NONLINEAR PREDICTIVE LATENT PROCESS MODELS FOR INTEGRATING SPATIO-TEMPORAL EXPOSURE DATA FROM MULTIPLE SOURCES.

Area of Science:

• Environmental Epidemiology
• Statistical Modeling
• Geospatial Analysis

Background:

• Accurate spatio-temporal prediction of environmental exposures is crucial in epidemiology.
• Traffic-related particle exposure studies often involve multiple data sources with varying spatial and temporal resolutions.
• Existing methods can be computationally intensive when dealing with nonlinear relationships and aggregated data.

Purpose of the Study:

• To develop an efficient Bayesian hierarchical joint model for spatio-temporal exposure prediction.
• To address computational challenges arising from nonlinearities and differing temporal scales in data sources.
• To improve the spatial coverage and reduce prediction error for environmental exposures like black carbon.

Main Methods:

• A Bayesian hierarchical framework with a joint model comprising submodels for each data source.
• Linearization of nonlinear components with respect to a latent process.
• Induction of sparsity in the latent process covariance matrix using compactly supported covariance functions.
• Development of an efficient Markov Chain Monte Carlo (MCMC) scheme tailored to the approximations.

Main Results:

• The proposed model effectively pools information from multiple sources, enhancing data coverage.
• Approximations and the efficient MCMC scheme make complex spatio-temporal models computationally tractable.
• The model successfully addresses temporal change of support problems, integrating daily and multiday data.

Conclusions:

• The developed Bayesian joint model offers an efficient and accurate approach for spatio-temporal exposure prediction.
• This methodology is particularly beneficial for environmental epidemiology studies involving traffic-related particles and diverse data.
• The study demonstrates the utility of the model in pooling data to maximize spatial coverage and minimize prediction error.