Loading field map
Spatial statistics is the branch of statistics concerned with the analysis of data that have locations in space. Its central challenge is that observations collected at nearby locations tend to be more similar than observations collected far apart—a property called spatial dependence. Standard statistical methods that assume independence among observations are therefore inappropriate for spatial data, and spatial statistics provides the theory and methods to model, estimate, and draw inferences from such dependent data.
The defining feature of spatial data is that the spatial arrangement of observations carries information. Two observations taken from the same location are not independent replicates; they are, in a sense, the same place. More generally, the closer two locations are, the more similar their values tend to be, a phenomenon known as Tobler's first law of geography: "Everything is related to everything else, but near things are more related than distant things." This spatial autocorrelation violates the independence assumption of classical statistics and, if ignored, leads to underestimated standard errors, inflated significance, and unreliable predictions.
Spatial statistics addresses three broad types of data: point-referenced data (also called geostatistical data), where measurements are taken at fixed locations; areal data (also called lattice data), where observations are aggregated over discrete regions; and point pattern data, where the locations of events themselves are the object of study. Each type raises distinct modeling questions and has developed its own set of methods.
Spatial statistics emerged from several distinct traditions. In the 1950s and 1960s, the South African mining engineer Danie Krige developed methods for estimating ore grades from spatially scattered borehole samples, leading to the theory of regionalized variables formalized by the French mathematician Georges Matheron. This tradition, known as geostatistics, remains central to spatial prediction. Independently, statisticians working with agricultural field trials and epidemiological maps developed methods for analyzing data aggregated over regions, leading to spatial autoregressive models. A third tradition, originating in ecology and geography, studied the spatial patterns of points—whether trees in a forest or cases of a disease—using methods from stochastic point process theory. These three traditions were brought together under a unified framework in the 1980s and 1990s, particularly through the work of statisticians such as Noel Cressie, whose 1993 book Statistics for Spatial Data provided a comprehensive synthesis.
Geostatistics addresses point-referenced data: measurements taken at a set of known locations, such as soil samples at monitoring wells or air pollution readings at monitoring stations. The core problem is spatial prediction—estimating the value at an unmeasured location from the measured values at nearby locations. The foundational tool is the variogram, which characterizes how the variance between two measurements depends on the distance and direction separating them. The variogram is estimated from the data and then used in kriging, a family of optimal linear predictors that give both a predicted value and a measure of prediction uncertainty.
Kriging is named after Danie Krige but was formalized by Matheron as the best linear unbiased predictor for a spatial random field. It is mathematically equivalent to Gaussian process regression, a connection that has become increasingly important as spatial statistics has merged with machine learning. The key assumption in geostatistics is that the spatial process is continuous and that the spatial dependence structure can be inferred from the data. Important limitations include the assumption of stationarity (that the spatial dependence structure is constant across the study region) and the computational burden of kriging with large datasets, which has motivated the development of approximation methods such as fixed-rank kriging and nearest-neighbor Gaussian processes.
Areal data consist of observations aggregated over discrete regions, such as counties, census tracts, or grid cells. Here the spatial structure is defined by neighborhood relationships—which regions share a border or are otherwise considered adjacent. The central modeling approach is the spatial autoregressive model, which extends classical regression by including a term that captures the average of neighboring values. The two main variants are the spatial lag model, where the dependent variable itself is spatially autocorrelated, and the spatial error model, where the error terms are spatially autocorrelated.
These models are estimated using maximum likelihood or Bayesian methods, and their interpretation requires care: a coefficient in a spatial lag model does not have the same meaning as in ordinary regression, because a change in one region propagates through the spatial system to affect its neighbors. This has led to the development of impact measures that decompose total effects into direct and indirect (spillover) components. A limitation of spatial autoregressive models is that they assume the spatial dependence structure is known and fixed by the neighborhood definition, which is often arbitrary. Misspecification of the neighborhood matrix can lead to biased estimates.
Point pattern analysis studies the locations of events—the positions of trees in a forest, the addresses of disease cases, the epicenters of earthquakes. The fundamental question is whether the observed pattern is completely spatially random (a homogeneous Poisson process) or exhibits clustering, regularity, or dependence on covariates. The key tools are summary functions such as Ripley's K-function, which compares the observed number of points within a given distance of each point to what would be expected under complete spatial randomness.
Point pattern models include the inhomogeneous Poisson process, where the intensity varies with location, and more complex models such as the Cox process (a Poisson process with a random intensity) and the Gibbs process, which models interactions between points. A major challenge is that point patterns are often influenced by an underlying spatial process that is not directly observed—for example, cases of a disease may cluster because the population at risk is clustered, not because the disease is contagious. This has motivated the development of methods that account for an inhomogeneous background intensity, such as the inhomogeneous K-function and case-control designs.
Bayesian methods have become increasingly influential across all branches of spatial statistics. The Bayesian approach treats all unknown quantities—including the spatial random field, model parameters, and predictions—as random variables with probability distributions. This provides a natural framework for incorporating prior information, quantifying uncertainty, and handling complex hierarchical models. In geostatistics, Bayesian kriging allows the variogram parameters to be estimated jointly with the spatial predictions, properly accounting for uncertainty in the dependence structure. In areal data, Bayesian hierarchical models, particularly the conditional autoregressive (CAR) model, are widely used for disease mapping, where the goal is to estimate the underlying spatial risk from observed counts in small areas.
The computational demands of Bayesian spatial models have been addressed by Markov chain Monte Carlo (MCMC) methods and, more recently, by integrated nested Laplace approximations (INLA), which provide fast approximate Bayesian inference for a large class of spatial models. The Bayesian approach has also facilitated the integration of spatial statistics with other modeling frameworks, such as spatio-temporal models and multivariate spatial models.
The three main data types—point-referenced, areal, and point pattern—are not always distinct. Point-referenced data can be aggregated into areal data, and the locations of point-referenced measurements can themselves be treated as a point pattern. The modeling approaches also overlap: geostatistical methods can be applied to areal data by treating the region centroids as point locations, though this ignores the spatial support of the observations. Conversely, areal models can be used for point-referenced data by discretizing the continuous space into a grid. The Bayesian framework provides a unifying language for all three traditions, and modern software implementations often allow users to switch between data types within a single modeling environment.
Contemporary spatial statistics is characterized by several developments. The increasing availability of large spatial datasets—from satellite imagery, GPS devices, and environmental sensors—has driven the development of scalable methods, including Gaussian process approximations, divide-and-conquer algorithms, and methods that exploit the sparsity of spatial covariance matrices. There is also growing integration with machine learning, particularly through deep Gaussian processes and spatial random forests, though these methods often sacrifice the inferential guarantees that are central to the spatial statistics tradition.
Another important trend is the extension of spatial methods to spatio-temporal data, where both location and time induce dependence. This includes models for data observed at fixed locations over time (spatio-temporal geostatistics), data aggregated over space and time (spatio-temporal areal models), and the trajectories of moving objects. The COVID-19 pandemic highlighted the importance of spatio-temporal models for infectious disease surveillance and forecasting.
Spatial statistics remains an active research field with applications in environmental science, epidemiology, ecology, economics, criminology, and many other domains. Its core contribution—providing principled methods for data that are not independent—is increasingly recognized as essential in an era of location-aware data collection. The field continues to grapple with fundamental challenges, including the modifiable areal unit problem (the sensitivity of results to the scale and shape of spatial aggregation), the ecological fallacy (inferring individual-level relationships from aggregate data), and the difficulty of causal inference from observational spatial data.