Form of prop_err and add_err from nlmixr?

[Dstix]

Is the output of prop_err or add_err in variance or standard deviation? In previous versions of nlmixr (nlmixr version 2.0.7) you could type fit$sigma and get the same prop_err value. In nlmixr2 version 2.0.8 the only place this appears is under fit$fixef[XX]. Is this correct? The reason I ask is I used the fit$sigma as input into my mrgsolve simulations and I want to make certain I correctly input variance or the standard deviation.

[mattfidler]

Hi @Dstix; The prop_err and add_err represents the standard deviation.

You could also directly simulate from rxode2 using the fit, this uses a fit$sigma as an identity matrix and uses the fixef[XX] as a scaling parameter for these values.

[mattfidler]

Yes the fit$omega is the variance-covariance matrix.

The fit$omegaR is the correlation matrix with the diagonals being the standard deviation of each component.

[mattfidler]

(I believe it is printed out in the fit object in case you want a quick place to figure it out)

[Dstix]

Ahh - I see it now:

 Full BSV covariance (fit$omega) or correlation (fit$omegaR; diagonals=SDs) 

This really helps. Thanks!

[mattfidler]

Thanks. I am not as conversant with rxODE having relied on mrgsolve for so long. Is this the same for fit$omega? Are the etas in the omega block in variance or standard deviation? I have been using these blocks in mrgsolve in the following form (where mod is the mrgsolve model):

`mod %<>% omat(dmat( fit$omega[1], # CL fit$omega[4], # VC 0, # Q 0 # VP ))

mod %<>% smat(dmat(fit$sigma))'

It seems OK (though I am not an expert in mrgsolve); The $sigma is fixed to one in nlmixr2 estimation for focei, so I believe you should still do the same thing in mrgsolve if you use this method.

Also if you use mrgsolve keep in mind that they use NOCB for covariates, whereas nlmixr2 uses LOCF for time-varying covarites. I think you can use LOCF with mrgsolve but I cannot remember how to specify it off-hand.

On the other hand the rxode2 model can now simply be the UI function you use in nlmixr2. (you can extract it from the fit by using as.function(fit), though it is easier to simulate from the fit directly since it will grab the right omega matrix for you directly (and specify the degrees of freedom needed for simulation with uncertainty if you take this approach)

[Dstix]

Yes, moving forward I will be using the rxSolve function. Much handier working with nlmixr models.