Re: A basic question on Canonical Correlation Analysis

Hi Gary -

happy, that you enjoyed my program.
Concerning the normalization:
it may be, that the difference between my
solution and published results is smaller,
if in that results they would use only
the significant orthogonal components. In
my case that were only two, and find the
canonical components only on that basis.
That would exclude PC3 und PC4, which have
nearly null-loadings, from the normalization
and give only slightly other results, with
difference to my solutione so small, that
one could easily overlook them.

The other way round it could be, that "thne
normalization" is already included in the
step, where I generate latent PC-components
with the variance 1 - since variance 1 for
each component may be termed "normalization"
since all sum-of-loadingssquares of the
PC's (which differ from 1, and are the
eigenvalues of the correlation-matrices
of each set separately) are replaced by
the "normalized" value of 1.

But still I do not know this exactly, but
I'm still confident, that my method includes
everything since I checked some examples
against the solution of SPSS (though not really
many and possibly no critical).

The reference, btw, is Stan Mulaik,
The Foundations of Factor Analysis, 1972,
Pg 417-420; chap 16-4, canonical correlation.

Yes- I'm interested in your further comments

sincerely -

Gottfried
.