[Paper Review] Spatial Statistics

Noel Cressie, Matthew T. Moores|arXiv (Cornell University)|May 15, 2021

Soil Geostatistics and Mapping4 citations

TL;DR

This paper presents a unified Bayesian hierarchical framework for spatial statistics, integrating uncertainty in spatial processes, locations, and parameters. It formalizes geostatistical, lattice, and point process models through conditional probability decomposition, enabling robust inference via posterior prediction and scalable computation.

ABSTRACT

Spatial statistics is an area of study devoted to the statistical analysis of data that have a spatial label associated with them. Geographers often refer to the "location information" associated with the "attribute information," whose study defines a research area called "spatial analysis." Many of the ways to manipulate spatial data are driven by algorithms with no uncertainty quantification associated with them. When a spatial analysis is statistical, that is, it incorporates uncertainty quantification, it falls in the research area called spatial statistics. The primary feature of spatial statistical models is that nearby attribute values are more statistically dependent than distant attribute values; this is a paraphrasing of what is sometimes called the First Law of Geography (Tobler, 1970).

Motivation & Objective

• To unify spatial statistical modeling under a single hierarchical framework that accounts for uncertainty in spatial processes, locations, and parameters.
• To formalize the role of the First Law of Geography—nearby locations are more dependent—within a probabilistic, Bayesian statistical model.
• To address computational challenges in spatial statistics by advocating spatial discretization and scalable inference methods for large datasets.
• To incorporate modern advances such as penalized complexity priors, surrogate likelihoods, and barrier models for nonstationarity and non-Gaussianity.
• To reframe classical geostatistical tools like kriging and the variogram within a modern Bayesian hierarchical modeling context.

Proposed method

• Formalizes spatial statistical models using the hierarchical decomposition: $[\mathbf{Z}, Y, D, \boldsymbol{\theta}] = [\mathbf{Z} \mid Y, D, \boldsymbol{\theta}][Y \mid D, \boldsymbol{\theta}][D \mid \boldsymbol{\theta}][\boldsymbol{\theta}]$, where $\mathbf{Z}$ is data, $Y$ is the spatial process, $D$ is the spatial index set, and $\boldsymbol{\theta}$ are parameters.
• Models spatial dependence through the First Law of Geography, encoded in the conditional distribution $[Y \mid D, \boldsymbol{\theta}]$, which induces stronger dependence between nearby locations.
• Applies Bayes' rule to compute the posterior $[Y, D \mid \mathbf{Z}]$ for inference, enabling predictive distribution updates from observed data.
• Uses spatial discretization to approximate continuous spatial domains $D^G$, enabling computation for geostatistical processes.
• Incorporates penalized complexity (PC) priors to favor simpler models and reduce overfitting in stationary process parameters.
• Employs surrogate models, emulators, and quasi-likelihoods to handle intractable likelihoods in non-Gaussian or point process models.

Experimental results

Research questions

• RQ1 How can spatial statistical models be unified under a single hierarchical Bayesian framework that accounts for uncertainty in processes, locations, and parameters?
• RQ2 How does the First Law of Geography manifest in a probabilistic spatial process model, and how is it encoded in the conditional dependence structure?
• RQ3 What computational strategies enable scalable inference in large spatial datasets, particularly for continuous spatial processes?
• RQ4 How can nonstationarity, anisotropy, and non-Gaussianity be modeled effectively within a Bayesian hierarchical framework?
• RQ5 How do modern tools like PC priors and surrogate likelihoods improve model selection and computational feasibility in spatial statistics?

Key findings

• The hierarchical Bayesian framework $[\mathbf{Z}, Y, D, \boldsymbol{\theta}] = [\mathbf{Z} \mid Y, D, \boldsymbol{\theta}][Y \mid D, \boldsymbol{\theta}][D \mid \boldsymbol{\theta}][\boldsymbol{\theta}]$ provides a comprehensive and coherent foundation for spatial statistical modeling.
• Spatial dependence through the First Law of Geography is naturally embedded in $[Y \mid D, \boldsymbol{\theta}]$, ensuring that nearby locations exhibit higher statistical dependence.
• Spatial discretization of continuous domains $D^G$ enables computationally feasible inference, making modern spatial statistics viable in large-scale applications.
• Penalized complexity (PC) priors effectively favor simpler models by penalizing complex parameter values, improving model parsimony and interpretability.
• Surrogate models and quasi-likelihoods provide practical solutions for intractable likelihoods in non-Gaussian and point process models, enabling broader application.
• The integration of geostatistical tools like kriging and the variogram within the hierarchical Bayesian framework offers a modern, probabilistically grounded interpretation of classical methods.