Multi-scale Geographically Weighted Regression

• MGWR is an advanced spatial regression technique that estimates variable-specific effects at distinct spatial scales to capture nonstationarity.
• It employs iterative back-fitting and adaptive bisquare kernels with AICc optimization to locally calibrate coefficients for robust spatial analysis.
• MGWR offers practical insights for epidemiology, urban economics, and environmental health, supporting targeted policy and decision-making.

Multiscale Geographically Weighted Regression (MGWR) is an advanced spatial modeling technique that extends traditional Geographically Weighted Regression (GWR) by allowing each covariate's effect to be estimated at its own distinct spatial scale. This enables more accurate modeling of spatial nonstationarity and local heterogeneity in the relationships between predictors and response variables, addressing critical limitations in single-scale spatial regression frameworks. MGWR is characterized by its robust mathematical formulation, rigorous bandwidth selection procedures, adaptive spatial weighting, and a proven capacity to disentangle spatial processes operating at multiple geographic resolutions.

1. Mathematical Formulation and Model Structure

MGWR generalizes the local regression paradigm of GWR by decomposing each coefficient surface into a function of both location and variable-specific spatial bandwidth. For an observation $i$ at coordinates $(u_i, v_i)$, the MGWR model is expressed as:

$$y_i = \beta_0\left(b_{0,i}\right) + \sum_{k=1}^p \beta_k\left(b_{k,i}\right) x_{ik} + \varepsilon_i$$

where:

• $y_i$ is the response (e.g., disease cases, economic metric) at spatial location $i$,
• $x_{ik}$ is the value of the $k^{th}$ covariate at $i$,
• $\beta_k(b_{k,i})$ is the locally estimated coefficient for covariate $k$, smoothed using its unique bandwidth $b_k$,
• $\varepsilon_i$ is an i.i.d. error term.

The model is calibrated via locally weighted least squares, with weights determined by an adaptive kernel (typically bisquare or tri-cube), where the effective bandwidth $b_k$ governs the spatial range used in estimating $\beta_k$. Large $b_k$ implies a near-global effect for covariate $k$, while small $b_k$ isolates localized relationships (Maiti et al., 2020).

2. Bandwidth Selection and Estimation Algorithms

The MGWR estimation process relies on simultaneous calibration of multiple bandwidths, one for each predictor and often the intercept. Bandwidth selection in published implementations utilizes an iterative back-fitting algorithm structured as follows:

• Initialize each $b_k$ (often from single-bandwidth GWR).
• For $k = 0, \ldots, p$, hold all other $\{\beta_j, b_j : j \neq k\}$ fixed and search for the $b_k$ minimizing the corrected Akaike Information Criterion (AICc).
• Update $\beta_k(b_{k,i})$ by weighted least squares given the new bandwidth.
• Repeat the process for all predictors and intercept until changes in AICc or bandwidths fall below a small tolerance (e.g., $10^{-6}$ or $10^{-5}$) (Maiti et al., 2020, Shabrina et al., 2019, Comber et al., 2020, Li et al., 2021).

AICc is the preferred metric for balancing model fit and complexity, especially in moderate sample sizes. Alternative criteria such as AIC, BIC, or cross-validation are supported but less commonly used.

3. Spatial Weighting Kernels and Adaptive Bandwidths

MGWR employs distance-decay kernels to weight observations when fitting local coefficients. The adaptive bisquare kernel is standard:

$$w_{ij}^{(k)} = \begin{cases} \left[1 - \left(\frac{d_{ij}}{b_k}\right)^2\right]^2 & d_{ij}<b_k \ 0 & d_{ij} \geq b_k \end{cases}$$

where $d_{ij}$ is the Euclidean (or great-circle) distance between locations $i$ and $j$, and $b_k$ represents a fixed number of nearest neighbors, not a metric radius, ensuring each local fit maintains comparable data density across heterogeneous spatial structures (Maiti et al., 2020, Murakami et al., 2017, Li et al., 2021). This scheme can be essential when modeling geographies with non-uniform population or sampling densities.

4. Model Diagnostics and Comparative Performance

MGWR models are evaluated using global and local goodness-of-fit metrics, spatial autocorrelation diagnostics, and collinearity checks:

• Global $R^2$ and adjusted $R^2$: MGWR often achieves strictly higher fit than single-scale GWR, OLS, and spatial error/lags models (e.g., $\mathrm{Adj}~R^2 \approx 0.969$ for COVID-19 cases in the US (Maiti et al., 2020), $R^2 = 0.77$ for food industry saturation in London (Shabrina et al., 2019), $R^2 = 0.631$ for household wealth in Bernalillo County (Okeke et al., 13 Oct 2025)).
• AICc: MGWR reports markedly lower AICc than GWR and global models, reflecting improved model parsimony despite increased parameterization.
• Local $R^2$: MGWR reveals substantial spatial variation, with "hot spots" of high model fit shifting dynamically in temporal studies (Maiti et al., 2020).
• Residual spatial autocorrelation: Moran's I is computed to assess whether spatial dependence remains after modeling; MGWR typically eliminates residual autocorrelation, as documented by residual Morans I dropping to near zero (Okeke et al., 13 Oct 2025).
• Collinearity: Local condition numbers, variance inflation factors, and variance decomposition proportions are computed for each moving window to ensure parameter stability (Maiti et al., 2020, Comber et al., 2020).

Confidence intervals for bandwidths are also mapped to assess estimation precision, and Cook's distance is used to flag influential observations in the local fits.

5. Empirical Applications and Multiscale Insights

MGWR's multibandwidth machinery has revealed nuanced spatial heterogeneity in diverse substantive settings:

• Epidemiology: In the US COVID-19 analysis, ethnicity, crime, and income showed distinct spatial scales, with migration effects and arson rates evidenced as highly localized drivers compared to income (near-global bandwidth) (Maiti et al., 2020).
• Urban economics: In London, MGWR showed Airbnb coefficients with broader-scale effects and hotel coefficients tightly localized, mapping industry dynamics otherwise obscured by single-scale GWR (Shabrina et al., 2019).
• Environmental health: In Bernalillo County, variables like income and length of residence operated at broad scales ($>200$ nearest neighbors), while disamenities (distance to hospitals, bus stops) and pollution had fine-scale localized effects ($<80$ neighbors), revealing trade-offs affecting household wealth at the neighborhood level (Okeke et al., 13 Oct 2025).
• Maritime safety: MGWR identified highly local (bandwidth $\approx4\%$ of sample) and sign-varying effects for "good visibility" on maritime accident consequences, which single-scale models entirely missed (Li et al., 2021).

The table below synthesizes empirical bandwidth findings:

Study	Covariate	Bandwidth (Neighbors)	Interpretation
US COVID-19 (Maiti et al., 2020)	Income, Migration	Large (near-global)	Gradual spatial change
	Arson, Domestic Mig.	Small	Highly localized
London F&B (Shabrina et al., 2019)	Airbnb	73	Smooth, city-wide
	Hotels	43	Localized zones
Bernalillo Wealth (Okeke et al., 13 Oct 2025)	Income, Parks	>200	County-wide
	Hospitals, Bus Stops	<80	Neighborhood scale
East China Sea (Li et al., 2021)	Visibility	50	Very local

6. Implementation Protocols and Best Practice Recommendations

MGWR is widely available in the R GWmodel and Python mgwr libraries, with published best practices emerging from empirical studies:

• Confirm the added value of multiscale over single-scale GWR by examining AICc reduction and local $R^2$ improvement.
• Employ adaptive kernels in data with strong spatial density variation.
• Calibrate bandwidths via AICc for moderate sample sizes; cross-validation is reserved for large $n$.
• Vigilantly check moving window collinearity, outliers, and residual autocorrelation; consider penalized or robust geographically weighted variants if diagnostics indicate instability or influential points (Comber et al., 2020).
• Always examine bandwidth confidence intervals; large bandwidths may suggest variables should be modeled globally.
• In time-series or dynamic settings, re-estimate MGWR for each time slice to capture spatiotemporal evolution (Maiti et al., 2020).

A plausible implication is that MGWR's variable-specific bandwidths provide not only higher explanatory power but also actionable geographic insight for policy targeting, intervention prioritization, and theoretical refinement in spatial analyses.

7. Theoretical and Practical Advances in Spatially Varying Coefficient Modeling

MGWR addresses two pivotal analytic challenges: instability and misspecification in spatially varying coefficient (SVC) modeling. Simulation experiments demonstrate that flexible-bandwidth GWR (FB-GWR/MGWR) and random effects eigenvector spatial filtering (RE-ESF) represent the most accurate and stable approaches, outperforming fixed-bandwidth SVC models across scenarios (Murakami et al., 2017). MGWR’s multiscale smoothing allows stabilization against local collinearity and prevents the over-/under-smoothing inherent in single-scale GWR, yielding interpretable and robust coefficient surfaces.

In summary, MGWR extends the frontier of spatial regression by decomposing spatial processes into their constituent scales, providing a transparent framework for modeling, diagnosing, and interpreting geographically nonstationary phenomena. Its methodological rigor, empirical success across domains, and formal implementation protocols position MGWR as the prevailing standard for contemporary spatially adaptive regression analysis.