# Geographically weighted regression for GPS tracking data?

The study I am working on has GPS locations at 4 hourly intervals of 5 lions in 3 National Parks over roughly 2 years in Namibia. I have been applying Geographically weighted regression analysis to the data. The study area is broken up into 11988 1 x 1km grid cells and for each grid cell I have calculated the frequency of lion occurrences. This count data I then log transformed to improve the skewness and kurtosis values of the data. The total number of grid cell with lion occurrences is 1460. For each grid cell in the study area I then calculated the distance to park boundary, distance to nearest settlement, distance to major road and distance to river. These are then the factors I then use as explanatory variables in the geographically weighted regression and the dependent variable is then the log transformed lion count.

Most of the literature I found using geographically weighted regression was usually for planned surveys with transects and systematic data collection, so if there is a survey point which has zero values it is included in the analysis.

Can I apply the same principal to the grid cells with zero value in the study area or should I be running the geographically weighted regression only on the grid cells which actually have lion readings as I did not confirm there is no lion activity in those grid cells which have zero GPS readings from the collar information?

I have run the GWR using all grid cells and lion only grid cells with a bandwidth of 125 neighbours, and the lion only analysis has a significant better AICc value.

If it is easier to include a table of my results I can do that too.

• I would think about treating your model as a Poisson rather than a Binomial process. You would have to systematically process the data to derive a density function that represents expected activity rather than the GPS logging interval. Please be aware of the notable issues and massive limitations of GWR. It is not considered valid outside of an exploratory context and does not account for 1st order variation. Thus, if you do not have a measurable nonstationarity in your data the results may be meaningless. Please look at methods such as Point Process models or nonparametrics. – Jeffrey Evans Jul 3 at 15:31