GWR settings: bandwidth parameter

[LaraSanders]

Hello,
My question is about the optimal number number of neighbours/distance to use with the GWR tool.  
I'm using GWR on a dataset with 311 area units.  When I use the adaptive kernel type and the AiCc bandwidth, the number of neighbours the program uses is 259 (out of 311 total).  There is not a lot of variation in the resulting coefficient maps.
The results are similar when I try other options (I've played with fixed kernel type and CV options).

So I tried setting the bandwidth parameter.  With 150 neighbours the residuals are not autocorrelated, and the coefficient maps show more variation. 
So now I'm wondering: should I use the AiCc/CV results, because my understanding is that ArcMap figures out the 'optimal' distance/number of neighbours? 
Or should I use the bandwidth that I set myself?  Are there some rules or guidelines that should be followed when setting my own bandwidth?
I'd really appreciate any help with this... 
Thanks

[SebastianSantibanez]

Hi,
Did you look at "Incremental Spatial Autocorrelation"?

This can give you a hint too
http://video.esri.com/watch/903/spatial-statistics-best-practices
I'm sure there is some reading material about that workshop but I didn't find it.

[JeffreyEvans]

Given the description of your problem it sounds like a local regression is not appropriate. The adaptive kernel distance is optimally derived using a cross validation. If it is choosing most of the available observations then it is telling you that there is very little spatial structure in your data. Ask yourself, why are you applying a GWR approach to begin with? Do you have measurable nonstationarity (2nd order autocorrelation) in your data? How autocorrelated is your dependent variable? It could be that a linear OLS regression or autoregressive model would work. You should specify a local regression to account for violation of specific assumptions. If your autocorrelation is a global (1st order) then local methods are not appropriate.

[LaraSanders]

Hi Jeffrey,
Thanks for your reply.  I'm applying GWR to a set of variables which I have identified as significant by using the Exploratory Regression tool to help me find a good model using OLS.  While the residuals from my OLS model do not exhibit significant clustering (determined by running the Global Moran's i tool), the Jarque Bera statistic is significant.
My understanding is that a significant JB can be due to non-stationarity in the explanatory variables, which was one reason why I wanted to run the GWR tool.  Indeed, when I run GWR, I can see that the coefficient for a couple of my explanatory variables goes from negative to positive across the study area. 
I'm unsure what you mean about having very little spatial structure in my data (with regard to GWR choosing most of the available observations in its calculations), I'd really appreciate it if you could clarify this point for me.

It seems to me that my model is a good candidate for GWR, but if you can see any big glitches (based on the limited amount of information I've provided!) I would be very grateful if you could point me in the right direction.