Geodetector

identifying and attributing spatial pattern in natural and social processes with software

                       

 

1. Introduction

2. Tutorial

3. Output

4. Download of the software, with example datasets

5. Citations

6. Bibliography

7. FAQs

8. Developer and Contact

 

1.        Introduction

Spatial Stratified Heterogeneity (SSH) is a phenomena that the within strata are more similar than the between strata. Examples of this include landuse types and climate zones in spatial data, seasons and years in time series, occupations, age groups, incomes strata. SSH occurs in all scales from universe to DNA, and has been studied since Aristotle time.

Geodetector, or Geographical Detector, is a statistical tool to measure SSH and to make attributions for or by SSH (Fig. 1):

(1) Measure and identity SSH among data;

(2) Test the coupling between two variables Y and X without assuming linearity of the association and with clear physical meanings; and

(3) Investigate the general interaction between two explanatory variables X1 and X2 and a response variable Y, without any specific form of interaction such as the assumed product in econometrics (Fig. 2).

Each of the above tasks can be accomplished using the Geodetector q-statistic:

 


Fig. 1. Principle of Geodetector q-statistic (Wang et al 2016)

(The bottom map, the color indicates the values of a population Y. The top map, the population Y is composed of L strata (h = 1, 2, …, L); the terms “stratification” and “partition” are equivalent, can be either classification or zonation. Between the two maps is the equation q(Y|{h}), in which the numerator is the summation of the within strata variance and the denominator is the pooled variance; N and s2 stand for the number of units and the variance of Y in a study area, respectively. [(N-L)q]/[(L-1)(1-q)] ~ F(L-1, N-L, g), where g is a non central parameter)

.

The strata of Y (red polygons in Fig.1) are a partition of Y, either by Y itself or by an explanatory variable X. X is a categorical variable or should be stratified if it is a numerical variable. The number of strata L might be 2-10 or more, according to prior knowledge or a classification algorithm. The “spatial” in “spatial stratified heterogeneity” can be either spatial in geoscience or in a broad mathermatical sense such as time and any attributes.

Interpretation of Geodetector q-statistic (Fig.1).

The value of q is strictly within [0, 1].

(1)  If Y is stratified by Y itself, then a q-statistic of 0 indicates that Y is absent of spatial stratified heterogeneity; a q-statistic of 1 indicates that Y is perfectly spatially stratified heterogenous; and a q-statistic of 100q% measures the degree of spatial stratified heterogeneity of Y.

(2)  If Y is stratified by an explanatory variable X, then a q-statistic of 0 indicates that there is no coupling between Y and X; a q-statistic of 1 indicates that Y is completely determined by X; and X explains 100q% of Y. Please note that the q-statistic measures the association between X and Y without assuming the linearity between X and Y.

Geodetector q-statistic can be used to understand spatial confounding, sample bias and overfitting.

(1)    Confounding can occur if a model is applied to a (spatial) stratified heterogeneneous population, leading to a misleading interpretation and statistical insignificance of the model outcome. This problem can be avoided by identifying SSH (by Geodetector q statistic) then modelling in the strata, separately.

(2)    A sample would be biased if a population is (spatial) stratified heterogeneous and the sample do not cover all strata. The problem can be solved by identifing (spatial) stratified heterogeneity (using Geodetector q statistic) then applying bias remedy models such as Heckman regression and Bshade method.

(3)    Local models aim to overcome heterogeneity but often suffer from overfitting and too many parameters to interpret. These problems can be avoided by modelling in strata or stratifying the outputs of a local model then interpreting the stratified parameters.

Functions of Geodetector:

(1)    The risk detector maps response variable Y in strata according to X;

(2)    The factor detector q-statistic measures the degree of spatial stratified heterogeneity of a variable Y if Y is stratified by itself; and the determinant power of an explanatory variable X on Y if Y is stratified by X;

(3)    The ecological detector identifies the difference in the impacts between two explanatory variables X1 and X2;

(4)    The interaction detector reveals whether the risk factors X1 and X2 (and more X, if applicable) have an interactive influence on a response variable Y (Fig.2).

 

Fig. 2. The General interaction between explanatory variables X1 and X2 impacting on a response variable Y: q(Y|X1X2).

 

back to the top ||

 

2.       Tutorial

The Geodetector software was developed using Excel and R, respectively. The tools are free of charge, freely downloadable, and easy to use, and were designed without any GIS plug-in components and with “one click” execution. Users can run the following demo, then simply replace the demo data in the software using your own data, click Run and get results ! We henceforth describe Excel Geodetector software. R users can download the R Geodetector software in the following section “Download of Geodetector Software and Example Datasets”.

As a demo, neural-tube birth defects (NTD) Y and suspected risk factors or their proxies Xs in villages are provided, including data for the health effect layers “NTD prevalence” and environmental factor layers, “elevation”, “soil type”, and “watershed”. Their field names are defined as Y and X1, X2, X3 respectively.

Step 1. Download the software and input your data in Excel

(1)  Download the Excel Geodetector software (In the following section “Software and Examples Data Download”), one click to download any one of the three Examples, unzip the downloaded file, you will find an Excel file (this is Geodetector software with an Example dataset!) and double click the Excel file, Fig. 3 and Fig. 5 appear. Fig. 3 is the format of the input data for the Geodetector: each row denotes a sample unit (e.g. a village); the 1st column record the response variable Y; the 2nd and following columns denote partitions of Y or factors X, the latter were partitioned according to the similarity within strata.

(2)  Input your data into the Excel Geodetector software in the format of Fig. 3. Then go to Step 2.

 

说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: image004

 

Fig. 3. Input data in Excel and the execution interface

(Note: Y should be numerical; X MUST be categorical, e.g. landuse types, seasons. If X is numerical, it should be transformed into a categorical variable, e.g. GDP per capita is stratified into 5 strata. At lease three sample units in each of the strata are required)

 

(3)  If your data is in GIS format, as shown in Fig. 4, you can use QGIS directly in Section 4, or you can transform the GIS data into Excel data as shown in Fig. 3.

 

 

Fig. 4. Data in GIS format

 

Step 2. Run Geodetector software

Only one operation interface was designed (Fig. 5). The function of the “Read Data” button is to load data; thus, when the button is clicked, all variables are listed in the “variables” list box. Then, disease and partition of Y or environmental factor variables are selected into their corresponding list boxes Y and X on the right of the interface. Finally, Geodetector is executed by clicking the “Run” button.

 

说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: UI

 

Fig. 5. User interface for Geodetector

back to the top ||

 

3.       Output

Geodetector outputs results from the risk detector, factor detector, ecological detector, and interaction detector in four Excel spreadsheets (Fig. 6).

 

说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: image014

 

Fig. 6. Interface for Geodetector results

 

In the “Risk detector” sheet (Fig. 7), result information for each environmental risk factor is presented in two tables. The first table gives the average disease incidence in each stratum of a risk factor, the name of which is written at the top left of the table. The second table gives the statistically significant difference in the average disease incidence between two strata; if there is a significant difference, the corresponding value is “Y”, else it is “N”.

 

risk-detector

 

Fig. 7. Results of risk detector

 

The Fig. 8 shows the output format of the q values for each environmental risk factor, as given in the “Factor detector” sheet. The table header gives the names of the environmental risk factors X(X1, X2, …, Xn), while the associated q values and their corresponding p values are presented in the row below.

 

factor-detector

 

Fig. 8. Results of factor detector

 

In the “Ecological detector” sheet (Fig. 9), results of the statistically significant differences between two environmental risk factors are presented. If Y(X1) (risk factor names in row) was significantly bigger than Y(X2) (risk factor names in column), the associated value is “Y”, while “N” expresses the opposite meaning.

ecological-detector

 

Fig. 9. Results of ecological detector

 

The format of the results for the interaction detector is shown in Fig. 10.Interaction relationships” below the table represent the interaction relationship for the two factors. The relationship is defined in a coordinate axis. It has 5 intervals, including “(-min(q(x), q(y)))”,“(min(q(x), q(y)), max(q(x), q(y)))”, “(max(q(x), q(y)), q(x) + q(y))”,“q(x) + q(y)”,“( q(x) + q(y),+∞)”, and the interaction relationship is determined by the location of q(xÇy) in the 5 intervals (see Table 1).

 

interaction-detector

 

Fig. 10. Results of interaction detector

 

Tab. 1. Interaction between Explanatory Variables (Xs)

 

Graphical representation

Description

Interaction

 

q(X1ÇX2) < Min(q(X1), q(X2))

 

Weaken, nonlinear

Min(q(X1),q(X 2))<q(X1Ç X2)<Max(q(X1)), q(X2))

 

Weaken, uni-

 

q(X1Ç X2) > Max(q(X1), q(X2))

 

Enhance, bi-

 

q(X1Ç X2) = q(X1)+ q(X2)

 

Independent

 

q(X1Ç X2) > q(X1)+ q(X2)

 

Enhance, nonlinear

Legend

说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: 说明: image029

back to the top ||

 

4.       Download of the software , with example datasets

The software was developed using Excel 2007, R and QGIS, respectively. It is completely free. You can click any one of the following links to download the Geodetector software. The first three are Geodetector software in Excel: (1) click one and unzip the file, an Excel file appears; (2) click the Excel file to start the Geodetector, you may exercise the demo data; then (3) input your own data to get your own results.

1: Geodetector Software in Excel, enclosed an Example of a Disease Dataset

2: Geodetector Software in Excel, enclosed an Example of a Toy Dataset

3: Geodetector Software in Excel, enclosed an Example of a NDVI Dataset

4: Geodetector Software in R

5: Geodetector Software in QGIS (please use google to access)

back to the top ||

 

5.       Citations. Geodetector can be cited as:

[1] Wang JF, Li XH, Christakos G, Liao YL, Zhang T, Gu X & Zheng XY. 2010. Geographical detectors-based health risk assessment and its application in the neural tube defects study of the Heshun region, China. International Journal of Geographical Information Science 24(1): 107-127.

[2] Wang JF, Zhang TL, Fu BJ. 2016. A measure of spatial stratified heterogeneity. Ecological Indicators 67: 250-256.

back to the top ||

6.       Bibliography

6.1 Featured articles using Geodetector

2016 Luo W, Jasiewicz J, Stepinski T, et al. 2016. Spatial association between dissection density and environmental factors over the entire conterminous United States. Geophysical Research Letters 43(2): 692-700.

2019 Yin Q, et al. 2019. Mapping the increased minimum mortality temperatures in the context of global climate change. Nature Communications 10: 4640.

2019 Zhang LQ, et al. 2019. Air pollution exposure associates with increased risk of neonatal jaundice. Nature Communications 10: 3741.

2020 Li JM, et al. 2020. Spatiotemporal trends and ecological determinants in maternal mortality ratios in 2,205 Chinese counties, 2010-2013: A Bayesian modelling analysis. PLoS Medicine 17(5): e1003114.

2021 Feng RD, et al. 2021. Urban ecological land and natural-anthropogenic environment interactively drive surface urban heat island: An urban agglomeration-level study in China. Environment International 157: 106857.

2021 Hu MG, et al. 2021. The risk of COVID-19 transmission in train passengers: an epidemiological and modelling study. Clinical Infectious Diseases 72(4): 604-610.

2021 Metaxas D, Qu H, Riedlinger G, Wu P, Huang Q, Yi J, De S. 2021. Deep learning-based nuclei segmentation and classification in histopathology images with application to imaging genomics. Computer Vision for Microscopy Image Analysis 185-201.

2021 Xu B, et al. 2021. Seasonal association between viral causes of hospitalized acute lower respiratory infections and meteorological factors in China: a retrospective study. Lancet Planetary Health 5: e154–63.

2022 Chen J, et al. 2022. Magnitudes and patterns of large-scale permafrost ground deformation revealed by Sentinel-1 InSAR on the central Qinghai-Tibet Plateau. Remote Sensing of Environment 268: 112778.

2022 Guo ZF, Boeing WJ, Xu YY, Borgomeo E, Liu D, Zhu YG. 2022. Data-driven discoveries on widespread contamination of freshwater reservoirs by dominant antibiotic resistance genes. Water Research. doi:https://doi.org/10.1016/j.watres.2022.119466.

2022 Yang JT, et al. 2022. Chain modeling for the biogeochemical nexus of cadmium in soil–rice–human health system. Environment International 167 (2022) 107424.

2023 Guo JH, Xu QS, Zeng Y, Liu ZH, Zhu XX. 2023. Nationwide urban tree canopy mapping and coverage assessment in Brazil from high-resolution remote sensing images using deep learning. ISPRS Journal of Photogrammetry and Remote Sensing 198 (2023) 115.

2023 Wang H, Wigneron JP, Ciais P, Yao YT, Fan L, Liu XZ, Li XJ, Green JK, Tian F, Tao SL, Li W, Frappart F, Albergel C, Wang MJ, Li SC. 2023. Seasonal variations in vegetation water content retrieved from microwave remote sensing over Amazon intact forests. Remote Sensing of Environment 285 (2023) 113409.

Lecture slides_20221122: Geodetector: Creating Randomness and Working with Heterogeneity in Big Data or Survey Data

 

6.2 Full list of the articles using Geodetector

Tab. 2. Articles using Geodetector [Numbered]

 

 

back to the top ||

 

2010

1.      Wang JF, Li XH, Christakos G, Liao YL, Zhang T, Gu X & Zheng XY. 2010. Geographical detectors-based health risk assessment and its application in the neural tube defects study of the Heshun region, China. International Journal of Geographical Information Science 24(1): 107-127.

2.      Liao YL,Wang JF, Wu JL, Driskell L, Wang WY, Zhang T, Gu X, Zheng XY. 2010. Spatial analysis of neural tube defects in a rural coal mining area.

International Journal of Environmental Health Research 20(6): 439-450.

back to the top ||

 

2011

3.      Hu Y, Wang JF, Li XH, Ren D, Zhu J. 2011. Geographical detector-based risk assessment of the under-five mortality in the 2008 Wenchuan earthquake,

China. PLoS ONE 6(6): e21427.

4.      Zou B, Wilson JG, Zhan FB, Zeng YN, Wu KJ. 2011. Spatial-temporal variations in regional ambient sulfur dioxide concentration and source-contribution analysis: A dispersion modeling approach. Atmospheric Environment 45: 4977-4985.

back to the top ||

 

2012

5.      Gajos M. 2012. Geoinformation technologies in biomedicine and health care: review of scientific journals. E. Piętka and J. Kawa (Eds.): ITIB 2012, LNCS

7339: 510–524.

6.      Li LF, Wang JF, Wu J. 2012. A spatial model to predict the incidence of neural tube defects. BMC Public Health 12: 951.

7.      Wang JF, Hu Y. 2012. Environmental health risk detection with GeogDetector. Environmental Modelling & Software 33: 114-115.

8.      刘彦随,     , 2012. 中国县域城镇化的空间特征与形成机理. 地理学报 67(8): 1011-1020.

Liu YS, Yang R.2012.Spatial characteristics and formation mechanism of the county urbanization in China. Acta Geographica Sinica 67(8): 1011-1020.

back to the top ||

 

2013

9.      Cao F, Ge Y, Wang JF. 2013. Optimal discretization for geographical detectors-based risk assessment. GIScience & Remote Sensing 50(1): 78-92.

10.     Li XW, Xie YF, Wang JF, Christakos G, Si JL, Zhao HN, Ding YQ, Li J. 2013. Influence of planting patterns on Fluoroquinolone residues in the soil of an intensive vegetable cultivation area in north China. Science of the Total Environment 458-460: 63-69.

11.     Lee WC. 2013. Assessing causal mechanistic interactions: a peril ratio index of synergy based on multiplicativity. PLoS ONE 8(6): e67424. doi:10.1371/journal.pone.0067424.

12.     Raghavan RK, Brenner KM, Harrington Jr JA, Higgins JJ, Harkin KR. 2013. Spatial scale effects in environmental risk-factor modelling for diseases. Geospatial Health 7(2): 169-182.

13.     Wang JF, Wang Y, Zhang J, Christakos G, Sun JL, Liu X, Lu L, Fu XQ, Shi YQ, Li XM. 2013. Spatiotemporal transmission and determinants of typhoid and paratyphoid fever in Hongta District, China. PLoS Neglected Tropical Diseases 7(3): e2112.

14.     Wang JF, Xu CD, Tong SL, Chen HY, Yang WZ. 2013. Spatial dynamic patterns of hand-foot-mouth disease in the People’s Republic of China. Geospatial

Health 7(2): 381-390.

back to the top ||

2014

15.     Bai HX, Ge Y, Wang JF, Li DY, Liao YL, Zheng XY. 2014. A method for extracting rules from spatial data based on rough fuzzy sets. Knowledge-Based Systems 57: 28-40.

16.     Hu Y, Gao J, Chi M, Luo C, Lynn H, Sun LQ, Tao B, Wang DC, Zhang ZJ, Jiang QW. 2014. Spatio-temporal patterns of schistosomiasis Japonica in lake and marshland areas in China