Online Program Home
  My Program
Legend:
CC = Baltimore Convention Center,    H = Hilton Baltimore
* = applied session       ! = JSM meeting theme

Activity Details


Register
CE_20C Mon, 7/31/2017, 8:30 AM - 5:00 PM H-Holiday Ballroom 4
Bayesian Modeling and Inference for High-Dimensional Spatial-Temporal Data (ADDED FEE) — Professional Development Continuing Education Course
ASA , Section on Bayesian Statistical Science
Bayesian hierarchical models have been widely deployed for analyzing spatial and spatio-temporal datasets commonly encountered in the fields of public health, forestry, ecology and climate sciences. However, with rapid developments in Geographic Information Systems (GIS) and related technologies, statisticians and data analysts frequently encounter massive spatial and spatio-temporal data that cannot be analyzed using these traditional approaches due to their heavy computing demands. In this course, we will present scalable Bayesian models and related estimation methods that can provide fast analysis of big spatial and spatio-temporal data using modest computing resources and standard statistical software environments such as R. We will begin with an introduction to the common types of geo-referenced spatial data, discuss easily available software packages for exploratory data analysis and subsequently provide an overview of statistical methods used to analyze them. We will briefly cover exploratory data analysis techniques like variogram fitting, basics of geo-statistical approaches like kriging, and Gaussian Processes. We will then highlight some of the computational issues experienced by Gaussian Process models when used to model large spatial data. In this context, we will present new scalable Bayesian models that can deliver fully model based inference for massive spatial data. Our discussion will focus on the newly developed Nearest Neighbor Gaussian Processes (NNGP) that yields massive computation gains and yet provide rich Bayesian inference for analyzing large univariate and multivariate spatial data. We will also present a comparative assessment of other related methods and strategies for large spatial data including low-rank models. Subsequently, we will move to the analysis of large spatio-temporal data. We will present Bayesian models and computational resources for analyzing massive spatio-temporal data sets including continuous-space, discrete-time models as well as continuous-space, continuous-time models that account for space-time interaction and can be used to interpolate at arbitrary spatial and temporal scales. We will also cover Bayesian models for large regionally aggregated (also referred to as areal) spatial and spatio-temporal data. Specifically, we will adapt Markov Random Fields based approaches such as the popular conditional autoregressive (CAR) models as well as some recent scalable alternatives based on directed acyclic graphs. Our discussion will also feature advanced topics like nonstationary models for large spatial data as well as MCMC-free Bayesian models for ultra high dimensional spatial data. We will demonstrate very practical implementation of these models using newly developed software in the form of packages in R and also in languages and Bayesian modeling environments such as BUGS, JAGS and Stan. While the course will primarily focus upon practical modeling, computing and data analysis, short course participants will benefit from some understanding of mathematical statistics and linear algebra at the undergraduate or advanced undergraduate level. We will not assume any significant previous exposure to spatial or spatiotemporal methods or Bayesian inference, although students with basic knowledge of the area will certainly face a gentler learning curve. All the computational tools and environments will also be introduced as necessary in the course.
Instructor(s): Abhirup Datta, Johns Hopkins University, Sudipto Banerjee, UCLA Fielding School of Public Health, Andrew O. Finley, Michigan State University
 
 
Copyright © American Statistical Association