Put the same numbers in different places on a map and you can change the scientific question without changing a single value.

That is the problem spatial statistics is built to handle. Nearby observations may resemble one another. Boundaries can create discontinuities. A hotspot can be genuine or an artefact of how areas were drawn. A smooth map can hide enormous local uncertainty. And a method that assumes independent observations can become overconfident when geography creates dependence.

Spatial statistics treats location, distance, neighbourhood and spatial support as part of the data-generating process rather than as decoration added after the analysis.

Quick Read

• Spatial data can represent points, lines, areas, grids or continuous fields, and those forms are not interchangeable.
• Spatial autocorrelation means observations can be related because of where they are located.
• A spatial weights structure defines which places count as neighbours and how strongly they are connected.
• Moran’s I and Geary’s C summarize forms of global spatial autocorrelation.
• Local indicators such as Local Moran can identify clusters and spatial outliers, but local multiple testing requires care.
• Geostatistics models continuous spatial dependence through covariance or variograms and can use kriging to predict at unsampled locations.
• Scale, boundaries, coordinate systems, sampling design and the modifiable areal unit problem can change results substantially.
• A spatially smooth prediction surface is not the same thing as a directly observed surface.

One-Sentence Answer

Spatial statistics works by modeling how observations depend on location, distance, neighbourhood and geographic support so that clustering, prediction and uncertainty reflect the structure of space rather than assuming every mapped observation is independent.

1. A Map Is Not Yet a Spatial Model

Mapping values can reveal patterns, but a map alone does not tell us whether apparent clusters are stronger than expected, whether neighbouring observations are statistically dependent or whether a smooth pattern arose from interpolation.

Spatial statistics adds an explicit model of spatial relationship. The central questions become: what is near what, how does similarity change with distance, what spatial unit is being measured, and what process could have generated the pattern?

2. Different Spatial Data Types Carry Different Meaning

• Point data record events or measurements at locations.
• Areal data attach values to polygons such as districts, census tracts or countries.
• Raster data represent space as cells or pixels.
• Line and network data live on roads, rivers, pipes or transport links.
• Geostatistical fields treat a variable as varying continuously through space even though it is observed only at sample locations.

Methods designed for one representation can be wrong for another. Distance along a river network is not necessarily straight-line distance. A polygon average is not a point measurement. A pixel is an area of support, not an infinitesimal location.

3. Spatial Support: What Area Does One Observation Represent?

Spatial support is the size and shape of the region over which a measurement applies. A temperature sensor gives something close to a point observation. A district poverty rate summarizes an area. A satellite pixel averages or integrates a signal over a footprint.

Changing support changes variance and correlation. Predictions at points and predictions for area averages therefore need different treatment. Ignoring support can make apparently precise comparisons scientifically mismatched.

4. Spatial Autocorrelation: Nearby Is Not Independent

Spatial autocorrelation occurs when values at nearby or connected locations are more similar or more dissimilar than expected under a spatially random arrangement.

Positive spatial autocorrelation produces clusters of similar values. Negative spatial autocorrelation produces neighbouring contrast. If spatial dependence is ignored, conventional standard errors can become too small because the dataset contains less independent information than the row count suggests.

5. “Neighbour” Must Be Defined

A spatial model needs a rule for connection. Polygon neighbours may share an edge, share any boundary point, fall within a distance band or belong to a fixed number of nearest neighbours. Point observations may be linked by distance decay. Network observations may follow actual paths.

There is no universally correct weights matrix. The neighbour definition is part of the scientific hypothesis. If disease spread follows commuting routes, administrative adjacency may be a poor model. If school access follows road travel time, Euclidean distance may be misleading.

6. Spatial Weights Turn Geography Into a Statistical Object

A spatial weights matrix records the connections among observations. An entry wᵢⱼ represents whether or how strongly location j influences the spatial neighbourhood of location i.

Weights can be binary, distance-based, row-standardized or constructed from domain-specific networks. The choice changes the spatial lag and can change test statistics and model estimates. It should therefore be reported, not hidden inside software defaults.

7. Moran’s I Measures Global Spatial Autocorrelation

Moran’s I compares deviations from the mean at each location with deviations at neighbouring locations under the chosen weights structure. Positive values generally indicate clustering of similar values; negative values indicate neighbouring dissimilarity relative to the reference expectation.

The R spdep documentation makes the dependence on the weights matrix explicit in the formula. That is important: Moran’s I is not a property of the attribute values alone. It is a property of values plus a definition of spatial relationship.

8. Geary’s C Emphasizes Neighbour Differences

Geary’s C is another global measure of spatial autocorrelation. Whereas Moran’s I emphasizes cross-products of deviations, Geary’s C is built more directly around squared differences between neighbouring values.

USGS documentation describes Geary’s contiguity ratio as a measure whose calculation depends on both the observed spatial data and a set of dyadic weights defined through proximity, distance or shared boundaries. Different statistics therefore illuminate different aspects of spatial pattern.

9. A Global Statistic Can Hide Local Structure

A whole region can have modest global autocorrelation while containing one strong hotspot, one coldspot and several spatial outliers. Global summaries average across all locations.

Local indicators of spatial association, often called LISA statistics, ask where clustering or local contrast occurs. Local Moran can identify high-high clusters, low-low clusters and locations whose value differs sharply from neighbouring values.

10. Local Hotspots Create a Multiple-Testing Problem

If a local statistic is tested at hundreds or thousands of locations, some apparently significant clusters will occur by chance. Permutation inference, false-discovery control, sensitivity analysis and replication across reasonable weights choices can all matter.

A colourful hotspot map can therefore look more certain than the underlying evidence. Statistical significance, practical magnitude and spatial stability should be read together.

11. Point Patterns Ask Whether Events Cluster in Space

When the data are event locations rather than attribute values attached to fixed areas, the question changes. Point-pattern analysis asks whether events are clustered, dispersed or compatible with a reference spatial process.

Raw point density can be deceptive when the population at risk is uneven. More shops, accidents or disease cases may appear where more people, roads or opportunities exist. A meaningful point-process model often needs an intensity surface or exposure process.

12. The Modifiable Areal Unit Problem

Aggregate the same underlying observations into different zones and statistical relationships can change. This is the modifiable areal unit problem, or MAUP.

Results can change because of scale, such as neighborhoods versus districts, or because of zoning, such as different ways of drawing boundaries at the same general scale. A correlation between district averages is therefore not a boundary-free fact about individual people.

13. Ecological Fallacy: Areas Are Not Individuals

An association between area-level variables does not automatically hold for individuals inside those areas. Districts with higher average income may have higher average exam performance without implying that every higher-income individual scores better than every lower-income individual.

Aggregation changes the unit of inference. Spatial analysis must keep the observational level explicit.

14. Geostatistics Models Continuous Spatial Dependence

Geostatistics is often used when a variable is conceived as a continuous field but measured at selected locations: soil concentration, groundwater level, mineral grade, rainfall or another environmental quantity.

USGS describes geostatistics as a family of methods for characterizing spatial dependence and predicting values at unsampled locations. The central object is usually a covariance or variogram model describing how similarity changes with separation.

15. The Variogram Describes How Difference Grows With Distance

An empirical variogram summarizes how squared differences between observations change as their separation distance increases. Nearby pairs often differ less than distant pairs when positive spatial dependence is present.

A fitted variogram commonly has concepts called the nugget, sill and range. The nugget represents very short-scale variation and measurement noise; the sill represents the plateau of variance under a bounded model; the range describes the distance over which spatial dependence effectively fades.

These are model parameters, not universal physical constants. Their meaning depends on sampling, support, transformation and the variogram family.

16. Anisotropy Means Direction Matters

Spatial dependence can differ by direction. Pollution may spread along prevailing wind. Groundwater follows geology. Travel follows a network. A process can therefore have a longer correlation range east-west than north-south.

Assuming isotropy when the process is directional can distort interpolation and uncertainty. Directional variograms and anisotropic covariance models are designed for this problem.

17. Kriging Predicts at Unsampled Locations

Kriging uses the modeled spatial dependence to construct weighted predictions at locations that were not directly sampled. Nearby observations usually receive more influence, but the exact weights depend on the entire covariance or variogram structure rather than on distance alone.

Ordinary kriging commonly assumes an unknown constant mean over the local domain and a specified spatial covariance structure. Universal kriging can incorporate a trend model. Cokriging can use correlated auxiliary variables.

18. Kriging Is Not Magic Interpolation

A smooth kriged surface can look like a high-resolution observation, but most pixels may never have been measured. The surface is a statistical prediction conditional on a fitted dependence model.

USGS notes that ordinary kriging is a minimum mean-square-error spatial estimator and that smoothing is inherent to such estimators. Fine-looking detail should therefore never be confused with fine-resolution evidence.

19. Prediction Uncertainty Is Part of the Map

Kriging can produce an uncertainty measure as well as a predicted value. Uncertainty is typically smaller near informative observations and larger far from them, though geometry, covariance parameters and local sampling density all matter.

A responsible spatial map should therefore ask two questions: what is predicted here, and how uncertain is that prediction? Showing only the mean surface hides half the model.

20. Spatial Regression Handles Dependence in Areal Outcomes

Regression with spatial data may leave spatially correlated residuals. Spatial-lag models allow outcomes to depend on neighbouring outcomes. Spatial-error models place dependence in the disturbance process. Conditional or simultaneous autoregressive models provide other representations.

The models imply different mechanisms. Choosing among them should be guided by how the process could propagate through space, not solely by which one minimizes an information criterion.

21. Geographic Variation in Relationships Is a Separate Question

Sometimes the issue is not residual spatial dependence but whether a relationship itself changes across geography. Local-regression approaches can explore spatially varying coefficients.

But local estimates can be unstable, highly correlated and sensitive to bandwidth. Apparent geographic variation can arise from unmeasured confounding, sparse data or multiple comparisons. Local models are exploratory tools unless their stability and interpretation are carefully validated.

22. Distance Depends on the Coordinate System

Latitude and longitude are angular coordinates on the Earth, not planar meters. Straight Euclidean calculations on raw degrees can be misleading, especially across large areas or near the poles.

Coordinate reference systems, geodesic distances and appropriate map projections matter because the statistical neighbourhood can depend on the distance calculation itself.

23. Edge Effects Change What “Neighbourhood” Means

A point near the boundary of the study region has unobserved space outside the window. A district on a national border may have neighbours across the border that are absent from the dataset. Point-pattern and local-spatial methods can therefore be biased near edges.

Edge correction, expanded study windows and explicit boundary assumptions help, but the scientific boundary itself may also matter. A coastline is different from an arbitrary administrative cut.

24. Preferential Sampling Can Bias Spatial Inference

Sampling locations are often not random. Sensors are easier to install near roads. Pollution is measured more often near suspected sources. Biodiversity observations cluster where observers travel.

If the probability of sampling depends on the underlying spatial process, conventional interpolation can be biased. USGS research on variogram estimation explicitly notes the challenge of preferentially located samples. Sampling design is therefore part of spatial inference, not a separate logistical detail.

25. Spatial and Temporal Dependence Can Interact

Rainfall, traffic, disease transmission and ecological populations vary in both space and time. A location can depend on nearby places, its own history and the histories of neighbouring places.

Spatiotemporal covariance models, dynamic state-space models and space-time variograms can represent these interactions. The challenge grows quickly because “near” now has both spatial and temporal dimensions.

26. Privacy Risk Also Has a Spatial Structure

Fine-resolution maps of people, households, rare conditions or small groups can reveal identities even when names are absent. Combining geography with other attributes can make re-identification easier.

Aggregation, masking, jittering, suppression and formal privacy methods can reduce risk, but each alters spatial information. The safest map is not always the most analytically useful map, so publication requires an explicit privacy-utility trade-off.

27. A Worked Intuition: School Results Across Neighbourhoods

Imagine mapping average examination results by neighbourhood. A cluster of high averages appears. There are several possible explanations: neighbouring areas share socioeconomic conditions; school catchments overlap; the zones were drawn around similar populations; the values are spatially smoothed; or random variation created a temporary pattern.

A responsible analysis would define the geographic unit, examine denominators, choose and justify a spatial neighbourhood, test global and local dependence, consider confounding, examine sensitivity to alternate zone systems and avoid inferring individual student behaviour from area averages.

The map is the beginning of the reasoning, not the conclusion.

28. A Responsible Spatial-Statistics Workflow

• Define the spatial object. Point, area, raster, line, network or continuous field?
• Define support and coordinate reference system.
• Map the raw observations and the sampling pattern.
• Choose a scientifically defensible neighbour or distance structure.
• Test global dependence before interpreting local patterns.
• Use local statistics with multiple-testing awareness.
• For geostatistics, inspect empirical variograms and directional structure.
• Cross-validate interpolation. Prediction surfaces need out-of-sample checks.
• Map uncertainty as well as predicted values.
• Test sensitivity to scale, zoning and weights choices.
• Keep area-level and individual-level inference separate.
• Assess privacy before publishing fine-grained outputs.

29. Common Failure Modes

• Treating mapped observations as independent merely because they occupy separate rows.
• Using a default weights matrix without explaining why it represents the process.
• Reading a global Moran’s I as proof that every part of the map clusters.
• Publishing local hotspot maps without considering multiple testing.
• Confusing interpolation with direct observation.
• Showing a kriged mean surface without an uncertainty surface.
• Ignoring MAUP and boundary sensitivity.
• Making individual claims from area-level correlations.
• Calculating distance on inappropriate coordinates.
• Ignoring preferential sampling.
• Fitting a spatial model to compensate for missing causal variables and then making causal claims.
• Using visually precise maps to imply statistically precise evidence.

30. How This Connects Across the eduKate Library

• How Maps and Geospatial Evidence Work — the broader owner for coordinates, projection, GIS, remote sensing, provenance and map interpretation.
• How Regression Models Work — the general regression route before spatial dependence is added.
• How Clustered and Multilevel Data Work — related dependence structures when observations belong to nested groups rather than geographic neighbours.
• How Data Visualisation Works — for honest visual encoding of spatial estimates and uncertainty.
• How Multiple Testing and Sequential Analysis Work — important when many local spatial tests are performed.
• How Time-Series Analysis Works — the neighbouring owner when dependence occurs through time rather than space, and a building block for spatiotemporal models.
• How Statistical Disclosure Control and Confidentiality Work — for the privacy consequences of publishing small-area geography.

31. Authoritative Sources and Further Reading

• GeoDa Center, Local Indicators of Spatial Association documentation — Local Moran, Local Geary, Getis-Ord statistics and permutation-based local inference.
• R spdep, Compute Moran’s I — formula and implementation details showing explicit dependence on spatial weights.
• R spdep, Local Moran’s I statistic — local spatial association and permutation variants.
• U.S. Geological Survey, A Statistical Evaluation of Non-Ergodic Variogram Estimators — discussion of variograms, kriging and preferentially located samples.
• U.S. Geological Survey, Merging Machine Learning and Geostatistical Approaches for Spatial Modeling of Geoenergy Resources, August 6, 2023 — modern overview of variogram-based geostatistical estimation and kriging.
• U.S. Geological Survey, Compensating for Estimation Smoothing in Kriging — why kriging predictions can be smoother than the underlying spatial field.

Final Idea

Spatial statistics begins with a simple correction to ordinary thinking: place changes information. Once location affects similarity, exposure, sampling or measurement support, geography becomes part of the model. The goal is not to make prettier maps. It is to stop a map from quietly turning dependence, interpolation, aggregation and boundary choices into facts that were never actually observed.