From: Leonid Gibiansky <*LGibiansky*>

Date: Tue, 15 Dec 2009 09:33:50 -0500

Pavel,

I do not see any justification for a proportional or exponential model:

no reasons to believe that error of measurement is proportional to the

value. I would try simple additive error model. In simulations, one can

truncate at 0 and 10.

The poster that I mentioned specifically discussed various approached to

the problem that you are trying to solve.

Thanks

Leonid

--------------------------------------

Leonid Gibiansky, Ph.D.

President, QuantPharm LLC

web: www.quantpharm.com

e-mail: LGibiansky at quantpharm.com

tel: (301) 767 5566

nonmem

*> Leonid,
*

*>
*

*> This is about visual analog scale. There are a lot of 0 and 1 values
*

*> (actually, VAS changes from 0 to 10 in this case, but it can be divided
*

*> by 10). There are articles, presentatione and dissertations which use
*

*> logit. So, I try diffrent transformations including logit.
*

*>
*

*> CV error works OK, but I still try to take care of the skewed
*

*> distribution.
*

*>
*

*> When I use exponential error, NONMEM transforms it into CV error.
*

*> Later, simulations do not make sense because NONMEM does not do the
*

*> same. Exactly as described in the nonmem6 manual.
*

*>
*

*> Thank you for the article. I'll keep digging.
*

*>
*

*> Pavel
*

*>
*

*>
*

*>
*

*> ----- Original Message -----
*

*> From: Leonid Gibiansky
*

*> Date: Tuesday, December 15, 2009 1:01 am
*

*> Subject: Re: [NMusers] $ERROR and LOGIT
*

*> To: nonmem *

*> Cc: nmusers *

*>
*

*> > Pavel,
*

*> > I am not sure what is the problem with the log-transformation of
*

*> > the
*

*> > data. log(x) = infinity only if x = infinity, do you have
*

*> > infinite
*

*> > observations in your data set? If not, then transformed data
*

*> > cannot be
*

*> > equal to infinity.
*

*> > log(x) = - infinity only if x=0
*

*> > do you have BQL observations coded as zeros? If so, you cannot
*

*> > use
*

*> > exponential error model. But you can either exclude BQLs (and
*

*> > use
*

*> > log-transformation) or treat them as BQLs (and still use
*

*> > log-transformation).
*

*> >
*

*> > Looks like your prediction F is between 0 and 1. I do not think
*

*> > that
*

*> > exponential error is appropriate for this type of data. Could
*

*> > you
*

*> > elaborate what exactly you are modeling? If this is indeed
*

*> > interval
*

*> > data, this poster can be relevant (Estimating Transformations
*

*> > for
*

*> > Population Models of Continuous, Closed Interval Data, Matthew
*

*> > M.
*

*> > Hutmacher and Jonathan L. French):
*

*> >
*

*> > http://www.page-meeting.org/default.asp?abstract=1463
*

*> >
*

*> > Thanks
*

*> > Leonid
*

*> >
*

*> > --------------------------------------
*

*> > Leonid Gibiansky, Ph.D.
*

*> > President, QuantPharm LLC
*

*> > web: www.quantpharm.com
*

*> > e-mail: LGibiansky at quantpharm.com
*

*> > tel: (301) 767 5566
*

*> >
*

*> >
*

*> >
*

*> >
*

*> > nonmem *

*> > >
*

*> > > Hello,
*

*> > >
*

*> > > NONMEM has the following property related to intra-subject
*

*> > variability:>
*

*> > > "During estimation by the first-order method, the exponential
*

*> > model and
*

*> > > proportional models give identical results, i.e., NONMEM
*

*> > cannot
*

*> > > distinguish between them." So, NONMEM transforms
*

*> > F*DEXP(ERR(1)) into F
*

*> > > + F*ERR(1).
*

*> > >
*

*> > > Is there an easy around it? / /I try to code the logit
*

*> > transformation.
*

*> > > I cannot log-transform the original data as it is suggested in
*

*> > some
*

*> > > publications including the presentation by Plan and Karlsson
*

*> > (Uppsala)
*

*> > > because many values will be equal to plus or minus infinity.
*

*> > Will
*

*> > > NONMEM "linearize" the following code:
*

*> > >
*

*> > > Z = DLOG((F+THETA(10))/(1-F+THETA(10)))
*

*> > > Y = DEXP(Z + ERR(1))/(1 + DEXP(Z + ERR(1)))
*

*> > >
*

*> > >
*

*> > >
*

*> > > Thanks!
*

*> > >
*

*> > > Pavel
*

*> > >
*

*> > >
*

*> > >
*

*> >*

Received on Tue Dec 15 2009 - 09:33:50 EST

Date: Tue, 15 Dec 2009 09:33:50 -0500

Pavel,

I do not see any justification for a proportional or exponential model:

no reasons to believe that error of measurement is proportional to the

value. I would try simple additive error model. In simulations, one can

truncate at 0 and 10.

The poster that I mentioned specifically discussed various approached to

the problem that you are trying to solve.

Thanks

Leonid

--------------------------------------

Leonid Gibiansky, Ph.D.

President, QuantPharm LLC

web: www.quantpharm.com

e-mail: LGibiansky at quantpharm.com

tel: (301) 767 5566

nonmem

Received on Tue Dec 15 2009 - 09:33:50 EST