A1633
Title: Efficient Bayesian computation for modeling dynamic counts
Authors: Yuko Onishi - Mitsubishi UFJ Morgan Stanley (Japan) [presenting]
Kaoru Irie - University of Tokyo (Japan)
Shonosuke Sugasawa - Keio University (Japan)
Abstract: Dynamic count data frequently appear in many scientific fields including finance, genomics, and social science. Although state space models based on the Poisson distribution are widely used, it is difficult to efficiently compute the posterior distribution due to the lack of conjugacy of normal distribution for the Poisson rate parameter even when the data is univariate. An efficient Bayesian computation method is proposed for multivariate Poisson state space models based on data augmentation. We use the negative binomial distribution as an approximation of the Poisson distribution and employ Polya-gamma data augmentation which enables us to compute the state variables efficiently. The proposed method is demonstrated through simulation and empirical studies.