Spatial regression

Thank you Roger,

In a paper (
http://www.ssc.wisc.edu/cde/methodology/statcore/Balleretal2001.pdf) that
uses spatial regression in homicide rates analysis (and discusses the
difference between spatial and lag models) there is a case where over time
there is a switch between the best model that fits the data, from a lag
model to an error one.

In this context ? a switch over time between an error to lag or vice versa ?
using just lm.LMtests function might not be very helpful in understanding
what is behind the change.

If, as an example, there is a change in better fit from an error to a lag
model what would be the steps to understand what caused the change.

So may question is: is it comparing spatial Durbin model (lagsarlm() type =
?mixed?) with spatial lag model (lagsarlm()) and also comparing Durbin model
with errorsarlm() a more structured approach?

Thanks again.



Adrian




On Sat, Jun 6, 2009 at 1:24 PM, Roger Bivand <Roger.Bivand at nhh.no> wrote:

 On Fri, 29 May 2009, Adrian Toti wrote:

Hi everyone,

In R, the function lm.LMtests is used to select which spatial model fits
better the data, an error or lag one. It looks to me that there are some
?limitations? using this function like being constraint to use only
row-standardized spatial weights and it does not support weighted lm
object
as argument. Also it is not helpful when both LM tests (LMerr and LMlag)
are
significant (let?s say their robust forms are not).

So my question is: is there another way or approach to select between a
spatial error and lag model?

As you appreciate, the two models are not nested, so an LR test is not
appropriate, and comparing the AIC values probably not either (although a
large difference in AIC should not be disregarded). Neither errorsarlm() not
lagsarlm() use weights - this is only possible in spautolm(), which is an
alternative implementation of the same model as errorsarlm(). So I'm not
sure how you could fit a weighted lag model.

There is a paper by Kelejian in Letters in Spatial and Resource Sciences in
2008 on a spatial J-test, but this relies on feasible GM and spatial TSLS
estimates, which most often also assume row standardisation and do not
permit weights. A recent paper by Burridge and Fingleton (from a meeting
earlier this week) discusses some alternatives:

http://sew2009.univ-fcomte.fr/english/Burridge_Fingleton.pdf

but these are too fresh to be coded and packaged yet (and are also in the
spatial econometrics view of the world).

Hope this helps,

Roger

Thanks for your time.



Adrian

       [[alternative HTML version deleted]]

--
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