Skip to content
Prev 26 / 29559 Next

conditional autoregressive models

The main catch would be to make sure that W is symmetric, since
the variance becomes (I - rW)-1 not the cross product as in
SAR (which is guaranteed to be symmetric, irrespective of W).
Also, the EGLS part would be slightly different, since what is
needed is X'(I - rW)X not X'(I - rW)'(I - rW)X.
More fundamentally though, one should make sure there is a good
reason to use CAR rather than SAR, they imply different correlation
structures but also different conceptual models.
L.
On Thursday, October 30, 2003, at 07:03 AM, Roger Bivand wrote: