Multiscale Moran's I
Also known as: multi-scale Moran's I, spatial correlogram Moran, Moran correlogram, multiscale spatial autocorrelation
Multiscale Moran's I extends the classic global Moran's I statistic by computing spatial autocorrelation across a series of distance lags or spatial scales. The resulting spatial correlogram reveals at which geographic scales clusters or dispersions of a variable exist, offering richer insight than a single summary statistic.
Key highlights
- Reveals the scale or scales at which spatial clustering, independence, or regularity occurs — information invisible to a single-scale statistic.
- The spatial correlogram provides an interpretable visual summary of multiscale spatial structure.
- Applicable to irregular point patterns and areal data alike, using distance-based or contiguity weights.
- Permutation-based inference is distribution-free and robust to non-normal residuals.
- Guides subsequent modelling decisions: the first zero-crossing of the correlogram suggests an appropriate bandwidth for kernel smoothing or GWR.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Multiscale Moran's I when you expect spatial autocorrelation to vary across distance scales and want to identify the characteristic scale of clustering — for example, in ecology (species distribution), epidemiology (disease spread), urban analysis (socioeconomic patterns), and environmental monitoring. It is especially valuable when a single global Moran's I returns a non-significant result that may mask scale-dependent structure. Do not use it when the study area is too small to support multiple meaningful distance classes, when data locations are very sparse (fewer than ~50 points), or when you need a causal model rather than a descriptive pattern summary.
Strengths & limitations
- Reveals the scale or scales at which spatial clustering, independence, or regularity occurs — information invisible to a single-scale statistic.
- The spatial correlogram provides an interpretable visual summary of multiscale spatial structure.
- Applicable to irregular point patterns and areal data alike, using distance-based or contiguity weights.
- Permutation-based inference is distribution-free and robust to non-normal residuals.
- Guides subsequent modelling decisions: the first zero-crossing of the correlogram suggests an appropriate bandwidth for kernel smoothing or GWR.
- Requires a sufficient number of observation pairs per distance class; sparse data at large lags produces unstable estimates.
- Multiple testing inflation: testing significance at each lag without correction leads to spurious findings.
- Sensitive to the choice of distance class width; too narrow creates noisy estimates, too wide obscures scale structure.
- Does not attribute clustering to any causal mechanism; it is a descriptive diagnostic, not a regression model.
- Assumes the spatial process is second-order stationary (constant mean and variance across space), which may not hold in heterogeneous landscapes.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does Multiscale Moran's I differ from Local Moran's I (LISA)?
Multiscale Moran's I computes a single global autocorrelation value at each of several distance scales, producing a profile (correlogram) of clustering strength across scales. Local Moran's I decomposes autocorrelation spatially — producing a value for each location at a fixed scale — to identify which specific places are hotspots or outliers. The two approaches answer different questions: multiscale Moran's I asks 'at what scale does clustering occur?' while LISA asks 'where does clustering occur?'
How should I choose the distance classes?
A common rule is to ensure at least 30 pairs of observations per lag class. Start at the minimum inter-observation distance and increment by the average nearest-neighbour distance. The maximum lag should not exceed half the largest dimension of the study area, beyond which the number of pairs drops steeply. Software such as R's spdep or Python's PySAL can compute suggested lag intervals automatically.
What does it mean when Moran's I changes sign across lags?
A positive I_d at short lags indicates that nearby locations share similar values (positive clustering). When I_d crosses zero at some lag d*, values at that separation are uncorrelated — d* approximates the range of spatial dependence. Negative I_d at larger lags indicates spatial regularity or inhibition, where dissimilar values alternate systematically.
Can I apply Multiscale Moran's I to regression residuals?
Yes, and this is a recommended diagnostic step. Running the correlogram on OLS residuals tests the independence assumption across scales. If significant positive autocorrelation appears at short lags, a spatial regression model (spatial lag, spatial error, or GWR) is warranted.
How do I handle edge effects in the correlogram?
Edge effects arise because locations near the boundary have fewer neighbours, biasing large-lag estimates. Common remedies include restricting analysis to pairs where both locations are at least d away from the boundary, using toroidal wrapping for regular grids, or applying guard zones around the study area boundary.
Sources
- 1.Moran, P. A. P. (1950). Notes on continuous stochastic phenomena. Biometrika, 37(1-2), 17-23.
- 2.Legendre, P., & Fortin, M.-J. (1989). Spatial pattern and ecological analysis. Vegetatio, 80(2), 107-138.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Multiscale Moran's I. ScholarGate. https://scholargate.app/spatial-analysis/multiscale-morans-i