permutations in localG_perm #124

Activity

1. rsbivand commented on Feb 1, 2023

   

   rsbivand

   on Feb 1, 2023

   Member

   More actions

   Thanks for very careful reading of the code and well-motivated points. I've added 1) to code not yet committed/pushed to github. With regard to 2), an R-only implementation is an order of magnitude slower - because it iterates over nsim, and sampling without replacement is in any case slower. I'm loking at implementing in C to speed this up a little. Should I commit/push what I have already done? May I ask whether I could add you as a contributor to the package, in which case a name and optionally ORCiD would be needed.
2. rsbivand

   added a commit that references this issue on Feb 1, 2023

   address #124 point 1)

   d81de31
3. rsbivand commented on Feb 1, 2023

   

   rsbivand

   on Feb 1, 2023

   Member

   More actions

   Thanks for very careful reading of the code and well-motivated points. I've added 1). With regard to 2), an R-only implementation is an order of magnitude slower - because it iterates over nsim, and sampling without replacement is in any case slower. I've added an implementation in C to speed this up a little (2x not 12x case with replacement). May I ask whether I could add you as a contributor to the package, in which case a name and optionally ORCiD would be needed?
4. rsbivand

   added 2 commits that reference this issue on Feb 1, 2023

   address 2) in #124: adding no repeats to cond_perm

   b4e574a

   address 2) in #124: adding no repeats to cond_perm

   bf2bdc8
5. JosiahParry commented on Feb 3, 2023

   

   JosiahParry

   on Feb 3, 2023

   Contributor

   More actions

   Good looks @rosieluain! To confirm sampling is now done without replacement at $i$ (which makes sense a geography can't occur twice)?.

   Regarding 1) this actualy came up in a convo with a colleague yesterday regarding the Local Moran. I'd love to understand this better. My understanding in using the permutation sample variance you are scaling by the expected variance under spatial randomness which to me feels a bit better.

   Though in practice they're essentially identical

   data(guerry, package = "sfdep")
   
   library(spdep)
   listw <- nb2listw(poly2nb(guerry$geometry))
   x <- guerry$crime_pers
   
   G <- localG(x, listw, return_internals = TRUE)
   Gv <- attr(G, "internals")[,3]
   
   
   Gp <- localG_perm(x, listw)
   Gpv <- attr(Gp, "internals")[,3]
   
   plot(Gpv, Gv, main = paste0("Corrlation: ", round(cor(Gpv, Gv), 2)))

   

   ^{Created on 2023-02-03 by the reprex package (v2.0.1)}
6. rosieluain commented on Feb 3, 2023

   

   rosieluain

   on Feb 3, 2023

   Author

   More actions

   @JosiahParry, good thought exercise!

   I think that when using the permutation sample variance, you are right that you are scaling by the expected value and the variance of the expected value, for a spatially random dataset. This might be ok if the data were spatially random; however, we don't actually know that the data are random. So I think the expected value and variance need to be estimated from the (non-randomized) dataset itself. And then the purpose of the permutations is just to compare to the dataset to determine if the data are statistically different from spatially random.

   What led me to identify this issue in the first place, is that if using the variance of the expected value from permutation, there is an issue when calculating G* for a location that does not have any neighbours: The variance is 0 (if sampling without replacement), so the calculation for G* fails when trying to divide by 0. If using the variance estimated from the data, a value for G* is still possible, but the p value obtained through permutations is 1. This makes sense to me, because the G* lets you see how this location compares to rest of the data, but p = 1 indicates there is no spatial structure within distance d (or there is not enough information to determine if there is spatial structure).
7. rosieluain commented on Feb 3, 2023

   

   rosieluain

   on Feb 3, 2023

   Author

   More actions

   @rsbivand, Thank you, but no need to add me as a contributor. Thank you for being so attentive and helping me work through things!

   👍React with 👍1Reacted by Roger Bivand
8. rsbivand commented on Feb 27, 2023

   

   rsbivand

   on Feb 27, 2023

   Member

   More actions

   Submission to CRAN to come 28 February morning CET - please indicate before then if any obvious impediments.
9. rsbivand commented on Feb 28, 2023

   

   rsbivand

   on Feb 28, 2023

   Last edited by rsbivand

   Member

   More actions

   "adegraphics" adeverse/adegraphics#14
   "adespatial" adeverse/adespatial#24
   "spData" Nowosad/spData#63
   "Voyager" pachterlab/voyager#3

   Submission: 11:16 CET (230228)
   Thanks, on its way to CRAN: 14:02 (230228)
10. rsbivand

    closed this as completedon Feb 28, 2023