A compound Dirichlet-Multinomial model for provincial level Covid-19 predictions in South Africa

De Waal, A.; De Waal, D.

Accurate prediction of COVID-19 related indicators such as confirmed cases, deaths and recoveries play an important in understanding the spread and impact of the virus, as well as resource planning and allocation. In this study, we approach the prediction problem from a statistical perspective and predict confirmed cases and deaths on a provincial level. We propose the compound Dirichlet Multinomial distribution to estimate the proportion parameter of each province as mutually exclusive outcomes. Furthermore, we make an assumption of exponential growth of the total cummulative counts in order to predict future total counts. The outcomes of this approach is not only prediction. The variation of the proportion parameter is characterised by the Dirichlet distribution, which provides insight in the movement of the pandemic across provinces over time.