M-SGWR: Multiscale Similarity and Geographically Weighted Regression

The first law of geography is a cornerstone of spatial analysis, emphasizing that nearby and related locations tend to be more similar, however, defining what constitutes “near” and “related” remains challenging, as different phenomena exhibit distinct spatial patterns. Traditional local regression models, such as Geographically Weighted Regression (GWR) and Multiscale GWR (MGWR), quantify spatial relationships solely through geographic proximity. In an era of globalization and digital connectivity, however, geographic proximity alone may be insufficient to capture how locations are interconnected. To address this limitation, we propose a new multiscale local regression framework, termed M-SGWR, which characterizes spatial interaction across two dimensions: geographic proximity and attribute (variable) similarity. For each predictor, geographic and attribute-based weight matrices are constructed separately and then combined using an optimized parameter, alpha, which governs their relative contribution to local model fitting. Analogous to variable-specific bandwidths in MGWR, the optimal alpha varies by predictor, allowing the model to flexibly account for geographic, mixed, or non-spatial (remote similarity) effects. Results from two simulation experiments and one empirical application demonstrate that M-SGWR consistently outperforms GWR, SGWR, and MGWR across all goodness-of-fit metrics.

💡 Research Summary

The paper introduces a novel spatial regression framework called Multiscale Similarity‑Geographically Weighted Regression (M‑SGWR). Traditional local regression models such as Geographically Weighted Regression (GWR) and its multiscale extension (MGWR) rely exclusively on geographic distance to define spatial weights, applying a single bandwidth (or variable‑specific bandwidths in MGWR) across all predictors. This approach ignores the growing importance of non‑geographic connections—digital networks, social ties, transportation routes—that can create strong relationships between distant locations.

M‑SGWR addresses this gap by constructing two separate weight matrices for each predictor: a geographic weight matrix (W_geo) derived from an adaptive bi‑square kernel based on Euclidean or alternative distance measures, and an attribute‑similarity weight matrix (W_attr) that quantifies how similar two observations are in the space of the specific predictor. A mixing parameter α_k (0 ≤ α_k ≤ 1) is introduced for each predictor k, and the final mixed weight matrix is W_k = α_k · W_geo^k + (1 − α_k) · W_attr^k. When α_k≈1 the predictor’s influence is governed mainly by geographic proximity; when α_k≈0 the influence is driven by attribute similarity; intermediate values capture mixed effects.

Crucially, both α_k and the bandwidth h_k for each predictor are estimated simultaneously through an iterative back‑fitting algorithm that mirrors the MGWR inference scheme. The algorithm updates variable‑specific projection matrices, computes local weighted least‑squares estimates, and refines α_k and h_k to minimize an information criterion (e.g., AICc). Standard errors are derived by propagating the residual variance through the projection matrices, allowing the construction of pseudo‑t statistics for local significance testing.

The authors evaluate M‑SGWR in three settings. (1) A controlled simulation where all predictors share the same spatial scale; here M‑SGWR performs on par with MGWR and better than plain GWR, confirming that the added flexibility does not harm baseline performance. (2) A more complex simulation where each predictor has a distinct spatial scale and differing degrees of attribute similarity. In this scenario, M‑SGWR’s variable‑specific α_k values correctly identify which predictors are geographically driven versus similarity‑driven, yielding substantial improvements in AICc, adjusted R², and RMSE over GWR, SGWR (single α), and MGWR. (3) An empirical case study using Pennsylvania data on housing prices, income, water availability, and sunlight exposure. The results show that water‑related coefficients exhibit high α (geographically dominated), while sunlight‑related coefficients have low α (similarity‑dominated), reflecting real‑world processes. Overall model fit and predictive accuracy are superior to the competing methods.

Key contributions include: (i) a formal integration of geographic and attribute‑based proximity into a unified local regression framework; (ii) variable‑specific mixing parameters that enable nuanced interpretation of how each predictor operates across spatial and non‑spatial dimensions; (iii) an algorithmic implementation that can be incorporated into existing GWR toolkits with moderate computational overhead. The paper also discusses limitations: simultaneous optimization of α_k and h_k increases computational demand, especially for high‑dimensional or large‑sample datasets, suggesting the need for parallelization or approximation techniques. Moreover, the construction of W_attr depends on the choice of similarity metric and scaling, which may affect results and warrants careful preprocessing.

Future research directions proposed include extending M‑SGWR to incorporate network‑based distances (e.g., airline routes, social media connections), adding temporal dynamics for spatio‑temporal non‑stationarity, and embedding the framework within a Bayesian hierarchical model to better quantify uncertainty. Overall, M‑SGWR represents a significant methodological advance for spatial analysts seeking to capture the multifaceted ways locations are related in an increasingly connected world.