[ADMB Users] parameterization of Cholesky factor

dave fournier davef at otter-rsch.com
Sat Nov 15 08:00:11 PST 2014

Actually using the choleski decomp to parameterize the covariance matrix 
is an old idea.

see Pinheiro and Bates.


But what is new in my approach is the particular parameterization used 
for the choleski decomp of the correlation
matrix. P and B suffer from an inability to deal with dependent 
variables.  This is the common R failing which
permeates the entire R approach to nonlinear parameter estimation.

I note that P and b remark that there are multiple choleski decomps of 
the covariance matrix leading to the same
covariance matrix and this may cause difficulties.  I believe that by 
starting with the correlation matrix and
then scaling it by   the std devs parameterized on the log scale leads 
to a unique parameterization.  So I think the entire thing
is superior to anything P and B did.

I suppose I should have published it but reviewers suck so bad they ruin 
my day.

More information about the Users mailing list