Spatial Multivariate Framework

• Spatial Multivariate Framework is a collection of statistical models designed to jointly analyze multiple spatially indexed variables by constructing valid joint covariance structures.
• It combines conditional covariance construction, deformation techniques, and latent variable linkages to manage mixed data types and varying spatial supports.
• The framework employs graphical, sparsity-driven, and Bayesian approaches to enhance scalability, precision, and interpretability in diverse applications like environmental and epidemiological studies.

A spatial multivariate framework encompasses the collection of statistical models, inference methods, and analytic strategies designed to jointly analyze multiple spatially indexed variables, capturing both their spatial correlation and mutual dependence. Contemporary spatial multivariate frameworks are motivated by complex scientific, environmental, epidemiological, and socio-economic processes in which multiple phenomena interact across space at varying scales and with varying data types and supports. These frameworks integrate developments from geostatistics, graphical models, functional data analysis, and spatial econometrics to provide interpretable, computationally efficient, and scientifically meaningful multivariate spatial models.

1. Theoretical Foundations and Covariance Constructions

Rigorous spatial multivariate modeling fundamentally depends on the construction of valid joint covariance structures among spatial fields. A canonical approach is the conditional construction of multivariate spatial covariance models, where for $p$ spatial fields $Y_1(\cdot),\ldots,Y_p(\cdot)$, the joint law is factorized recursively as:

$$[Y_1(\cdot),\ldots,Y_p(\cdot)] = \prod_{q=1}^p [Y_q(\cdot)\mid Y_{1:(q-1)}(\cdot)]$$

At each stage, a conditional mean is expressed as a sum of spatial integrations over prior fields, involving interaction kernels $b_{qr}(s,v)$, while conditional variances are modeled by valid univariate spatial covariance functions $C_{q|1:(q-1)}(s,u)$. The resulting joint covariance matrix over all variables and sites is built iteratively, and this procedure guarantees nonnegative-definiteness provided the univariate covariances and kernels are valid and integrable (Cressie et al., 2015).

This approach generalizes to networks of spatial fields corresponding to DAGs, wherein causality, directionality, or parsimony can be imposed by the choice of graph structure. Symmetry and asymmetry in cross-covariance can be resolved and compared empirically, as in meteorological applications where conditional modeling elucidates directional dependence (e.g., temperature vs. pressure forecast errors).

Alternatively, deformation models construct nonstationary and asymmetric multivariate spatial covariance by warping the input domain via injective maps (possibly learned through deep compositions), then applying standard stationary symmetric covariance models in the warped space. Asymmetry is achieved by using process-specific deformations (Vu et al., 2020).

2. Flexible Models and Data Types

Spatial multivariate frameworks must accommodate various data types and structures, including:

• Mixed-type (continuous, count, binary, ordinal) responses via latent multivariate Gaussian processes; here, each marginal response is linked through an appropriate exponential family model to latent fields, with a separable cross-covariance structure (e.g., $C(s,s') = K_\phi(s,s')\,\Sigma$), and identifiability achieved through constraints such as fixing marginal variances for certain data types (Mukherjee et al., 15 Oct 2025).
• Matrix-variate responses and incomplete or spatiotemporal data: Matrix-variate models with separable (Kronecker) or deformation-induced covariance allow for the joint modeling of multiple variables across locations and time, with real-time imputation of missing data via data-augmentation MCMC (Bulhões et al., 22 Nov 2025).
• Mixed-type spatial hybrid processes: A general framework represents all spatial data—point process, lattice, marked point—via multivariate, possibly function-valued, random fields subjected to a Markov property encoded by the spatial dependence graph model (SDGM). Conditional independence among components is examined in the frequency domain via the inverse spectral density, allowing for robust exploratory analysis across data types (Eckardt et al., 2019, Eckardt et al., 2023).
• Compositional and functional data are handled using isometric log-ratio transformations for the former and signature or FPCA expansions for the latter, integrated into spatial autoregressive or regression models that guarantee compositional constraints and exploit the functional structure (Wang et al., 2018, Frévent et al., 2024, Eckardt et al., 18 Jul 2025).

3. Graphical and Sparsity-Driven Approaches

For high-dimensional and large-scale spatial data, sparsity in the joint precision matrix (inverse covariance) is crucial for scalability. The cross-Markov random field (cross-MRF) framework constructs maximally sparse precision matrices by enforcing a two-stage conditional independence structure:

1. A DAG among the $p$ component fields, with cross-component dependencies parameterized via regression kernels,
2. Markov (local neighborhood) dependence across spatial locations, and after moralization, among components at each site.

The result is a block-sparse joint precision matrix amenable to parallel generation and efficient inference, with computable asymmetric cross-covariance structure and explicit mapping onto scientifically motivated conditional independence graphs (Chen et al., 2024).

Similarly, Bayesian matrix-normal frameworks exploit the Kronecker or low-rank structure of the covariance, enabling non-iterative inference at massive scale (Zhang et al., 2020). Tile low-rank (TLR) approximations further accelerate computation on modern parallel architectures, while preserving inferential and predictive accuracy (Salvaña et al., 2020).

4. Joint Modeling, Fusion, and Regionalization

Integrated spatial multivariate frameworks support fusion and regionalization tasks:

• Process-based fusion allows joint analysis of geostatistical, areal, and point-pattern data via a latent-process coregionalization structure, using change-of-support operators to align data of various types and resolutions (Wang et al., 2019).
• Regionalization methods based on aggregation error leverage the multivariate CAGE (MVCAGE) criterion, which quantifies the expected squared difference between point and areal summaries of vector-valued fields via multivariate Karhunen–Loève decomposition. This metric, rigorous through a null-MAUP theorem, guides spatial clustering to minimize aggregation error, with extensions for practical computation via generating basis functions and spatially constrained clustering algorithms (Daw et al., 2023).

5. Specialized Frameworks and Extensions

Spatial multivariate frameworks also encompass:

• Bayesian spatial illness-death survival models with multivariate Leroux, MCAR, or GMRF priors for transition-specific random effects, yielding coherent and interpretable risk maps and posterior transition probabilities at regional level (Llopis-Cardona et al., 2022).
• Extreme-value scenarios modeled via multi-factor copulas or skew-$t$ processes, capturing all combinations of tail-period dependence and enabling joint assessment of spatial and cross-process extremal events, supporting spatial risk estimation under flexible and fully Bayesian uncertainty quantification (Gong et al., 2022, Hazra et al., 2018).
• Survey, ordinal, or block-structured responses via latent-variable models with multivariate spatial CAR effects, supporting estimation and mapping of sets of correlated indicators (e.g., public health surveys) jointly at high spatial resolution, with robust posterior inference (Beltrán-Sánchez et al., 28 Jul 2025).
• Efficient small area estimation for multivariate spatial settings via variational autoencoded spatial Fay-Herriot models, allowing reusable emulators of spatial priors (e.g., MCAR, GMS-FH) and reducing computational costs by several orders of magnitude in high dimensions (Wang et al., 18 Mar 2025).

6. Spatially-Constrained Multivariate Analysis

In community ecology, spatially constrained multivariate analysis targets the simultaneous extraction of variable covariation and spatial autocorrelation. Several strategies have been formalized:

• Between-class analysis (BCA) with a priori partitions emphasizes between-region contrasts.
• Redundancy analysis with polynomial spatial trends (PCAIV-POLY) or Moran’s eigenvector maps (PCAIV-MEM) incorporates smooth spatial predictors or eigenfunctions capturing patterns from the spatial weights matrix.
• MULTISPATI directly optimizes the product of explained variance and spatial autocorrelation (generalized Moran’s coefficient), providing axes that jointly maximize both criteria (Dray et al., 2012).

The choice among these approaches reflects the structure of spatial information, balance of flexibility vs. robustness, and the scale of spatial effects under study.

7. Implementation, Model Comparison, and Applications

Contemporary spatial multivariate frameworks are implemented using modern computational techniques—MCMC, INLA, Hamiltonian Monte Carlo, variational inference, tile-low-rank or approximate nearest-neighbor GPs—combined with domain-specific modeling choices (e.g., factorial or Kronecker structures for multidimensional random effects, covariate incorporation, process-based fusion). Model identification, selection (via AIC, cross-validation, DIC), inference about causal directionality, and posterior predictive checks are essential in practice.

Empirical studies demonstrate that joint spatial multivariate modeling consistently improves predictive and inferential performance over separate univariate or marginal models, especially in high-dimensional, heterogeneous, or mixed-type regimes (Cressie et al., 2015, Vu et al., 2020, Mukherjee et al., 15 Oct 2025, Wang et al., 2019, Frévent et al., 2024, Beltrán-Sánchez et al., 28 Jul 2025, Daw et al., 2023). These frameworks are critical in environmental prediction, disease mapping, socio-economic modeling, and survey inference.

---

For a survey of the conditional approach and its DAG representation, see (Cressie et al., 2015). For deep deformation frameworks, see (Vu et al., 2020). For scalable high-dimensional mixed-type inference, see (Mukherjee et al., 15 Oct 2025). For Kronecker and regionalization-based approaches, see (Daw et al., 2023, Zhang et al., 2020). For graphical Markov and cross-MRF sparsity, see (Chen et al., 2024, Eckardt et al., 2019). For function-valued point process analysis, see (Eckardt et al., 2023). For spatially constrained multivariate analysis, see (Dray et al., 2012).