Forecasting and Uncertainty in Modeling the 2014-2015 Ebola Epidemic in West Africa

Abstract

The 2014-2015 Ebola epidemic in West Africa is the largest ever recorded, with over 27,000 cases and 11,000 deaths as of June 2015. The public health response was challenged by difficulties with disease surveillance (particularly in more remote regions), which impacted subsequent analysis and decision-making regarding optimal interventions. We developed a stage-structured model of Ebola virus disease (EVD). A key feature of the model is that it includes a generalized correction term accounting for factors such as the fraction of cases reported and fraction of the population at risk (e.g. due to contact patterns, interventions, spatiotemporal spread, pre-existing immunity, asymptomatic cases, etc.). We generated a range of short term forecasts for Guinea, Liberia, and Sierra Leone, which we then validated using subsequent data. We also used the model to examine the uncertainty in the relative contributions to transmission by the different stages of infection (early, late, and funeral). We found that a wide range of forecasted trajectories fit approximately equally well to the early data. By including the correction factor term, however, the best-fit models correctly forecasted EVD activity for all three countries, both individually and for all countries combined. In particular, the models correctly forecasted the slow-down and stabilization in Liberia, as well as the continued exponential growth in Sierra Leone through November 2014. Parameter unidentifiability issues hindered estimation of the relative contributions of each stage of transmission from incidence and deaths data alone, which poses a challenge in determining optimal intervention strategies, and underscores the need for additional data collection. Even with these limited data, however, it is still possible to accurately capture and predict the epidemic dynamics by using a simplified correction term that approximately accounts for the complex underlying factors driving disease spread.

Model distinguishability and inference robustness in mechanisms of cholera transmission and loss of immunity

“…This is to be expected in the very early part of the epidemic, during the exponential growth phase, as during this phase the epidemic is largely linear on a log-scale and so can be explained by only two parameters. This idea is borne out in other real-time forecasting efforts-for example, efforts to forecast the trajectory of the 2014 Ebola epidemic in West Africa met with difficulty [62][63][64][65]. We note that our results illustrate how agreement across a range of models may not guarantee accuracy; however once the epidemic peak was observed, all models tended to converge on similar and more accurate forecasts.…”