Re-estimate NONMEM model parameter values in Phoenix - Issue

Dear Community,

I have an issue in re-estimating parameter values in Phoenix of a 2-CMT model that was initially developed using NONMEM.

The NONMEM model:

  • 2-CMT (ADVAN3 TRANS4)

  • IIV on Vc, Q, Vp and an additional baseline parameter

  • No IOV

  • Combined error model

  • OMEGA block

  • FOCE (with “INTER”)

I brought everthing to Phoenix NLME using PML. I re-estimated the parameter values and I constantly obtain parameter values (robust!) which are somewhat different from the NONMEM results. The Phoenix model diagnostics in general look very good. The IPREDs of obtained with Phoenix and NONMEM are comparably good! The differences are big when comparing PRED vs. IVAR of Phoenix vs. NONMEM. I would really like to know where the differences in the estimations come from!

To give you an impression on the differences in parameter values:

NONMEM Phoenix

baseline parameter: 0.8 0.83

CL: 0.04 0.03

Vc: 4.6 27.2

Vp: 7.9 49

Q: 0.18 0.33

error add: 0.011 (variance?) 0.094 (standard deviation?)

error prop: 0.0029 (variance?) 0.095 (standard deviation?)

The OMEGA then shows more differences. Although the values of the IIVs (diagonal) are somewhat similar, the covariances are very different!

NONMEM:

positive ; IIV_Baseline
negative positive ; IIV_Vc
positive negative positive ; IIV_Q
negative positive negative positive ; IIV_Vp

Phoenix:

positive ; IIV_Baseline
positive positive ; IIV_Vc
negative positive positive ; IIV_Q
negative negative positive positive ; IIV_Vp

May be the differences come from the missing “INTER” option in Phoenix?

What could be other reasons?

Thank you very much in advance - Every thought/idea is very much appreciated!

Best

Daniel

Dear Daniel,

sorry for the delay in response. I saw that you have sent your question to support. We will deal with that immediately. If there is anything more general in your case, we will post a response to the forum as well.

Bernd

Note that in Phoenix FOCE-ELS always use interaction you cannot deactivate it.

I will be good if support post back solutions in here ? will be good to know if any code to prevent flip flop of parametrs was done in nonmem so we can compare

regards,

Samer

Hi ,

Almost same question for a basic Advan 2 model 267 ID, 1874 DV (Cl, Ka, V) with IIV on all and omega-block between V and Cl.

  • First some difficulty to get minimization (LL out),

  • Second impossible to get covariance step (option Sandwich or Hessian)

  • Third: Parameters much different. Had to constrain by Lower/upper to get comparable

I have some difficulty to understand which parameters to play on in Run options tab? It would be educational for both Phoenix and NONMEM. Does anyone produce some interesting doc recently ?

Maybe first of all: what may have hindered covariance step?

Tk.

Gilles

Hi Gilles, are you able to post a Phoenix project that illustrates your problem ?

and corresponding NM run/results?

Simon.

a few more comments from a colleague specific to your questions;

  1. some difficulty to get minimization (LL out),

  2. impossible to get covariance step (option Sandwich or Hessian)

need a project really to help here

  1. Parameters much different. Had to constrain by Lower/upper to get comparable

That could be local identifiability problem, but again to be able to clarify, we would need a project to look at

  1. I have some difficulty to understand which parameters to play on in Run options tab?

this will depend on the engine chosen

  • Try to turn on MAP-NP start. FO-family is very sensitive to initial estimates.

  • Another option is to use Adaptive Gaussian quadrature (generalization of the Laplacian method)

  • When Laplacian with FOCE Hess is selected with NAGQ=1, the resulting method is the same as FOCE ELS and very similar to NONMEM’s FOCE engine with interaction.

  • When FOCE Hess is not selected and NAGQ=1, it is similar to NONMEM’s Laplacian engine.

  • When NAGQ is set to a value greater than 1, the method has no NONMEM equivalent, and the quality of the likelihood approximation is improved over simple Laplacian or FOCE ELS. > Try to increase N AGQ to 3 or 5.

  1. Maybe first of all: what may have hindered covariance step?

The problem is usually in the inversion of the second derivative matrix of the -2*Log Likelihood function. There could be not enough information to get estimates for all parameters.