RE: computation of exponential error model for RUV
Hi,
I agree with Jakob except the following:
(1) Altering the value of a prediction, changes the value of the objective
function value. So DEL should only be used if absolutely necessary.
(2) The current code changes the prediction for every observation. DEL
should only be set if a prediction is actually zero.
(3) If used, DEL should be set equal to 10E-16 (approximately machine zero).
This will prevent large changes in the OBJ value when a predicted value changes
from 0 to a very tiny number such as 10E-15.
Please see suggested code below.
Best regards,
Luann Phillips
$ERROR
;first oral dose F is always zero. Other dose records can also have F=0.
;set DFLAG (dose flag) to prevent log(0)
;Dose records do *not* change the OBJ value so the value of DFLAG does not
matter.
DFLAG=0
IF(AMT>0)DFLAG=1
;set a concentration flag (CFLAG). Conc records *do* change the OBJ value.
;change this prediction by as small a value as possible.
;For records with CFLAG=1, check the surrounding data for errors. Is the time
since previous dose
;excessively long for the compound (incorrect sample date or previous dose
date)? Does it make sense
;to have an observable concentration at that time point? Etc.
;Using values greater than 10E-16, can result in very large changes in the OBJ
when a covariate or other
;model change causes the F=0 record to change to a non-zero value which can be
as tiny as 10E-15 or less.
CFLAG=0
IF(F==0)CFLAG=10E-16
IPRED=LOG(F+DFLAG+CFLAG)
W=SIGMA(1,1)
IRES=DV-IPRED
IWRES=IRES/W ;exact IWRES for log or additive error model only
Y=IPRED+EPS(1)
Quoted reply history
From: [email protected] <[email protected]> On Behalf Of
Jakob Ribbing
Sent: Monday, July 19, 2021 12:03 PM
To: Guidi Monia <[email protected]>
Cc: [email protected]
Subject: Re: [NMusers] computation of exponential error model for RUV
Dear Monia,
Yes, that is correct.
So if one were to use this parameterization for FOCE then one would effectively
get a proportional (symmetric) error distribution, exactly according to what
you suggested:
Y=F*(1+EPS(1))
For this reason, the standard approach in NONMEM (for FOCE and exponential
RUV), is to log-transform both sides, i.e.like this:
$ERROR (OBSERVATION ONLY)
DEL=0.0000001
IPRED=LOG(F+DEL)
Y=IPRED+EPS(1)
So additive on the log-transformed scale is exactly the error model you would
like to use.
Maybe that would be a solution for your comparison?
Best wishes
Jakob
On 19 Jul 2021, at 17:34, Guidi Monia
<[email protected]<mailto:[email protected]>> wrote:
Dear colleagues,
We would like to compare the NONMEM predictions with those obtained by a
Bayesian TDM software for models describing the residual unexplained
variability with exponential errors.
We need to know if NONMEM performs a first order Taylor expansion of the
exponential error when data are fitted by the FOCE method:
Y=F*EXP(EPS(1)) -> Y= F*(1+EPS(1)).
Could someone help with this?
Thanks in advance
Monia
Monia Guidi, PhD
Pharmacometrician
Service of Clinical Pharmacology | University Hospital and University of
Lausanne
Center of research and innovation in Clinical Pharmaceutical Sciences |
University Hospital and University of Lausanne
BU17 01/193
CH-1011 Lausanne
email: [email protected]<mailto:[email protected]>
tel: +41 21 314 38 97
CHUV
centre hospitalier universitaire vaudois
<image001.png>
This communication is confidential and is only intended for the use of the
individual or entity to which it is directed. It may contain information that
is privileged and exempt from disclosure under applicable law. If you are not
the intended recipient please notify us immediately. Please do not copy it or
disclose its contents to any other person.
Any personal data will be processed in accordance with Pharmetheus' privacy
notice, available https://pharmetheus.com/privacy-policy/.