MGWR_Software_Comparison

Regression-based local modeling through multiscale geographically weighted regression (MGWR) and its predecessor, geographically weighted regression (GWR), has proved enormously popular in studies of spatial processes across a wide array of disciplines. However, three major software packages exist to calibrate MGWR models, and we show here that the calibration results are not always consistent across the three packages so care must be taken in selecting a particular package and it is imperative that any empirical analysis employing MGWR cite the software used to generate the results described therein.This paper examines and compares major functionalities, model specification options, and performance attributes in the following software packages: MGWR 2.2.1 (we discuss both the graphical user interface (GUI) version of this software and its Python equivalent and note where results digress); MGWR in ArcGIS Pro 3.2.2; and GWmodel 2.3.1 in R. We employ three simulated data sets and two empirical data sets to evaluate calibration consistency in terms of optimal bandwidth estimation, local parameter estimation, diagnostic interpretation, scalability, and computing time.In addition to the comparison of the three software packages, we undertake two further examinations which have generic applicability across the three packages. The first is an analysis of how sensitive MGWR calibration results are to the selection of a minimum bandwidth in the search for covariate-specific optimal bandwidths. The second is an examination of the stability of the model calibration results with respect to the order in which the covariates are added to the model. In both these comparisons we use the Python version of MGWR 2.2.1 which proves to be the most efficient of the three packages in terms of scalability and computing time.Our findings from the simulated data sets indicate reassuring consistency in the surfaces of the parameter estimates across the three algorithms but there are substantial differences in computing time required with GWmodel being virtually unusable with even medium-sized data sets. When using both empirical data sets, we uncover variations in the results across the three packages both in terms of optimal bandwidths and other model diagnostics. Given the prevalence of the use of MGWR in geographic information science, it is therefore imperative that users be aware of these differences and that they cite the software they use in any empirical application. The results of our two further investigations into the performance of the three routines suggest that the lower bound of the bandwidth search should be lowered in MGWR 2.2.1 and that the order in which the covariates are entered into the model can affect the results of all three algorithms so it recommended that henceforth all MGWR calibrations should enter the covariates in order of descending influence. We suggest how this can be done.

Authors

Publication Details

Journal
Figshare
Published
2026-11-21
DOI
https://doi.org/10.6084/m9.figshare.27229503
Primary Topic
Spatial and Panel Data Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

MGWR_Software_Comparison

Alexander Stewart Fotheringham, Chen-Lun Kao
Figshare
Spatial and Panel Data Analysis
article

MGWR_Software_Comparison

Alexander Stewart Fotheringham, Chen-Lun Kao
article en

Abstract

Regression-based local modeling through multiscale geographically weighted regression (MGWR) and its predecessor, geographically weighted regression (GWR), has proved enormously popular in studies of spatial processes across a wide array of disciplines. However, three major software packages exist to calibrate MGWR models, and we show here that the calibration results are not always consistent across the three packages so care must be taken in selecting a particular package and it is imperative that any empirical analysis employing MGWR cite the software used to generate the results described therein.This paper examines and compares major functionalities, model specification options, and performance attributes in the following software packages: MGWR 2.2.1 (we discuss both the graphical user interface (GUI) version of this software and its Python equivalent and note where results digress); MGWR in ArcGIS Pro 3.2.2; and GWmodel 2.3.1 in R. We employ three simulated data sets and two empirical data sets to evaluate calibration consistency in terms of optimal bandwidth estimation, local parameter estimation, diagnostic interpretation, scalability, and computing time.In addition to the comparison of the three software packages, we undertake two further examinations which have generic applicability across the three packages. The first is an analysis of how sensitive MGWR calibration results are to the selection of a minimum bandwidth in the search for covariate-specific optimal bandwidths. The second is an examination of the stability of the model calibration results with respect to the order in which the covariates are added to the model. In both these comparisons we use the Python version of MGWR 2.2.1 which proves to be the most efficient of the three packages in terms of scalability and computing time.Our findings from the simulated data sets indicate reassuring consistency in the surfaces of the parameter estimates across the three algorithms but there are substantial differences in computing time required with GWmodel being virtually unusable with even medium-sized data sets. When using both empirical data sets, we uncover variations in the results across the three packages both in terms of optimal bandwidths and other model diagnostics. Given the prevalence of the use of MGWR in geographic information science, it is therefore imperative that users be aware of these differences and that they cite the software they use in any empirical application. The results of our two further investigations into the performance of the three routines suggest that the lower bound of the bandwidth search should be lowered in MGWR 2.2.1 and that the order in which the covariates are entered into the model can affect the results of all three algorithms so it recommended that henceforth all MGWR calibrations should enter the covariates in order of descending influence. We suggest how this can be done.

Figshare
Spatial and Panel Data Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon