Spatial Statistics

Spatial statistics is the branch of statistics that deals with data whose values are tied to locations and whose nearby values tend to resemble one another.

7 frameworks, from 1950 to today · Statistics · Mathematics

still a working view ended · each leads to its entry below

Overview

Spatial statistics is the branch of statistics that deals with data whose values are tied to locations and whose nearby values tend to resemble one another12. That single fact—that geography breaks the usual assumption that observations are independent—is what makes the field necessary2. If you sample air pollution at monitoring stations, house prices across a city, or disease cases across a country, a station's reading tells you something about its neighbours' readings32. Standard statistics treats that as a nuisance or a violation; spatial statistics treats it as the central object of study2. Its history is a sequence of attempts to describe that dependence, to use it for prediction, and to test whether apparent patterns could have arisen by chance.

Early mapping and the roots of spatial thinking

Long before formal theory, practical people noticed spatial dependence. John Snow's 1854 map of cholera deaths around the Broad Street pump in London is the canonical early example of using the location of cases to reason about a cause, though Snow's argument was primarily about water supply rather than a statistical model of spatial association. In the nineteenth and early twentieth centuries, cartographers and geologists developed interpolation by hand: drawing contours through scattered measurements, a practice that already assumed nearby points are more alike than distant ones.

The first formal statistical treatment of spatial dependence came from agronomy. In the 1920s and 1930s, R. A. Fisher and colleagues at Rothamsted Experimental Station developed the analysis of field trials, where plots laid out in a grid are subject to soil variation that is itself spatially structured. Fisher's randomization and blocking designs were, in part, a way to neutralize spatial dependence so that treatment effects could be estimated cleanly. The complementary question—how to model that dependence—came later.

Geostatistics: from mining to kriging

The decisive turn came from mining. In the 1950s, Danie Krige, a South African mining engineer, noticed that when estimating the grade of gold ore from borehole samples, the standard formulas gave unreliable answers because samples close together carried redundant information. He developed empirical weighting schemes that down-weighted clustered samples. In the 1960s, the French mathematician Georges Matheron formalized Krige's intuition into what he named the theory of regionalized variables, and the method of optimal spatial prediction became known as kriging in Krige's honour.

The core idea is the variogram: a function describing how the dissimilarity between two measurements grows with the distance separating them4. From an estimated variogram, one can compute the best linear unbiased prediction at an unobserved location—a weighted average of nearby observations, with weights chosen to minimize prediction variance. This made it possible to produce not just a map but a map with error bars. Geostatistics spread from mining into hydrology, soil science, and environmental monitoring, and by the 1980s it was a standard tool in the earth sciences. The 1990s saw it absorbed into mainstream statistics through the work of Noel Cressie, whose 1991 book Statistics for Spatial Data reframed kriging as a special case of linear mixed models, connecting it to the general theory of best linear unbiased prediction.

Point patterns: asking whether events cluster

A second strand concerned events that occur at points—trees in a forest, disease cases in a city, crimes on a street network. The question is whether the points are more clustered, more regular, or more random than chance would produce. The foundational tool was the nearest-neighbour distance distribution, and in the 1950s and 1960s ecologists and statisticians developed tests based on it. The key advance was the Ripley's K-function, introduced by Brian Ripley in 1976, which summarizes clustering across all distance scales rather than at a single nearest-neighbour distance. A related line, the analysis of quadrat counts (counting events in grid cells and testing against Poisson variation), had been used since the 1920s but was limited by the arbitrary choice of cell size.

Point pattern analysis became a mature field through the 1980s and 1990s, with the development of edge corrections (because points near a study boundary have fewer neighbours than interior points), inhomogeneous models (where the underlying intensity varies across space), and space-time extensions. The 2000s brought spatstat, an open-source R package by Adrian Baddeley and Rolf Turner, which made the full toolkit accessible and is now the standard reference implementation.

Areal and lattice models: when space is a set of regions

A third tradition dealt with data aggregated to regions—census tracts, counties, countries—where each unit has a value and neighbours are defined by adjacency. The central problem is that neighbouring regions tend to have similar values, and ignoring this inflates apparent significance. The Moran's I statistic, introduced by Patrick Moran in 1950, measures global spatial autocorrelation: whether values are more similar among neighbours than expected under randomness. The Geary ratio (1954) offered a complementary measure. Local versions—local indicators of spatial association (LISA), developed by Luc Anselin in 1995—identified where clustering occurs, not just whether it exists.

For modelling, the key development was the conditional autoregressive (CAR) and simultaneous autoregressive (SAR) models, which specify a region's value as a function of its neighbours' values plus noise. These were formalized in the 1970s and 1980s, drawing on earlier work in econometrics and geography. The spatial lag and spatial error models became standard in spatial econometrics, a field Anselin helped consolidate in the late 1980s and 1990s, which applies these structures to economic data—house prices, voting, regional growth—where spillovers between neighbouring units are substantively meaningful rather than a nuisance.

Bayesian hierarchical modelling and the smoothing revolution

The 1990s brought a major unification. Bayesian hierarchical models allowed spatial dependence to be represented as a latent random effect with a structured prior—typically a CAR or a Gaussian process—nested within a larger model for the observed data. The Besag–York–Mollié (BYM) model, introduced by Julian Besag, Jeremy York, and Annie Mollié in 1991, became the workhorse for disease mapping: it decomposes each region's relative risk into a spatially structured component and an unstructured heterogeneity component, so that maps of estimated risk are smoothed toward neighbouring values without erasing genuine local spikes.

This framework was computationally demanding, and its spread depended on Markov chain Monte Carlo (MCMC) methods, which became practical in the 1990s. The later development of integrated nested Laplace approximation (INLA) by Håvard Rue and colleagues in 2009 made Bayesian spatial models fast enough for routine use, and the SPDE approach (stochastic partial differential equations) gave a computationally efficient bridge between Gaussian processes and the Markov random fields used in areal models.

Spatially varying coefficients: letting relationships change across space

A parallel development asked not just where values are high but whether the relationship between variables changes across space. Geographically weighted regression (GWR), introduced by Stewart Fotheringham, Chris Brunsdon, and Martin Charlton in the mid-1990s, fits a separate regression at each location, weighting nearby observations more heavily. It produces maps of local coefficients—showing, for example, that the effect of income on house prices differs between neighbourhoods. GWR has been widely used in geography and urban analysis, though it has also drawn criticism for collinearity among local estimates and for the difficulty of inference. Bayesian spatially varying coefficient models, developed in the 2000s, offer an alternative with formal uncertainty quantification.

Machine learning and the present

The most recent turn has been the incorporation of machine learning methods for spatial prediction. Tree-based ensembles, Gaussian process regression, and neural networks can capture nonlinear spatial relationships that kriging and linear models miss, and they scale to large datasets. The tension is that these methods are often black boxes with no built-in notion of spatial dependence; researchers have responded by building spatial features (neighbourhood averages, distances to covariates) into the models, by using random forests with spatial cross-validation to avoid over-optimistic error estimates, and by combining geostatistical covariance structures with flexible learners. Gaussian process regression itself, long central to spatial statistics, has been absorbed into machine learning as a general-purpose Bayesian nonparametric method, so the boundary between the fields is now porous.

Today the main approaches coexist: geostatistical kriging for continuous surfaces, point process models for event locations, areal models and spatial econometrics for regional data, Bayesian hierarchical models for smoothing and uncertainty, and machine learning for flexible prediction at scale. The open questions are familiar in shape but hard in practice: how to do honest inference when the spatial dependence structure is itself estimated, how to combine mechanistic and data-driven models, how to handle massive datasets without losing the uncertainty quantification that made the field worth having, and how to move between the many scales at which spatial processes operate.

References

  1. Spatial statistics, Wikipedia
  2. Spatial analysis (Spatial auto-correlation, Spatial dependence), Wikipedia
  3. Geostatistics (Background), Wikipedia
  4. Variogram, Wikipedia

The frameworks of Spatial Statistics

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1950-Present

    Areal and lattice models

    How do neighbouring regions influence each other's values?

    Fly to its tower in Nousfarer

  2. 1950-Present

    Point pattern analysis

    Are events clustered, regular, or random in space?

    Fly to its tower in Nousfarer

  3. 1960-Present

    Geostatistics

    How can we predict values at unobserved locations?

    Fly to its tower in Nousfarer

  4. 1988-Present

    Spatial econometrics

    How do economic spillovers between neighbouring units work?

    Fly to its tower in Nousfarer

  5. 1991-Present

    Bayesian hierarchical spatial modelling

    How can we smooth noisy spatial estimates with uncertainty?

    Fly to its tower in Nousfarer

  6. 1996-Present

    Geographically weighted regression

    Do relationships between variables change across space?

    Fly to its tower in Nousfarer

  7. 2010-Present

    Machine learning for spatial prediction

    Can flexible learners capture nonlinear spatial relationships?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Areal and lattice modelsPoint pattern analysisGeostatisticsSpatial econometricsBayesian hierarchical spatial modelli…Geographically weighted regressionMachine learning for spatial predicti…

Beyond the field

Where Spatial Statistics's ideas came from in other fields, and what they led to there: the strongest 12 of 24 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Spatial Statistics is a planet in Nousfarer

Its city has a tower for each of these 7 frameworks. Land, walk in, and climb one concept at a time.

Land on Spatial Statistics

Related fields

Fields its ideas are linked to

More in Statistics

See where Spatial Statistics sits on the Map of Human Knowledge