[R-sig-Geo] Choosing the right nb2listw-style

[Breitbach, Nils]

Dear Roger, dear Community,

thanks for the detailed answer to my question(s)! It helped a lot, but on the other hand cause some more uncertainties for me.

Why do some functions for the Moran's I test require detailed knoledge of creating neighbours and weighting them like the moran.test() function while others do not like the Moran.I() function? The moran.test() function from the spdep package requires a listw object, where you have to create a weighted neighbourhood matrix with all the preliminary considerations, while the Moran.I() function sipmly works with a ad hoc created distance matrix (e.g. done with the dist() function) as it was done here: http://www.ats.ucla.edu/stat/r/faq/morans_i.htm (even though Moran.I expects "a matrix of weights" looking closer to the help files for the function). Can the distances itself be considered as weights? The method with the Moran.I() function seems much simpler and requires a lot less knoledge and preliminary considerations ...

Thanks for help in this difficult field!
Regards,

Nils


________________________________________
Von: Roger Bivand [Roger.Bivand at nhh.no]
Gesendet: Montag, 23. August 2010 08:15
Bis: Breitbach, Nils
Cc: r-sig-geo at stat.math.ethz.ch
Betreff: Re: [R-sig-Geo] Choosing the right nb2listw-style

On Sun, 22 Aug 2010, Breitbach, Nils wrote:

> Dear Community,
>
> to be able to evaluate the spatial autocorrelation within may data I am
> forced with the question of how to correctly choose the neighbours for
> my data. My study plots (points) are not evenly distributed over my
> study area (approx. 30 x 30 km in size) and also the land-use type is
> not evenly distributed over the study area and therefore I want to
> evaluate the spatial autocorrelation for this data set. To finally
> calculate the Moran's I or plot variograms/correlograms I now need to
> calculate the neighbourhood relationships of my study plots. For the
> given characteristic of my data (especially their non-even distribution)
> I am now somewhat uncertain about the right style (W, B, C, U or S) of
> the nb2listw() object that suits best for my kind of data.
>
> Can anyone recommend the "right" style for my kind of data?

Typically, varying sub-discipline communities have different prefered
flavours, both of the neighbour list object, for using general weights (or
not) - including inverse distance weighting, and for using
row-standardisation (W), raw (binary or general - B), or standardised raw
(C - sum to n, U - sum to 1). There isn't a tradition dor using variance
stabilising (S) although there probably should be. It seems sensible to
see what others in your field use, and choose among those. The same for
schemes for finding the neighbours to start with. Using an approach which
is unusual in your field will attract referees' attention to your choice -
they will want to know why you are doing something different. Since you
are in ecology, look at papers using weights there, and unless you can see
that the modal scheme is suboptimal for you, go with the stream.

Note however that some of the graph-based neighbour schemes advanced early
on by Sokal in ecology are little used, and probably deserve more
exposure, especially when the distances between observations differ a lot
- leading to observations in dense parts of the study area having many
neighbours in schemes using a distance threshold. Try to think about the
plausibility of the science in the implied spatial process - could
observations realistically influence each other at that distance? It may
not matter if the weights are only "mopping up" unwanted spatial
autocorrelation, but if the dependencies have a substantive
interpretation, it isn't wise to imply mutual dependence that isn't
scientifically plausible (think of natural boundaries that organisms
cannot "cross" as well as distances). But no, no "right" scheme as such -
it's up to you! Pay attention to the inhomogeneity of your setting too, as
it may induce apparent dependency if not modelled.

Hope this helps,

Roger

>
> Thanks for help!
> Regards,
>
> Nils
>
> _________________________________________________________
>
> Nils Breitbach, Dipl.-Biol.
> Institut für Zoologie, Abt. 5: Ökologie
> J.-J.-Becher-Weg 13
> Johannes Gutenberg-Universität
> 55128 Mainz
> Germany
>
> phone: +49 6131 39-22718
> fax:   +49 6131 39-23731
> WWW: www.community-ecology.uni-mainz.de/126_ENG_HTML.php
> _________________________________________________________ x k,
> _______________________________________________
> 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