[Openspace] minimum distance sampling and spatial autocorrelation

[Julia Koschinsky]

Victor,

If your new deforested surface variable is continuously 
distributed, you should be fine.

I'm not sure about the details of how you constructed the 
correlogram: e.g., if you use Moran's I as your Y, are you 
using different distance band weights matrices? Note that if 
you are comparing Moran's I values with different matrices, 
you want to you the standardized Moran's I values (deduct 
the mean and divide by the standard deviation, which are 
reported when you right-click on the Moran scatterplot and 
go to permutations) instead of the unstandardized Moran's I 
values. This is because unstandardized Moran's I values are 
not comparable between different weights matrices.

The idea with the sampling design is that you identify a 
distance (e.g., may be 60km in your case) within which 
observations are basically substitutable so that a sample 
point is representative of the ones within the same 
distance. By using only uncorrelated sample points, you 
address the problem of spatially autocorrelated points. I am 
sending you Griffith's article in a separate message for 
personal use.

If you want to use a spatial model in GeoDa instead of the 
spatial sampling design, you could regress your deforested 
surface variable on your independent variables using a 60km 
distance band weights matrix.

Feel free to send me the graph with more info on how you 
implemented it if the above information did not address all 
of your questions.

Best regards,
Julia 

---- Original message ----
>Date: Tue, 13 Jun 2006 13:09:19 -0600
>From: Victor Hugo Ramos WCS <vhramos at wcs.org>  
>Subject: Re: [Openspace] minimum distance sampling and 
spatial autocorrelation  
>To: koschins at uiuc.edu
>Cc: openspace at sal.uiuc.edu
>
>Julia:
>
>First, thank you for your fast response.  I think that 
GEODA is a 
>wonderful tool.
>
>Second, I think that I found a workarond on the problems 
related to SAC 
>and binary dependent variables.  What I did was to instead 
of using 
>Deforested/Forested data, I extracted deforested surfaces 
from 25 pixel 
>windows and then used that data in the SAC analysis.  The 
analysis was, 
>apparently correct, and I built a correlogram with MORAN 
(Y) and 
>distance between pairs of points (X) that shows that there 
is 
>significant SAC until the distance reaches 60 km, and then 
goes to 
>significant negative values to go up again to positive 
significant 
>values until it reaches the upper limits of the maximun 
distance between 
>points.  The distance where there is no significant SAC is 
useless for 
>me in terms of sampling because my study area is relatively 
small and 
>I´m not going to be able to make enough sampling units to 
develop my 
>model.  I´m trying to access the sampling paper that you 
recomended me, 
>but in between that, giving what I have explained before 
can you give 
>some general advice on the following questions:
>
>Is the workaround that I implemented correct?
>Do you have any recomendations on sampling design giving 
the behavior of 
>Moran as I explained before?  Can I send you the graph in 
order to 
>better visualize the information?
>
>Any help is greatly appreciated.
>
>Victor Hugo Ramos
>Wildlife Conservation Society
>Guatemala, Central America
>
> this, I have a couple of questions:Julia Koschinsky wrote:
>
>>Victor,
>>
>>You raise two issues that have to be addressed with 
software 
>>other than GeoDa: 1) Controlling for spatial 
autocorrelation 
>>through sampling design and 2) running spatial 
>>autocorrelation tests with a binary dependent variable.
>>
>>1) On the first issue, an example of a recent reference is:
>>
>>Griffith, Daniel A. (2005). "Effective Geographic Sample 
>>Size in the Presence of Spatial Autocorrelation," Annals 
of 
>>the Association of American Geographers, Vol. 95, 
December, 
>>pp. 740-. 
>>
>>2) On the 2nd issue, some of the references on our site on 
>>spatial probit include:
>>
>>http://sal.uiuc.edu/courses/se/pdf/w13_probit_notes.pdf
>>http://sal.uiuc.edu/courses/se/pdf/w13_probit_out.pdf
>>http://sal.uiuc.edu/users/anselin/papers/hood.pdf (pp. 8-9)
>>
>>You might be able to find experimental code on spatial 
>>probit in R or Python.
>>
>>Julia
>>
>>---- Original message ----
>>  
>>
>>>Date: Tue, 06 Jun 2006 08:39:57 -0600
>>>From: Victor Hugo Ramos WCS <vhramos at wcs.org>  
>>>Subject: [Openspace] minimum distance sampling and 
spatial 
>>>    
>>>
>>autocorrelation  
>>  
>>
>>>To: openspace at sal.uiuc.edu
>>>
>>>How do I use GEODA in order to lay out a sampling design 
>>>    
>>>
>>that avoids 
>>  
>>
>>>spatial autocorrelation giving the following conditions:
>>>
>>>
>>>
>>>-         I want to build a logistic regression model 
that 
>>>    
>>>
>>is going to 
>>  
>>
>>>use deforestation as dependent variable and a number of 
>>>    
>>>
>>environmental 
>>  
>>
>>>and physical layers as independent variables
>>>
>>>-         Deforestation is coded 1 for deforested and 0 
for 
>>>    
>>>
>>not deforested
>>  
>>
>>>-         I just want to test spatial autocorrelation on 
>>>    
>>>
>>deforestation
>>  
>>
>>>-         The main value coming from the test for spatial 
>>>autocorrelation should be a minimum distance between 
samples
>>>
>>>
>>>
>>>Thanks in advance for your help
>>>
>>>Victor Ramos
>>>_______________________________________________
>>>Openspace mailing list
>>>Openspace at sal.uiuc.edu
>>>http://sal.uiuc.edu/mailman/listinfo/openspace
>>>    
>>>
>>
>>  
>>
>