Multiscale Spatial Autocorrelation
Multiscale Spatial Autocorrelation Analysis · Also known as: multi-scale spatial autocorrelation, scale-decomposed spatial autocorrelation, multiscale Moran analysis, MSA
Multiscale spatial autocorrelation extends classical spatial autocorrelation analysis by computing and comparing autocorrelation statistics (such as Moran's I) across a range of spatial scales simultaneously. This reveals at which geographic distances or resolutions spatial clustering or dispersion is strongest, providing a richer picture than a single global measure.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use multiscale spatial autocorrelation when you suspect a phenomenon operates at multiple spatial scales simultaneously (e.g., urban heat islands nested within regional climate patterns) or when you need to identify the characteristic spatial grain of a process before fitting a spatial regression model. It is particularly valuable in ecology, epidemiology, and urban studies where scale-dependency is theoretically motivated. Do not use it as a substitute for a properly specified spatial regression model, and avoid it when the dataset is very small (fewer than ~50 observations) since scale-specific statistics become unstable. It is also unsuitable when spatial units are irregular and extremely heterogeneous in size, as the scale decomposition becomes ambiguous.
Strengths & limitations
- Reveals the dominant spatial scale(s) of a process rather than collapsing autocorrelation to a single global measure.
- Supports theory-driven scale selection for subsequent spatial regression or interpolation models.
- Works with both regular and irregular spatial arrangements via Moran eigenvector maps.
- Provides a visual correlogram that communicates scale structure clearly to non-specialist audiences.
- Compatible with multivariate extensions (e.g., cross-correlograms between two variables across scales).
- Results depend strongly on the choice of distance bands or eigenvector thresholds, requiring domain knowledge for defensible parameterization.
- Multiple simultaneous tests inflate Type I error rates; correction methods reduce power, especially at fine scales with few pairs.
- Computationally intensive for very large datasets with many candidate scales or high-resolution eigenvector decompositions.
- Does not directly estimate causal or predictive relationships; it is an exploratory and diagnostic tool.
Frequently asked
How is multiscale spatial autocorrelation different from a standard Moran's I correlogram?
A standard Moran's I correlogram computes autocorrelation for each distance lag using ad hoc distance bins. Multiscale approaches using eigenvector maps (PCNM/MEM) decompose spatial variation into a mathematically orthogonal set of scale components, avoiding arbitrary bin choices and enabling variance partitioning across scales.
What software can I use?
The R packages adespatial and vegan implement MEM/PCNM-based multiscale analysis. GeoDa and spdep in R compute spatial correlograms. Python users can use PySAL's esda module for correlogram-style analysis.
How many distance lags or eigenvectors should I include?
The maximum distance lag is conventionally set to at most half the maximum pairwise distance in the study area to ensure sufficient pairs at each lag. For eigenvector maps, retain eigenvectors associated with positive spatial autocorrelation (positive eigenvalues) unless there is reason to model negative autocorrelation.
Can I use this approach with non-continuous outcomes?
Yes. Correlograms and MEM-based decompositions can be applied to count data (using appropriate autocorrelation coefficients such as join-count statistics) and binary variables, though interpretation requires care and permutation-based inference is preferred.
Does finding significant multiscale autocorrelation imply I need a spatial regression model?
Not necessarily. Significant autocorrelation in the raw variable may disappear after controlling for covariates. Run the multiscale analysis on regression residuals: if autocorrelation persists across scales, a spatial model or scale-specific covariates are warranted.
Sources
- Borcard, D., & Legendre, P. (2002). All-scale spatial analysis of ecological data by means of principal coordinates of neighbour matrices. Ecological Modelling, 153(1-2), 51-68. DOI: 10.1016/S0304-3800(01)00501-4 ↗
- Csillag, F., & Kabos, S. (2002). Wavelets, boundaries, and the spatial analysis of landscape pattern. Ecoscience, 9(2), 177-190. DOI: 10.1080/11956860.2002.11682704 ↗
How to cite this page
ScholarGate. (2026, June 3). Multiscale Spatial Autocorrelation Analysis. ScholarGate. https://scholargate.app/en/spatial-analysis/multiscale-spatial-autocorrelation
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Geographically Weighted RegressionSpatial analysis↔ compare
- Local Indicators of Spatial AssociationSpatial analysis↔ compare
- Local Spatial AutocorrelationSpatial analysis↔ compare
- Moran's ISpatial analysis↔ compare
- Multiscale Geographically Weighted RegressionSpatial analysis↔ compare
- Spatial AutocorrelationSpatial analysis↔ compare