RE: Why does covariance fail?
Dear Dr Sale,
I agree with what you, Leonid et al. wrote and wanted
to add two comments and ask a question.
1) There may be situations, when one cannot apply the
bootstrap to obtain confidence intervals. Such situations
may occur more often in PD than in PK. Assume there
are e.g. two experimental runs of some PD profile that
are recorded under say 10 different experimental conditions.
Each set of experimental conditions provides information
on some of the model parameters, but no single experimental
condition provides information on all parameters simultaneously.
A bootstrap stratified by experimental condition seems most
appropriate. However, sampling with replacement from two
replicates makes little sense, as two replicates are not
representative of the whole population of profiles for each
experimental condition.
2) If you plan to implement an algorithm like: If $COV in
NONMEM fails with error NNN, then the model still passes
if condition ZZZ is fulfilled. This sounds specific to NONMEM.
Other algorithms / programs like the MC-PEM algorithm
(e.g. in S-Adapt) seem to always provide standard errors
and - as far as I am aware - WinBugs naturally provides
credibility intervals.
My question: Are you planning to devise a program independent
strategy for accepting / rejecting / revising models for that
meaningful confidence intervals are difficult to obtain?
Thank you & best regards
Juergen
-----------------------------------------------
Juergen Bulitta, PhD, Post-doctoral Fellow
Pharmacometrics, University at Buffalo, NY, USA
Phone: +1 716 645 2855 ext. 281, [EMAIL PROTECTED]
-----------------------------------------------
-----Ursprüngliche Nachricht-----
Von: "Mark Sale - Next Level Solutions" <[EMAIL PROTECTED]>
Gesendet: 29.01.08 19:55:14
An: undisclosed-recipients;
CC: [email protected]
Betreff: RE: [NMusers] Why does covariance fail?
Leonid
Thanks, that makes sense. Overall, I think we have better tools now for
pretty much everything that the covariance step is supposed to do (give
confidence in the model (should be done with PPC or NPDE), provide SEE and
estimation correlations (better done with bootstrap)) etc, mostly because we
have faster and parallel computers. I'm hoping to provide some justification
for the (very, very rare of course) occasions when I like a model that fails a
covariance step. Ideal, I could say "if the covariance step error is XXX, then
the model is still OK if it passes this test". Your sampling (I thi!
nk technically it is called hyper cube sampling) to generate a multi
dimensional likelihood profile makes sense.
Mark Sale MD
Next Level Solutions, LLC
www.NextLevelSolns.com
919-846-9185
Quoted reply history
-------- Original Message --------
Subject: Re: [NMusers] Why does covariance fail?
From: Leonid Gibiansky <[EMAIL PROTECTED]>
Date: Tue, January 29, 2008 12:37 pm
To: Mark Sale - Next Level Solutions <[EMAIL PROTECTED]>
Cc: [email protected]
Hi Mark
I am pretty confident (although I do not have a proof) that all (or all
reasonably good, meaning that you invested some time trying to make them
good) models with failed covariance steps are not the saddle points but
over-parametrized models with degenerate direction(s). Some exception
could be related to the problems with the odd-type data where I have
less experience, so let's restrict the discussion by continuous-type
data. The check should be pretty easy. I would just evaluate OF in
100-1000 random points in the vicinity of the solution (automated
similar to the bootstrap runs). It is better be random rather than
univariate (parameter by parameter) to investigate all possible
parameter-space directions.
If you have bootstrap results, it could be used instead of this new
check since it is very unlikely that all bootstrap runs would stick to
the saddle-point rather than move along the gradient to the lower minimum.
Answer seems obvious to me (not a saddle points) but it would be
interesting to see a more definite results.
Please, update if you see anything interesting
Thanks
Leonid
P.S. As far as I remember, this message:
PARAMETER ESTIMATE IS NEAR ITS BOUNDARY
THIS MUST BE ADDRESSED BEFORE THE COVARIANCE STEP CAN BE IMPLEMENTED
can be given if some of the OMEGA or SIGMA elements (including the
off-diagonal terms) are close to zero. You can block this case, ICON
distributed the patch for it, see archives.
see also
http://www.cognigencorp.com/nonmem/current/2007-July/0335.html
--------------------------------------
Leonid Gibiansky, Ph.D.
President, QuantPharm LLC
web: www.quantpharm.com
e-mail: LGibiansky at quantpharm.com
tel: (301) 767 5566
Mark Sale - Next Level Solutions wrote:
>
> I'm thinking of doing a somewhat formal analysis of the meaning of a
> failed covariance step. Some years ago Stu Beal explained that (as I
> recall), if the covariance step fails you cannot be sure that the
> minimum isn't a saddle point, which makes sense to me, and is consistent
> (I think), with the common message from NONMEM
> R MATRIX ALGORITHMICALLY SINGULAR
> AND ALGORITHMICALLY NON-POSITIVE-SEMIDEFINITE
> R MATRIX IS OUTPUT
> 0COVARIANCE STEP ABORTED
>
>
> I'm also finding one in NONMEM VI that I don't recall from NONMEM V, and
> I don't know what it means:
> ERROR RMATX- 1
>
> Then there are messages! that seem to be related to conditional estimates:
> NUMERICAL HESSIAN OF OBJ. FUNC. FOR COMPUTING CONDITIONAL ESTIMATE
> IS NON POSITIVE DEFINITE
> MESSAGE ISSUED FROM COVARIANCE STEP
>
> and version VI of NONMEM will refuse to even try the covariance step for
> various reasons:
> PARAMETER ESTIMATE IS NEAR ITS BOUNDARY
> THIS MUST BE ADDRESSED BEFORE THE COVARIANCE STEP CAN BE IMPLEMENTED
> even, it seems when the parameter estimate is no where near the boundary.
>
> I'm thinking of looking at these various reasons that the covariance
> step fails, and seeing if any of them mean anything WRT whether the
> model is "good", by some objective measure (PPC, NPDE, predictive check).
> My question is, is there any way to formally test whether the failure is
> due to a saddle point in the objective function surface? My
> understanding of the current search algorithm used by NONMEM is that it
> is very, very robust WRT saddle points. So, I suspect that the vast
> majority of the failures are not due to a saddle, but rather just a
> fairly flat surface, with near 0 first and second derivatives, causing
> numerical problems inverting it, rather than actually being a saddle
> point. If the surface is just fairly flat, not a saddle, then I think
> that the answer is not "wrong", just not especially good, therefore
> other simulation based tests of "goodness" might be just fine.
> I suspect that you could test whether it is a saddle point by trying a
> slightly different value for the parameter (e.g., "minimum" is 10, so
> try 9.9 and 10.1 and see if the OBJ is better, in each dimension. Would
> this work?
>
> thanks
> Mark
>
>
> Mark Sale MD
> Next Level Solutions, LLC
> www.NextLevelSolns.com http://www.NextLevelSolns.com
> 919-846-9185
>
>