Projections of epidemic transmission and estimation of vaccination impact during an ongoing Ebola virus disease outbreak in Northeastern Democratic Republic of Congo, as of Feb. 25, 2019

Worden, Lee; Wannier, Rae; Hoff, Nicole A.; Musene, Kamy; Selo, Bernice; Mossoko, Mathias; Okitolonda-Wemakoy, Emile; Muyembe Tamfum, Jean Jacques; Rutherford, George W.; Lietman, Thomas M.; Rimoin, Anne W.; Porco, Travis C.; Kelly, J. Daniel

The performance of our short-term prediction models can be assessed on the outbreak to date. The relatively simple auto-regression model we used performed more consistently on the Projections of an EVD outbreak in Northeastern DRC range of partial data sets used for scoring than the more complex stochastic simulation model did. The stochastic model has tended to produce tighter prediction distributions that are prone to extreme failure when they get it wrong, while the auto-regression model's predictions are more tolerant of unpredictable outcomes. Conversely, the stochastic model outperforms the auto-regression when it gets the prediction right. It should also be noted that because the stochastic simulation model is based on mechanistic knowledge of the transmission process generating the outbreak while the regression model is a purely statistical inference from past outbreaks, it may be that if conditions emerge that are substantially different from past outbreak conditions, the mechanistic model may produce sensible predictions where a purely statistical model fails. Because the other included models produce only final outbreak size projections, they can not be fully evaluated before the outbreak has ended. However, we can note that our early projections of final size (S9 and S10 Figs) fell below the counts that have been observed as of February. In other words, our forecasts based on early reports and on an assumption that past outbreaks can be used to forecast the present one were more optimistic than warranted by subsequent events. Events to come may shed light on whether this outbreak is qualitatively different than the past ones we have used to construct forecasts, perhaps due to the impact of conflict conditions on the outbreak.