Back to the Top
I observe that the CV%, Stderr, CI estimates for population typical =
values only
appear sometimes and do not understand (1) what controls when they are
calculated and (2) how to force the program to calculate them. Yes, they =
may not
be the most important values calculated but some people want them. I =
have
upgraded this July to the most recent version of the total Phoenix =
suite. I
think that I am running NONMEM 6.1.
=20
Thanks -
=20
Richard=
Back to the Top
The following message was posted to: PharmPK
Hi Richard,=20
You should know we have a dedicated user forum; =
www.pharsight.com/extranet,
where you can get tips and discuss modeling and Pharsight products with =
other
users. One advantage of this forum is you can attach files and images =
so people
can look at your project to see what might be going wrong, so it may be =
worth
joining that too.
An initial guess as to the cause of the problem should be found in the =
textual
output labeled "Core output" and "Core status"
You should look at the return code to understand why you are not getting
Stderr's etc. consistently, these will differ for different engines e.g. =
the
case below 1,2 and 3 are all acceptable, the lower numbers being better.
Naive pooled engine exit code =3D 1
0 =3D evaluation only, no optimization attempted
1 =3D terminated with gradient small solution very probably optimal
2 =3D terminated with stepsize small solution probably optimal
3 =3D no lower objective value can be found solution probably optimal
4 =3D max iterations in optimization at final main iteration - solution =
obj value
probably near optimal but one or more theta values not resolved or near =
bound=20
5 =3D final iteration failed: too many large steps function may be =
unbounded
6 =3D max number of main iterations reached solution requires father =
optimization
try increasing iteration limit=20
other - optimization aborted with error
You may also see a message regarding the issues with the Hessian matrix =
in
calculating Stderr- I have found that refining my initial estimates e.g. =
taking
the last final estimates as the new initial estimate sometimes resolves =
this
problem, it would be interesting to see a project file exhibiting the =
issues you
are seeing.
Simon.
PS the current version for Phoenix is 6.3, which includes the NLME =
engine,
version 1.2, this is the one that includes the new QRPEM algorithm. You =
can
check your version by clicking Help> About Phoenix
PPS if you look at =
http://www.pharsight.com/training/training_upcoming.php you
can see the next NLME course in the US is 7-9 Aug at UNC Eshelman School =
of
Pharmacy, Chapel Hill, NC, US and you can sign up via this page.
Best regards,=20
Simon.=20
--
Simon.Davis.aaa.certara.com
Senior Scientific Consultant
Pharsight- A Certara(tm) Company=
Back to the Top
The following message was posted to: PharmPK
Hi Richard,
The covariance matrix of the fixed effects determines stderr (and hence =
CV% and
CI) for the typical population values. After the estimation step is =
complete,
the covariance matrix (of the fixed effects, don't confuse with the =
covariance
of the random effects which is usually denoted as "omega") is =
approximated from
the inverse of the approximated information matrix. The problem is that =
if the
information matrix is not invertable or is not positive definite, then =
the
covariance step will fail. This is more likely to happen when the =
likelihood
surface is not "bowl shaped" at the estimate. Maybe there is a =
discontinuity,
or the surface is flat in one or more dimensions or is a very curved in =
a couple
dimensions. Also, estimates at the bounds (of theta) can cause =
problems. This
applies, in general, to most nonlinear mixed effects modeling tools.
In Phoenix NLME, the overall result sheet has a return code. The return =
codes
are: 1 - estimate is sitting at the bottom of a "bowl"; 2 - the estimate =
is
"near" (within tolerance) of the bottom of the bowl; 3 - the estimate =
couldn't
be improved upon. The probability of success with the covariance step =
decreases
with increasing return codes 1, 2 and 3.
There are some things you can try to get the covariance step to succeed:
-if the return code is not 1, try rerunning the estimation from the =
current
solution. Sometimes that will improve the final solution slightly.
-if the solution is at a bound, widen the bound or freeze the parameter.
-maybe the model is overparameterized, or insensitive to a parameter? =
See if
you can eliminate parameters.
-change the stepsize for the numerical derivatives used in the standard =
error
calculation. There's an option for that in Phoenix NLME on the "Run =
Options"
tab.
After examining those options if the covariance step is still a problem =
it may
be that the model is "too" nonlinear or discontinuous near the estimate. =
This
is fairly common and not too worrying result for complicated models. =
There is
still the option of bootstrapping.
I hope this helps,
Jason Chittenden
Director, Product Development
Certara(tm)=20
Implementing Translational Science
5625 Dillard Drive, Suite 205, Cary NC 27518
Certara: The name behind the names you know
Tripos - Simcyp - Pharsight
www.certara.com=
Want to post a follow-up message on this topic?
If this link does not work with your browser send a follow-up message to PharmPK@boomer.org with "Phoenix NLME Pop Theta CV% Estimates" as the subject | Support PharmPK by using the |
Copyright 1995-2011 David W. A. Bourne (david@boomer.org)