I am currently about to run a spatial regression (OLS, Lag, Error etc) for my dataset. My dataset contains 1000 randomly generated points throughout the utilisation distribution for my animals (so, there are more or less high and low values attached to whether an animal goes to this area frequently). Since the species that I study has a very specific core range (where most of its time is spent), there will be hotspots (used often) and cold spots (used infrequently). I hope to determine whether their utilisation distribution is primarily in response to predation risk, the location of "competitive" individuals (i.e. other animals) as well as food availability.

In so far, I ran an OLS, Lag Model (there is spatial auto-correlation in my response) and an Error model in the package spdep in R. The Error model usually has the lowest AIC. In addition, I made spatial weights based off of euclidean distance in Geoda.

One thing that has really been bugging me since I have yet to find a guide on this, and I am rather new with spatial analytics involves well...choosing the right weights (i.e. K neighbour selection, queen, rook, distance). It should be noted that since I am dealing with 1000 random points for this analysis...and, they are throughout the landscape (for the animals home range) there are obvious "hot spots" of clusters for high and low values. Attached is a photo from geoda that show where these values are. enter image description here

Anyways, I apologise for the daft question. However, given the fact that I have the following:

  1. animal home range data (0.1-99) that were then made into 1000 points (my response)
  2. these 1000 points show clear clusters (primarily at feeding and sleeping sites)
  3. predictors (attached to these points through arcgis) including food availability and predation.

I am truly concerned about the appropriate spatial weights to use given this. Digging into the literature hasn't really helped and to be honest, I am beginning to become more confused than informed!

closed as too broad by Vince, Andre Silva, Jeffrey Evans, Dan C, xunilk Jun 13 '18 at 20:47

Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. Avoid asking multiple distinct questions at once. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

  • 1
    You are in a very common trap here. Basically, you are jumping around to different software, with different statistical options, in a piecemeal attempt to find something that sticks to the wall. Stop, take a deep breath and tell us "what is your question". That is to say, simply state your hypothesis and elicit advice on what the best statistical approach would be in addressing said hypothesis and then start reading the relevant lit. When you try to do it the other way is when you end up where your are now. You should not need ArcGIS and GeoDa for what you have described thus far, so why? – Jeffrey Evans Jun 12 '18 at 22:12
  • Thank you for your honest response. My hypotheses are that my focal animals utilisation will be negatively impacted by predation risk and postively impacted by food availability. As stated, I currently have 1000 points with the appropriate data attached (one point has ud info, predator info, and food info). Well, I did read a few articles that touched up upon the methods (which I am using) yet they fail to address the appropriate spatial weights to utilise. ArcGIS was used for creating and extracting the point data and GeoDa was used to create spatial weights. Both were then put into R. – Redskies421 Jun 12 '18 at 22:18
  • 1
    Please Edit the question in response to requests for clarification. It's not fair to those who would answer to need to mine the comments for critical information. This question is far too broad for our "Focused question / Best answer" model. Please take the Tour to better understand our purpose. – Vince Jun 12 '18 at 23:20

I am going to tackle this because it is likely going to be closed as "too broad" and I doubt that you will get a satisfactory answer if migrated to something like Cross Validated. Although, there is an important underlying spatial component here that is often neglected. That issue is the introduction of autocorrelation that is then treated as an observed spatial process.

The primary reason that you are not seeing the application of spatial methods in this specific type of habitat utilization modeling is that it is not valid to do so. Because you are pulling randomization's from a set of volume contours any observed spatial structure is purely a result of the underlyig density function and as such is entirely conditional on the volume contours. This is not a real observed spatial process.

This is a similar issue that the ESRI "collect events" tool introduces. Because you need a numeric response to model global or local autocorrelation and often all that is at hand are discrete events, ESRI built a tool that overlays a fishnet lattice, counts events within each cell and the then represents the centroid of each grid cell as an independent observation. Autocorrelation is then derived for this synthesized data. Following this method you are, in effect, imposing autocorrelation following the systematic aggregation scheme and not any observed spatial process underlying the original event data. If you stop and think about all of the possible aggregation schemes that could be applied and how a simple change of something like grain could effect the resulting statistic, you can start to see how arbitrary the results can be.

It is quite unclear as to how you are specifying your model and how you are deriving autocorrelation. Commonly, this type of statistical model is defined as a binomial process, representing use verses available, and solved via a GLM-logit. Given this type of data, the only real valid autocorrelation statistic would be a joins-count. A global Moran's-I or local LISA statistic, based on data representing [0,1] would be invalid as well as the spatial weights. If you are using the marked kernel density as the dependent variable in a Poisson process model, to put it simply, you cannot. The only real spatial approach I can recommend would be an autologistic model solved via an MCMC or perhaps a mixed effects model.

If you have any further questions please feel free to email me (my contact info is avalible in my profile).

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