Research articleEvaluating the frequency of asymptomatic Ebola virus infection
Abstract
The potential for asymptomatic infection from Ebola viruses has long been questioned. Knowing the proportion of infections that are asymptomatic substantially changes the predictions made by mathematical models and alters the corresponding decisions based upon these models. To assess the degree of asymptomatic infection occurring during an Ebola virus disease (EVD) outbreak, we carried out a serological survey in the Djera district of the Equateur province of the Democratic Republic of the Congo affected by an Ebola outbreak in 2014. We sampled all asymptomatic residents (n = 182) of 48 households where at least one case of EVD was detected. To control for potential background seroprevalence of Ebola antibodies in the population, we also sampled 188 individuals from 92 households in an unaffected area with a similar demographic background. We tested the sera collected for anti-Ebola IgG and IgM antibodies at four different dilutions. We then developed a mixture model to estimate the likely number of asymptomatic patients who developed IgM and IgG responses to Ebola antigens in both groups. While we detected an association between medium to high titres and age, we did not detect any evidence of increased asymptomatic infection in the individuals who resided in the same household as cases.
This article is part of the themed issue ‘The 2013–2016 West African Ebola epidemic: data, decision-making and disease control’.
1. Introduction
The 2013–2016 Ebola virus disease (EVD) epidemic in West Africa has shaken the public health community worldwide into better understanding this disease [1,2].The Democratic Republic of the Congo (DRC) has endured the most outbreaks of EVD of any country since the virus was first discovered there in 1976 [3]. In 2014, concurrent to the epidemic in West Africa, the Equateur province in the north of the DRC experienced an outbreak of the same species, Zaire ebolavirus, which resulted in 68 cases and 49 deaths [3]. It has been shown, however, that the two outbreaks were not linked [4].
One of the questions that arose during the Ebola outbreaks was whether there is asymptomatic infection. The answer is essential to understand the transmission dynamics of the disease. In particular, knowing the proportion of infections that are asymptomatic can improve the accuracy of mathematical model predictions of EVD transmission. More accurate modelling can help guide policy-making regarding interventions during future EVD outbreaks and strengthen our understanding of the transmission dynamics in past outbreaks such as the one in West Africa. Moreover, knowing the number of asymptomatic cases can also be used to calculate a more accurate vaccination threshold (the proportion of the population that needs to be vaccinated to control the disease), which could be lower because individuals that have seroconverted are already likely to be immune prior to vaccination [5].
Seroprevalence surveys of anti-Zaire ebolavirus IgG antibodies have been carried out in several countries in central Africa. Northeastern Gabon has been affected by four consecutive outbreaks (1994, twice in 1996 and 2001). In 1996, the seroprevalence was 10.2% (24/236, with one positive Ebola survivor). The relative risk of being seropositive if individuals had been in contact with an EVD case was 3.21% (95% CI = 1.53–6.75) [6]. Twenty-four asymptomatic close contacts of EVD cases were also tested and 11 were positive [7]. In the 1997 outbreak, 979 of the 2533 residents of eight villages were tested. The seroprevalence was much lower, at 1.4% (4% of the individuals found positive were Ebola survivors) [8]. In a more complete study also set in Gabon between 2005 and 2008, 4349 individuals from 220 randomly selected villages covering all ecological regions in the country (forest, grassland, savannah and lakeland) were tested; 15.3% tested positive. The seroprevalence ranged from 19.4% in the forest areas to 2.7% in the lakeland regions. No significant differences in seroprevalence were observed between villages affected by the outbreaks and those that were unaffected [9]. Even in areas where no EVD outbreaks have been reported, such as in the rainforest of the Central African Republic, 5.3% of tested individuals were seropositive with a significantly higher prevalence in the Aka pygmy population (13.2%) than in non-pygmies (4.2%; p < 0.03) [10].
In the DRC, during the 1995 outbreak, the seroprevalence in forest workers in the vicinity of Kikwit, the epicentre of the outbreak, was 2.2% (9/414). One hundred and sixty-one volunteers from Kikwit's surrounding villages where no EVD was reported were also recruited. In this population, 9.3% were positive [11]. A further seroprevalence study was carried out in Tandala (DRC) following the discovery of one case of EVD. In this instance, the seroprevalence was estimated at 7% [12]. More recently, in sera collected in 2002 in the Watsa region (DRC) in the Efé pygmy population, the seroprevalence was estimated at 18.7% [13].
While providing a useful picture of seroprevalence against EVD in various populations, most of these studies did not look at the question of asymptomatic infection, i.e. whether seropositive individuals are likely to have been infected by EVD and if so, at which frequency asymptomatic infection occurs. Consequently, a recent meta-analysis reviewed existing evidence and estimated 27.1% of EVD infections to be asymptomatic. However, the studies reviewed used a heterogeneous set of enzyme-linked immunosorbent assay (ELISA) techniques that are problematic to compare in the absence of standardization and are difficult to interpret because of the risk of bias by false-positive results [14].
In our study, we sought to ascertain the extent of asymptomatic infection during the 2014 EVD outbreak in the DRC. For this, we compared the seroprevalence among asymptomatic household contacts of cases to the seroprevalence in households in a control (unaffected) population nearby. The rationale was that, if asymptomatic infection occurred, we should detect an excess seroprevalence in the affected households. We analysed thus the presence of anti-Ebola antibodies IgM and IgG by ELISA in the sera of individuals residing in the villages comprising five health areas (two affected by the outbreak and three unaffected) in the health district (zone de santé) of Boende in the former Equateur province. In order to overcome known difficulties in interpreting ELISA results using cut-off based methods, we developed a Bayesian mixture model that takes better account of the available information present in the different ELISA dilutions to discriminate between positive and negative sera.
2. Material and methods
(a) Epidemiological survey
Individuals residing in the same households as an EVD confirmed or probable case (as defined in [3]) and individuals in households in neighbouring unaffected villages were contacted by local nurses. We planned to obtain informed consent from approximately 200 individuals between the Lokolia and Watshikengo health areas (which were affected by the outbreak) and 200 between the Bokoto, Lisafu and Efoku health areas (unaffected). These five health areas (aires de santé) were of similar demographic background (in terms of size, ethnicity, local ecosystem and professional activity). Consent was taken from individuals and, after signing of the appropriate forms approved by the Congolese (Ecole de Santé Publique, Université de Kinshasa) and UK (London School of Hygiene and Tropical Medicine) ethics boards, approximately 5–10 ml of participants' blood were drawn via venipuncture by phlebotomy-trained personnel. Patient identifiers that permitted the linkage of the seroprevalence data to epidemiological data (family name, first name, postname/other name, age, sex, village and occupation) were collected by the researcher. Several patient identifiers were collected because recollection of age is often problematic and, occasionally, individuals are known by several names.
(b) Serology
IgM antibodies start to develop as early as 2 days after the onset of symptoms and are most commonly present 10–29 days after the onset of symptoms [15]; IgG antibodies develop as early as 6 days after symptom onset although they more commonly occur after 19 days. IgG antibodies to Z. ebolavirus have been detected more than 11 years after infection, making them excellent markers for infection [16].
Seroprevalence was determined at the Institut National de Recherche Biomédicale (INRB) in Kinshasa, DRC. Collection of blood samples occurred between 3 December 2014 and 9 January 2015, almost immediately after the Ebola outbreak was declared over by the Congolese authorities on 21 November. The time window chosen allowed for the immune response to develop (most cases occurred in August–September 2014 [3]) while maximizing the chance to find household contacts present at the time of the outbreak.
Blood samples were collected using a serum tube, increased silica act clot activator and silicone-coated interior (BD worldwide®). Samples were kept at room temperature for 30 min and then centrifuged at 3500 r.p.m. for 5 min to separate the serum from blood cells. All samples collected in the affected area (Lokolia and Watshikengo) were stored in liquid nitrogen until their shipment to the National Laboratory. Samples collected in the non-affected area were stored at −20°C then transferred in cool box for shipment to the National Laboratory. Upon arrival at the National Laboratory, all samples were immediately transferred to a −80°C freezer where they were stored until the ELISA analysis.
The detection of Z. ebolavirus IgG and IgM antibodies was performed using an ELISA kit and method designed and supplied by the Public Health Agency of Canada [17]. One of the samples collected from the first suspected Ebola cases during the outbreak in Boende (polymerase chain reaction negative, IgM positive and IgG positive) was used as the positive control for each test. Rows A–D of a 96-well ELISA plate were coated with sucrose purified Z. ebolavirus glycoprotein (ZEBOV GP) diluted at 1 : 500 (positive wells), and rows E–H were coated with vesicular stomatitis virus (VSV) lysate, also diluted at 1 : 500 (negative wells). The plate was stored overnight at 4°C. The following morning, the coated plate was removed and washed three times with phosphate-buffered saline (PBS) + 0.1% Tween; 100 µl of diluent (5% milk in PBST) were added to each row, loading the bottom half of the plate first, then changing tips and adding the same to the top half of the plate. While milk-PBST was sitting in the plate, a dilution plate was set up where the sera were diluted at 1 : 25. The positive control was in column 1 and the negative control in column 2. The diluted serum from each sample (plus positive and negative control) were loaded into their corresponding wells on the ELISA plate at dilutions of 1 : 200, 1 : 400, 1 : 800 and 1 : 1600. Tips were changed when moving from the antigen-positive side of the plate to the antigen-negative side. After an hour-long incubation, the plate was washed three times and dried. A secondary goat anti-human-horseradish peroxydase (HRP) was used for the detection of IgG antibodies and a secondary goat anti-human IgM antibody/HRP conjugate was used for the detection of IgM antibodies. Both were diluted at 1 : 2000; 100 µl were pipetted into each well, again starting with the negative-antigen side first. The plate was incubated at 37°C for 1 h. After incubation, the plate was washed three times and dried; 100 µl of a mixed A and B solution of TMB (3,3′,5,5′-tetramethylbenzidine) substrate was applied to each well, starting again with the negative-antigen side of the plate. The plate was incubated at 37°C for 30 min and then read at 405 nm.
(c) Interpretation of serological results
Four pairs of optical density (OD) measurements were obtained per serum sample for each of the two measured antibodies (IgG and IgM). Each pair was composed of a ‘positive’ well reading (containing ZEBOV GP) and a ‘negative’ well reading (containing VSV lysate). Readings from the positive wells were expected to be positively correlated with readings from their corresponding negative wells (which served as controls) as they were processed in parallel. As a result, high ODs for positive wells had to be interpreted conditionally on the results of their associated negative wells.
Different methods based on cut-off points [18] are currently used to assess the positivity of serum samples. The first method (commonly used for diagnostics) simply computes the difference between the OD readings of the corresponding positive and negative antigen wells for a sample at a given dilution (referred to subsequently as the difference method). At a given dilution, a difference of 0.95 or higher is interpreted as positive, a difference between 0.5 and 0.95 is considered equivocal and a value lower than 0.5 is negative. For research ELISAs (which are usually run in triplicate), if the average sample OD is more than 2× the average OD of the corresponding negative wells + 2× the standard deviation of those negative wells then the sample is considered positive at that dilution. If not, it is considered negative at that dilution. Finally, a third method calculates the ratio of the reading of the positive well over the negative well. A ratio above 3 indicates positivity at a particular dilution while a result between 2 and 3 is considered equivocal. If the ratio is lower than 2, the sample is considered negative (at that dilution).
To assess the likely ‘final’ positivity of a serum sample, the results obtained at different dilutions must be compared. Typically, a sample is considered positive if it is positive for at least two different dilutions. The higher the number of dilutions at which the sample is positive, the more certain the result. In diagnostics, a rerun can be carried out if the evidence for positivity is weak.
(d) Bayesian mixture model
To facilitate the rigorous characterization of the status of a given serum, we developed a multivariate mixture model of two components. Recall that for each individual we have a four-dimensional sample with the result for each dilution (difference between the positive and negative wells at the corresponding dilution). The model assumed that both seropositive and seronegative samples produced readings, which are assigned to two quadrivariate normal distributions having different means (positive samples having a higher mean for each dilution) but the same covariance matrix. The common covariance matrix is not necessary in our modelling framework but seemed like a reasonable assumption given the sparsity of the data regarding the second component (negative samples). Note that the only restriction we impose is that the means of one component are larger than the means of the other component. This seems like a natural assumption and facilitates avoiding the well-known problem of label switching in mixtures [19].
This method made possible the simultaneous inference for the probability that a sample is positive as well as the means, the variance and the correlations of each of the two (four-dimensional) components of the mixture. Using Bayesian inference, it is thus possible to sample from the joint posterior density of these parameters, taking all sources of uncertainty into account. This task was completed using the BUGS language [20]. Code and data have been made available in the electronic supplementary material.
(e) Factors associated with high titres
We implemented various logistic regression models using different combinations of age, location, sex and residing in the same household as an EVD case as predictors and the presence of negative, equivocal or positive titres as the dependent variable. For each dilution, due to the very small number of titres classified as positive, to increase the chance of detecting a significant association, we subsequently grouped samples classified as positive with the equivocal samples (referred to subsequently as ‘medium to high titres’). We proceeded iteratively by dropping the less informative predictors (while keeping the EVD case household category of interest). The Akaike information criterion (AIC) was used to select the most appropriate model.
3. Results
(a) Epidemiological and serological survey
We sampled 182 individuals from 48 households in the affected area and 188 from 92 households in the unaffected area. While we obtained sera from all asymptomatic individuals residing in the households of EVD cases who were present during the outbreak, we did not obtain sera from complete households in the unaffected area. Nevertheless, sera from individuals across all age groups were sampled. The proportion of women was higher in the unaffected area (57%) than in the affected area (46%), while the proportion of younger individuals (under 20) was similar in both locations (40% in the affected area and 43% in the unaffected area; table 1). Only the difference in gender was significant (two-sample test for equality of proportions with Yates' continuity correction, p = 0.0385).
| affected area (%) | non-affected area (%) | total (%) | |
|---|---|---|---|
| female | 84 (46.2) | 108 (57.4) | 192 (51.9) |
| age <20 | 73 (40.1) | 80 (42.8) | 153 (41.5)a |
| IgM > 0.5 at 1/200 | 16 (8.8) | 10 (5.3) | 26 (7) |
| IgM > 0.5 at 1/400 | 10 (5.5) | 5 (2.7) | 15 (4.1) |
| IgM > 0.5 at 1/800 | 2 (1.1) | 3 (1.6) | 5 (1.4) |
| IgM > 0.5 at 1/1600 | 1 (0.5) | 3 (1.6) | 4 (1.1) |
| IgM > 0.95 at 1/200 | 6 (3.3) | 3 (1.6) | 9 (2.4) |
| IgM > 0.95 at 1/400 | 1 (0.5) | 2 (1.1) | 3 (0.8) |
| IgM > 0.95 at 1/800 | 0 (0) | 0 (0) | 0 (0) |
| IgM > 0.95 at 1/1600 | 0 (0) | 0 (0) | 0 (0) |
| IgM positive | 1 (0.5) | 1 (0.5) | 2 (0.5) |
| IgG > 0.5 at 1/200 | 21 (11.5) | 24 (12.8) | 45 (12.2) |
| IgG > 0.5 at 1/400 | 9 (4.9) | 11 (5.9) | 20 (5.4) |
| IgG > 0.5 at 1/800 | 6 (3.3) | 6 (3.2) | 12 (3.2) |
| IgG > 0.5 at 1/1600 | 3 (1.6) | 6 (3.2) | 9 (2.4) |
| IgG > 0.95 at 1/200 | 6 (3.3) | 8 (4.3) | 14 (3.8) |
| IgG > 0.95 at 1/400 | 2 (1.1) | 3 (1.6) | 5 (1.4) |
| IgG > 0.95 at 1/800 | 1 (0.5) | 2 (1.1) | 3 (0.8) |
| IgG > 0.95 at 1/1600 | 2 (1.1) | 2 (1.1) | 4 (1.1) |
| IgG positive | 2 (1.1) | 3 (1.6) | 5 (1.4) |
| total | 182 | 188 | 370 |
(b) Serological testing
Readings from positive wells were strongly correlated with readings from negative wells (r = 0.81). As only single measurements for each dilution were obtained, the difference method appeared to be more adequate than the methods based on ratios as it discriminates sera with truly higher ODs for the positive well from sera where both ODs in the negative and positive wells were very low. In addition, the method is more sensitive for highly positive titres (figure 1). Consequently, in the rest of the manuscript, we assessed positivity using the difference method. The positivity results (using the ‘equivocal’ and ‘positive’ threshold of, respectively, 0.5 and 0.95 alongside the combined value) for the four dilutions and the two antibodies measured are given in table 1. Most sera appear to be negative both for IgM (0.5% positive) and IgG (1.4% positive) both in the group of individuals residing in EVD case households and in the group of individuals residing in the unaffected villages. While some of the IgG positive signals appeared strong as they were conserved through several dilutions, the IgM positives were more likely to be false positives. Indeed, at dilutions 1/800 and 1/1600, no sera presented IgM titres above 0.95. The positivity was similar for both arms of the study and decreased with increased dilution (table 1). A cross dilution comparison can be used to estimate which samples were likely to be true positives as illustrated in figure 2.
Figure 1. Two possible interpretations of a measurement for a given dilution. The ratio method (light green and grey area) assumes that the measurement is positive if the optical density (OD) of the positive well is three times higher than the OD of the negative well. The difference method (pink and grey area) assumes that the sample is positive if the difference in OD between the positive and negative well is higher than a certain threshold (here 0.95). Samples in the grey region are positive for both methods. Blue points are results from the affected area while red points are results from the control (unaffected) area. Figure 2. Red points represent the optical density difference distribution of the sera for 1 : 400 (x-axis) and 1 : 800 (y-axis) dilutions of the IgG antibodies. Lines represent the two thresholds (0.5 equivocal and 0.95 positive) for interpretation. Only three samples are positive at both dilutions. The area inside the red ellipse indicates the area of strong density of negative samples (the ellipse represents the 95% credible interval of the posterior of the negative component of the mixture model) and the area inside the green ellipse indicates a strong density of positive samples (95% credible interval of the posterior of the positive component of the mixture model).
(c) Bayesian mixture model
The Bayesian mixture model provided a novel method to assess seropositivity. The mixture model suggests that there is a lower percentage of sera that are positive for IgG (0.8%) than the percentage assessed by the difference method (1.4%) and allows us thus to discriminate better the positive. The mixture model is better at discriminating the true values due to the simultaneous consideration of all the dilutions (including their correlation structure) through a multivariate likelihood. Hence, the assessment of the probability that a sample belongs to the positive or negative component is based upon all the available data and explicitly acknowledges each source of uncertainty, thus giving more efficient results. In comparison, the difference model drops information by first categorizing the sample into ‘negative’, ‘equivocal’ and ‘positive’ for each dilution and then makes a decision based on the resulting four dilutions. In figure 2, we can see how the mixture model can be used to highlight the regions where the samples are likely to be positive or negative, and then give a probability that the sample is positive or not.
(d) Factors associated with high titres
The suitable model with the lowest AIC included only age and residing in the same household as an EVD case as predictor variables. Only age appeared to be a significant predictor of medium to high IgG titres (defined as titre above the equivocal cut-off) at all considered dilutions but the highest one (1/1600) while residing in the same household as an EVD case was not significantly associated with medium to high titres at any dilution (table 2). For example, at a 1/400 dilution, for any additional year in age of a patient, the odds of obtaining medium to high IgG titres for a particular serum increased by 3.5% (p = 0.00473). To illustrate the logistic regression results we show (figure 3) the proportion of individuals with antibodies above the 0.5 threshold, stratified by age and whether they were members of a EVD case household.
Figure 3. Stratification of individuals with equivocal or positive IgG titres at 1 : 200, 1 : 400, 1 : 800 and 1 : 1600 dilutions by contacts with cases status (a) and age (b). Vertical lines indicate 95% CI using binomial distributions. (Online version in colour.)
| outcome | predictor | estimate | s.e. | p-value | AIC |
|---|---|---|---|---|---|
| IgG(1/200) > 0.5 | age | 0.01783 | 0.00884 | 0.0437* | |
| contact of case | −0.19481 | 0.32298 | 0.5464 | 275.53 | |
| IgG(1/400) > 0.5 | age | 0.03493 | 0.01236 | 0.00473** | |
| contact of case | −0.33529 | 0.47163 | 0.47714 | 153.52 | |
| IgG(1/800) > 0.5 | age | 0.03671 | 0.01556 | 0.0183* | |
| contact of case | −0.12504 | 0.59602 | 0.8338 | 106.4 | |
| IgG(1/1600) > 0.5 | age | 0.02767 | 0.01789 | 0.122 | |
| contact of case | −0.80152 | 0.72211 | 0.267 | 87.388 |
4. Discussion
We estimate the extent of asymptomatic infection during an outbreak of Z. ebolavirus using an epidemiological and a household-based serological survey. The study was carried out shortly after the 2014 DRC EVD outbreak was officially declared over, which made the study challenging due to fears related to the disease. The study compared an affected area with a matched unaffected (control) area. These areas were remote with difficult access and no laboratory facilities. While we identified individuals with positive (i.e. with high titres at different dilutions) sera in both arms of the study, we did not find evidence of a higher IgG positivity in the group in the affected area compared to the group in the unaffected area. This suggests that either asymptomatic infection is occurring at very low numbers or not at all. We showed, though, an association between higher titres of IgG and age, and designed an original algorithm to infer the positivity of samples. This algorithm is more rigorous and robust than the existing methods of assessment.
Our study suffers, however, from limitations. One of the most important limitations of this study was the absence of sera from survivors to compare to the samples with high antibody titres. Whilst one serum sample from a survivor collected during the outbreak was used as a positive control (both positive for IgG and IgM), the IgG values are difficult to compare as IgG antibodies need time to plateau post-infection. Another limitation is the absence of duplicates in the ELISA, resulting in a high level of uncertainty in the OD results, which makes them difficult to interpret. A more expensive study design that involves running experiments in duplicates or triplicates and taking the mean OD would probably lead to a lower variance in the results. Additionally, the surveying team encountered hostility towards the end of the project, which hampered the original design of matching the households in the affected area to the households in the unaffected area. Although a similar number of sera were collected in the control group, the household structure design was disrupted. We believe, nonetheless, that this study provides a valuable control group of a similar demographic and age. Finally, while residency in the household of a case is a proxy for exposure, residency in the unaffected village did not preclude individuals from being exposed per se as individuals in both villages could have been in contact with each other. This could be further analysed using mathematical models with two levels of mixing [21].
The small number of IgG positive samples (5) render the results difficult to interpret. Overall, our results are lower than most seroprevalence studies of background antibody levels (1.4% of individuals positive for IgG). The potential reasons for the variability between studies are numerous. Some asymptomatic infection might result from low virulence virus emerging from the animal reservoir (thought to be fruit bats [22,23]), or from exposure to cross-reactive viruses. This risk will be modulated by the environment surrounding the different populations [24] and the type of activities they carry out (e.g. hunting in the forest). Finally, the sensitivity and specificity of the different ELISA tests might play a role in the heterogeneity of results. The recent outbreak in West Africa will certainly provide opportunities to study the serological antibody response following different degrees of exposure to the virus and permit more robust conclusions.
Despite its limitations, this study has served to highlight the importance of serology in understanding transmission and has provided insights into the interpretation of serological results. For example, the first dilution (1/200) does not appear to be specific enough while the highest (1/1600) appears to be insensitive. Intermediate dilutions (1/400 and 1/800) appear to be more informative (table 1). We have also shown an association between age and higher IgG titres. This could be explained by environmental exposure over a lifetime to cross-reactive agents. We also showed that the odds of observing medium to high titres of IgG were not significantly higher in the group of asymptomatic individuals that resided in the same household as EVD cases than in the control group, indicating that asymptomatic infection in this outbreak occurred at very low numbers or not at all.
We designed a method that uses the information from all dilutions in the ELISA to classify the samples into positive and negative. This is more rigorous than the current methods, which are based on arbitrary areas of the OD measurement pair plane. This method could potentially be applied to other infections where quantification of asymptomatic infection is important and interpretation of ELISA-based assay results is similarly difficult to interpret.
Evidence suggests that some Ebola infections may be asymptomatic, with seroprevalence estimates between 1.4 and 19.4% [6–12]. A recent systematic review and meta-analysis [14] also estimated that 27.1% (95% CI, 14.5%–39.6%) of EVD infections were asymptomatic. Following these estimates, we should expect to find positive samples in both arms of the study and between 1 and 4 positive samples more in the affected households than in the control population (as we observed seven secondary transmission events in the 48 households). We do not observe this as the number of IgG positives we found is 1 for the affected area versus 2 in the control area (2 versus 3, respectively, if using the cut-off methods rather than the mixture model estimate). However, the studies used for the meta-analysis do not use control populations, and the specificity and interpretation of ELISA assays can be misleading (as shown in our study).
5. Conclusion
In agreement with other studies, we have detected background exposure to Ebola viruses. Asymptomatic infection in the household of cases did not seem to play an important role in Ebola transmission in this setting as no evidence of asymptomatic infection linked to the 2014 outbreak was found in our study. We propose a Bayesian mixture model as a novel method to determine the seropositivity of samples.
Ethics
This study was carried out following two ethical approvals from the LSHTM ethical committee (LSHTM Ethics Ref: 8785) and the School of Public Health of the University of Kinshasa (ESP/CE/056/14). All individuals sampled have provided informed consent to participate in the study.
Data accessibility
Data and code for the Bayesian mixture model are available in the electronic supplementary material.
Authors' contributions
Pl.M., M.B., A.R., N.D., W.J.E. and J.-J.M. conceived and designed the study. Pl.M., I.N. and P.M. collected and carried out the testing of the serological samples; M.B., N.D. and W.J.E. analysed the data. Pl.M., M.B., A.R. and N.D. drafted the manuscript which was subsequently revised critically by all authors. All authors approved the final version of the article.
Competing interests
We have no competing interests. W.J.E. is a Guest Editor of this issue.
Funding
This study was funded by the Fischer Family Trust (A.R.) and the National Institute for Health Research Health Protection Research Unit in Immunisation at the London School of Hygiene and Tropical Medicine in partnership with Public Health England (M.B.). The views expressed are those of the authors and not necessarily those of the funders. This study was also made possible by the generous support of the American people through the United States Agency for International Development (USAID) Emerging Pandemic Threats Program-2 PREDICT-2 (Cooperative Agreement no. AID-OAA-A-14-00102). The contents are the responsibility of the authors and do not necessarily reflect the views of USAID or the United States Government.
Acknowledgements
We are grateful to Dr Passy Bosomba, medical district officer of the Boende Health District, for his support during our fieldwork, Ioannis Ntzoufras from the Athens University of Economics and Business for assistance with the multivariate mixture model, Stefan Flasche and Judith Glynn from the London School of Hygiene and Tropical Medicine for useful discussions, Geoff Soule and Gary Kobinger at Public Health Canada for advice on the serological assay and reagents. We were also allowed to use the Metabiota laboratory to conduct our analysis when the INRB had problems with the electricity.
