Power Function Model

NONMEM Topic 34

Keywords: ADVAN3, TRANS3

Topic started by: Valerie Cosson  - 12 Dec 94

I would like to know if some users had problems with using the power
function model.
I have tried to fit my data with the power function model. My data are
quite simple : twenty-three 15min IV infusion full profile. The NM-TRAN
control file is the following :

$PROB TWO COMPARTMENTS / ALL SUBJECTS : IV
$INPUT PRN ID TIME DV AMT RATE EVID
$DATA
$SUBROUTINE ADVAN3 TRANS3
$PK
       TVCL=THETA(1)
         CL=TVCL*EXP(ETA(1))
        TVV=THETA(2)
          V=TVV*EXP(ETA(2))
        TVQ=THETA(3)
          Q=TVQ*EXP(ETA(3))
      TVVSS=THETA(4)
        VSS=TVVSS*EXP(ETA(4))
         S1=V
$THETA (0,80) (0,20) (0,90) (50,70) (0,0.5,1)
$OMEGA 0.5 0.5 0.5 0.5
$ERROR
      W=F**THETA(5)
      Y=F+W*EPS(1)

      IPRED=F
      IRES=DV-IPRED
      IWRES=IRES/W
$SIGMA 0.25
$ESTIMATION MAXEVALS=7000 PRINT=5 METHOD=0 NOABORT POSTHOC
$COVARIANCE
$TABLE PRN ID CL V Q VSS IPRED IWRES NOPRINT FILE=
$TABLE PRN ID TIME ETA1 ETA2 ETA3 ETA4
$SCATTER PRED VS DV UNIT
$SCATTER PRED VS TIME
$SCATTER WRES VS PRED
$SCATTER IWRES VS IPRED

NONMEM is running on VAX station.
I get get the following error message :
Job terminated with error status %MTH-F-UNDEXP, undefined exponentiation!/
user PC !XL.

Could you tell me if there is any problem with the power function model, or
if I am not using it correctly?

Response from: LSheiner (lewis@c255.ucsf.EDU) and Alison - 12 Dec 94

Alison adds ...

Most likely the only records with the prediction F =0 are the infusion
dose event records.  By default, the $ERROR block is evaluated (i.e.,
the ERROR subroutine is called) with every event record.  It is
possible to prevent the $ERROR block from being evaluated with
non-observation events.  Include:
   $ERROR
     CALLFL=0
     .. etc...

Only if the F=0 records are unavoidable *must* one of your two coding
solutions be used.  (Unavoidable situations include models having ALAG
and observations that fall prior to the lagged dose time, or when some
individuals have observations so far out in time that compartment
amounts have underflowed to 0.)

---------------
I would add that I don't really like the power function model
not only because it gets us into trouble at zero, but because
it is unrealistic *near* zero.  There is always a lower limit of
detection. I have found that if (i) one uses the additive plus
proportional error model, (ii) sets values less than the lower limit
as equal to that lower limit, and (iii) sets the SD of the
additive part of the error equal to the lower limit, then
everything behaves just about right.  Thus, in the following code,
just fix theta(5)=lower detection limit.

$ERROR
      W = THETA(5)*EPS(1) + THETA(6)*F*EPS(2)
      Y=F+W

      IPRED=F
      IRES=DV-IPRED
      IWRES=IRES/W
$SIGMA 1 FIX 1 FIX

End of Topic - 30 May 95