From: "Rik Schoemaker" <rs@chdr.nl>
Subject: full covariance matrix
Date: Mon, 11 Jan 1999 22:44:32 +0100
Dear fellow nmusers,
If I were to analyse a model say with 3 parameters and use a constant CV inter-individual error model:
P1 = THETA(1)*EXP(ETA(1))
P2 = THETA(2)*EXP(ETA(2))
P3 = THETA(3)*EXP(ETA(3))
and I would like to estimate the full covariance matrix:
$OMEGA BLOCK(3) .1 .01 .1 .01 .01 .1
then what is the meaning of the off-diagonal estimates (the ¡covariances¢) and can they be translated back to for instance correlation coefficients?
Can anyone comment on whether I *should* use a full covariance matrix (provided the data can support all the extra parameters)?
Thanks a lot,
Rik Schoemaker
Centre for Human Drug Research,
Leiden, Netherlands
rs@chdr.nl
full covariance matrix
From: Lewis Sheiner <lewis@c255.ucsf.edu>
Subject: Re: full covariance matrix
Date: Mon, 11 Jan 1999 22:19:24 -0800
Rik,
In general, you "should" use a full cov matrix, as any other choice is a modeling choice that says you know that certain random effects are uncorrelated. You would have to have a scientific basis for making such an assertion.
Off diagonal terms can be thought of as correlations, when transformed to by normalizing to the the cv's of the corresponding diagonal elements - that is cov(a,b)/sd(a)/sd(b) ... the fact that they are CV's doesn't change this ...
LBS.