Skip to content

Local indicators for categorical data - spatial analysis

3 messages · Guilherme Ottoni, JLong, Roger Bivand

Original
Guilherme Ottoni
# 
I think the last e-mail was not ideally configured for the list, so
I'm re-sending it. Sorry if it duplicates.


Hi listers,

I'm in the final sprint at my thesis in poverty and I would like to
try to find clusters, like LISA's clusters, for vulnerability to
poverty. However, my data is categorical, 3 categories, and LISA suits
better in continuous data.

I have been searching about any software or script which executes the
join-count test for 3 categories at individual level. In other words,
a LISA ,or LICD as Boots named it in the paper Developing Local
Measures of Spatial Association for Categorical Data.

All I have found so far is the Join-Count statistic for the whole data.

Does anyone know if something has already been done at this area?

PS: sorry for the grammar errors.

Guilherme
JLong
# 
Hi Guilherme,

Send me an email to jed.long at st-andrews.ac.uk
I think I have some software/code that Barry Boots developed way back then
somewhere on my machine...

Cheers,
Jed




-----
Jed Long
Lecturer in GeoInformatics
Department of Geography & Sustainable Development
University of St Andrews, UK
--
View this message in context: http://r-sig-geo.2731867.n2.nabble.com/Local-indicators-for-categorical-data-spatial-analysis-tp7587775p7587776.html
Sent from the R-sig-geo mailing list archive at Nabble.com.
Roger Bivand
#
On Wed, 11 Feb 2015, JLong wrote:

            

      
      
      
    

        
LICD are not implemented, but you can get part of the way (just the 
tabulations) by:

library(spdep)
data(oldcol)
HICRIME <- cut(COL.OLD$CRIME, breaks=c(0,35,80), labels=c("low","high"))
names(HICRIME) <- rownames(COL.OLD)
lw <- nb2listw(COL.nb, style="B")
joincount.multi(HICRIME, lw)

for a global baseline, and

Vis <- lapply(1:length(HICRIME), function(i) listw2star(lw, i, "B",
   n=length(HICRIME), D=NULL, a=NULL))
ljcs <- t(sapply(Vis, function(x) { res <- joincount.multi(HICRIME, x,
   zero.policy=TRUE); res[-nrow(res),1]}))

There are warnings because inference fails (as it should), but the tally 
is the same, with:

apply(ljcs, 2, sum)/2

giving the counts of joins by type, dropping double-counting in the 
summation. listw2star() is used in localmoran.sad() and elsewhere to 
generate observation-wise weights objects. This isn't an implementation of 
LICD in that the subregion for each i is simply its set of neighbours, not 
a moving window, but at least provides a basis for classification.

As with other LISA, it is important to remove the effects of covariates, 
so that an inferential approach might be based on GLM residuals (just a 
speculation).

Hope this helps,

Roger