2

I have two datasets that I would like to analyze together. The first dataset has bird nest locations in neighborhoods (nest). The second dataset has bird feeder locations in neighborhoods (food). At each X,Y location of food I also have information about the amount of food (extent of the resource). Each of these are point shapefiles in ArcGIS.

I would like to create a smoothed surface (food raster) of food so I can extract a food index value for the location of each nest in my dataset. The food index should depend on the nearness of a birdfeeder to that nest, and the extent of food at that birdfeeder.

I used Getis Ord Gi* to generate zscores for each of my food locations (weighting each location based on extent of resource), and then used these values in IDW analysis. However, the Getis Ord Gi* zscores are both positive and negative because they compare the food value at each location to the mean food value at all of the feeders within the neighborhood. The result is that some of my feeders turn out to be "coldspots" in the resulting IDW raster.

What I'd like is a food raster where all of the food locations are "hot" (to varying degrees dependent on extent) and unmeasured locations are by definition "colder" than these (because I know there are not feeders there).

Does anyone know of a way to do this (e.g. constrain the Zscores so they are all positive or otherwise specify that the measured points are all "hot")?

Or do I need to use an alternate analysis altogether?

I have tested KDE on these data as well but would prefer IDW if possible because multiple fields of KDE are chosen at the "whim" of the researcher. I think I may have a difficult time defending choices of bandwidth, etc. with my datasets.

2
  • Sorry, but you will run into the same sample bias issues with point pattern statistics as with kernel density, perhaps even more so. You may want to consider a Kriging approach, or even a spatial glm. Commented Nov 12, 2014 at 21:50
  • I do not seem to run into the same issues using kernel density estimates- in that method, all of the feeders appear "hot" with cold spots where there are no feeders. I was simply trying to avoid using that method because of the need for user-defined parameters that I may not have the science to inform (e.g. there isn't a lot of information out there about the sphere of influence of a bird feeder on the species I am examining).
    – Jenn
    Commented Nov 14, 2014 at 14:12

2 Answers 2

0

As commented by @JeffreyEvans:

you will run into the same sample bias issues with point pattern statistics as with kernel density, perhaps even more so. You may want to consider a Kriging approach, or even a spatial glm.

0

you will run into the same sample bias issues with point pattern statistics as with kernel density, perhaps even more so. Keep in mind that an assumption of PPA is that you are representing the population and not a sample.

The parameterization of the kernel function can have profound effects on the resulting density estimate. Results can be influenced by the type of kernel (eg., Gaussian, Uniform) as well as the bandwidth. It would be quite difficult to detangle the signal being provided by the KDE from the underlying spatial process and would require considerable additional modeling, such as a Monte Carlo across multiple bandwidths and kernel specifications, to support any inference.

These are also being modeled as independent spatial processes and are not directly comparable. To make them comparable you need some sort of formalized spatial regression model or a cross-correlation type analysis. This is why "cross" functions (eg., Cross-K, bivariate LISA) were created in the first place. However, due to violations in PPA, a spatial regression approach seems in order.

Please, also note that you are attempting to represent a quantification of a spatial process and not a spatial process directly. This means that you have two components of model influence in relation to scale inference, bias and error. Simply put, you have a model driving a model where underlying error in the data and decisions on parameters could profoundly change outcomes.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.