[R-sig-Geo] Is GWR in ggwr function with poisson distribiution the same as Geographically Weighted Poisson Regression (GWPR) disused by Nakaya 2005

[Roger Bivand]

On Fri, 28 Dec 2012, Roosbeh Nowrouzian wrote:

> Dear list members:
>
> I am wondering if Geographically weighted Regression(GWR) applied in
> Generalized Linear Models(GLM) framework with defined family="poisson" in
> ggwr.sel and ggwr functions is the same as Geographically Weighted Poisson
> Regression (GWPR) disused by Nakaya et al. 2005 for disease association
> mapping.

Since GWR3 and GWR4 are closed-source implementations, nobody knows 
whether they implement Nakaya et al. 2005 either. If you read the code of 
ggwr(), you'll see that it uses GW on glm() in a way that is probably not 
dissimilar to that proposed. But you'd need to check carefully on a test 
data set to see whether the local coefficients agree - probably glm() is a 
more robust GLM than that in GWR, certainly the OLS in GWR used to be 
based on very old matrix inversion code, something nobody serious does 
nowadays, because it is not numerically stable. You should also be 
concerned about overdispersion.

Try rather GLMM or GAMM unless you have an unmeasurable non-stationarity - 
if it is measurable, re-specify the model to include it and forget GWR.

Roger

>
> Thanks
>
> Roosbeh Nowrouzian
> PhD Student
> University of Florida
> Department of Civil and Coastal Engineering
>

-- 
Roger Bivand
Department of Economics, NHH Norwegian School of Economics,
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