|
Activity Number:
|
137
|
|
Type:
|
Invited
|
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistics and the Environment
|
| Abstract - #302954 |
|
Title:
|
An Adaptive Predictive Process Modeling Approach for Large Spatial-Temporal Data Sets
|
|
Author(s):
|
Andrew Finley*+ and Sudipto Banerjee and Alan E. Gelfand
|
|
Companies:
|
Michigan State University and The University of Minnesota and Duke University
|
|
Address:
|
Department of Forestry, East Lansing, MI, 48824,
|
|
Keywords:
|
spatial-temporal ; MCMC ; predictive process ; Gaussian process ; reversible jump ; Big-N problem
|
|
Abstract:
|
We consider the setting where the number of locations yielding observations is too large to fit the desired hierarchical spatial or spatial-temporal random effects models using Markov chain Monte Carlo methods. This problem is exacerbated in spatial-temporal and multivariate settings where many observations occur at each location. Sacrificing model richness, especially in second order model properties, is undesirable when quantifying and propagating uncertainty is of central interest. Using several large spatial data sets that exhibit complex dependence structures we illustrate the use of an adaptive predictive process to maintain the richness of desired hierarchical modeling specifications. The predictive process is knot-based leading to questions regarding knot design, which we address through optimal allocation and placement of knots in space and time.
|
