I'm running a binomial GWR in both MGWR and GWR4, and sometimes models get a high number of neighbors, sometimes even close to N-1. Is that a problem?
Should I interpret the number of neighbors relative to number of observations as a metric of model quality?
Also, should I be worried when GWR requires a large number of neighbors to have inversible matrices?
Hi!
When the average bandwidth of your GWmodel is close to the global one, you need to go back and consider whether the model is suitable for your research. In this case, this model is actually very close to a linear regression. Considering the computational complexity of GWmodels, usually their AICc value will be higher than the result of simply running an OLS, so they are not the preferred model at this time.
However, bandwidth only reflects the scale of geographical process, so I don't think the size of bandwidth can be a measure of model quality. You can have a look at: Article On the notion of ‘bandwidth’ in geographically weighted regr...
.Hope it helps.
Sui Zhang, thanks for the answer and suggestion of article. For anyone with the same question, I recommend the article above, as it really gives a good understanding on the topic.
The bandwidth is the same in both global and local, for softwares such as MGWR or GWR4.
The other metrics, like percent deviance explained either improve at lot bit or have little to no difference, depending on the model. So due to that and the more rich results, GWLR is preferable.
Since you pointed out that AICc might be somewhat biased, do you recommend that I adopt other criterion? Like BIC or CV.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R