Loading field map
When data points are collected across a landscape, a brain scan, or a city map, the fundamental assumption of classical statistics—that observations are independent—breaks down. Locations close together tend to be more alike than locations far apart, a property known as spatial dependence. Spatial statistics is the branch of statistics that builds models and methods specifically to account for this dependence, turning what would otherwise be a violation of assumptions into the central object of study. The field has developed four major frameworks, each addressing a different kind of spatial data: continuous fields, discrete areal units, event locations, and data that also unfold through time.
The first framework to crystallize was geostatistics, which emerged from mining and petroleum engineering in the 1960s. The practical pressure was straightforward: given a few drill-core samples, how do you predict the grade of ore at an unsampled location? Geostatistics treats the spatial phenomenon as a continuous random field—a surface of values that varies smoothly across space. The key innovation was the variogram, a function that describes how the variance between two measurements grows with the distance between their locations. Using the variogram, the method of kriging (named after South African engineer Danie Krige) produces optimal linear predictions at unsampled points, along with a measure of prediction uncertainty.
Geostatistics remains a leading framework today, especially in earth sciences, environmental monitoring, and natural resource estimation. Its strength lies in interpolation and uncertainty quantification for continuous variables. Later frameworks did not replace geostatistics; rather, they coexisted with it by addressing data types that geostatistics was not designed for. Geostatistics assumes a continuous underlying surface, but many spatial problems involve data aggregated into discrete zones or patterns of points.
By the 1970s, researchers working with data aggregated into regions—counties, census tracts, grid cells—needed a different approach. In lattice data, the spatial locations are fixed polygons or pixels, and the question is how to model the dependence between neighboring areas. The foundational contribution was the conditional autoregressive (CAR) model, which specifies the value at each region as conditionally dependent on its neighbors. This framework is closely tied to Markov random fields and is often implemented in a Bayesian hierarchical setting.
Lattice data analysis differs from geostatistics in a fundamental way: it does not assume a continuous field that can be interpolated anywhere. Instead, it models the correlation structure among a fixed set of areal units. This makes it the natural framework for disease mapping, social epidemiology, and spatial econometrics. The CAR model and its extensions (e.g., the intrinsic CAR and Besag-York-Mollié model) remain workhorses in public health and criminology. The framework coexists with geostatistics by occupying a distinct data regime—discrete regions rather than continuous surfaces.
Also emerging in the 1970s, spatial point process analysis addresses a third type of data: the locations of events themselves. Instead of measuring a value at every location, the data are a set of points—tree positions in a forest, earthquake epicenters, crime incidents. The central question is whether the points exhibit clustering, regularity, or complete spatial randomness. The key tool is the Ripley's K-function (and its derivative, the L-function), which compares the observed distribution of inter-point distances to what would be expected under a Poisson process.
This framework differs from both geostatistics and lattice analysis in that the locations are random, not fixed. The focus is on the stochastic process generating the points, not on interpolating a continuous field or modeling dependence among fixed regions. Spatial point process analysis has been extended to include covariates (via the Cox process and Gibbs processes) and to handle marked points (where each event carries additional attributes). It remains active in ecology, seismology, and epidemiology, often used alongside geostatistics when both event locations and continuous covariates are present.
By the 1990s, the growing availability of repeated measurements over time—satellite imagery, climate records, disease surveillance—pushed spatial statistics to incorporate the temporal dimension explicitly. Spatiotemporal statistics extends all three earlier frameworks to data that vary in both space and time. The distinctive shift is that dependence must now be modeled across two dimensions: locations close in space and time are expected to be more similar than those far apart in either dimension.
Spatiotemporal models often build on geostatistical foundations, using space-time variograms and spatiotemporal kriging. For lattice data, dynamic CAR models and spatiotemporal autoregressive models allow the spatial dependence structure to evolve over time steps. For point patterns, spatiotemporal point processes (e.g., the space-time Poisson process and the epidemic-type aftershock sequence model) capture how event clusters propagate. A major practical pressure driving this framework was the need for real-time disease surveillance and environmental monitoring, where predictions must be updated as new data arrive.
Spatiotemporal statistics did not replace the earlier frameworks; it absorbed and extended them. Today, it is the fastest-growing area within spatial statistics, with applications in climate science, air quality monitoring, epidemiology, and remote sensing. Its relationship to the other frameworks is one of transformation: it takes the core ideas of geostatistics, lattice analysis, and point process analysis and adds a temporal dimension, often requiring new computational strategies (such as dimension reduction and Gaussian process approximations) to handle the massive datasets involved.
The four frameworks are not competing paradigms but a division of labor based on data type. Geostatistics is the method of choice when the goal is interpolation of a continuous field from sparse samples. Lattice data analysis is used when data come pre-aggregated into discrete regions and the focus is on spatial autocorrelation and areal smoothing. Spatial point process analysis is applied when the locations themselves are the data and the interest is in clustering or regularity. Spatiotemporal statistics extends all three when time is a factor.
Despite their different starting points, the frameworks share a core commitment: they all model spatial dependence explicitly rather than treating it as a nuisance. They also increasingly borrow tools from each other. For example, geostatistical methods are now routinely used to smooth lattice data, and point process models often include geostatistically interpolated covariates. The main disagreement among practitioners is not about which framework is correct but about how to handle computational scalability and model choice in large datasets.
Spatial statistics has deep ties to statistical learning, particularly through kernel methods and Gaussian processes. Gaussian process regression, a staple of modern machine learning, is essentially a geostatistical kriging model with a flexible covariance function. Bayesian learning frameworks are widely used in spatial statistics for hierarchical modeling, especially in lattice data analysis and spatiotemporal settings. Conversely, spatial statistics has influenced statistical learning by contributing structured covariance models and methods for handling non-Euclidean spaces.
Another important connection is to survival analysis. Spatial survival models combine frailty models (random effects that account for unobserved heterogeneity) with spatial correlation structures. For example, in epidemiology, the time until disease onset may depend on both individual risk factors and the spatial location of exposure. These models typically use a lattice or geostatistical framework to capture spatial dependence in the baseline hazard or frailty term. The integration of spatial and survival methods remains an active research frontier, especially in cancer registry studies and environmental health.
Today, all four frameworks agree on the fundamental principle that spatial dependence must be modeled, not ignored. They also agree on the value of hierarchical Bayesian approaches for quantifying uncertainty and incorporating prior information. Where they disagree is on the most appropriate representation of space: continuous random fields (geostatistics) versus discrete adjacency structures (lattice) versus stochastic point processes (point patterns). This disagreement is not a conflict but a reflection of the different data types each framework was designed for. The leading frameworks coexist because real-world spatial problems rarely fit neatly into one category; a single study might involve interpolating a continuous pollutant surface (geostatistics), mapping disease rates by county (lattice), and analyzing the clustering of case locations (point process), all within a spatiotemporal framework. The future of the field lies in hybrid models that combine elements of all four traditions, often powered by advances in computation and machine learning.