Pseudo p-value calculation and standardization

[GoogleCodeExporter]

Measures of spatial autocorrelation compute permutation-based pseudo p-values. 
While comparing Getis and Ord local G stats across PySAL and GeoDa, two issues 
have appeared. 
1. PySAL computes standardized pseudo p-values, in addition to original 
p-values. 
    There is a question of why we do this. 
     If there is an obvious reason for this standardization, 
     how about dropping this calculation across all edsa modules? 
2.  GeoDa computes pseudo p-values as follows. This info came from Mark.
     For the case of local Gi:
     1) compute Gi for all i.
     2) When # permutations is 499, compute 499 permuted Gi by drawing random samples of neighbors without including i
     3) Keep track of every time a permuted Gi is greater than the observed Gi and call this #larger
     4) If #permuations/2 > #larger, then p-value is (1 + #larger)/(1 + #permutations)
         Otherwise, p-value is (1 + #permutations - #larger)/(1 + #permutations)

     Step 4 above is different from how PySAL works. 
     How about making necessary changes across PySAL esda modules?

Original issue reported on code.google.com by mhwa...@gmail.com on 28 Jun 2011 at 7:17

[GoogleCodeExporter]

In the discussion of issue 1, I meant "If there is not an obvious reason for 
this standardization". 

Original comment by mhwa...@gmail.com on 29 Jun 2011 at 1:56

[GoogleCodeExporter]

on 1) the logic was to use the empirical distribution of the statistic under 
the null to estimate the moments and use them in the z-transform in place of 
the analytical moments. i don't think we should drop it.

on 2) we should be aligned with the treatment in GeoDa on this so making 
changes in this regard would be a good thing.

Original comment by sjsrey on 29 Jun 2011 at 6:04

[GoogleCodeExporter]

Issue 1 dropped and Issue 2 addressed

Original comment by mhwa...@gmail.com on 8 Jul 2011 at 5:35

• Changed state: Fixed