Dear NMusers:
I have two questions regarding model fitting.
1. FOCE vs. FOCE with INTERACTION. I have a rich data from phase I study. Drug was administered by iv infusion. I used a one-compartment model with nonlinear clearance (Michaelis-Menten kinetics) to fit this data. And I tried both FOCE and FOCE with INTERACTION. The FOCE method generated a reasonable fit, while FOCE with INTERACTION generated a biased prediction (underpredict) of concentration. I thought FOCE with INTERACTION usually generate better result than FOCE. Does this mean my model is just not good enough? I used a proportional plus additional residual error model.
2. I also tried to fit log transformed data, but in the PRED vs. DV plot, the points at lower concentrations are much more scattered than those at higher concentrations. And this forms a trend that points are getting closer and closer to the line as the concentration goes up. Does that mean log transformation of my data is not appropriate or something is wrong with my residual error model? The concentration ranges from 2 ng/ml to 1600 ng/ml. The residual error model I used is listed as below:
$ERROR
CALLFL=0
IPRED=-3
IF(F.GT.0)IPRED=LOG(F); to avoid LOG(0)run-time error
Y=IPRED+EPS(1)
Any suggestion will be highly appreciated!
Huali
Questions on FOCE and log transformed data
Dear NMusers:
I have two questions regarding model fitting.
1. FOCE vs. FOCE with INTERACTION. I have a rich data from phase I study. Drug
was administered by iv infusion. I used a one-compartment model with nonlinear
clearance (Michaelis-Menten kinetics) to fit this data. And I tried both FOCE
and FOCE with INTERACTION. The FOCE method generated a reasonable fit, while
FOCE with INTERACTION generated a biased prediction (underpredict) of
concentration. I thought FOCE with INTERACTION usually generate better result
than FOCE. Does this mean my model is just not good enough? I used a
proportional plus additional residual error model.
2. I also tried to fit log transformed data, but in the PRED vs. DV plot, the
points at lower concentrations are much more scattered than those at higher
concentrations. And this forms a trend that points are getting closer and
closer to the line as the concentration goes up. Does that mean log
transformation of my data is not appropriate or something is wrong with my
residual error model? The concentration ranges from 2 ng/ml to 1600 ng/ml. The
residual error model I used is listed as below:
$ERROR
CALLFL=0
IPRED=-3
IF(F.GT.0)IPRED=LOG(F); to avoid LOG(0)run-time error
Y=IPRED+EPS(1)
Any suggestion will be highly appreciated!
Huali
Huali,
A quick note on item number 2. If the model is predicting F=0, the selection of IPRED=-3 could be altering the fit of the model.
Try the following:
$ERROR
CALLFL=0
FLAG=0
IF(AMT.NE.0) FLAG=1 ;set flag=1 for dose records
;prevents log of 0 for dose records only
;changing IPRED (or F) for dose records does not change the computation
;of the objective function value.
IPRED=LOG(F+FLAG)
W=1 ;additive error model
IRES=DV-IPRED
IWRES=RES/W
Y=IPRED +EPS(1)
Changing IPRED (or F) on concentration records alters the computation of the objective function value. This should only be used as a last resort.
If you actually predict a zero for a concentration record, I suggest evaluating the data first. Does the data make sense or is there an error in sample collection time or dose times (especially check for a missing dose or an incorrect ADDL value)?
If everything is good with the data, then you may not have any other option than to alter the predicted concentration. If this is the case, then I suggest testing different values of IPRED using your code. Run the model using IPRED=-3 then IPRED=-4 then IPRED=-5, etc. until two runs have the same MVOF (Since the log(0)=-infinity, IPRED=-3 may not be small enough). I would then use the smallest IPRED that you tested to minimize the impact of changing a predicted concentration on your modeling results.
Regards,
Luann Phillips
Director PK/PD
Cognigen Corporation
Huali Wu wrote:
> Dear NMusers:
>
> I have two questions regarding model fitting. 1. FOCE vs. FOCE with INTERACTION. I have a rich data from phase I study. Drug was administered by iv infusion. I used a one-compartment model with nonlinear clearance (Michaelis-Menten kinetics) to fit this data. And I tried both FOCE and FOCE with INTERACTION. The FOCE method generated a reasonable fit, while FOCE with INTERACTION generated a biased prediction (underpredict) of concentration. I thought FOCE with INTERACTION usually generate better result than FOCE. Does this mean my model is just not good enough? I used a proportional plus additional residual error model. 2. I also tried to fit log transformed data, but in the PRED vs. DV plot, the points at lower concentrations are much more scattered than those at higher concentrations. And this forms a trend that points are getting closer and closer to the line as the concentration goes up. Does that mean log transformation of my data is not appropriate or something is wrong with my residual error model? The concentration ranges from 2 ng/ml to 1600 ng/ml. The residual error model I used is listed as below: $ERROR
>
> CALLFL=0
> IPRED=-3
> IF(F.GT.0)IPRED=LOG(F); to avoid LOG(0)run-time error
> Y=IPRED+EPS(1)
>
> Any suggestion will be highly appreciated!
>
> Huali
Actually with CALLFL=0 in the $ERROR block, this (guarding against F=0
at dosing records with FLAG variable) is not needed as the code in
$ERROR is called only at observation records.
Regards,
Katya
-------------------
Ekaterina Gibiansky
Senior Director, PKPD, Modeling & Simulation
ICON Development Solutions
[email protected]
Quoted reply history
-----Original Message-----
From: [email protected] [mailto:[email protected]]
On Behalf Of Luann Phillips
Sent: Wednesday, February 25, 2009 11:57 AM
To: Huali Wu
Cc: [email protected]
Subject: Re: [NMusers] Questions on FOCE and log transformed data
Huali,
A quick note on item number 2. If the model is predicting F=0, the
selection of IPRED=-3 could be altering the fit of the model.
Try the following:
$ERROR
CALLFL=0
FLAG=0
IF(AMT.NE.0) FLAG=1 ;set flag=1 for dose records
;prevents log of 0 for dose records only
;changing IPRED (or F) for dose records does not change the computation
;of the objective function value.
IPRED=LOG(F+FLAG)
W=1 ;additive error model
IRES=DV-IPRED
IWRES=RES/W
Y=IPRED +EPS(1)
Changing IPRED (or F) on concentration records alters the computation of
the objective function value. This should only be used as a last resort.
If you actually predict a zero for a concentration record, I suggest
evaluating the data first. Does the data make sense or is there an error
in sample collection time or dose times (especially check for a missing
dose or an incorrect ADDL value)?
If everything is good with the data, then you may not have any other
option than to alter the predicted concentration. If this is the case,
then I suggest testing different values of IPRED using your code. Run
the model using IPRED=-3 then IPRED=-4 then IPRED=-5, etc. until two
runs have the same MVOF (Since the log(0)=-infinity, IPRED=-3 may not be
small enough). I would then use the smallest IPRED that you tested to
minimize the impact of changing a predicted concentration on your
modeling results.
Regards,
Luann Phillips
Director PK/PD
Cognigen Corporation
Huali Wu wrote:
> Dear NMusers:
>
> I have two questions regarding model fitting.
> 1. FOCE vs. FOCE with INTERACTION. I have a rich data from phase I
> study. Drug was administered by iv infusion. I used a one-compartment
> model with nonlinear clearance (Michaelis-Menten kinetics) to fit this
> data. And I tried both FOCE and FOCE with INTERACTION. The FOCE
method
> generated a reasonable fit, while FOCE with INTERACTION generated a
> biased prediction (underpredict) of concentration. I thought FOCE
> with INTERACTION usually generate better result than FOCE. Does this
> mean my model is just not good enough? I used a proportional plus
> additional residual error model.
> 2. I also tried to fit log transformed data, but in the PRED vs. DV
> plot, the points at lower concentrations are much more scattered than
> those at higher concentrations. And this forms a trend that points are
> getting closer and closer to the line as the concentration goes up.
Does
> that mean log transformation of my data is not appropriate or
something
> is wrong with my residual error model? The concentration ranges from 2
> ng/ml to 1600 ng/ml. The residual error model I used is listed as
below:
>
> $ERROR
> CALLFL=0
> IPRED=-3
> IF(F.GT.0)IPRED=LOG(F); to avoid LOG(0)run-time error
> Y=IPRED+EPS(1)
>
> Any suggestion will be highly appreciated!
>
> Huali