Re: Problem with estimating sigma when using M3 method
Hi Hiba,
I agree that often the issues should be found in the data rather than the model. I recommend checking the data with NMcheckData from the R package called NMdata. It scans for a long list of potential issues, some that will make Nonmem fail, some that won't. If it finds issues, they will be returned in a data.frame with reference to row numbers and ID's so you can easily identify the root cause. If you look at ?NMcheckData you may identify arguments you can specify to add to the list of checks the function can run.
Having said this, a data/model issue can also be that your data poorly supports estimation of parts of your model (practical identifiability). NMcheckData won't help you identify such issues.
NMdata: https://philipdelff.github.io/NMdata/
NMcheckData manual: https://philipdelff.github.io/NMdata/reference/NMcheckData.html
An example with a few arguments that activate additional checks:
res.checks <- NMcheckData(mydata,covs="WEIGHTBL",cols.num="WEIGHT",col.usubjid="USUBJID") Here, NMcheckData will (in addition to a bunch of other checks) see if WEIGHTBL exists numeric, non-na and unique within subjects, WEIGHT exists and is numeric and non-NA, and that ID is unique against USUBJID and vice versa. (Obviously, Nonmem can't read USUBJID if it contains characters, but you could still keep it to the right in the dataset for reference). See the manual above for more options.
Best,
Philip
Quoted reply history
On 2023-05-26 11:06 AM, Leonid Gibiansky wrote:
> Yes, SIGMA should be fixed to 1 (do not try anything else, it has to
> be done correctly in the code first, and then we should worry about
> how to make it work)
>
> For combined error, expression is
> W = SQRT(W1**2 + W2**2) (squares in both terms)
>
> Do not worry about error 134, this is harmless, and you can fix it any
> time after you get your model right. Add UNCONDITIONAL MATRIX=S to the
> $COV step.
>
> For PARAMETER ESTIMATE IS NEAR ITS BOUNDARY try to add
> NOSIGMABOUNDTEST NOOMEGABOUNDTEST NOTHETABOUNDTEST
> to $est record
>
> Most of the time, numerical difficulties come from the problems with
> the data, so it makes sense to clean the data set first as much as
> possible.
>
> Leonid
>
> On 5/26/2023 9:30 AM, Hiba Sliem wrote:
>
> > Hi
> >
> > I already tried fixing the value of sigma to 1, the covariance step isn't implemented when I do that.
> >
> > If I try fixing it to 0.144 the minimization isn't successful.
> >
> > I also tried a combined error model like this:
> > LOQ=0.1
> > IPRED = F
> > W1 = THETA(8)*IPRED
> > W2 = THETA(9)
> > W = SQRT(W1**1 + W2**2)
> > DEL = 0
> > IF(W.EQ.0) DEL = 1
> > IRES = DV - IPRED
> > IWRES = IRES/(W + DEL)
> > DUM = (LOQ -IPRED)/(W + DEL)
> > CUMD = PHI(DUM)
> > IF(BLQ.EQ.0) THEN
> > F_FLAG=0
> > Y= IPRED +W*ERR(1)
> > ELSE
> > F_FLAG=1
> > Y=CUMD
> > MDVRES = 1
> > ENDIF
> >
> > In which case I get a PARAMETER ESTIMATE IS NEAR ITS BOUNDARY error message When trying to fix Sigma in the combined model I have a MINIMIZATION TERMINATED
> >
> > DUE TO ROUNDING ERRORS (ERROR=134) message.
> >
> > My dataset has a lot of predose samples and washouts between different periods, is it possible the issue comes from my dataset?
> >
> > Regards
> >
> > -----Original Message-----
> > From: Leonid Gibiansky <[email protected]>
> > Sent: Friday, 26 May 2023 14:51
> > To: Hiba Sliem <[email protected]>; [email protected]
> >
> > Subject: Re: [NMusers] Problem with estimating sigma when using M3 method
> >
> > [You don't often get email from [email protected] . Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
> >
> > you should fix
> >
> > $SIGMA
> > 1 FIX
> >
> > as you are already estimating the SD using THETA(8).
> >
> > Leonid
> >
> > On 5/26/2023 4:57 AM, Hiba Sliem wrote:
> >
> > > Hello
> > >
> > > I'm fairly new to nonmem, I'm currently trying to model a phase 1
> > >
> > > study with BLQ values, while the run was successful with no error message, my residual error has a %rse >70 and a confidence interval that includes
> > >
> > > zero.
> > >
> > > Here's my code:
> > >
> > > $ERROR
> > >
> > > LOQ=0.1
> > >
> > > IPRED = F
> > >
> > > SD = THETA(8)*IPRED
> > >
> > > DEL = 0
> > >
> > > IF(SD.EQ.0) DEL = 1
> > >
> > > IRES = DV - IPRED
> > >
> > > IWRES = IRES / (SD + DEL)
> > >
> > > DUM = (LOQ -IPRED) / (SD + DEL)
> > >
> > > CUMD = PHI(DUM) + DEL
> > >
> > > IF(BLQ.EQ.0) THEN
> > >
> > > F_FLAG=0
> > >
> > > Y= IPRED +SD*ERR(1)
> > >
> > > ELSE
> > >
> > > F_FLAG=1
> > >
> > > Y=CUMD
> > >
> > > MDVRES = 1
> > >
> > > ENDIF
> > >
> > > $EST METHOD=1 INTERACTION LAPLACIAN PRINT=5 MAX=9999 SIG=3 SLOW
> > > NUMERICAL MSFO=*.msf
> > >
> > > $SIGMA
> > >
> > > 0.38 ;[P] sigma(1,1) (estimated in a previous model)
> > >
> > > Furthermore, when trying to fit this model to my phase 2 dataset,
> > > covariance step fails when I implement it.
> > >
> > > Any suggestions are welcome
> > >
> > > Thank you