Indexing and partitioning the spatial linear model for large data sets.

Area of Science:

• Geostatistics
• Spatial Statistics
• Computational Statistics

Background:

• Analyzing large spatial datasets presents computational challenges for standard geostatistical methods.
• Existing approaches like spatial basis functions and covariance tapering have limitations.
• Spatial indexing offers a promising alternative, akin to composite likelihood methods.

Purpose of the Study:

• To develop a unified framework for four key geostatistical goals using spatial indexing.
• To demonstrate the efficiency and accuracy of spatial indexing for large-scale spatial linear models.
• To extend spatial indexing to non-Euclidean data structures like stream networks.

Main Methods:

• Developed a spatial indexing framework to create block covariance structures and nearest-neighbor predictions.
• Utilized exact inference for fixed effects under the block covariance construction.
• Employed simulations to validate methods and compare performance against other popular techniques.

Main Results:

• Spatial indexing achieved accurate covariance parameter estimation, fixed effects inference, and kriging predictions.
• Spatially compact partitions (around 50 samples) and 50 nearest neighbors were found to be optimal.
• The method demonstrated appropriate confidence and prediction interval coverage.
• Significant speed improvements were observed, reducing analysis time from days to minutes for large datasets.

Conclusions:

• Spatial indexing offers a computationally efficient and accurate solution for large-scale spatial linear modeling.
• The framework effectively handles point and block kriging, fixed effects estimation, and covariance parameter estimation.
• Its ability to extend to non-Euclidean topologies like stream networks highlights its versatility and broad applicability.