Hierarchical factor models for large spatially misaligned data: a low-rank predictive process approach.

Area of Science:

• Environmental statistics
• Geostatistics
• Statistical modeling

Background:

• Analyzing large-scale geographically referenced data presents challenges in modeling associations and spatial dependencies.
• Spatial misalignment, where variables are not observed across all locations, complicates joint modeling.
• High dimensionality in both outcome vectors and spatial locations necessitates dimension reduction techniques.

Purpose of the Study:

• To propose a class of hierarchical low-rank spatial factor models for jointly modeling numerous geographically referenced outcomes.
• To capture associations among variables and the strength of spatial association for each variable.
• To address spatial misalignment and reduce dimensionality in large spatial datasets.

Main Methods:

• Development of hierarchical low-rank spatial factor models merging latent variable and low-rank spatial process concepts.
• Stochastic selection of latent factors without complex algorithms, using identifiability characterizations.
• Implementation of a Markov chain Monte Carlo algorithm for estimation, handling spatial misalignment and missing values within a Bayesian predictive framework.

Main Results:

• The proposed model effectively captures variable associations and spatial dependencies in large, high-dimensional spatial data.
• The methodology successfully addresses spatial misalignment and enables recovery of missing data distributions.
• Demonstrated utility through simulation experiments and analysis of air pollutant data in California.

Conclusions:

• Hierarchical low-rank spatial factor models provide a flexible and computationally feasible framework for complex spatial data analysis.
• The developed Bayesian approach effectively estimates model parameters and handles missing data in spatially misaligned settings.
• The methodology offers a robust tool for environmental statistics and geostatistics, particularly for large-scale spatial outcome modeling.