[R-sig-Geo] RE : question about autocov_dist function in "spdep" package

[Julien Beguin]

Thank you Roger for your fast response and Jenn for cc:ing your message about autocov_dist in spdep. I am far to be a specialist in spatial statistic and I know that autologistic models have been criticised in inferential context (e.g. Dorman et al. ). Several authors (in ecology, at least) also found it useful in prediction context (Betts et al.). I guess that the debate might not be over...
 
My objective is not to trust results of autologistic model without criticism but rather to compare them with other approaches (e.g. Moran eigenvector, spatial regression, etc...). 
 
Sorry for my ignorance but what means:

"general weights sum to zero". Is autocovariate term (not) calculated as follows ? A = sum(wij*yj)/ sum(wij) where wij=1/euclidian distance between i and j (for inverse distance). It is maybe a stupid question but how could weights sum to zero?

Also, it is not false to say that I am asking to include a constant vector of zeros, but only for points that have no neighbour in a single region. Imagine a study area composed of 5 regions containing each N presence/absence points with associated covariates. If I calculate the autocovariate term for each region separately (with constant nbs) and that for one region (not the others) the minimum distance between pair of points < nbs value, it will crash for that region. It might be that spatial autocorrelation vary from region to region, and that in a specific region, A value based on a constant nbs among regions would be zero (because of no neighbour) for that region.   

A GLMM with fixed factors: Y~ covariates + Autocovariate; and random factors = 1|Region/Autocovariate could account for that, no?

Thank you again for your time,

Julien 
_______________________________________
De : Roger Bivand [Roger.Bivand at nhh.no]
Date d'envoi : 25 juin 2010 15:55
À : Julien Beguin
Cc : r-sig-geo at stat.math.ethz.ch
Objet : Re: [R-sig-Geo] question about autocov_dist function in "spdep" package

On Fri, 25 Jun 2010, Julien Beguin wrote:

> Dear list member,
>
> I am using the autocov_dist function in spdep on binary point data set
> to estimate the autocovariate to be used in autologistic regression
> (session info at the end):
>
> coords<-as.matrix(cbind(datafile$X_COORD, datafile$Y_COORD))
> ac1 <- autocov_dist(datafile$binary_variable, coords, nbs=400, type = "inverse", zero.policy=TRUE)
>
> It works fine with large value of nbs. When nbs value is below the neibhorhood distance of some pairs of point I get warnings because of empty neighbors:
>
>> Warning messages: 1: In autocov_dist(datafile$binary_variable, coords,
>> nbs = 400, type = "inverse", :

>> With value 400 some points have no neighbours
>> 2: In nb2listw(nb, glist = gl, style = style, zero.policy = zero.policy) :
>> zero sum general weights
>
> It seems, but I might be wrong, that points with no neighbour have an
> autocov_dist value of zero. Since it is what I want, it is ok for me.

No, the general weights sum to zero, as it says, you are using inverse.

> But if I decrease nbs value below the smallest neighbourhood distance
> among pairs of points in my data set, I get an error message:
>
>> Error in nb2listw(nb, glist = gl, style = style, zero.policy = zero.policy) :
>>  No valid observations
>> In addition: Warning messages:
>> 1: In autocov_dist(datafile$binary_variable, coords, nbs = 100, type = "inverse",  :
>> With value 100 some points have no neighbours
>> 2: In nb2listw(nb, glist = gl, style = style, zero.policy = zero.policy) :
>>  zero sum general weights
>
> My question is as follows: why do autocov_dist function require that at
> least one neighbor pair exists? and why not returning zero value for all
> points when no any point has a neighbor?
>
> This might appear to be a strange question but in my case I estimating
> autocov_dist for different spatial block in which the minimum distance
> between points vary from block to block, but still I would like a commun
> nbs value for all block.

It does indeed seem very strange, as you are asking to include a constant
vector of zeros, which will alias the constant. The function itself is a
really bad idea, and is included only to show (see Dormann et al 2007)
that using it is inferior to all other methods examined there. Trying to
include a second zero constant doesn't seem well-founded, to put it
mildly.

Hope this helps,

Roger

>
> Thank you for your help,
>
> Julien Beguin
> --------------------
> Ph.D. student
> Laval University
> Québec, Canada
>
> SESSION INFO:
>
> R version 2.10.1 (2009-12-14)
> i386-pc-mingw32
> locale:
> [1] LC_COLLATE=French_Canada.1252  LC_CTYPE=French_Canada.1252
> [3] LC_MONETARY=French_Canada.1252 LC_NUMERIC=C
> [5] LC_TIME=French_Canada.1252
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base
> other attached packages:
> [1] spdep_0.5-4         spam_0.20-3         coda_0.13-4
> [4] deldir_0.0-12       maptools_0.7-29     foreign_0.8-40
> [7] nlme_3.1-96         MASS_7.3-4          Matrix_0.999375-33
> [10] lattice_0.17-26     boot_1.2-41         sp_0.9-60
> [13] compositions_1.01-1 robustbase_0.5-0-1  tensorA_0.31
> [16] rgl_0.91
> loaded via a namespace (and not attached):
> [1] grid_2.10.1  tools_2.10.1
> _______________________________________________
> 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