Recent advances in interpretable model-based geostatistics for analyzing complex spatial data

737 

Topic-Contributed Paper Session 

Changwoo Lee (1), Michele Peruzzi (3), Toryn Schafer (2), Bokgyeong Kang (1), Changwoo Lee (1), Michele Peruzzi (3), Mary Lai Salvaña (4), Xiaotian Zheng (5)

(1) Duke University, N/A, (2) Texas A&M University, N/A, (3) University of Michigan, N/A, (4) University of Connecticut, Storrs, CT, (5) University of Wollongong, N/A

Toryn Schafer  
Texas A&M University

Michele Peruzzi  
University of Michigan

Changwoo Lee  
Duke University

Bokgyeong Kang  
Duke University

Changwoo Lee  
Duke University

Michele Peruzzi  
University of Michigan

Mary Lai Salvaña  
University of Connecticut

Xiaotian Zheng  
University of Wollongong

Model-based geostatistics is a subfield of spatial statistics that seeks to understand data through assumed underlying stochastic models. As spatial data become more complex – often large-scale, multivariate outcomes, and possessing intricate dependence structures – there is a growing need for statistical models that balance interpretability, flexibility, and scalability. These 3 ingredients are key to providing a richer understanding of spatial data and allowing for a broad range of applications.

The proposed session, consisting of 5 early career researchers, presents recent methodological advances to address unique challenges in model-based geostatistics. Attendees will gain exposure to new methodological developments, novel computational strategies, and models with better interpretability that can be applied across various domains, including but not limited to cancer proteomics, ecology, and seismology.

Titles and summaries:

1) Joint spatiotemporal modeling of zooplankton and whale abundance in a dynamic marine environment (Bokgyeong Kang)
We develop an innovative joint species distribution model that integrates a point pattern model for whale distribution and a geostatistical model for zooplankton abundance, linked through a latent conditional-marginal specification. By incorporating novel data fusion techniques across multiple data sources, our approach better captures species distributions than independent models. We demonstrate its effectiveness through simulations and apply it to Cape Cod Bay data.

2) Marginally interpretable spatial logistic regression with bridge processes (Changwoo Lee)
We propose a new class of spatial logistic models that preserves both population-averaged and site-specific interpretations. We show how the proposed process achieves interpretability and scalability with the scale mixture of normal representation and illustrate with childhood malaria prevalence data in Gambia.

3) Inside-out cross-covariance for spatial multivariate data (Michele Peruzzi)
We introduce a novel cross-covariance function for modeling multivariate spatial data. Our Inside-Out Cross-covariance (IOX) enables direct marginal inference, flexible covariance parameterization, easy prior elicitation, flexible dimension reduction, and the ability to handle outcomes with unequal smoothness. IOX leads to scalable models for noisy vector data as well as flexible latent models, complementing Vecchia approximations and related methods.

4) A spatial Bayesian causal network for cascading disasters modeling of seismic events (Mary Lai Salvana)
This study proposes a spatial extension to the naive Bayesian causal networks wherein the nodes are now treated as spatial random variables with certain spatial covariance structures. Rather than assume conditional independence given the parent nodes, the proposed spatial BCN leverages spatial information to model the dynamics between colocated and/or neighboring disasters.

5) Generalized chi-squared processes for spatial pairwise-outcomes (Xiaotian Zheng)
Considerably less attention has been devoted to constructing spatial processes indexed by pairs of locations. Such processes are essential for statistical inference from pairwise outcomes arising from comparisons between possible pairs of spatially-dependent ensembles. We propose a new class of generalized chi-squared processes and embed it within a hierarchical framework that enables model-based inference for incomplete data.

International Society for Bayesian Analysis (ISBA) ³

Section on Bayesian Statistical Science ¹

Section on Statistics and the Environment ²