Two questions

[mattfidler]

Hi @andrewhooker,

I have two questions:

• I tried implementing a residual error where the sigma is fixed to one and the additive error is defined in the ff model. When running evaluate_design the residual components are not being estimated in the fim:

$fim
               tcl         tv add.err
tcl     30.2591311 -0.2562828       0
tv      -0.2562828 40.0014869       0
add.err  0.0000000  0.0000000       0

$rse
     tcl       tv  add.err 
 3.76520 30.95288       NA 

Warning message:
  The following parameters are not estimable:
  add.err
  Is the design adequate to estimate all parameters? 

Is fixing the add.err component equivalent to what is done?

• When using a mixed log-normal and normal distribution, is it adequate to simply do the transformation of f. In dynamic transformation of both sides in NONMEM there is an objective function adjustment. Is that needed?

Thanks

The answers to these two questions will help me produce an interface between nlmixr2 and PopED.

[mattfidler]

For dynamic transform of both sides, I suppose I could add the objective function correction using a new ofv that relies on ofv_fim

https://github.com/andrewhooker/PopED/blob/master/R%2Fofv_fim.R

But I don't think it is needed.... I'm not confident enough in the methodology to really know.

In general I cannot find the "y" value that would lend itself to the dTBS

[mattfidler]

Is fixing the add.err component equivalent to what is done?

Looking through the examples this is a clear no. I guess the fixed sigma trick that works in NONMEM does not work here.

[mattfidler]

I also noticed that the prediction functions return DV so there must be some sort of DV prediction. So dynamic transform of both sides won't work 100% by simply transforming the f value.

[mattfidler]

I also noticed there is an example (example 8) that sort of has this transformation built in for log-values. For now, in the case of dynamic transform of both sides, I handle log-normal for the example 8 case.

In this case:

• f function is untransformed
• err function takes the untransformed and converts the values

This may be sufficient. However, it makes me wonder a bit about the standard errors of the estimate -- They seem way to small if I calculate them, so I think it is still an issue.

[andrewhooker]

Hi Matt,

I'll take a look!

Andy

[mattfidler]

Thanks Andy