B1601
Title: Fitting higher order state-space models using hidden Markov methodology
Authors: Takis Besbeas - Athens University of Economics and Business (Greece) [presenting]
Abstract: Count time series are frequently encountered in a variety of scientific disciplines, including ecology, biology and public health. In addition to autocorrelation, which may exceed order one, overdispersion and zero-inflation may be present in such a series. To accommodate these features, a number of researchers have proposed a flexible class of dynamic models in the state-space framework coupled with Monte Carlo Expectation Maximization (MCEM) algorithms based on the particle filter for parameter estimation. We propose a new method for model fitting based on hidden Markov model methodology. The method involves a discretisation technique of the underlying state-space together with an approach for transforming a higher-order state-space into an equivalent first-order. The proposed approach is simpler to both implement and compute, and opens the way to efficient model selection. We illustrate the practical utility of the method using an application from public health pertaining to the diagnosis coding of severe disease.
