NONMEM Users Network Archive

Hosted by Cognigen

Re: var-cov matrix issue?

From: Ethan Wu <ethan.wu75>
Date: Fri, 27 Feb 2009 09:22:29 -0800 (PST)

Dear Kyun,   in the current I only have diagonal element   usua=
lly I only used distribution of fixed-effects parameters to account for the=
 uncertainty of this level (trial level), then for each sampled fixed-effec=
ts parameter, I would just use the estimated ETA to simulate patients withi=
n each trial. I never practice like you suggested, i.e. using the var-co=
v matrix to sample Omega/Sigma too  and I realize that I don't have a =
strong rationale for opposing this approach  so I wonder if others woul=
d share some insights into this  __________________________=
______ From: "kyunseop.bae
n Wu <ethan.wu75

ur model has only diagonal elements in OMEGA matrix, you don't need to ca=
re about the following - positive definiteness of OMEGA matrix during sim=
ulation. If you use VAR-COV output of NONMEM for the simulation, it me=
ans you generate THETAs, OMEGA matrix and SIGMA matrix from the MVN distr=
ibution of VAR-COV matrix. OMEGA and SIGMA matrix are always positiv=
e definite in nature. However, if you generate full-block OMEGA matrix=
 using MVN (multi-variate normal) of VAR-COV matrix, many of generated =
OMEGA matrix will not be positive definite. One way to avoid this is t=
est positive definiteness and discard in-adequate ones. This may hel=
p you. Regards, Kyun Seop Bae MD PhD _________________=
_______________ From: Ethan Wu [mailto:ethan.wu75
Thursday, February 26, 2009 4:05 AM To: kyunseop.bae
ers
ear Kyun, thanks for your help. I don't know if I understand this one =

ive definite. " Could you explain a bit more? _________________=
_______________ From: "kyunseop.bae
> To: Ethan Wu <ethan.wu75
: Tuesday, February 24, 2009 3:07:30 PM Subject: RE: [NMusers] var-cov ma=
trix issue? Hi, Ethan, I think your question can be reduced wheth=
er pseudo-inverse matrix can be used instead of inverse matrix. I do n=
ot know quite different cases, but I suppose it can be used. To be mor=
e adequate answer in your context, MATRIX=R option could be more appro=
priate, if you use VAR-COV matrix output for simulation under normal dist=
ribution assumtion, If your data supports normal distribution assump=
tion, MATRIX=R option will not give much difference in SEs. Default V=
AR-COV output in NONMEM is a kind of sandwich estimate, which is thought =
to be more robust (a little larger) than inverse Fisher's information mat=
rix (given MATRIX=R option). Some caution is necessary to simulate o=
mega matrix that is alwasy positive definite. This may help you. =


maxnm.com [mailto:owner-nmusers
ent: Tuesday, February 24, 2009 11:09 AM To: justin.wilkins

? Hi Justin, only ETA was estimated with high SE but, again, I gu=
ess it came back to the question: how trustful it is if such error messag=
e appears ________________________________ From: "justin.wilki=
ns

-cov matrix issue? Dear Ethan, Algorithmically singular matri=
ces are often a sign that that your model is ill-conditioned in some way;=
 I would be careful in how I used the variance-covariance matrix in this =
scenario, and especially for simulation. Are there any parameters that ar=
e being estimated with particularly high standard errors? This might sugg=
est overparamaterization. Not sure how helpful this is! Best =

rmacology) CHBS, WSJ-027.6.076 Novartis Pharma AG Lichtstrasse 35 C=
H-4056 Basel Switzerland Phone: +41 61 324 6549 Fax: +41 61 324 3039=

in.wilkins
vartis on 2009/02/24 07:15 PM ----- Ethan Wu <ethan.wu75

    To     nmusers


y objective was to use nonmem var-cov output  for approximation of dist=
ribution of parameters for performing a simulation. if such error mess=
age occur, is the var-cov matrix  still OK to use? -- I know that bett=
er way to figure out distribution of parameters is to do bootstrap, but g=
iven limited time I have.....   thanks   "0MINIMIZATION SUC=
CESSFUL NO. OF FUNCTION EVALUATIONS USED:  331 NO. OF SIG. DIGITS IN =
FINAL EST.:  3.3 ETABAR IS THE ARITHMETIC MEAN OF THE ETA-ESTIMATES,=

.. ETABAR:  0.11E-02 SE:      0.23E-01 P VAL.:  0.96E+00=

RIANCE MATRIX SET TO RS*R, WHERE S* IS A PSEUDO INVERSE OF S 1 " =
Received on Fri Feb 27 2009 - 12:22:29 EST

The NONMEM Users Network is maintained by ICON plc. Requests to subscribe to the network should be sent to: nmusers-request@iconplc.com.

Once subscribed, you may contribute to the discussion by emailing: nmusers@globomaxnm.com.