[R-sig-Geo] How to estimate a SDM with IV method?
Roger.Bivand at nhh.no
Tue Jul 13 21:33:45 CEST 2010
On Tue, 13 Jul 2010, Paolo Veneri wrote:
> I have a cross section of 158 spatial units. Robust LM tests suggest me to
> estimate a Spatial Durbin Model (SDM). However, given the non-normality of
> OLS residuals, the presence of
> heteroschedasticity and the presence of one endogenous explanatory
> variable, I should not use the classic "lagsarlm" function, since I should
> not use ML methods.
> The point is that I did not find any package in R that estimate a SDM with
> IV method (hence correcting for endogeneity and heteroschedasticity).
> Would you suggest any strategy?
You are quite correct that the structuring of the stsls (and equivalent
heteroskedastic version in the sphet package) makes it effectively
impossible to fit a Spatial Durbin model. Even if one tries (using higher
lags by hand), the results are typically numerically unstable. So you are
left with ML - I would not worry about the distribution of the residual
too much, but analysis of outliers from the influence measures of the
linear fit might suggest a missing variable, or possibly a dummy that
could releive som heteroskedasticity. You could check this out on the
linear model first, and then use ML to fit the improved SDM model.
The only SDM with heteroskedastic errors that I know of is the Bayesian
approach in the Matlab toolbox, documented in LeSage and Pace (2009). As
I'm sure you know, you need to take the impacts of the RHS variables into
account, rather than interpreting the SDM coefficients - this is provided
for in impacts() methods in spdep for the SDM model.
Hope this helps,
> R-sig-Geo mailing list
> R-sig-Geo at stat.math.ethz.ch
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
More information about the R-sig-Geo