Simulations using naive pooled engine

Phoenix Certara Forums Modelling and Simulation
1

Dear Certara Team,

I guess it’s a simple question, but I can’t solve it currentlly.

I have a standard 2 comp PK model and do naive pooled fitting. I then copy the model, accept all params and switch to simulation. For this I created a dosing sheet using 3 regimens (ADLL function) plus dummy sheet for simulation of these 3 regimens.

If I now simulate #replicates:50 using a sim table I get 50 profiles for each regimen that do not make sense at all. See attached.

  1. What could be the problem here?

  2. What is the naive pooled engine using for these simulations, as there are no random effects? Only the residual error that was determined during fitting?

Thanks

Frank

Capture.PNG
Capture.PNG563×324 31.5 KB

2

Hi Frank,

I will need more information before I can help.

It is not clear what you are trying to do and whether you want to simulate with or without residual error.

It will also depend on what variable you asked to output in the table e.g. C = IPRED = PRED ( in your cae) or Cobs which has the residual error.

Naive pooling will still use the dose information and each dose will have a different curve.

Can you share the project ? what is your residual error model and the stdev value ?

Samer

3

[attachment=2782:Structure.pdf]

Thanks Samer,

Unfortunately I cannot upload the file. I made some screenshots of the settings - maybe that will help you to understand what I am doing? The error is very high (0.5in this test project, multiplicative error. But still - as I asked for C as output it shoud not make use of that residual error if I understand your post correctly? For every cohort (i.e. regimen I am simulating), where does the huge variability of the 50 simulation samples come from then?

Thanks

Frank

Structure.pdf (68.9 KB)

4

Hi Frank,

It will be really helpful if you can use a dummy data set and send the projects from the pdf what I see is that you are using a multiplicative error model with 50 % error which is huge

you can double check if you mapped the addl correctly by using the initial estimates tab you should see multiple doses/up and downs in the profiles

when you run with sim/Pred check and request C you are getting individual predictions (in this case naive pooled prediction since there is no random effects) without residual error (IPRED)

why are you using naive pooling and not a real population engine

can you share the list of doses/addl that you are using ?

5

Hi Samer,

I upload a dummy phx file with data so it is easier to follow what my problem is. The file is easy to understand: first I do a 2 comp fit on data wíth the naive pooled engine (the fit is quite bad, but does not matter as I can illustrate my problem). Then I copy the model object and so a simulation on 2 scenarios with the addl function and again using the naive pooled enigne. I ask for a sim table with C and 10 iterations per id. Then I export the sim table and import it back to plot it. Now where does the variance in the 10 profiles per id come from? Should not all 10 be the same per id as there are no random effects?

Thanks

Frank

dummy.phxproj (1.14 MB)

6

dummysm.phxproj (1.53 MB)

The fitting engine is naive pooled but I am not sure if the simulation engine will zero out the omegas better to have near zero values to make sure that you are simulation without between subject variability.

Another option is not to use Population and use sorting by ID.

Samer

7

Hi Samer and Frank,

That’s working a little bit different. From the User’s guide:

When the Predictive Check run option is selected, the model is first fit with 0 iterations (regardless of N Iter in the user
interface) in order to generate the data needed to fill the standard results (such as Residuals, Theta,
etc.). Predictive check is based on the resulting model fit estimates of theta, sigma, and omega. It also uses
the estimates of eta(s) for each subject. <…> Both eta (individual variation) and epsilon (observation error) are sampled and used to create a population prediction.
As you see the simulation engine won’t zero out the omegas, but will try to fit the model in accordance with resulting model fit estimates. So it is possible to turn off the random effects, PHX will fit the model for simulationwithout them in population mode too.

8

Hi mittyright and Samer,

Thanks for solving the problem. I switched to textual and removed all the random effects from the modfel text, that are there by default even when doing naive pooled modeling. Then doing a simulation (Sim/Pred Check) as expected gives the same result (i.e. time conc curve) for each replicate.

Thanks for the help

Frank