Efficient and Flexible Hierarchical Algorithms Using the Nimble Software Package

• Activity Number: 565
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Daniel Turek*+
• Companies: University of California, Berkeley
• Keywords: Hierarchical models ; Numerical algorithms ; MCMC ; Statistical programming ; Statistical software

Abstract

In the Bayesian paradigm, hierarchical models are generally analysed using some flavour of MCMC. The available mainstream MCMC software packages (e.g. WinBUGS) function largely as black-box algorithms, and afford few options for algorithmic customization. As a case study, we consider capture-recapture models common in statistical ecology, in which the hierarchical formulation requires specifying a large number of latent states. Mainstream MCMC sampling of the parameters of interest is hampered by sampling of these latent states, which may lead to poor mixing and slow convergence. To improve performance, we consider use of the new NIMBLE software package. The modular nature of NIMBLE permits general combinations of numerical algorithms, such as MCMC, EM, or sequential MC. We demonstrate how a hybrid "MCMC-filtering" algorithm can be specified using NIMBLE, combining MCMC sampling for the parameters of interest and direct filtering for calculation of likelihood components. We present results using this MCMC-filtering algorithm on a capture-recapture dataset, and discuss the generality of algorithmic specification in NIMBLE.