[R-sig-Geo] Help with interpreting correlog results, package ncf

[Beazley, Lindsay]

Dear Roger Bivand and list members,

I am looking for advice on how to interpret a correlogram of species
richness data collected from photographic transects of the seabed. I am
an amateur in R, and am completely new to spatial analysis, so please
forgive the naivety of my issue.

I have four photographic transects consisting of photos at varying
distances apart. Photos within transects are not at equal distances
apart, and the transects themselves also vary in distance from one
another. Three of these transects are relatively close together (within
a 26 km diameter), while the fourth transect is roughly 100 km away from
the first three. The total distance between the two farthest photos is
138 km. 

Using glm I have fit a poisson distribution with log link to determine
the influence of four environmental variables on species richness. This
was fit using all photos across all transects. I am interested in
determining whether there is spatial autocorrelation present in the
residuals of this model. I used correlog( ) in package 'ncf' to create a
Moran's I correlogram. The code I used is: 

cor1<-correlog(data$Longitude, data$Latitude, residuals(glm1),
increment=1, latlon=TRUE, resamp=1000). 

Latlon was set to true as my coordinates are geographical. I set
increment=1 as I understood each distance class would represent 1 km. I
then plotted the correlogram using plot(cor1$correlation[1:30]). There
are 30 distance classes. The correlogram is attached to this email. I am
confused on how to interpret the results of the correlogram. I have 30
distance classes, although the scale of my data is roughly 138 km,
therefore each distance class does not equal 1 km as I expected. The
output of cor1 (attached) shows 30 distance classes, although they are
labelled 1-8, 19-26, 111-119, and 134-138. I'm assuming the distances
classes are broken up like this because the photos are not contiguous?
Should the autocorrelation therefore be assessed within transects
instead of across them? 

My second question relates to significance testing. I set resamp to 1000
and generated the p values for each distance class. Is this a suitable
method for testing the significance of the autocorrelation at each
class? In this case my first distance class is significant (p=0.028).
>From reading other list postings there doesn't appear to be a method to
test for significant autocorrelation in glm fits (although there is for
lm fits). 

Thanks for any advice you can provide.

Cheers,

Lindsay

 <<Correlogram_SpeciesRichness.pdf>>  <<Cor1.txt>> 




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20130130/c837a499/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Correlogram_SpeciesRichness.pdf
Type: application/pdf
Size: 2512 bytes
Desc: Correlogram_SpeciesRichness.pdf
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20130130/c837a499/attachment.pdf>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: Cor1.txt
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20130130/c837a499/attachment.txt>