[R-sig-Geo] Re:multivariate moran & lisa for spdep

[Nicholas Lewin-Koh]

Hi,
I remember in the Wartenberg paper there being a big problems;
1) The moran matrix is not symetric
2) The matrix can be non-positive definite, and I think not even
non-negative definite.

I believe Robin Reich worked out a bivariate Moran's I, and I think there
is a
publication in Environmental and Ecological Statistics, I don't remember
which issue.
I worked on the probelm a bit and came up with a combination grand-tour
and Moran Scatter plot.
In otherwords, project into 1-dimension using a random orthonormal basis,
calculate Moran's I on 
the projected data. The derivatives can be easily derived (quadratic
form) and the function can be optimized ocver projections to
maximize/minimize Moran's I. I had some other thoughts on this at the
time, but I would have to dig out my notes. There is an interactive
implementation of this in java nested in the ORCA libraries.

Also along these lines William Christensen did some really nice work on
spatial factor analysis:

Christensen, W. F., and Amemiya, Y. (2001).  "Generalized shifted-factor
analysis method for multivariate geo-referenced data," Mathematical
Geology, 33, 801-824

Christensen, W. F., and Amemiya, Y. (2002).  "Latent variable analysis of
multivariate spatial data," Journal of the American Statistical
Association, 97, 302-317.

I think ther solution was in the cases where the spatial covariance was
not pd they nudged the matrix until it was. I think aysymptotically there
is no bias.

Anyway if i am wrong about the multivariate Moran's I, I am sure Luc will
set me straight.

My 2c

Nicholas