[R-sig-Geo] biparametric model

[eaeff at mtsu.edu]

Roger,
Thanks for your response--as always (and I've lurked here a while), you were very helpful. 
There are three contexts I can think of in which two weight matrices would be desirable:
1. For a gravity model, where each observation is a link between source region and destination region, one spatially lagged variable for the source and one spatially lagged variable for the destination (this is the original Brandsma and Ketellapper application).
2. There is a recent paper by Donald J. LaCombe (“Does Econometric Methodology Matter? An Analysis of Public Policy Using Spatial Econometric Techniques”, Geographical Analysis, vol. 36, no. 2, April 2004, pp 105–118.) in which he uses two weight matrices to disentangle border effects.
3. There is a literature in anthropology and cross-cultural research, pretty much fully developed by the mid-1980s, in which “Galton's Problem” is addressed by introducing two kinds of spatial dependence: via physical proximity and via cultural proximity. (e.g., Malcolm M. Dow, Michael L. Burton, Douglas R. White, Karl P. Reitz. (1984) “Galton's Problem as Network  Autocorrelation.“ American Ethnologist. 11(4):754-770 )

My own research interest is actually the third option above. 

best,
Anthon


<Anthon Eff>
Please forgive what may be a naive question. I'm working on a model that requires two different weight matrices, as in the "biparametric" model introduced by Brandsma and Ketellapper in 1979. I haven't had any luck finding a way to do this in R. Any suggestions?

<Roger Bivand>
The spatial error model fitting functions spautolm() and errorsarlm(), and the spatial lag model fitting function lagsarlm() only fit a single set of spatial weights, by optimising in one dimension.

It would be possible to generalise them to optimise in more than one dimension, but is it justified? Is there a very clear behavioural model that requires the fitting of more than one spatial regression coefficient that is driving thw question? Is it going to be practical to fit more than one coefficient on a probably rather flat surface?

The slm() function in S-Plus SpatialStats module can do this for the error model if need be, but there is a case to be made for why it is necessary, unless there is a clear behavioural model.

Roger