Hierarchical Bayesian Functional Data Models with Model-Based Time-Frequency Predictors

• Activity Number: 41
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Wen-Hsi Yang*+ and Christopher Wikle and Scott Holan and Mark Wildhaber
• Companies: University of Missouri-Columbia and University of Missouri and University of Missouri and U.S. Geological Survey
• Keywords: Bayesian hierarchical models ; Functional data ; Time-frequency

Abstract

Time-frequency analysis has become increasingly recognized as an important component in many scientific investigations. In this domain, crucial aspects of the underlying process of interest often become apparent that would otherwise go undetected. Moreover, time-frequency representations can often serve as powerful predictors in modeling complex processes. In order to utilize these representations within a statistical modeling framework, we propose a class of fully hierarchical Bayesian time-frequency functional data models that incorporate time-frequency representations as predictors. Importantly, our approach extracts features of the original signal that improve prediction rather than explaining variation. Finally, we illustrate the effectiveness of our approach through simulation and by applying our model to predict spawning success of shovelnose sturgeon.