[R-sig-Geo] Is there a lagsarglm?

[Roger Bivand]

On Fri, 2 May 2008, Bjarke Christensen wrote:

> Hi,
>
> I can't seem to find any implementation of spatial probit models in R.
> Perhaps my limited understanding of spatial statistics is preventing me
> from seeing what should be obvious, but it does not seem that the methods
> in geoRglm are of help in that sense. Can anyone help?

Well, contributions are always welcome! This is near the area of GLMMs 
which is seriously hard for most cases. Even if models can be fit, there 
is no agreement on how to put distributions around the fitted 
coefficients, or on how to assess their significance (MCMC being a 
computational route that at least gives something, if the assumptions in 
the prior hold).

You are thinking of a model with the spatially lagged limited dependent 
variable on the RHS, right? Are you sure of your specification?

Could you try tuning a regular glm() fit (or gam() with a spline term in 
the observation coordinates) - is the dependent variable "well behaved" - 
is its distributional shape as expected (zero-inflation, dispersion, etc)? 
Do you need case weights? Is it a rate - could it be modelled directly 
with a log offset (like epidemiologists do)? If you can establish where 
the spatial dependence is coming from, you may find that you can get to a 
residual with no marked patterning (but testing isn't obvious either).

I imagine that you have looked at the Matlab toolbox, but that is pursuing 
a specific route, one liked by econometricians, but one that say 
Schabenberger & Gotway probably would leave until after checking out the 
alternatives. There is so little use of spatial "probit" - in glm() for 
the binomial family, the default link is "logit" not "probit" - so we 
really don't know how credible results from implementations are.

Sorry not to be more help, but getting to a glm(..., 
binomial(link="probit")) with clean residuals will be more robust that 
one-off "spatial" code (as would be a gam() with an s(x,y) term - if it 
isn't significant, there probably isn't a spatial story left).

Roger


>
> Venlig hilsen
> Sydbank
>
> Bjarke Christensen
> økonom
>
> Forretningsprocesser - Modeller
> Peberlyk
> DK-6200 Aabenraa
> Direkte tlf. 74 36 44 72
> Telefax
> bjarke.christensen at sydbank.dk
> _______________________________________________
> R-sig-Geo mailing list
> R-sig-Geo at stat.math.ethz.ch
> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no