Skip to content
Prev 27735 / 29559 Next

Spatial Autocorrelation Estimation Method

On Thu, 7 Nov 2019, Robert R wrote:

            
I think we are getting more clarity. You do not know the location of the 
lettings beyond their zipcode. You do know the boundaries of the zipcode 
areas, and can create a neighbour object from these boundaries. You then 
want to treat all the lettings in a zipcode area i as neighbours, and 
additionally lettings in zipcode areas neighbouring i as neighbours of 
lettings in i. This is the data structure that motivated the 
spdep::nb2blocknb() function:

https://r-spatial.github.io/spdep/reference/nb2blocknb.html

Try running the examples to get a feel for what is going on.

I feel that most of the variability will vanish in the very large numbers 
of neighbours, over-smoothing the outcomes. If you do not have locations 
for the lettings themselves, I don't think you can make much progress.

You could try a linear mixed model (or gam with a spatially structured 
random effect) with a temporal and a spatial random effect. See the HSAR 
package, articles by Dong et al., and maybe 
https://doi.org/10.1016/j.spasta.2017.01.002 for another survey. Neither 
this nor Dong et al. handle spatio-temporal settings. MRF spatial random 
effects at the zipcode level might be a way forward, together with an IID 
random effect at the same level (equivalent to sef-neighbours).

Hope this helps,

Roger