you can't be that confident of the parameter estimates, what sort of CV% were you getting in your fitting of a 2com model?  are you sure 1 com would not be sufficient, maybe if you have data from other studies it would be more informative to fit them all in one model?

 

  Simon.

 

Attached Thumbnails

  • mul.png

Attached Files



#4 Simon Davis

Simon Davis

    Advanced Member

  • Administrators
  • 1,316 posts

Posted 21 November 2022 - 08:15 AM

Joy, I think it is because with multiplicative the error is proportional to the conc, whereas additive is assumin uniform.  take a look at this post and perhaps that clarifies it for you

 

https://support.cert...-option/?p=3116



#5 joybaker

joybaker

    Advanced Member

  • Members
  • PipPipPip
  • 45 posts

Posted 23 November 2022 - 10:20 AM

Hi Simon,

 

Thanks for the link, I read through and learn posts with similar topics. I still have two questions to make sure I am on the right track to understand them.

 

In this post (CMultStdev - Phoenix WNL basics - Certara Forums) why "you cannot use more than one epsilon in each observe statement" for a Add+Mul error model? Just like Y=F +EPS(1) + F*EPS(2), to report eps1 for additive and eps2 for multiplicative varible and their corresponding sigma1 and sigma2. 

 

So based on the answers, are Add+Mul error model and Mix Ratio error model basically the same error model? For Add+Mul model, it reports CmultStdev as sigma2 and sigma() as sigma1 for additive error, whereas in Mix Ratio, it only reports CMixRatio, which equals to Sigma2/Sigma1. So you can only know the ratio, which is less interpretable than Add+Mul model?.

And since they represent sigma, so both CmultStdev and CMixRatio have to be positive?

 

Thank you

Joy, I think it is because with multiplicative the error is proportional to the conc, whereas additive is assumin uniform.  take a look at this post and perhaps that clarifies it for you

 

https://support.cert...-option/?p=3116







Also tagged with one or more of these keywords: VPC, error model

0 user(s) are reading this topic

0 members, 0 guests, 0 anonymous users