Dynamic linear model and SARIMA: a comparison of their forecasting performance in epidemiology

One goal of a public health surveillance system is to provide a reliable fo recast of epidemiological time series. This paper describes a study that us ed data collected through a national public health surveillance system in t he United States to evaluate and compare the performances of a seasonal aut oregressive integrated moving average (SARIMA) and a dynamic linear model ( DLM) for estimating case occurrence of two notifiable diseases. The compari son uses reported cases of malaria and hepatitis A from January 1980 to Jun e 1995 for the United States. The residuals for both predictor models show that they were adequate tools for use in epidemiological surveillance. Qual itative aspects were considered for both models to improve the comparison o f their usefulness in public health. Our comparison found that the two fore casting modelling techniques (SARIMA and DLM) are comparable when long hist orical data are available (at least 52 reporting periods). However, the DLM approach has some advantages, such as being more easily applied to differe nt types of time series and not requiring a new cycle of identification and modelling when new data become available. Copyright (C) 2001 John Wiley & Sons, Ltd.