The magnitude and pace of the current Ebola outbreak is unprecedented and requires tools to assess the future scope of the epidemic, as well as the efficacy of intervention tools and strategies. Unfortunately, our understanding of Ebola transmission dynamics is incomplete and data on the present outbreak are limited. Consequently, we present our forecasts as estimates, and cannot provide well-constrained certainties or likelihoods to any of the predicted outcomes.
Data: Ebola data for Guinea, Liberia, and Sierra Leone were compiled from World Health Organization’s Disease Outbreak News and Situation Reports . Total cases and deaths were used to train these model forecasts. The data include confirmed, probable, and suspected cases and may therefore decrease between measurements should some of the unconfirmed cases (i.e. the probable and suspected cases) be excluded after testing. Additionally, delays in reporting may lead to temporal imprecision of both incidence and mortality data and an underestimation of outbreak growth.
Methods: The model used to generate these forecasts contains a stochastic component that allows the force of transmission to vary through time. This variability is intended to emulate the spatial-temporal variability of Ebola transmission dynamics within country due to changes in intervention, containment and social practices. Three scenarios are forecast using the optimized model:
Data Revision: Cumulative incidence and mortality estimates, as provided by the WHO, underwent a sharp correction upon release of the October 29, 2014 WHO Situation Report. In that report, cumulative incidence rose precipitously (relative to the prior week) and cumulative mortality decreased in some countries. These changes occurred due to more detailed examination of patient records, which, for Liberia in particular, increased the proportion of confirmed Ebola cases. While cumulative incidence and mortality levels changed in that report, it remained unclear how to distribute those changes through time.
The newer reports, beginning with the November 12, 2014 WHO Situation Report, partially resolve this issue by providing a patient database. of weekly case counts (confirmed and probable) distributed through time for Guinea, Liberia and Sierra Leone. These data provide time series of case incidence for each country; however, as stated in the situation report, 'data for the most recent weeks are less complete [in the patient database] than in the weekly situation reports'. In addition, no distributed time series of weekly mortality have yet been posted.
For the forecasts posted here, we train our model using weekly case counts from the patient database; however, for the most recent
three four weeks we use data from the weekly situation reports instead. The weekly mortality counts were updated to equal the new cumulative mortality count while keeping the relative proportion of a given week constant. For example, before revision, if week w of the outbreak had n deaths that accounted for x% of cumulative mortality, after revision, week w still accounts for x% of the revised cumulative mortality but with m deaths
An archive of our forecasts made prior to the October 29, 2014 revision can be accessed here..
Retrospective 'no change' forecasts for Guinea made with these updated data tend to overestimate future cases, while the 'improved' scenario underestimate future cases. For Liberia, both the 'no change' and 'improved' forecasts have predicted too many future cases and deaths in past weeks, but the more recent 'improved' forecasts (last 4 weeks) are more in line with observed outcomes. For Sierra Leone, the 'improved' forecasts have been in line with observed outcomes whereas the 'no change' forecasts have predicted aggressive, spurious exponential growth. A number of factors may be contributing to this forecast error. Specifically, the model is a gross simplification of much more complicated spatially heterogeneous transmission processes, the majority of which are not represented. In addition, shifting data biases, in particular rates of under-reporting, may be corrupting model optimization. Alternately, intervention and control effectiveness may be on the rise. Discussion of these and other issues in relation to an alternate model structure can be found here.
Mean estimates of cumulative infections (upper-pane), cumulative mortality (middle-pane), and the number presently infectious (lower-pane) are shown below. The number infectious also represents the beds needed should all persons seek medical attention upon becoming symptomatic. The shaded region around the forecasts shows the interquartile range. A forecast horizon of 6 weeks is displayed. Hover on a data point to look at values. Use the Fit Cutoff slider at the top right to base estimates on a different observation cutoff date and Country selection box to see estimates for a different country. Use the Intervention selection box to see the forecasts under different scenarios. Use the Data tab for a table of forecast errors.
Last updated on December 17, 2014 using data reported through December 14, 2014 in Guinea and Sierra Leone and December 9 in Liberia.