Comparing the performance of SARIMA and dynamic linear model in forecasting monthly cases of mumps in Hong Kong

Abstract

Background To provide a reliable forecast of a disease is one of the main purpose of public health surveillance system. Basic information obtained from data collection can provide the nature knowledge of and the history pattern of a disease. In public health surveillance system, a lot of data are time series, especially for infectious diseases. SARIMA method and DLM method are both applicable tools for time series data analysis. Hong Kong has a relative low mumps prevalence. And the prevalence followed an increasing trend until 2004and kept stable after 2006. However, outbreaks may be also occurred occasionally in developed countries.

Method This paper constructs SARIMA models and DLM models of monthly cases of mumps in Hong Kong based on 7 different modeling periods respectively. Then these models were used to predicting the mumps cases in each corresponding forecasting period. The forecasting performance of SARIMA models and DLM models are compared with visualization of the predicting values and three forecasting error measures: MAD, MSE, and MAPE. A forecasting of mumps cases during 2013. 07 and 2014.06 will be made with the method with better forecasting performance of mumps cases in Hong Kong

Result For intervals 2009. 01 to 2009. 02, 2011. 01 to 2011. 12, and 2012. 01 to 2012. 12, the forecasts of DLM models have smaller forecasting error measures and are more closely to the real observed values. And the visualization predicting values of SARIMA and DLM models are closely for forecasting intervals 2008 and 2010, where SARIMA forecasts own smaller forecasting error measures. Compare with that based on fitting period 1997 to 2012, the forecasts obtained by the SARIMA model based on fitting period 2006 to 2012 are more close to the real observations. Both SARIMA models and DLM models based on fitting period 1997 to 2003 underestimate the observed value of 2004. 05 to 2004. 12.

Conclusion DLM modeling method presents a better performance on forecasting the monthly cases of mumps in Hong Kong. And DLM method is more appropriate to be applied on the analysis of time series with count data and the research of diseases with small counts. And both SARIMA and DLM method are appropriate for analyses based on long time trend. But they are not appropriate to be applied as short time monitor tools. From the result of time series decomposition analysis result the mumps cases had a seasonal pattern, and shows that between July and the next January, the seasonal impact will contribute to the increase of case number of mumps. So it is highly suggest to recommend people under risk to practice more prevention measures to protect them against mumps infectious during that period.