Local Indicators of Spatial Association (LISA)

• Local Indicators of Spatial Association (LISA) are spatial statistical tools that decompose global measures like Moran's I to quantify local autocorrelation.
• They classify observations into distinct clusters (high-high, low-low) or outliers (high-low, low-high) to facilitate local spatial analysis.
• LISA methods use spatial weights and permutation tests to assess significance, with applications spanning economics, epidemiology, ecology, and medical imaging.

Local Indicators of Spatial Association (LISA) are a class of spatial statistics designed to measure the degree, nature, and significance of spatial autocorrelation at local, rather than solely global, scales. LISA statistics decompose global indices of spatial association, such as Moran’s I, providing spatially explicit diagnostics that distinguish between local spatial clusters, outliers, and randomness. Their adoption is foundational in exploratory spatial data analysis (ESDA), spatial econometrics, and spatial epidemiology, among other domains.

1. Mathematical Formulation and Key Principles

Local Moran’s I, the canonical LISA, quantifies the contribution of each observation to the overall spatial autocorrelation diagnosed by global Moran’s I. For a univariate variable $X = [x_1, ..., x_n]$ with mean $\bar x$, and a spatial weight matrix $W = [w_{ij}]$, the local statistic is most commonly given as:

$I_i = \frac{z_i}{m_2}\,\sum_{j=1}^n w_{ij}\,z_j$

where:

• $z_i = (x_i - \bar x)/s$ (z-score normalization, with $s^2 = (1/n)\sum_{k=1}^n (x_k - \bar x)^2$)
• $m_2 = \sum_{k=1}^n z_k^2 / n$: second moment
• $w_{ij}$: spatial weights that encode proximity or adjacency (Mason et al., 2024).

With row-normalized weights ($\sum_j w_{ij} = 1$), an equivalent, widely-used formulation is:

$I_i = \frac{z_i\,\sum_j w_{ij}z_j}{n-1}$

This normalization reflects the effective degrees of freedom at each site.

2. Relationship to Global Moran’s I

Global Moran’s I measures overall spatial autocorrelation by aggregating cross-products over all pairs:

$$I = \frac{n}{S_0}\;\frac{\sum_{i,j} w_{ij}(x_i - \bar x)(x_j - \bar x)}{\sum_i (x_i - \bar x)^2}, \quad S_0 = \sum_{i,j} w_{ij}$$

LISA statistics "disaggregate" this quantity spatially. The global index can be recovered as a weighted average (or sum) of the local indicators:

$I \propto \sum_{i=1}^n I_i$

when weights and normalization are compatible (Mason et al., 2024, Chen, 2016). Each $I_i$ quantifies the extent to which the value at location $i$ and the values at neighboring sites jointly contribute to the overall spatial autocorrelation.

3. Interpretation, Classification, and Statistical Testing

The sign and magnitude of local Moran’s $I_i$ are diagnostic:

• $I_i > 0$: observation $i$ and its neighbors exhibit similar z-signs (local clusters: high-high or low-low)
• $I_i < 0$: observation is an outlier relative to its neighbors (high-low or low-high, i.e., spatial outlier)

Local LISA diagnostics thus classify each observation into one of four categories:

• High-High (HH): $z_i > 0$, $\sum_j w_{ij}z_j > 0$: local hot spot
• Low-Low (LL): $z_i < 0$, $\sum_j w_{ij}z_j < 0$: local cold spot
• High-Low (HL): $z_i > 0$, $\sum_j w_{ij}z_j < 0$: high-value spatial outlier
• Low-High (LH): $z_i < 0$, $\sum_j w_{ij}z_j > 0$: low-value spatial outlier (Mason et al., 2024)

Empirical significance of each $I_i$ is established via conditional random permutation tests. For each location $i$, the neighborhood is randomly permuted many times (typically 999 or more), and the pseudo-p-value is the fraction of permuted $I_i$ values exceeding the observed magnitude. Multiple testing adjustments are critical for interpreting statistical significance due to the large number of local tests (Mason et al., 2024).

4. The Construction and Role of Spatial Weights

The spatial weights matrix $W=[w_{ij}]$ is foundational to LISA statistics:

• It is typically symmetric ($w_{ij}=w_{ji}$), with zero diagonal ($w_{ii}=0$), and constrained so that each row sums to unity (row-standardization) or the total weights sum to $S_0$.
• Common choices include binary adjacency (rook/queen contiguity), k-nearest neighbors, or inverse distance (Mason et al., 2024).

The sensitivity of LISA statistics to the specification of $W$ is substantial; results may vary markedly based on neighborhood definitions, with direct implications for the spatial clusters and outliers detected.

5. Visualization and Diagnostic Tools

Advanced visual analytic techniques have been developed for LISA, reflecting their complexity and interpretive richness:

• Moran Scatterplot: Plots standardized value $z_i$ vs. the spatial lag $\sum_j w_{ij}z_j$. Trend-line slope approximates the global index. The quadrant location provides immediate categorization (HH, LL, HL, LH) (Chen, 2016, Mason et al., 2024).
• Dual-Density Plot: Visualizes the distribution of z-scores and local neighbor values, spatial lags, permutation distributions of $I_i$, and empirical significance. Encodings (point size, color) communicate weight and cluster type (Mason et al., 2024).
• Network Scatterplot and Lag Radial Plot: Explore the relational network among local statistics and the geographic orientation of local clusters, augmenting traditional spatial maps.

Integrated dashboards linking these displays via interactive brushing and detail-on-demand facilitate multiscale diagnosis of spatial association structure within large geostatistical datasets (Mason et al., 2024).

6. Theoretical Properties, Extensions, and Limitations

• Quadratic Form and Eigenstructure: Local Moran’s indices arise as diagonal entries of $z\,z^\top W$. The eigenstructure of $W$ underpins both global and local statistics (Chen, 2016).
• Distribution Under Spatial Randomness: Under the null hypothesis, $E[I_i]=0$ (after standardization and suitable normalization), but in finite samples $E[I]\approx -1/(n-1)$ (Mason et al., 2024).
• Limitations and Considerations:
  • The choice of $W$ heavily impacts results—artifacts can arise with poorly matched neighborhood definitions (Mason et al., 2024).
  • Non-Gaussianity in X, low neighborhood variance, or sparse graphs complicate the interpretation of $I_i$.
  • Multiple testing issues are acute—naively interpreting clusters without controlling Type I error can be misleading.
  • LISA is strictly first-moment based; it detects only linear association in values across space.

Extensions of LISA include bivariate forms, multiple scales, and separate treatment for categorical variables, though the basic local cross-product structure remains foundational.

7. Applications and Empirical Paradigms

LISA is operationalized in diverse settings:

• Regional science: Identification of economic hot-spots and cold-spots.
• Epidemiology: Detection of disease clusters beyond environmental randomness.
• Ecology: Mapping of biodiversity hot-spots.
• Spatial machine learning: Auxiliary-task learning using multi-resolution LISA to enhance neural networks with inductive biases for spatial autocorrelation (Klemmer et al., 2020).
• Radiomics and medical imaging: Segmentation of pathological tissue via slice-wise LISA calculations (Ryan et al., 2018).

Empirical workflows typically consist of exploratory spatial data transformation, construction and standardization of $W$, computation of LISA statistics, permutation-based significance testing, and visualization using scatterplots, mapped clusters, and interactive diagnostics (Mason et al., 2024).

---

LISA statistics provide a rigorous, mathematically grounded mechanism to analyze spatial autocorrelation at local scales. They extend the interpretive range and resolution of global autocorrelation indices, are sensitive to spatial weight formulations, and require careful statistical interpretation, especially when assessing local significance. Their development and integration with modern geostatistical visualization and modeling paradigms have made them core instruments in spatial quantitative analysis (Mason et al., 2024, Chen, 2016).