Towards realtime spatiotemporal prediction of district level meningitis incidence in sub-Saharan Africa

Within an area of sub-Saharan Africa termed 'the meningitis belt', meningococcal meningitis epidemics are a major public health concern. The epidemic control strategy that is currently utilized is reactive, such that a vaccination programme is initiated in a district once a predefined weekly incidence threshold has been exceeded. We report progress towards the development of an early warning system based on statistical modelling of district level weekly incidence data. Four modelling approaches are considered and their forecasting performances are compared by using weekly epidemiological data from Niger for the period 1986-2007. We conclude that the models under consideration are advantageous in different situations. The three-state Markov model described in which observed incidence is categorized according to policy-defined thresholds gives the most reliable short-term forecasts, whereas the dynamic linear model proposed, using log-transformed weekly incidence as the response variable, gives more reliable predictions of annual epidemics.