.: Prof. Jean D. Opsomer, Department of Statistics, Colorado State University :.
.: Bayesian Semiparametric Regression and Prediction for Circular Data :.
Lunedì 9 Giugno, ore 12
Dipartimento di Scienze Politiche, Aula 201
Università degli Studi di Perugia
Abstract.
The impetus for this work was the need to predict time-of-day distributions for departures of recreational anglers from fishing sites along the East coast of the US, based on a very large survey dataset of fishing trips. After explaining the background of this application, regression models based on the projected normal distribution are developed, and Bayesian estimation and prediction are considered. We first describe simple parametric models and then expand this to semiparametric mixtures of models, to account for observed heterogeneity of trips. Because of the size of the dataset, traditional Markov Chain Monte Carlo methods are impractical here, and are replaced by variational Bayes, a relatively new and promising class of tools to obtain numerical approximations to posterior distributions. Along the way, we develop a number of improvements to variational methods that are more generally applicable. Finally, we describe the application of these methods to the original survey dataset.
