Dynamic modeling for persistent event-count time series

We present a method for estimating event-count models when the data is generated from a persistent time-series process. A Kalman filter is used to estimate a Poisson exponentially weighted moving average (PEWMA) model. The model is compared to extant methods (Poisson regression, negative binomial regression, and ARIMA models). Using Monte Carlo experiments, we demonstrate that the PFWMA provides significant improvements in efficiency. As an example, we present an analysis of Pollins (1996) models of long cycles in international relations.