gaike

autocorrelation question

Discussion created by gaike on Jan 23, 2014
In my analysis, I have calculated elephant fence breaks per 100m  (along a 55km electrified fence).  Because of the skewed distribution of fence breaks /100m (many sections with 0 breaks , one section with 38 breaks).  i have binary coded the outcome variable (i..e. each 100m section) to have breaks/no breaks

I have generated a number of independent variables to predict the occurrence of  breaking (e.g distance to crops, water, voltage of the fence etc).  However I have found  (using a Runs test and in ArcGIS) that the occurrence of breaks/no breaks (and break count per 100m)  is spatially autocorrelated.  Therefore the predictive ability of my logistic regression model may be compromised: elephants may simply break near to where they have previously broken.  I feel I need to capture this in my model.

Similar analyses assessing the occurrence of crop raids/ no crop raids in a landscape (represented as a raster) have addressed this problem by generating an autocovariate term which is the euclidean distance weighted mean of all 8 surrounding cells in the raster. The autocovariate term is then treated as an independent variable in the regression model

However I am not dealing with a raster but with lengths of a vector. Can I calculate the autocovariate to introduce into my logistic regression model by calculating the euclidean distance weighted mean of the break counts of the two neighbouring sections?

If so, how do I do this in ArcGIS?

Thanks in advance,
George Aike

Outcomes