[R-sig-Geo] Help for spdep package

[Roger Bivand]

On Tue, 18 Dec 2012, Saman Monfared wrote:

> Dear all.
> I try to fit conditional auto regressive models (CAR and  SAR) in package spdep.
> Also, I have fitted some other models like GLM, Empirical Bayes and ...
>
> My program to CAR and SAR is:
>
> esar1f1 <- spautolm(IMR.m~ 0+PCFP+Sup+B.F,data =data1,
> listw=nb2listw(nb6, style="W"), family="SAR", method="full", verbose=TRUE)
>
> summary(esar1f1)
>
> esar1f2 <- spautolm(IMR.m~ 0+PCFP+Sup+B.F,data =data1,
> listw=nb2listw(nb6, style="W"), family="CAR", method="full", verbose=TRUE)
>
>> esar1f1
>
> Call:
> spautolm(formula = IMR.m ~ 0 + PCFP + Sup + B.F, data = data1,
>    listw = nb2listw(nb6, style = "W"), family = "SAR", method = "full",
>    verbose = TRUE)
>
> Coefficients:
>      PCFP        Sup        B.F     lambda
> 0.1711382 -0.2262700  0.2414336 -1.7946991
>
> Log likelihood: -56.24405
>
>
> What is lambda?? I couldn't find a good refrences to understand tese model!!!

See the help page (lambda is said to be the ML autoregressive coefficient 
about half way down), and certainly read Waller & Gotway (it may be that 
they term the spatial coefficient rho, but lambda is used here for 
consistency with errorsarlm()).

> How can I predict this model for new data??

There are no predict methods provided - but analysis of the model will 
show that \hat{y} is X \beta, isn't it?

> How can I cross validate results of CAR and SAR??
>

They are not nested processes, but as Clif & Ord, Ripley and the others 
show, a CAR and a SAR may be expressed in terms of one-another.

By the way, your lambda may me maxing out at the bottom of its range, so 
be careful, very strong negative autocorrelation is very unlikely.

Roger

>

-- 
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