- Hi,

I have been trying to compare the autocorrelation of a dataset, using both spatial distance and other non spatial metrics. Yes, non spatial metrics (I'm a demographer).

1) I have a first problem when I want to compute the distance using more than two non spatial variables (ie, coordinates): I don't know of any software that would let me compute autocorrelation or a semivariogram over various distance ranges*when the distance is to be computed from more than two coordinates*. I'd rather avoid computing distances and autocorrelation with my usual statistical software.

Any suggestion ?

2) Once I do that, I have a further problem comparing spatial and non spatial autocorrelation: distances are follow different metrics and distributions are differently shaped. The idea is to contrast spatial proximity with "proximity" measured with other variables. To compare these distances, should I sort my pairs of observations into distance quantiles (the first 100 pairs, etc.), into standardized distance (with maximum range =100 or average distance=100)?

Any idea on that?

Thanks

CZG

Christophe Z. GuilmotoDemographe, IRDCEIAS-EHESS54, Boulevard Raspail75006 Paris FranceTél.: 06 67 19 87 10

Christophe Z Guilmoto wrote:Hi,

- there is software that lets you calculate variograms in three dimensions

I have been trying to compare the autocorrelation of a dataset, using both spatial distance and other non spatial metrics. Yes, non spatial metrics (I'm a demographer).

1) I have a first problem when I want to compute the distance using more than two non spatial variables (ie, coordinates): I don't know of any software that would let me compute autocorrelation or a semivariogram over various distance ranges*when the distance is to be computed from more than two coordinates*. I'd rather avoid computing distances and autocorrelation with my usual statistical software.

Any suggestion ?

- there is open source software that you can modify for your purpose

- you could use multidimensional scaling to approximate your higher

dimensional space with a lower (2? 3?) dimensional one.

calculation in feature space is always sensitive to scaling, as is calculation of

2) Once I do that, I have a further problem comparing spatial and non spatial autocorrelation: distances are follow different metrics and distributions are differently shaped. The idea is to contrast spatial proximity with "proximity" measured with other variables. To compare these distances, should I sort my pairs of observations into distance quantiles (the first 100 pairs, etc.), into standardized distance (with maximum range =100 or average distance=100)?

Any idea on that?

distances in space-time. In the geostatisics realm I don't know of applications

of the approach you suggest. In machine learning, people use covariance

kernels in feature space, and like in geostatistics the problem is the inference

of a suitable model: isotropy is an illusion when scales don't match naturally.

Last idea: attend geoENV in Neuchatel, which starts tomorrow,

and try to talk to as many people as you can.

Best regards,

--

Edzer

- Dear Christophe,

Maybe you could use Mantel statistics. You first need a relevant multivariate

distance measure (see e.g. Legendre & Legendre 198? "Numerical Ecology"), then

you create a distance matrix with distances in simillarity (m x n). Finally the

geographical distance matrix (m x n) is correlated with your multivariate

distance matrix through Monte Carlo randomisations. Another reference here is

'isolation by distance' in Sokal & Rohlf 199? "Biometry".

Jakob

Quoting Christophe Z Guilmoto <guilmoto@...>:

> Hi,

--

> I have been trying to compare the autocorrelation of a dataset, using both

> spatial distance and other non spatial metrics. Yes, non spatial metrics

> (I'm a demographer).

>

> 1) I have a first problem when I want to compute the distance using more

> than two non spatial variables (ie, coordinates): I don't know of any

> software that would let me compute autocorrelation or a semivariogram over

> various distance ranges when the distance is to be computed from more than

> two coordinates. I'd rather avoid computing distances and autocorrelation

> with my usual statistical software.

> Any suggestion ?

>

> 2) Once I do that, I have a further problem comparing spatial and non

> spatial autocorrelation: distances are follow different metrics and

> distributions are differently shaped. The idea is to contrast spatial

> proximity with "proximity" measured with other variables. To compare these

> distances, should I sort my pairs of observations into distance quantiles

> (the first 100 pairs, etc.), into standardized distance (with maximum range

> =100 or average distance=100)?

> Any idea on that?

>

> Thanks

>

> CZG

>

>

>

>

> Christophe Z. Guilmoto

> Demographe, IRD

> CEIAS-EHESS

> 54, Boulevard Raspail

> 75006 Paris France

> Tél.: 06 67 19 87 10

Jakob Petersen

Research Technician

School of Biological Sciences

M. Trimmer laboratory (1.05)

Queen Mary

University of London

Mile End

London

E1 4NS

United Kingdom

Tel +44 (0)20 7882 3200

Fax +44 (0)20 8983 0973

Directions:

http://www.qmul.ac.uk/contact/directions.shtml

Map:

http://www.qmul.ac.uk/contact/mileend.shtml

We are in lab 1.05 on the 1st floor in building 24. The main entrance is to the

East.