Geographically Weighted Regression Analysis

• Geographically Weighted Regression is a spatial analysis technique that estimates local relationships by allowing model parameters to vary over geographic space.
• It captures spatial non-stationarity through adaptive bandwidth and kernel functions, enabling precise modeling of diverse phenomena.
• Advanced extensions, including multiscale, robust, Bayesian, and neural approaches, improve fit and uncertainty quantification in complex spatial datasets.

Geographically Weighted Regression Analysis (GWR) is a spatial statistical methodology for modeling and exploring spatial heterogeneity in relationships between a response variable and a set of covariates. By allowing model parameters to vary continuously over geographic space, GWR provides a direct means to capture the non-stationarity that is often present in environmental, socio-economic, and many other geographically indexed phenomena. This approach has spawned a rich family of extensions, including robust, multiscale, Bayesian, and neural-network-augmented frameworks, as well as numerous software implementations in R and Python.

1. Mathematical Foundations and Model Specification

At its core, Geographically Weighted Regression models the response at each location $(u_i, v_i)$ as a linear combination of covariates with location-dependent coefficients:

$$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$

Given $n$ observations with coordinates $\{(u_i, v_i)\}_{i=1}^{n}$, GWR estimates the local parameter vector $\beta(u_i, v_i)$ at each location via a locally weighted least squares problem:

$$\widehat{\beta}(u_i, v_i) = \left( X^T W_i X \right)^{-1} X^T W_i y$$

where:

• $X$ is the $n \times (p+1)$ design matrix,
• $y$ is the $n \times 1$ response vector,
• $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$0 is a spatial weight matrix for target location $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$1.

The weight $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$2 is obtained from a kernel function $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$3 that decays with spatial distance $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$4 between locations $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$5 and $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$6, and $$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$7 is the bandwidth controlling the spatial extent of the local fit (Gollini et al., 2013, Fotheringham et al., 2024).

Common Kernel Choices

Kernel	Formula	Support
Gaussian	$$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$8	global
Bi-square	$$y_i = \beta_0(u_i, v_i) + \sum_{k=1}^p \beta_k(u_i, v_i) x_{ik} + \varepsilon_i, \qquad \varepsilon_i \sim N(0, \sigma^2)$$9 if $n$0, 0 otherwise	compact
Tri-cube	$n$1 if $n$2, 0 otherwise	compact
Box-car	1 if $n$3, 0 otherwise	compact

Adaptive bandwidths (e.g., choosing $n$4 so that each $n$5 includes a fixed number of nearest neighbors) are often preferred for uneven sampling patterns (Gollini et al., 2013, Comber et al., 2020).

2. Bandwidth Selection, Model Diagnostics, and Implementation

Correctly specifying the bandwidth $n$6 is critical, as it determines the bias–variance trade-off: small $n$7 yields high spatial resolution but large variance, while large $n$8 approaches the global model. Two widely used selection criteria are:

• Leave-one-out Cross-Validation (CV): $n$9.
• Corrected Akaike Information Criterion (AICc):

$\{(u_i, v_i)\}_{i=1}^{n}$0

where $\{(u_i, v_i)\}_{i=1}^{n}$1 is the local residual variance and $\{(u_i, v_i)\}_{i=1}^{n}$2 is the “hat matrix.”

Adaptive and fixed bandwidths can both be optimized by grid search or information criteria. Local multicollinearity is identified using condition numbers or local VIF, and can be mitigated by locally compensated ridge regression (Gollini et al., 2013, Comber et al., 2020). Outlier resistance is possible via robust objective functions or re-weighting strategies (Sugasawa et al., 2021).

Practical implementation is facilitated by open-source packages, e.g., GWmodel (R) (Gollini et al., 2013) and mgwr (Python) (Fotheringham et al., 2024, Li et al., 2021).

3. Model Extensions: Multiscale, Robust, Bayesian, and Hybrid GWR

Multiscale GWR (MGWR)

Classical GWR assumes all covariates operate at the same spatial scale. The MGWR extension assigns each coefficient $\{(u_i, v_i)\}_{i=1}^{n}$3 its own bandwidth $\{(u_i, v_i)\}_{i=1}^{n}$4, capturing the reality that different processes diffuse over different spatial extents:

$\{(u_i, v_i)\}_{i=1}^{n}$5

Estimation proceeds via iterative backfitting and bandwidth selection for each coefficient (Fotheringham et al., 2024, Li et al., 2021, Comber et al., 2020).

Robust GWR

To address sensitivity to outliers, robust GWR substitutes the squared-error criterion with alternatives such as M-estimation, iteratively re-weighted least squares, or more formal $\{(u_i, v_i)\}_{i=1}^{n}$6-divergence-based objectives. Adaptively robust GWR automatically tunes both robustness and spatial smoothness parameters, incorporates a robust cross-validation criterion, and supplies robust standard error estimates and local outlier detection via influence measures (Sugasawa et al., 2021).

Bayesian GWR

Fully Bayesian formulations estimate spatially varying coefficients and the bandwidth jointly, enable spatial variable selection (e.g., via spike-and-slab priors), and produce posterior uncertainty intervals for local estimates. Fused-lasso priors can induce spatial smoothness among coefficients—particularly effective in sparse or irregular spatial designs—outperforming both classical GWR and Gaussian-penalized Bayesian GWR in mean squared error and uncertainty quantification (Ma et al., 2020, Sakai et al., 2024, Liu et al., 2021).

Advanced Extensions: Neural and Boosted GWR

Hybrid GWR/neural network models (e.g., AGWNN, GNNWR) integrate local spatial weighting within deep learning architectures, relaxing the local linearity constraint and learning nonlinear, spatially heterogeneous relationships, while preserving or even improving interpretability and predictive accuracy (Cao et al., 1 Apr 2025, Wang et al., 2022). Ensemble boosting of GWR (GWRBoost) recursively adds locally weighted linear models, optimizing via gradient boosting and preserving local coefficient surfaces for interpretation (Wang et al., 2022).

4. Generalizations: Attribute Distance, Multivariate, Survival Data

Standard GWR only accounts for spatial proximity. Covariate-distance weighted regression (CWR) augments the kernel weights with similarity in selected covariates:

$\{(u_i, v_i)\}_{i=1}^{n}$7

where $\{(u_i, v_i)\}_{i=1}^{n}$8 is a (possibly high-dimensional) attribute distance. CWR has shown significant improvements in predictive accuracy for real-estate and other heterogeneous domains (Chu et al., 2023).

Generalized GWR frameworks have been formulated for survival analysis via geographically weighted Cox regression, estimating location-specific hazard ratios using local kernel weighting and specialized information criteria for bandwidth selection (Xue et al., 2019). Modular Bayesian GWR variants can also handle generalized linear models via partial power posteriors (Liu et al., 2021).

5. Application Workflow and Interpretation

A canonical GWR analysis follows these steps (Comber et al., 2020, Fotheringham et al., 2024, Kilgarriff et al., 2020):

1. Exploratory Data Analysis: Spatial visualization, global regression, residual diagnostics (e.g., Moran's I, Geary's C for spatial autocorrelation).
2. Kernel and Bandwidth Selection: Choose kernel family (bisquare, Gaussian), distance metric (Euclidean/great-circle), and bandwidth (CV or AICc minimization).
3. Model Fitting: Estimate local coefficients at each observation.
4. Model Diagnostics: Map coefficients, local $\{(u_i, v_i)\}_{i=1}^{n}$9, local t-statistics/p-values, and standardized residuals. Check for spatial variation in fit and for regions of inflated multicollinearity.
5. Interpretation: Visualize and interpret coefficient surfaces to reveal spatially non-stationary relationships. Apply clustering or segmentation on coefficient vectors for zonation or typology mapping (Sarjou, 2021).
6. Prediction and Uncertainty: For new locations, compute kernel-weighted predictions and, in Bayesian variants, credible intervals (Sakai et al., 2024).

Special care is required for local collinearity (via local VIF/condition index), bandwidth overfitting, and edge-effects. Where the density of observations is highly uneven, fused-lasso or adaptive Bayesian approaches are recommended for estimation stability (Sakai et al., 2024).

6. Empirical Performance and Theoretical Properties

Comparative studies consistently find GWR and its variants capable of detecting and mapping spatial non-stationarity that global and spatial-error models cannot (Kilgarriff et al., 2020, Li et al., 2021, Namadi et al., 2024). Multiscale and hybrid approaches improve fit, localize effects at appropriate scales, and reduce residual spatial autocorrelation to near-zero in well-calibrated settings. Theoretical results establish the local linear estimator (GWLE) as asymptotically more efficient in local MSE than multidimensional-kernel variable coefficient models, due to the explicit spatial weighting and manageable bandwidth complexity (Yuan, 2018).

Advanced estimation schemes, including robust, Bayesian, neural, or boosting-based techniques, further reduce estimation and prediction error, effectively handle outliers, and provide interpretable, spatially resolved parameter surfaces suitable for spatial policy, epidemiology, environmental modeling, and urban analytics (Wang et al., 2022, Cao et al., 1 Apr 2025, Sakai et al., 2024, Li et al., 2021, Chu et al., 2023).