Analyzing count data using a time series model with an exponentially decaying covariance structure

Count data appears in various disciplines. In this work, a new method to analyze time series count data has been proposed. The method assumes exponentially decaying covariance structure, a special class of the Mat\'ern covariance function, for the latent variable in a Poisson regression model. It is implemented in a Bayesian framework, with the help of Gibbs sampling and ARMS sampling techniques. The proposed approach provides reliable estimates for the covariate effects and estimates the extent of variability explained by the temporally dependent process and the white noise process. The method is flexible, allows irregular spaced data, and can be extended naturally to bigger datasets. The Bayesian implementation helps us to compute the posterior predictive distribution and hence is more appropriate and attractive for count data forecasting problems. Two real life applications of different flavors are included in the paper. These two examples and a short simulation study establish that the proposed approach has good inferential and predictive abilities and performs better than the other competing models.