Estimating undetected Ebola spillovers

Abstract

The preparedness of health systems to detect, treat, and prevent onward transmission of Ebola virus disease (EVD) is central to mitigating future outbreaks. Early detection of outbreaks is critical to timely response, but estimating detection rates is difficult because unreported spillover events and outbreaks do not generate data. Using three independent datasets available on the distributions of secondary infections during EVD outbreaks across West Africa, in a single district (Western Area) of Sierra Leone, and in the city of Conakry, Guinea, we simulated realistic outbreak size distributions and compared them to reported outbreak sizes. These three empirical distributions lead to estimates for the proportion of detected spillover events and small outbreaks of 26% (range 8-40%, based on the full outbreak data), 48% (range 39-62%, based on the Sierra Leone data), and 17% (range 11-24%, based on the Guinea data). We conclude that at least half of all spillover events have failed to be reported since EVD was first recognized. We also estimate the probability of detecting outbreaks of different sizes, which is likely less than 10% for single-case spillover events. Comparing models of the observation process also suggests the probability of detecting an outbreak is not simply the cumulative probability of independently detecting any one individual. Rather, we find that any individual's probability of detection is highly dependent upon the size of the cluster of cases. These findings highlight the importance of primary health care and local case management to detect and contain undetected early stage outbreaks at source.

Fig 1. Diagram of stages of analysis (solid lines and numbered steps at right) and data inputs at each stage (italics and dashed lines).

Parallel analyses were conducted for data from the full outbreak across West Africa, in Sierra Leone only (Western Area), and in Guinea only (Conakry).

Fig 2

A. Estimated probabilities of cluster detection, by cluster size and observation function. The geometric observation function (green) is based on the cumulative distribution function of a geometric distribution (with a single parameter), while the logistic function (orange) is a generalized logistic function with two parameters. B. Ranges, 95% quantiles, interquartile ranges, and medians of AICc values for observation models applied to each dataset, by linking function used.

Fig 3

A. Ranges of secondary infection distributions of all three datasets. B. Estimated numbers of all small outbreaks and spillover events (solid) and single-case spillover events (white) for each dataset. The dashed line indicates all observed EVD outbreaks and the dotted line indicates all observed single-case spillover events. C-E. Medians and ranges of estimated true (observed and unobserved) small EVD outbreaks by outbreak size for the full West African outbreak (C), Western Area, Sierra Leone only (D), and Conakry, Guinea only (E). The colored areas indicate ranges of from all 500 parameter sets. The dashed lines and grey areas indicate estimated probabilities of observing an outbreak of each size (median and full ranges from 500 parameter sets). Points marked with ‘X’ indicate the number of observed outbreaks of each size. Outbreaks of 15 cases or larger are pooled.