Spatial Heterogeneity Explanation(GOZH & LESH)

The GOZH(geographically optimal zones-based heterogeneity) model generates the optimal spatial zone based on the binary classification of the decision tree and then calculates the power of determinants. The LESH(locally explained stratified heterogeneity model) based on GOZH model and combined with additive shapely theory to reasonably allocate variable interaction’s power of determinants. In this vignette, we use ndvi data in gdverse package to demonstrate the spatial heterogeneity explanation based on GOZH and LESH model.

Load data and package

library(tidyverse)
library(gdverse)

data("ndvi")
head(ndvi)
## # A tibble: 6 × 7
##   NDVIchange Climatezone Mining Tempchange Precipitation   GDP Popdensity
##        <dbl> <chr>       <fct>       <dbl>         <dbl> <dbl>      <dbl>
## 1    0.116   Bwk         low         0.256          237. 12.6      1.45  
## 2    0.0178  Bwk         low         0.273          214.  2.69     0.801 
## 3    0.138   Bsk         low         0.302          449. 20.1     11.5   
## 4    0.00439 Bwk         low         0.383          213.  0        0.0462
## 5    0.00316 Bwk         low         0.357          205.  0        0.0748
## 6    0.00838 Bwk         low         0.338          201.  0        0.549

Univariate power of determinants detection

gozh.uvi = gozh(NDVIchange ~ ., data = ndvi)
gozh.uvi
## ***   Geographically Optimal Zones-based Heterogeneity Model       
##                 Factor Detector            
## 
## |   variable    | Q-statistic | P-value  |
## |:-------------:|:-----------:|:--------:|
## | Precipitation | 0.87255056  | 4.52e-10 |
## |  Climatezone  | 0.82129550  | 2.50e-10 |
## |  Tempchange   | 0.33324945  | 1.12e-10 |
## |  Popdensity   | 0.22321863  | 3.00e-10 |
## |    Mining     | 0.13982859  | 6.00e-11 |
## |      GDP      | 0.09170153  | 3.96e-10 |
plot(gozh.uvi)

Variable interaction detection

gozh.bi = gozh(NDVIchange ~ ., data = ndvi, type = 'interaction')
gozh.bi
## ***   Geographically Optimal Zones-based Heterogeneity Model       
##                 Interaction Detector         
## 
## |    Interactive variable     |    Interaction     |
## |:---------------------------:|:------------------:|
## |    Climatezone ∩ Mining     |    Weaken, uni-    |
## |  Climatezone ∩ Tempchange   |    Weaken, uni-    |
## | Climatezone ∩ Precipitation |    Enhance, bi-    |
## |      Climatezone ∩ GDP      |    Enhance, bi-    |
## |  Climatezone ∩ Popdensity   |    Enhance, bi-    |
## |     Mining ∩ Tempchange     |    Enhance, bi-    |
## |   Mining ∩ Precipitation    |    Weaken, uni-    |
## |        Mining ∩ GDP         |    Enhance, bi-    |
## |     Mining ∩ Popdensity     |    Enhance, bi-    |
## | Tempchange ∩ Precipitation  |    Enhance, bi-    |
## |      Tempchange ∩ GDP       | Enhance, nonlinear |
## |   Tempchange ∩ Popdensity   |    Enhance, bi-    |
## |     Precipitation ∩ GDP     |    Enhance, bi-    |
## | Precipitation ∩ Popdensity  |    Enhance, bi-    |
## |      GDP ∩ Popdensity       |    Weaken, uni-    |
plot(gozh.bi)

Variable interaction contribution

lesh.m = lesh(NDVIchange ~ ., data = ndvi, cores = 6)
lesh.m
## ***       Locally Explained Stratified Heterogeneity Model         
## 
## |    Interactive variable     |    Interaction     | Variable1 SPD | Variable2 SPD |
## |:---------------------------:|:------------------:|:-------------:|:-------------:|
## |    Climatezone ∩ Mining     |    Weaken, uni-    |  0.75353265   |  0.06776285   |
## |  Climatezone ∩ Tempchange   |    Weaken, uni-    |  0.64437728   |  0.17691822   |
## | Climatezone ∩ Precipitation |    Enhance, bi-    |  0.39405554   |  0.48986045   |
## |      Climatezone ∩ GDP      |    Enhance, bi-    |  0.79843017   |  0.05246998   |
## |  Climatezone ∩ Popdensity   |    Enhance, bi-    |  0.74240657   |  0.11069841   |
## |     Mining ∩ Tempchange     |    Enhance, bi-    |  0.10161351   |  0.31023743   |
## |   Mining ∩ Precipitation    |    Weaken, uni-    |  0.05886173   |  0.81368883   |
## |        Mining ∩ GDP         |    Enhance, bi-    |  0.12735177   |  0.09306564   |
## |     Mining ∩ Popdensity     |    Enhance, bi-    |  0.13123771   |  0.21760488   |
## | Tempchange ∩ Precipitation  |    Enhance, bi-    |  0.16187198   |  0.73291613   |
## |      Tempchange ∩ GDP       | Enhance, nonlinear |  0.35277116   |  0.08443737   |
## |   Tempchange ∩ Popdensity   |    Enhance, bi-    |  0.28786726   |  0.15633619   |
## |     Precipitation ∩ GDP     |    Enhance, bi-    |  0.84089496   |  0.04445297   |
## | Precipitation ∩ Popdensity  |    Enhance, bi-    |  0.79267181   |  0.09507756   |
## |      GDP ∩ Popdensity       |    Weaken, uni-    |  0.06828443   |  0.15493420   |
plot(lesh.m, pie = TRUE, scatter = TRUE)

Compared to GOZH Interaction Detector, LESH only has a decomposition of the interactive contribution of variables, and the rest remains consistent.

gdverse supports modifications to the default ploting results, such as adding subfigure annotations and adjusting the size of the text on the x-y axis:

plot(lesh.m, pie = TRUE, scatter = TRUE,
     pielegend_x = 0.98, pielegend_y = 0.15) +
  patchwork::plot_annotation(tag_levels = 'a',
                             tag_prefix = '(',
                             tag_suffix = ')',
                             tag_sep = '',
                             theme = theme(plot.tag = element_text(family = "serif"))) &
  ggplot2::theme(axis.text.y = element_text(family = 'serif',size = 15),
                 axis.text.x = element_text(family = 'serif',size = 15,
                                            angle = 30,vjust = 0.85,hjust = 0.75),
                 axis.title = element_text(family = 'serif',size = 15))

And you can only look at the contribution part of the variable interaction:

plot(lesh.m, pie = TRUE, scatter = FALSE)

By accessing the concrete result through lesh.m$interaction, which returns a tibble.

lesh.m$interaction
## # A tibble: 15 × 8
##    variable1    variable2 Interaction Variable1 Q-statisti…¹ Variable2 Q-statisti…²
##    <chr>        <chr>     <chr>                        <dbl>                  <dbl>
##  1 Climatezone  Mining    Weaken, un…                 0.821                  0.140 
##  2 Climatezone  Tempchan… Weaken, un…                 0.821                  0.333 
##  3 Climatezone  Precipit… Enhance, b…                 0.821                  0.873 
##  4 Climatezone  GDP       Enhance, b…                 0.821                  0.0917
##  5 Climatezone  Popdensi… Enhance, b…                 0.821                  0.223 
##  6 Mining       Tempchan… Enhance, b…                 0.140                  0.333 
##  7 Mining       Precipit… Weaken, un…                 0.140                  0.873 
##  8 Mining       GDP       Enhance, b…                 0.140                  0.0917
##  9 Mining       Popdensi… Enhance, b…                 0.140                  0.223 
## 10 Tempchange   Precipit… Enhance, b…                 0.333                  0.873 
## 11 Tempchange   GDP       Enhance, n…                 0.333                  0.0917
## 12 Tempchange   Popdensi… Enhance, b…                 0.333                  0.223 
## 13 Precipitati… GDP       Enhance, b…                 0.873                  0.0917
## 14 Precipitati… Popdensi… Enhance, b…                 0.873                  0.223 
## 15 GDP          Popdensi… Weaken, un…                 0.0917                 0.223 
## # ℹ abbreviated names: ¹`Variable1 Q-statistics`, ²`Variable2 Q-statistics`
## # ℹ 3 more variables: `Variable1 and Variable2 interact Q-statistics` <dbl>,
## #   `Variable1 SPD` <dbl>, `Variable2 SPD` <dbl>

Use lesh.m$spd_lesh to access the SHAP power of determinants:

lesh.m$spd_lesh
## # A tibble: 6 × 2
##   variable      spd_theta
##   <chr>             <dbl>
## 1 Precipitation    0.218 
## 2 Climatezone      0.176 
## 3 Tempchange       0.0482
## 4 Popdensity       0.0262
## 5 Mining           0.0158
## 6 GDP              0.0115