4295.0 - Bayesian spatial statistical modeling on the cloud: a web tool for analyzing geospatial HIV datasets at small-area levels
Program: Applied Public Health Statistics
Author: Yusuf Ransome
See all authors and presenters →
Abstract
The Bayesian approach has been increasingly applied to analyze geospatial health datasets at small-area levels (e.g., census tracts and zip codes). Both posterior sampling (e.g., Markov chain Monte Carlo, MCMC) and approximation (e.g., Integrated Nested Laplace Approximation, INLA) algorithms have been developed to implement Bayesian models through software and tools such as WinBUGS, Stan, NIMBLE, and R-INLA. While these software and tools make Bayesian statistical modeling much more accessible than two decades ago, its application in analyzing small-area level geospatial health data is still limited to researchers with expertise in programming such as R, Python, and BUGS. An interactive platform that supports Bayesian statistical modeling, especially for spatial and spatiotemporal datasets and in the Web setting, is still lacking. Our study fills this gap by developing a Web R-shiny tool for implementing Bayesian spatial and spatiotemporal models. Models fitted via both MCMC (i.e., NIMBLE) and INLA are supported in the tool, with the former more flexible in addressing issues such as data censoring while the latter more computationally efficient. The tool leverages the cloud computing paradigm for Bayesian model computation, a feature especially benefitting MCMC-based models by making model convergence more time-efficient. The tool is used for analyzing publicly accessible geospatial HIV datasets from AIDSVu.org and is highly responsive to the Ending the HIV Epidemic in the U.S. (EHE) initiative. It will be distributed to and used by different stakeholders including local health departments at jurisdictions prioritized by EHE for reducing HIV infections and improving health equity.