'ValueError: kth out of bounds' in MGWR fitting about mgwr HOT 3 CLOSED

@DKchenliyuan

Thanks for reporting the issue. What I see here is that you are using a fixed bi-square kernel. We had some internal discussions previously and found that the combinations of fixed-bisquare and adaptive-gaussian sometimes have some fitting issues. So we would suggest to use either adaptive-bisquare or fixed-gaussian kernel. I think the reason is that a fixed-bisquare kernel may contain too few points (plus the weights drop to zero at bandwidth) at some locations when the spatial distribution of the points are very irregular. If you changed to a gaussian kernel, it would run ok. Please see the attached screenshot.

Regarding the bandwidth exceeds the largest pair-wise distance, this is possible and expected when the relationship approaches global. The upper bound of the bandwidth search range in mgwr and GWmodel may be different. Here we use the twice of the maximum pairwise distance as the upper limit. See this line of code here. That may explain why the bandwidth is mgwr is almost twice of the one reported in the Comber paper (I guess Lex was using GWmodel)?

from mgwr.

@Ziqi-Li
Thanks for your comment. It helps me a lot.
An additional question:
I find mgwr seems to have no functions to directly or conveniently fit the Mixed GWR model (such as gwr.mixed or gwr.multiscale(...bws0=c(Inf, 100, 100, Inf)...) in GWmodel). Is that right?

from mgwr.

@DKchenliyuan Yes, the bandwidth cannot be forced to be inf yet. Please see this issue for some discussions #4. It is not something hard to do, so we could probably add this in somewhere in the summer.

I'm closing the issue, if anything else pops up, feel free to reopen it.

from mgwr.