Global Moran’s I in Spatial Analysis

• Global Moran’s I is a quadratic-form statistic that measures global spatial autocorrelation, indicating clustering when positive, randomness near zero, and dispersion when negative.
• It is computed using a spatial weight matrix and standardized data, with its interpretation supported by spectral bounds and regression-based eigenvalue methods.
• The index underpins spatial econometric modeling, offering applications in permutation testing, multi-scale analysis, and connections to information theory through entropy measures.

Global Moran’s I is a quadratic-form statistic that serves as the classical global index of spatial autocorrelation for a univariate attribute distributed over a spatial domain. Its rigorous mathematical construction, spectral properties, relationships to alternative spatial statistics, and role in spatial econometric modeling make it a foundational tool in modern spatial analysis.

1. Definition and Mathematical Structure

Let $x = (x_1, \ldots, x_N)^\top$ be a real-valued variable observed on $N$ spatial units, with sample mean $\bar x$ and (possibly sample or population) variance $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$. Fix a spatial-weight matrix $W = [w_{ij}]$ with zero diagonal ($w_{ii}=0$) and $\sum_{i,j} w_{ij} = 1$ (global normalization). The z-score standardized vector is $z = (x - \bar x \mathbf{1})/\sigma_x$, satisfying $z^\top \mathbf{1} = 0$, $z^\top z = N$.

The global Moran’s I is defined as

$N$0

which, under global normalization and $N$1, reduces to $N$2 (Chen, 27 Aug 2025).

Moran’s I quantifies the degree of spatial correlation: $N$3 signals clustering of similar values, $N$4 anti-clustering (checkerboard-like alternation), and $N$5 spatial randomness.

2. Spectral Properties, Bounds, and Interpretation

Moran’s I can be interpreted as a Rayleigh quotient, with attainable values determined by the spectral properties of $N$6. For symmetric $N$7, extremal values satisfy

$N$8

where $N$9 and $\bar x$0 are the smallest and largest eigenvalues of $\bar x$1 (projected onto the zero-mean subspace) (Maruyama, 2015, Chen, 2022). If $\bar x$2 is globally normalized, the “absolute” theoretical bounds are $\bar x$3, although in practice the achievable range is $\bar x$4, intersected with $\bar x$5 (Chen, 2022). Row-standardization of $\bar x$6 can enforce $\bar x$7, but general weighting schemes (including irregular degree graphs or distance-based weights) may yield I-values outside the unit interval (Duchin et al., 2021).

A monotone transformation of $\bar x$8 to $\bar x$9 rescales it into $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$0, where $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$1 is computed from the spectral bounds, yielding a direct interpretational comparability to Pearson’s correlation (Maruyama, 2015).

3. Structural Decomposition and Connection with Other Indices

Chen introduced an exact decomposition of global Moran’s I in terms of the Getis–Ord indices and the size correlation function (Chen, 27 Aug 2025):

$\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$2

where

• $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$3 is the normalized size vector,
• $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$4 is the size-correlation function,
• $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$5 is the global Getis–Ord index,
• $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$6, $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$7 is the sum of local Getis–Ord indices.

This structural decomposition reveals that global spatial autocorrelation depends on four distinct components: the global Getis–Ord index, the sum over local Getis–Ord indices, the number of spatial elements $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$8, and $\sigma_x^2 = \sum_{i=1}^N (x_i - \bar x)^2/N$9—the correlation structure of the normalized attribute. The local Getis–Ord indices are equivalent to gravity-model–based potential indices, establishing the association of Moran’s I with spatial interaction: weak spatial interaction implies weak global autocorrelation (Chen, 27 Aug 2025). The relationship is strictly nonlinear, and the interplay between $W = [w_{ij}]$0 and $W = [w_{ij}]$1 (and, implicitly, the gravity model) governs the magnitude and sign of $W = [w_{ij}]$2.

4. Algebraic, Regression, and Eigenvalue Perspectives

Moran’s I admits multiple algebraic forms:

• Quadratic form: $W = [w_{ij}]$3;
• Outer-product eigen formulation: $W = [w_{ij}]$4 (so $W = [w_{ij}]$5 is a nonzero eigenvalue of $W = [w_{ij}]$6);
• Inner-product formulation: $W = [w_{ij}]$7, providing regression-model–based interpretations (Chen, 2022, Chen, 2022).

Regression techniques can compute $W = [w_{ij}]$8 as the slope of the least-squares fit of $W = [w_{ij}]$9 on $w_{ii}=0$0, either with or without intercept, and thus directly connect autocorrelation measurement with spatial autoregressive models. The inner-product equation $w_{ii}=0$1 is formally the inverse of the simplest noise-free spatial autoregressive (SAR) model $w_{ii}=0$2, with the explicit relationship $w_{ii}=0$3 in that limiting case (Chen, 2022). This reconciliation aligns spatial autocorrelation estimation and the SAR coefficient within a unified algebraic framework and clarifies stability domains for SAR estimation via eigen-analysis (Chen, 2022, Chen, 2022).

5. Weight Matrix Choices and Their Implications

The choice and normalization of the weight matrix $w_{ii}=0$4 are fundamental to the behavior of global Moran’s I (Chen, 2016, Duchin et al., 2021). Standard options include:

• Binary adjacency (e.g., rook or queen contiguity in planar maps),
• Row-standardized versions ($w_{ii}=0$5) so that rows sum to one,
• Inverse distance decay ($w_{ii}=0$6 or $w_{ii}=0$7),
• Graph Laplacian ($w_{ii}=0$8), which inverts the interpretation of I,
• Doubly-stochastic matrices ($w_{ii}=0$9), e.g. Metropolis–Hastings random-walk kernels for robust boundedness $\sum_{i,j} w_{ij} = 1$0 and interpretable Markov variance-retention properties (Duchin et al., 2021).

The spectrum of $\sum_{i,j} w_{ij} = 1$1 and its degree of regularity (homogeneity of neighbor counts) directly control the attainable range of $\sum_{i,j} w_{ij} = 1$2 and the stability of its interpretation (Duchin et al., 2021, Maruyama, 2015). Irregular graphs can yield arbitrarily extreme $\sum_{i,j} w_{ij} = 1$3 values for suitably concentrated attribute vectors.

6. Statistical Inference and Information-Theoretic Interpretation

Under the null hypothesis of spatial randomness (random permutations of spatial labels), the expected value is $\sum_{i,j} w_{ij} = 1$4 (Arthur, 2024, Tillé et al., 2017). Analytical variance expressions exist, depending on $\sum_{i,j} w_{ij} = 1$5 and higher moments of $\sum_{i,j} w_{ij} = 1$6. Both parametric (normal approximation) and permutation-based inference for $\sum_{i,j} w_{ij} = 1$7 are standard: the latter matches the exact reference distribution under exchangeability (Kashlak et al., 2020, Wang et al., 2021). Permutation-free closed-form p-value bounds using concentration inequalities have been developed for computational efficiency at scale (Kashlak et al., 2020, Wang et al., 2021).

Recent developments connect Moran’s I to self-information (surprisal) and entropy. The probability of observing a given value $\sum_{i,j} w_{ij} = 1$8 under random spatial arrangement defines $\sum_{i,j} w_{ij} = 1$9—yielding an information-theoretic measure of spatial surprisingness. High autocorrelation (high $z = (x - \bar x \mathbf{1})/\sigma_x$0) corresponds to high self-information and low entropy, i.e., the data are compressible and “surprising” given the underlying spatial arrangement (Wang et al., 2024). This provides a bridge between spatial statistics and information theory, and motivates loss functions in spatial learning or ecological inference.

7. Extensions, Applications, and Analytical Generalizations

Analytical generalizations include scale-dependent spatial autocorrelation functions (SACF), constructed by parametrizing $z = (x - \bar x \mathbf{1})/\sigma_x$1 with a spatial lag $z = (x - \bar x \mathbf{1})/\sigma_x$2 (e.g., “relative step” function) (Chen, 2020). This yields cumulative and density SACFs, analogous to temporal autocorrelation functions, and admits partial SACF versions via spatial Yule–Walker equations. The resulting framework allows multi-scale and fractal analysis of spatial autocorrelation, with implications for identifying correlation dimensions in geographical systems (Chen, 2020).

Empirical studies demonstrate the versatility of global Moran’s I: from quantifying the tempo-spatial clustering of infectious disease (Deeb, 2020), measuring spatial balance in sample surveys (Tillé et al., 2017), to providing a null-preserving resampling algorithm that fixes observed autocorrelation for hypothesis testing (Arthur, 2024). The index is foundational for exploratory spatial data analysis, spatial panel model diagnostics, and for distinguishing clustering, randomness, and dispersion in heterogeneous spatial systems (Chen, 2016, Wang et al., 2021).

---

References

(Chen, 27 Aug 2025): "Structural Decomposition of Moran's Index by Getis-Ord's Indices" (Duchin et al., 2021): "Measuring Segregation via Analysis on Graphs" (Wang et al., 2024): "Probing the Information Theoretical Roots of Spatial Dependence Measures" (Arthur, 2024): "A General Method for Resampling Autocorrelated Spatial Data" (Chen, 2016): "New Approaches for Calculating Moran's Index of Spatial Autocorrelation" (Chen, 2020): "An Analytical Process of Spatial Autocorrelation Functions Based on Moran's Index" (Maruyama, 2015): "An alternative to Moran's I for spatial autocorrelation" (Deeb, 2020): "Spatial autocorrelation and the dynamics of the mean center of COVID-19 infections in Lebanon" (Chen, 2022): "Spatial autocorrelation equation based on Moran's index" (Chen, 2022): "Deriving two sets of bounds of Moran's index by conditional extremum method" (Kashlak et al., 2020): "Computation-free Nonparametric testing for Local and Global Spatial Autocorrelation with application to the Canadian Electorate" (Wang et al., 2021): "Local Statistics for Spatial Panel Models with Application to the US Electorate" (Chen, 2022): "Derivation of an Inverse Spatial Autoregressive Model for Estimating Moran's Index" (Tillé et al., 2017): "Measuring the spatial balance of a sample: A new measure based on the Moran's I index"