The impact of control strategies and behavioural changes on the elimination of Ebola from Lofa County, Liberia

1. Introduction

The epidemic of Ebola virus disease (EVD) in West Africa, has, at the time of writing, caused more than 11 000 deaths reported from more than 28 000 cases [1]. The outbreak was first identified in March 2014 in South Western Guinea [2–4]. By 30 March 2014, the infection had spread across the border, and the first cases were reported from Liberia, more specifically from Lofa County near the northern border with Guinea. Following this first small cluster of cases, in early June, a large outbreak was observed in Lofa, with rapid growth in case numbers between early July and late August. The outbreak peaked in August before declining in September and decreasing to zero cases of EVD by early November [5,6].

Ebola spreads through contact with bodily fluids that continue to be infectious after death [7]. This makes finding and isolating infected patients and preventing transmission at funerals crucial components of the public health response [8]. Ebola treatment centres (ETC) are specifically designed to provide isolation beds for infected patients, with staff trained to minimize the risk of onward transmission to themselves or others by wearing protective gear and practising the highest levels of hygiene. Infection at funerals can be prevented by conducting burials where only those wearing appropriate equipment touch the body of the deceased.

The exponential rise in case numbers in the summer of 2014 in Lofa County and elsewhere triggered a large-scale international response of humanitarian assistance to West Africa. In Lofa County, this led to a rapid expansion in ETC bed capacity, from 10 beds in June 2014 to about 100 by mid-August. Beyond its humanitarian contribution through the provision of appropriate care for a greater number of patients, providing a sufficient number of beds can contribute to curbing an epidemic by removing cases from the community and isolating them, thereby preventing further spread.

As much of transmission of Ebola occurs in (not necessarily nuclear) families, communities and at funerals, behaviour change can have a great impact on the course of an outbreak. This makes it important to understand the cultural contexts and processes that can trigger such change as well as its impact on the transmission [9]. Experiences in earlier outbreaks of Marburg virus disease in Angola [10] and EVD in Uganda [11] and in Sierra Leone during the current outbreak [12], emphasize the importance of health promotion interventions, in particular the need to understand the local context and culture, and work closely with community leaders to adapt the intervention to the setting [13,14]. World Health Organization (WHO) Standard Operating Procedures for controlling Ebola and Marburg virus epidemics highlight the importance of a comprehensive, integrated approach [15]. Around the same time as the expansion of ETC bed capacity, a suite of community-based health promotion and social mobilization activities were launched in Lofa County. To what extent these activities and ETC bed expansion contributed to end the epidemic is unclear.

Mathematical modelling of EVD epidemics has previously provided insights into the different routes of Ebola transmission [16,17] and the potential role of different control strategies, including interventions based on case management and safe handling of dead bodies as well as community-based interventions [18–23]. A recent modelling study has suggested that the spread of awareness and consequent self-initiated behavioural changes may have played a substantial role in turning the epidemic in Lofa County [24], but did not include a full account of the public health response. Another study has suggested that community transmission of Ebola in Liberia was reduced by 40–80% over the course of the outbreak. The provision of hospital beds, on the other hand, has been estimated to have averted more than 50 000 cases in Sierra Leone [25]. A recent study has found that control of the Ebola epidemic in Liberia was achieved through reductions in community transmission [26].

Here, we aim to characterize the epidemiological features of, and the public health response to, the EVD outbreak in Lofa County in 2014, and assess the role of the multiple components of intervention in the control of the outbreak. To tease out the impact of different clinical and community-based interventions, and to estimate the relative contribution of these to containing the epidemic in Lofa County, we fitted a mathematical model of Ebola virus transmission with a time-varying transmission rate to the series of ETC admissions. This allowed us to fully account for the progression of EVD in infected individuals and associated delays, and thus to pinpoint when declines or increases in transmission rates must have occurred to produce the observed time series of cases.

2. Methods

(b) Transmission model

A stochastic compartmental model of Ebola virus transmission was constructed and parametrized using detailed information obtained from the line list of ETC admissions. The model extended the classic susceptible–exposed–infectious–recovered (SEIR) model as follows: at time t, susceptibles (S) experience a (time-varying) force of infection λ(t). Upon infection, susceptibles move into the exposed (E) compartment, representing those infected but not yet infectious. After the incubation period, people in the exposed class move into one of two infectious (I) classes (with rate ρ), depending on whether they are healthcare seeking (I_H, for (H)ospital, where these individuals may ultimately arrive) or not (I_C, for (C)ommunity, where these individuals will remain). With time-varying probability h_t, individuals leaving the E compartment enter the I_H compartment where they remain in the community until they get admitted to the ETC (H_R or H_D, with time-varying rate κ_t), recover (R_C) or die (D_H, both with rate γ) and get a safe burial (B_H). With (time-varying) probability (1 − h_t), they enter the I_C compartment where they remain until they recover (R_C) or die (D_C, both with rate γ) and get an unsafe burial (B_C, at rate α). Individuals from the I_H compartment who get admitted to the ETC are split into those (H_D) who eventually die (D_H) and get a safe burial, and those (H_R) who recover (R_H), with the relative probabilities of entering the H_D or H_R compartments given by the case fatality rate c. Hospital admissions (into the H_R and H_D compartments) occur with multiplicative gamma-distributed noise (mean 1, standard deviation 0.1), a convenient choice for a rate that is subject to random variation [30,31]. A schematic illustration of the model is given in figure 1.

• Download figure
• Open in new tab
• Download PowerPoint

In the model, when infectious people are split into those that seek healthcare (I_H) and those that do not (I_C), the ones that would seek healthcare (i.e. are in I_H) but do not get to the ETC (because they die or recover before) are safely buried. The ones who would not seek healthcare (i.e. are in I_C) are not safely buried and remain infectious (in D_C) until they are buried (B_C). While we did not have any data to inform this choice, the rationale is that these might be part of resistant communities who would not call for a safe burial. Note that, as healthcare seeking changes during the course of the outbreak, so does the proportion of dead that is safely buried.

Because we had reports that no patient arriving at the ETC had ever been rejected in Lofa, we did not include any effects of ETC capacity and occupancy in fitting the model to the epidemic trajectory. When considering counterfactual scenarios later, however, we did include one where ETC capacity was limited and patients rejected if it was full (see section ‘Counterfactual scenarios’ below).

Because a focus of this work was on timing, we used the patient data to parametrize the model with realistic transition time distributions. Transition times in standard compartmental models can be made to follow gamma distributions with integer shape parameters, that is Erlang distributions, by putting compartments in sequence [32]. We fitted gamma distributions to the observed distributions of the admission delay, the length of stay of survivors and the time from onset to death of casualties, with rounded-down and rounded-up shape parameters using maximum-likelihood estimation with the R [33] package fitdistrplus [34], and chose the one with the greater likelihood as an estimate of the shape parameter of the Erlang distribution corresponding to each transition. The compartments were then split into the number of subcompartments given by the shape parameter, and transition rate multiplied with the same shape parameter, to recover the correct mean transition times. Where individuals were subject to multiple Erlang-distributed transitions (in I_H), they would have an index for each of these transitions. The index with respect to any transition was maintained under other transitions such that, for example, people admitted to the ETC late in their infectious period would proceed to death quicker than those who were admitted early in the infectious period.

We assumed that the symptomatic period coincided with the infectious period, as suggested by a study of a previous Ebola outbreak where infection was found to have been transmitted by ill patients both in early and late stages of illness [35]. We further assumed that infectiousness was the same for those who proceeded to recovery or death, and did not change after death until buried (unless safely buried), the simplest assumption in the absence of any data to inform this matter. We did, however, consider a modified model where transmissibility increased late in the infectious period and in an unsafe funeral when calculating the effect of isolation on the reproduction number, as it has been suggested that patients might become more infectious during later stages of their infection [19,35].

We included the time-varying admission delay by weekly varying the rate of admission κ_t according to the observed mean (figure 2). The case fatality rate was extracted from the line list as 67%, and the population size of Lofa County set to N = 270 114 [36]. The incubation period was set to 9.4 days with shape parameter 2 [37]. The average period of infectiousness after death for those unsafely buried was taken to be 1 day [17]. All delays and their corresponding parameters are summarized in table 1.

• Download figure
• Open in new tab
• Download PowerPoint

Table 1.Delays in the model, the corresponding rate parameter, the mean of the delay (the reciprocal of the rate) and the shape parameter of the Erlang distribution governing transitions. The parameter means are given in days.

Collapse

delay	rate parameter	mean	shape	source
incubation period	ρ	9.4	2	WHO Ebola response team [37]
infectious period	γ	7.8	3	data
admission	κt	variable	2	data
ETC stay (survivors)	θ	13.0	10	data
unsafe burial	α	1	1	assumption

In order to track changes in transmission owing to behavioural or other changes in the community, we assumed that both the proportion seeking healthcare and the transmission rate could change over time; the extent and direction of change was estimated during the model fitting process. Healthcare seeking was modelled as a sigmoid function,

2.1

where h_early and h_late are the proportion of people seeking healthcare early and late in the epidemic, respectively, τ is an inflection point of the sigmoid and ε is the speed of change. This functional form ensured that healthcare seeking was either monotonically increasing or decreasing, as we would expect attitudes to healthcare to either improve or worsen during the course of the outbreak as a result of behavioural changes and health promotion activities, but not undergo rapid jumps.

The time-varying transmission rate β_t, on the other hand, was assumed to be more likely to undergo sudden changes if there were, for example, superspreading events or spatial or other heterogeneities in transmission. We therefore modelled it as a weekly varying random variable that followed a normal distribution with given mean and standard deviation σ

2.2

Assuming that no transmission occurs in an ETC or during a safe burial, the force of infection at any time is

2.3

Assuming that there was no depletion of susceptibles (S ≈ N), the time-varying community reproduction number (i.e. the average number of infections caused by an infected person who does not seek healthcare) is

2.4

the two parts of the sum representing transmission before and after death. For those seeking healthcare and isolated (and, if necessary, safely buried), this changes to approximately:

2.5

An exact calculation of this reproduction number calculated using a next-generation matrix [38] is given in the electronic supplementary material. It deviates slightly from the one in equation (2.5) because of the discrete nature of the multiple sequential compartments introduced to obtain Erlang-distributed transmission times.

The ratio between the two reproduction numbers is approximately

2.6

We calculated a range of values for this ratio using bootstrapped parameter values estimated from the line list of cases and using the exact formula for R_H given in the electronic supplementary material.

The overall reproduction number at any time is

2.7

Incident ETC admissions were aggregated at a weekly level, and reporting assumed to follow a Poisson distribution centred around weekly incidence. The likelihood of each trajectory was obtained by multiplying the probability densities of observation at each time point, and the likelihood of a given set of parameter values was estimated using a particle filter [39]. The model was fitted to the time series of ETC admissions of confirmed and probable EVD cases using a Bayesian approach. Metropolis–Hastings Markov chain Monte Carlo in combination with the estimated likelihoods from the particle filter (a method called pMCMC) was used to sample parameter combinations and state trajectories from the posterior distribution of the parameters [40]. The filtered trajectories of the transmission rate β_t could then be used to track the behaviour of the transmission rate over time [31,41], and we used this to calculate the corresponding time-varying overall reproduction number (the average number of secondary cases generated by each case). Generally, if at any time the overall reproduction number is greater than 1, case numbers will typically rise whilst if it is less than 1, case numbers will generally decline.

The initial conditions and the parameters of the time-varying rates were estimated during the fitting process: the number of infectious people in Lofa County on 26 May 2014 (one week before the beginning of the time series of ETC admissions), the mean and standard deviation of the transmission rate, and the four parameters governing the proportion of the population seeking healthcare (h_early, h_late, ε and τ).

The stochastic elements of the model were updated at every week-long time step, and the model run as ordinary differential equations between time steps. The fitting procedure was implemented using libbi [42] with the rbi [43] and rbi.helpers [44] R packages on a high-performance computing cluster. After affirming that the method could recover the time-varying transmission rate from simulated data, we fitted the model using 100 parallel chains of 10 000 iterations each in the Markov chain Monte Carlo sampler. Convergence and mixing was confirmed visually. All code and data are available online at http://github.com/sbfnk/ebola_lofa, and instructions for reproducing the modelling results given in the electronic supplementary material.

3. Results

Between 18 March and 9 November 2014, a total of 466 confirmed and probable cases of EVD were identified (453 confirmed, 13 probable), of which 396 were ETC admissions (385 confirmed) and 70 community deaths (68 confirmed). The peak incidence of 67 reported new confirmed and probable cases of EVD occurred during the week of 18–24 August (58 patients admitted to the ETC, nine community deaths). This was followed by a steady decline to one confirmed case admitted to the ETC and no confirmed community death during the week of 20–26 October (figure 2). The number of community deaths reported (and thus provided a safe burial) increased from the end of June to a peak in the period from mid-September to mid-October, but these were progressively less likely to be owing to EVD. The proportion of community deaths confirmed positive remained near 90% until mid-August, then declined thereafter. Similarly, the proportion of ETC admissions that were confirmed EVD positive declined from 90% in mid-June to around 50% in mid-September. A detailed description of the outbreak and the response can be found in box 1.

Males (48%) and females (52%) were equally affected by the outbreak. Most (63%) of the cases were older adults (greater than or equal to 40 years of age), a third (33%) were young adults (15–39 years of age), whereas a minority were children (less than 15 years of age; 2%). Overall, 310 (67%) of the admitted patients with confirmed or probable EVD died of the disease. Within Lofa County, the case fatality rate did not differ by district of residence, between men and women, nor did it change significantly over time. Of the confirmed and probable cases of EVD, 21 (5%) were healthcare workers, of whom 14 (67%) died, six recovered and one had an unknown outcome.

The epidemic first affected the western districts of Foya, Kolahun and Vahun, and later the eastern districts of Voinjama, Zorzor and Salayea, Quardu Boundi (figure 3). No cases of local origin were seen in Foya, Kolahun or Vahun districts after mid-September, whereas local cases persisted in the eastern districts until the end of October. During the final weeks of October, more than half of the cases of EVD were exposed outside of Lofa County, in either Monrovia or Guinea.

• Download figure
• Open in new tab
• Download PowerPoint

The delay between symptom onset and ETC admission increased from a median of 3 days in June (50% of cases being admitted within 2–4 days), to 5 in July (50% between 3 and 7 days), before dropping again to 3 in August (50% between 2 and 5 days), September (50% between 2 and 5 days) and October (50% between 2 and 4 days). Weekly changes in the delay to admission can be seen in figure 2.

Our model reproduced the trend in hospital admissions between June and November (figure 4). Because of correlations between the parameters, they were relatively poorly identified: the mean estimate for the initial number of infected people was 34 (interquartile range, IQR, 9–43, 95% credible interval, CI, 3–146, uniform prior range 0–200) and was negatively correlated with the average transmission rate. The mean estimate of the mean community reproduction number was 1.5 (IQR 0.7–2.0, 95% CI 0.1–3.9, uniform prior range 0–10), itself positively correlated with the early proportion of people seeking healthcare. That is, for more people seeking healthcare a higher transmission rate was required to create the observed data, as a larger proportion of the population was being isolated. The mean estimate for the weekly standard deviation in the transmission rate (on the scale of the community reproduction number R_C) was 4.5 (IQR 3.1–5.6, 95% CI 1.8–9.0, uniform prior range 0–10).

• Download figure
• Open in new tab
• Download PowerPoint

The estimated community reproduction number (i.e. the average number of new cases caused by each case if there were no beds available to isolate them) shows a pattern of multiple sub-outbreaks followed by brief periods of no transmission, combined with an overall decline (figure 2f). The longest sustained period of transmission was in the five weeks between 30 June and 4 August, during which the reproduction number can be seen to steadily increase. The subsequent peaks in mid-August and early and late October are significantly smaller. After the first week of October, more than 78% of posterior samples of the reproduction number were below 1.

We estimated that healthcare-seeking behaviour increased during the course of the outbreak, from a median estimate of 36% (IQR 16–67, 95% CI 3–96) of the population seeking healthcare at the beginning of June to a median estimate of 87% (IQR 72–96, 95% CI 26–100) at the end of October. By comparing the late to early healthcare-seeking behaviour on a sample-by-sample basis, our median estimate suggests that individuals were 2.0 (IQR 1.3–4.3, 95% CI 1.0–21) times as likely to seek healthcare late in the epidemic as early. With the parameters in our model, isolation in the ETC reduced the reproduction number of those that seek healthcare to a fraction r = 0.38 (median estimate, IQR 0.37–0.39, 95% CI 0.37–0.40), or by 66%. If individuals were assumed to be twice as infectious during the later stage of infection and burial, the effect was even stronger, with r = 0.33 (median estimate, IQR 0.32–0.34, 95% CI 0.31–0.35) or a 67% reduction of the reproduction number of those seeking healthcare (implying safe burial).

Aggregated simulated trajectories of total incidence (in the community and the ETC) under the scenarios of (i) no bed expansion (and rejection of patients if no beds were available) and (ii) no change in healthcare-seeking behaviour from early in the outbreak are shown in figure 5. The exact number of cases and deaths prevented depended on the exact proportion of healthcare seeking, which was only weakly identified in our model. Comparing the scenarios on a (posterior) sample-by-sample basis, we found that keeping capacity at 10 beds (the official capacity) increased the overall number of cases by a factor of 4.7 (median, IQR 2.1–14, 95% CI 1.1–59), while keeping it at 40 (the number of patients for which the 10 bed ETC was used during the period of overcrowding in July to early August) increased the number of cases by a factor of 1.4 (median, IQR 1.2–1.8, 95% CI 1.0–2.8). Keeping healthcare-seeking behaviour as it was at the beginning of the outbreak increased the number of cases by a factor of 1.9 (median, IQR 1.2–5.2, 95% CI 1.0–109).

• Download figure
• Open in new tab
• Download PowerPoint

4. Discussion

Lofa County was the first region in Liberia to experience the 2013–2015 EVD outbreak in West Africa, and it was struck severely. A rapid rise in cases over the summer of 2014 was followed by a fall in cases towards the end of August 2014, and control by the end of October. The behaviour of the epidemic in Lofa mirrors what has been observed across many regions of West Africa, and a better understanding of this behaviour can yield important lessons for investigating the epidemic elsewhere. Here, we analysed the impact of different clinical and community-based interventions that were put in place at different times, in an attempt to pinpoint what exactly turned the epidemic around.

The main clinical intervention was the expansion of the ETC in Foya from a capacity of 10 (although in fact used for up to 40 patients during periods of overcrowding) to 100. We have estimated that the provision of isolation enabled through the expanded ETC reduced the reproduction number of those who entered the ETC by about two-thirds and thus made a strong contribution to the substantial decline in case numbers observed in the three weeks following the opening of the expanded ETC. Without the expansion, the outbreak would have continued to grow for weeks longer, and we estimated that hundreds of lives were potentially saved by this measure.

The national government imposed a state of emergency in Liberia on 6 August. By closing markets, schools and restricting public gatherings and movement across international borders, this measure may have altered patterns of contact in communities and contributed to preventing geographical spread, but possibly not to a reduction in disease transmission, much of which occurs among family members [51]. At the same time, the delay between the development of symptoms and hospitalization of cases increased when case numbers were growing rapidly before stabilizing after the ETC was expanded, and declining further as the outbreak was well under control. Key observers have suggested that the increase observed in June and early July was an indicator of prevailing community attitudes at the time: denial of the EVD outbreak, and reluctance to present for care because of the stigma associated with the disease. With the ETC operating at capacity, this could explain why it took longer to admit patients. More prompt presentation for care from August onwards may reflect a combination of greater availability of ETC beds, more positive attitudes toward treatment, and greater community awareness of the disease, however, one cannot ascertain the relative contribution of these factors.

From the modelling analysis, it appeared that the expansion of the ETC capacity, occasional interruptions of transmission and improvements in healthcare-seeking behaviour all contributed to mitigating the epidemic and eventually stopping it. Two features stood out in the time-varying behaviour of the transmission rate: one is that the epidemic in Lofa appeared to have consisted of several subepidemics which intermittently stopped before starting again. This is consistent with geographically heterogeneous dynamics where outbreaks flare up and are subsequently controlled by a combination of local behavioural change and targeted intervention, but potentially trigger outbreaks elsewhere. The second feature is the overall decline in the reproduction number from August, leading to eventual extinction. This decline has been previously observed using a model of behavioural change that did not include the detailed information of the outbreak which we included here [24]. It is impossible to establish with certainty the reasons for the decline from the available data, but a strong role of concurrent efforts of the provision of clinical care and isolation and community engagement is consistent with our findings.

The exact extent of healthcare-seeking behaviour was difficult to establish from fitting to hospital admission data alone, which did not contain any information on how many cases may still have been occurring in the community but went unreported. Still, our model suggests that it did appear to improve between the beginning and the end of the epidemic. Interestingly, much of the increase in healthcare-seeking behaviour appeared to happen before the start of health-promotion activities, indicating that attitudes may have changed substantially before the onset of these activities. In this context, it is important to bear in mind that the time course of healthcare-seeking behaviour was constrained by the particular structure of the model we chose. While the transmission rate was allowed to vary randomly and therefore increase and decrease at any point in time, we assumed that healthcare-seeking behaviour could only change in one direction over the course of the outbreak. If we had allowed similar flexibility in the way we modelled healthcare-seeking behaviour to the way we modelled the changes in the transmission rate, it might not have been possible to pinpoint these changes. The behaviour we found was consistent with observations that overall attitudes towards health institutions improved during the course of the epidemic. It helps explain the very sharp rise of cases admitted to the ETC in July (which was stronger than could be explained with a constant transmission rate), and it aligns with the fact that the outbreak was eventually ended relatively rapidly. Without further information on community behaviour, however, it remains difficult to establish these relationships with certainty.

More generally, it would be desirable in the course of future outbreaks to implement concurrent community surveillance to directly monitor total deaths and their causes, and measure a limited number of key social variables. Assessing all-cause mortality within the community could provide an estimate of the number of suspect EVD cases in the community that did not present to the ETC for care or were not provided a safe community burial. This would allow ongoing measurement of healthcare- and safe-burial-seeking behaviour, including an estimate of the proportion of EVD in the community not being identified and properly managed. Similarly, if one had direct information on public attitudes (e.g. perception of the ETC, EVD and its transmission, willingness to seek healthcare if ill), such information could assist in not only more accurately understanding and modelling the role of these factors in the evolution of the outbreak, but also in tailoring social mobilization efforts to increase their effectiveness in particular communities. While such studies are clearly secondary to immediate efforts to mitigate transmission and care for the sick, they could provide crucial information for evaluating and tailoring the ongoing outbreak response and planning for future such events.

Ethics

The research was approved by the London School of Hygiene and Tropical Medicine Research Ethics Committee (reference number 8627).

Data accessibility

All data are available at http://github.com/sbfnk/ebola_lofa.

Author's contributions

S.F., I.G., A.C., A.J.K., W.J.E. and B.R. devised the study; S.F. analysed the data; S.F. and B.R. wrote a first draft of the paper; all authors contributed to critical interpretation of the results and revision of the manuscript.

Competing interests

We have no competing interests.

Funding

S.F., A.C., A.J.K. and W.J.E. were supported by the Research for Health in Humanitarian Crises (R2HC) Programme, managed by Research for Humanitarian Assistance (grant no. 13165). S.F., A.C. and A.J.K. were supported by the UK Medical Research Council (MRC) Career Development Awards in Biostatistics (S.F., fellowship MR/K021680/1; A.C., fellowship MR/J01432X/1; A.J.K., fellowship MR/K021524/1). R.M.E. and W.J.E. were supported by the Innovative Medicines Initiative 2 (IMI2) Joint Undertaking under grant agreement EBOVAC1 (grant 115854). The IMI2 is supported by the European Union Horizon 2020 Research and Innovation Programme and the European Federation of Pharmaceutical Industries and Associations.