Consistency With the Results of Existing GDM R Packages

This vignette discusses the consistency of results between the factor detector in the gdverse package and existing R packages of GDModels (i.e., geodetector and GD packages). Use the NTDs data from gdverse package as the demo;

NTDs = gdverse::NTDs[,1:4]

res1 = gdverse::gd(incidence ~ ., data = NTDs)
res1
##                 Factor Detector            
## 
## | variable  | Q-statistic |   P-value   |
## |:---------:|:-----------:|:-----------:|
## | watershed |  0.6377737  | 0.000128803 |
## | elevation |  0.6067087  | 0.043382244 |
## | soiltype  |  0.3857168  | 0.372145486 |

res2 = GD::gd(incidence ~ ., data = NTDs)
res2
##    variable        qv         sig
## 1 watershed 0.6377737 0.000128803
## 2 elevation 0.6067087 0.043382244
## 3  soiltype 0.3857168 0.372145486

# The geodetector package do not support the `tibble`
res3 = geodetector::factor_detector("incidence",
                                    c("soiltype","watershed","elevation"),
                                    as.data.frame(NTDs))
res3
## [[1]]
##          q-statistic   p-value
## soiltype   0.3857168 0.3632363
## 
## [[2]]
##           q-statistic      p-value
## watershed   0.6377737 0.0001169914
## 
## [[3]]
##           q-statistic    p-value
## elevation   0.6067087 0.04080407

The q-statistic calculations for all variables in the three packages are consistent, but there are slight differences in the results of the q-values. Among them, gdverse is consistent with the GD package, and there are differences with the geodetector package. This is caused by the inconsistent choice of the non-central F-distribution parameters of the p-value for the q-statistic; when there is only one sample in a certain stratification, it cannot calculate the variance and therefore does not contribute to the q-statistic calculation. The gdverse and GD packages use the same strategy, which is to directly remove these single-sample stratas, but the geodetector package calculates the total sample size and stratification number directly before data processing, so it causes a slight difference in the estimation of the p-value for the q-statistic. In actual problems, this situation occurs less frequently. We believe that using the actual number of samples and stratifications participating in the calculation is more prudent, so we chose the same processing strategy as the GD package.