M-SVC: Moran’s Spatially Varying Coefficient Model

• M-SVC is a spatial regression framework that uses Moran eigenvectors as basis functions to express multiscale, spatially varying relationships while mitigating spurious correlations.
• The method decomposes each coefficient into a global mean and spatially varying random effects, with smoothness controlled by the parameter α and efficient eigenvector truncation.
• By integrating spatial and nonspatial components, M-SVC provides scalable, likelihood-based inference with low bias and high accuracy even for massive datasets.

Moran’s Eigenvector Spatially Varying Coefficient (M-SVC) modeling is a spatial regression framework that employs Moran eigenvectors as basis functions to represent the spatial structure in coefficient surfaces. By leveraging eigenvectors that summarize patterns of positive spatial autocorrelation, M-SVC achieves flexible, multiscale, and computationally efficient modeling of spatially varying relationships, providing interpretable, likelihood-based inference even for massive datasets and circumventing the spurious correlation and scalability limitations of conventional SVC methodologies (Murakami et al., 2016, Murakami et al., 2018, Murakami et al., 2020, Murakami et al., 2024).

1. Model Formulation

The M-SVC framework considers observations $y = (y(s_1), \ldots, y(s_N))'$ at $N$ spatial locations $s_i$ with $K$ covariates $x_k(s_i)$, and models the response as: $$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$ Each coefficient surface $\beta_k(s)$ is decomposed via a Moran eigenvector basis: $$\beta_k = b_k\, 1_N + E\, y_k, \qquad y_k \sim N(0, \tau_k^2 \operatorname{diag}(\lambda_1^{-\alpha_k}, ..., \lambda_L^{-\alpha_k}))$$ where $b_k$ is the global mean, $E$ is the $N$0 matrix of Moran eigenvectors (corresponding to positive eigenvalues $N$1 of the doubly centered proximity matrix $N$2), $N$3 is a vector of random effects, $N$4 governs spatial variance, and $N$5 modulates smoothness through differential shrinkage of fine-scale eigenvectors.

An extended S&NVC (spatial & non-spatial variation) model augments each $N$6 with nonspatial (covariate-driven) variation: $N$7 This decomposition enables simultaneous modeling of spatially autocorrelated and nonspatial structure within each coefficient (Murakami et al., 2020).

2. Construction and Interpretation of Moran Eigenvectors

Moran eigenvectors arise from the eigendecomposition of the centralized spatial proximity matrix: $N$8 with $N$9, $s_i$0 the spatial weights, and positive eigenvalues $s_i$1 dictating the scale of spatial dependency. Retaining the first $s_i$2 eigenvectors with $s_i$3, each eigenvector $s_i$4 represents a distinct, orthogonal spatial map pattern of positive autocorrelation.

The spatial smoothness for each $s_i$5 is controlled by the parameter $s_i$6: higher $s_i$7 values lead to smoother, large-scale coefficient surfaces by shrinking lower-$s_i$8 eigenvector contributions, whereas small $s_i$9 permit high-frequency (fine-scale) variation (Murakami et al., 2016).

3. Estimation and Computational Acceleration

Estimation proceeds via linear mixed-effects modeling and type-II restricted maximum likelihood (REML), yielding joint estimation of fixed coefficients and Moran-basis random effects. The core mixed-model is: $K$0 with $K$1 encoding block-diagonal combinations of covariates and Moran eigenvectors, and $K$2 block-diagonal with spatial-variance scaling. Parameters $K$3 are estimated by optimizing the restricted log-likelihood: $K$4 where $K$5 (Murakami et al., 2020, Murakami et al., 2016).

Three computational accelerations enable scalability:

1. Rank-reduction: Nyström-type approximations provide low-rank Moran eigenvector bases with $K$6, reducing eigendecomposition cost from $K$7 to $K$8 (Murakami et al., 2018).
2. Pre-compression: Sufficient statistics (matrix inner products) summarizing all $K$9-dependence are computed once, allowing likelihood evaluation and maximization without iterating over $x_k(s_i)$0, making run-time independent of sample size (Murakami et al., 2018).
3. Sequential/blockwise maximization: Parameter updates exploit block-matrix identities, reducing per-parameter update cost to $x_k(s_i)$1 (Murakami et al., 2018, Murakami et al., 2024).

Sub-model aggregation strategies (e.g., via spatial clustering and generalized product-of-experts (gPoE) aggregation) further scale M-SVC to massive data, where multiple local ESF-based submodels are estimated in parallel and aggregated to yield global SVC surfaces (Murakami et al., 2024).

4. Addressing Spurious Correlation in SVC Estimation

Standard SVC models, which express all $x_k(s_i)$2 coefficient surfaces over a shared Moran eigenvector basis, tend to produce artificially correlated coefficient estimates when covariates are themselves spatially clustered. This collinearity inflates variance and produces “spurious correlation” between $x_k(s_i)$3 and $x_k(s_i)$4 even if the underlying $x_k(s_i)$5 and $x_k(s_i)$6 are conditionally independent.

By contrast, the S&NVC (M-SVC) model includes covariate-specific nonspatial variation, supplying each coefficient surface with its own set of (potentially nonspatial) basis functions. This separate modeling of covariate-driven (1D) and spatial components sharply reduces spurious cross-correlation, yielding nearly unbiased inference about true functional relationships—even when spatial association among covariates is high (Murakami et al., 2020).

5. Monte Carlo and Empirical Evidence

Simulation studies systematically compare M-SVC, pure SVC, geographically weighted regression (GWR), and standard ESF models:

• M-SVC maintains low root mean squared error (RMSE) for coefficient surfaces even under high spatial autocorrelation and small sample sizes (as low as $x_k(s_i)$7), while pure SVC and GWR methods suffer accuracy loss as covariate spatial dependence increases (Murakami et al., 2020, Murakami et al., 2016).
• M-SVC achieves bias under 5% and best RMSE in high autocorrelation regimes. Computational timings show M-SVC scales linearly or near-linearly with $x_k(s_i)$8 (e.g., 835s for $x_k(s_i)$9, $$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$0), outperforming both GWR and standard ESF at large $$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$1 (Murakami et al., 2018, Murakami et al., 2024).
• Type I error rates in inferring cross-$$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$2 correlation are controlled at nominal levels only by M-SVC; pure SVC procedures can experience false-positive rates above 50% (Murakami et al., 2020).

6. Large-Scale and Applied Implementations

The M-SVC framework is implemented in the R package spmoran (including parallelized submodel aggregation in addLearn_local), supporting both modest and massive spatial data. Typical analysis proceeds by extracting Moran eigenvectors (meigen), partitioning the space into clusters (spatClust), fitting local ESF models (addLearn_local), aggregating SVC estimates by gPoE, and visualizing coefficient surfaces (Murakami et al., 2024).

Applied analyses—such as Japanese residential land prices—demonstrate the interpretability and scale-adaptivity of M-SVC surfaces:

• For railway distance, flood risk, and land price, M-SVC uncovers both broad spatial trends and localized nonlinearities (e.g., differential flood effects, accessibility plateaus) that GWR or global ESF cannot resolve (Murakami et al., 2020, Murakami et al., 2018, Murakami et al., 2016, Murakami et al., 2024).
• In large-$$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$3 cases ($$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$4), sub-model aggregation avoids oversmoothing, localizes effects, and dramatically reduces run-times compared to GWR or classical ESF (Murakami et al., 2024).

7. Limitations and Extensions

• Selection of the eigenvector truncation $$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$5 entails a bias-variance tradeoff, with too-small $$y(s_i) = \sum_{k=1}^K x_k(s_i)\, \beta_k(s_i) + \varepsilon(s_i), \qquad \varepsilon \sim N(0, \sigma^2 I_N)$$6 over-smoothing fine-scale features. Multiresolution approaches or hybrid GWR-M-SVC models can attenuate this (Murakami et al., 2018).
• For interpretability, it is recommended to apply SVC structure selectively, with most covariates held constant or with spatial variation fitted only for key predictors (Murakami et al., 2018).
• The sub-model aggregation is parallelizable but introduces design choices regarding clustering and overlap; model selection can be guided via marginal likelihood or BIC (Murakami et al., 2024).
• Extensions include spatio-temporal SVCs, hierarchical spatial models, and integrating multilevel spatial random effects via the same computational acceleration strategies (Murakami et al., 2018).

M-SVC provides a theoretically grounded, scalable, and empirically validated solution for spatially varying coefficient regression, robust to spatial confounding and suitable for multi-resolution spatial inference (Murakami et al., 2016, Murakami et al., 2018, Murakami et al., 2020, Murakami et al., 2024).