Drivers of African Filovirus (Ebola and Marburg) Outbreaks

Abstract

Outbreaks of African filoviruses often have high mortality, including more than 11,000 deaths among 28,562 cases during the West Africa Ebola outbreak of 2014–2016. Numerous studies have investigated the factors that contributed to individual filovirus outbreaks, but there has been little quantitative synthesis of this work. In addition, the ways in which the typical causes of filovirus outbreaks differ from other zoonoses remain poorly described. In this study, we quantify factors associated with 45 outbreaks of African filoviruses (ebolaviruses and Marburg virus) using a rubric of 48 candidate causal drivers. For filovirus outbreaks, we reviewed >700 peer-reviewed and gray literature sources and developed a list of the factors reported to contribute to each outbreak (i.e., a “driver profile” for each outbreak). We compare and contrast the profiles of filovirus outbreaks to 200 background outbreaks, randomly selected from a global database of 4463 outbreaks of bacterial and viral zoonotic diseases. We also test whether the quantitative patterns that we observed were robust to the influences of six covariates, country-level factors such as gross domestic product, population density, and latitude that have been shown to bias global outbreak data. We find that, regardless of whether covariates are included or excluded from models, the driver profile of filovirus outbreaks differs from that of background outbreaks. Socioeconomic factors such as trade and travel, wild game consumption, failures of medical procedures, and deficiencies in human health infrastructure were more frequently reported in filovirus outbreaks than in the comparison group. Based on our results, we also present a review of drivers reported in at least 10% of filovirus outbreaks, with examples of each provided.

Introduction

Zoonotic disease is widely recognized as a major threat to human health (Jones et al. 2008, Smith et al. 2014, Murray et al. 2015, Han et al. 2016), and the overall frequency of zoonotic outbreaks appears to be increasing over time (Smith et al. 2014). Among zoonotic diseases that have caused large outbreaks (>10,000 cases) in modern times (Stephens et al. 2021), outbreaks of African filoviruses (Ebolavirus, Marburg virus) are among those with the highest mortality. For example, during the 2014–2016 Zaire Ebolavirus outbreak, human-to-human transmission caused at least 11,000 fatalities before it was controlled (Kamorudeen et al. 2020), and other outbreaks of Ebola have had fatality rates of up to 90% (Kuhn 2008).

Thus far, only one outbreak of Marburg virus of more than 100 cases has been documented (Kuhn 2008, Berger 2017). The case fatality rate of 88% during this outbreak (Red Cross 2006, Ajelli and Merler 2012) suggests that Marburg virus could pose a threat to human health similar to Ebolavirus were it to spread more widely. Given high mortality in filovirus outbreaks in some threatened wild species, including gorillas (Bermejo et al. 2006) and chimpanzees (Formenty et al. 1999), filovirus outbreaks are also an important topic for wildlife conservation. In addition, the frequency of filovirus outbreaks appears to be either stable or increasing over time (Fig. 1).

FIG. 1.

Frequency of filovirus (Ebola and Murburg) outbreaks over time, based on 45 outbreaks included in this study (see Materials and Methods section for definitions). The last study year included is 2017. At the time of publication, based on CDC and WHO data, there have been at least seven subsequent Ebola outbreaks and one Marburg outbreak.

Much research has focused on understanding the factors that contribute to filovirus outbreaks (reviewed in Kuhn 2008, Alexander et al. 2015, Jacob et al. 2020). Socioeconomic factors such as wild game consumption (Georges et al. 1999, Leroy et al. 2009), armed conflict (Maurice 2000, Stanturf et al. 2015), inadequate health systems (Shoman et al. 2017) and poverty (Heymann et al. 1999, Fallah et al. 2015), as well as ecoenvironmental factors such as the transition between wet and dry seasons (Schmidt et al. 2017) and seasonal variation in bat abundance (Leroy et al. 2009, Amman et al. 2012), have been reported to play a role in starting and spreading outbreaks.

Most work to date has focused on the drivers of individual outbreaks (e.g., WHO 1978a, 2005, Francesconi et al. 2003, Chevalier et al. 2014). A number of studies have also considered how spatial variation in factors such as forest loss (Olivero et al. 2017b), host biodiversity (Olivero et al. 2017a), and seasonal fluctuations in temperature and rainfall (Schmidt et al. 2017) affects filovirus spillover risk. However, even for these and closely related factors, the proportion of outbreaks, in which each has played a role as a proximate trigger, has not been quantified.

It is also unknown whether the factors that most often contribute to filovirus outbreaks are similar or dissimilar to those that tend to drive outbreaks of other zoonotic pathogens. For example, a recent study considered a random sample of outbreaks of zoonotic pathogens from a global database, and found that wild-game hunting (including the consumption, capture and processing of wildlife) was rarely a proximate driver of outbreaks, being documented as a contributing factor in only four of 300 outbreaks investigated (Stephens et al. 2021). However, wildlife hunting was associated with spillover events that caused several Ebola outbreaks (Rizkalla et al. 2007, Kurpiers et al. 2016, Gryseels et al. 2020), including at least one cluster of cases with more than 200 fatalities (Leroy et al. 2004).

Stephens et al. (2021) hypothesized that due to the transmission dynamics of filoviruses, wildlife hunting is a much more frequent driver of outbreaks of filoviruses compared to most zoonotic pathogens. However, no study has quantified the proportion of filovirus outbreaks in which wildlife hunting or consumption played a role, and thus whether the frequency is significantly higher than that of other zoonoses. Additional untested hypotheses are that the drivers of filovirus outbreaks tend to more closely mirror those of other viruses, and/or other viruses that cause hemorrhagic fevers, than those of zoonotic pathogens in general.

In this study, we considered nearly every confirmed outbreak of African ebolaviruses (including Zaire ebolavirus, Sudan ebolavirus, Tai Forest ebolavirus, and Bundibugyo ebolavirus; we refer to disease caused by any of these species as Ebola) and Marburg marburgvirus with human cases before 2018 (Fig. 2). We conducted a review of both peer-reviewed literature (Kuhn 2008, Alexander et al. 2015, Jacob et al. 2020) and gray literature such as ProMED e-mails (Yu and Madoff 2004) describing each outbreak. The reported drivers of outbreaks were scored with the rubric used in a previous global study of zoonotic outbreaks (Stephens et al. 2021), which includes 48 potential demographic, environmental, and socioeconomic drivers of outbreaks (Table S1 in the Supplemental Materials).

FIG. 2.

Origin, based on location of patient zero, of African filovirus outbreaks included in study. In cases where multiple outbreaks from the same year and country were combined, indicates patient zero of earliest outbreak. Background map generated by ggmap (Kahle and Wickham 2013) from data at OpenStreetMap.

We compared the drivers of filovirus outbreaks with those of 200 outbreaks chosen randomly from a global database of 4463 bacterial and viral zoonotic outbreaks to quantify the degree to which the drivers of filovirus outbreaks differ from those of a random sample of outbreaks of other zoonotic pathogens. We also performed analyses excluding bacterial pathogens to test whether the driver profile of filovirus outbreaks is more similar to that of other viral pathogens compared with the broader sample, including both viruses and bacteria. The latter analysis was performed using both completely random viral outbreaks, and a random sample of additional nonfiloviral hemorrhagic fever outbreaks. Finally, based on our results, we present a review of factors reported in at least 10% of filovirus outbreaks, providing examples of how each factor can contribute to outbreaks and citing relevant case studies.

Materials and Methods

Here, we outline research materials and methods. See “Supplemental Materials and Methods” in the Supplemental Materials for additional details and rationale. We compared the drivers of 45 outbreaks of African filoviruses (Ebolavirus and Marburg) to 200 background examples chosen at random from a global database of outbreaks of zoonotic viral and bacterial pathogens (Stephens et al. 2021). Our definition of a “zoonotic” pathogen follows that of Stephens et al. (2021), which includes diseases classified as zoonotic by working groups of the CDC (2021a), UK Health Ministry (UK Public Health England 2019), and the Pan American Health Organization (Dubinsk 2005).

We scored outbreaks drivers using the rubric of Stephens et al. (2021), a binary rubric of 48 potential drivers based on factors discussed in reviews and syntheses of zoonotic outbreak literature (e.g., Oaks et al. 1992, Lederberg et al. 2003, Ceddia et al. 2013, Gottdenker et al. 2014) (Table S1 in the Supplemental Materials). For filovirus outbreaks, we reviewed >700 sources, including reviews (Kuhn 2008, Alexander et al. 2015, Berger 2017, Jacob et al. 2020), peer-reviewed sources cited in reviews, and gray literature such as ProMED e-mails (Yu and Madoff 2004), Morbidity and Mortality Weekly Reports (CDC 2021b) and WHO reports (e.g., WHO 1978b, 2005, 2012).

For each outbreak, a given driver was scored as either (0) not reported as contributing to an outbreak or (1) reported as contributing to an outbreak by at least one source. The drivers of background outbreaks were scored as described in Stephens et al. (2021).

All analyses were conducted in R version 4.1.2 (R Core Team 2020). We conducted two main sets of analyses: (1) omnibus tests for significant differences between the overall driver profiles of filovirus and background outbreaks and (2) analyses testing for differences in the frequency with which individual drivers were reported. Analyses were conducted using all filovirus and background outbreaks (N = 245), comparing filovirus outbreaks to outbreaks of other viruses from the completely random background (N = 80), and comparing them to randomly selected outbreaks of other viral hemorrhagic fevers (N = 90). Omnibus tests were conducted using contingency table analysis, a permutation test of independence implemented in coin version 1.4-0 (Hothorn et al. 2008).

Omnibus tests were repeated using predictors that appeared in 3% of rows (19 of 48 drivers considered), 5% of rows (13 drivers), and only the 4 most common drivers (Table 1), following Stephens et al. (2021). We also used multiple correspondence analysis (MCA), implemented with FactoMineR version 2.4 (Lê et al. 2008), to visualize differences in the driver profiles of outbreaks of filoviruses and other zoonotic pathogens (i.e., as an exploratory analysis of qualitative patterns). MCA is the counterpart of principal components analysis designed for identifying orthogonal axes of variation among multivariate distributions of discrete variables (Di Franco 2016). Analyses of individual drivers were conducted using both chi-squared and Firth's bias-reduced logistic regression analyses.

Table 1.

Permutation Tests Comparing Filovirus Outbreaks to Background Zoonotic Outbreaks

Covariates	N	χ2	p	Covariates	N	χ2	p
Drivers scored in >3% outbreaks				Four most common drivers
None	245	165.937	<0.0001	None	245	150.572	<0.0001
GDP, phone lines, internet users	224	55.596	<0.0001	GDP, phone lines, internet users	224	55.590	<0.0001
Population, population density, latitude	241	8.311	0.0039	Population, population density, latitude	241	8.311	0.0039
All but internet use	241	8.311	0.0039	All but internet use	234	8.288	0.0040
All covariates	241	8.311	0.0039	All covariates	224	8.299	0.0040
Drivers scored in >5% outbreaks
None	245	164.065	<0.0001
GDP, phone lines, internet users	224	55.595	<0.0001
Population, population density, latitude	241	8.311	0.0039
All but internet use	234	8.288	0.0040
All covariates	224	8.299	0.0040

Covariates, indicates which of six sample bias covariates were included in each model. N: number of rows included (complete case analysis); GDP, gross domestic product; phone lines and internet users are all per capita.

The latter is a method for data that exhibit “separation” issues (Firth 1993, Heinze and Schemper 2002, Heinze and Ploner 2003), where all or most rows of a predictor are invariant for one of the response classes. This was common in our data because many drivers were rarely scored as reported to contribute. Bias-reduced logistic regression was implemented in logistf version 1.24 (Heinze et al. 2020).

Both permutation tests and logistic regression analyses were performed including and excluding six country level covariates (listed below) that have been shown to bias global outbreak data in previous studies (reviewed in Smith et al. 2014, Stephens et al. 2021). We refer to these six covariates collectively as the “sample bias” covariates, although we acknowledge that some may show functional relationships with outbreak risk (see discussion in the Supplemental Materials and Stephens et al. 2021). Each covariate was scored for the country and year in which each outbreak originated. We generally compared five models: (1) no sample bias covariates, and including (2) three measures of expected reporting and detection capabilities (internet use, phone lines and gross domestic product [GDP]), (3) three measures of expected disease diversity and host availability (latitude, human population, human population density), (4) all measures but internet use (which was largely 0 before 1990), and (5) all six sample bias covariates. This research did not included human subjects or data from identifiable individuals and was not subject to IRB review.

Results

Omnibus tests showed that drivers associated with African filovirus outbreaks differed significantly from those of the background outbreaks (Table 1 and Table S2 in the Supplemental Materials) and other viral hemorrhagic fevers (Table S4 in the Supplemental Materials), reflecting that some drivers were more common in filovirus outbreaks compared with controls, while others were more common in background outbreaks (Table 2 and Supplemental Tables S3 and S5). This pattern persisted regardless of whether all (Table 1) or only viral (Tables S2 and S4 in the Supplemental Materials) outbreaks were included in the comparison set, although the specific drivers that were included varied somewhat between the three sets of models.

Table 2.

Frequency of Drivers in Filovirus Versus Background Outbreaks

	Filovirus	Filovirus	Background	Background
	No.	%	No.	%	χ2	p
Human-animal contact	37	82.22	21	10.50	100.65	<0.001
Bushmeat	16	35.56	1	0.50	64.59	<0.001
Local livestock production	0	0	35	17.50	7.81	0.005
Industrial livestock production	0	0	10	5.00	1.24	0.265
Human population density	5	11.11	6	3.00	3.90	0.482
Food contamination a	13	28.89	96	48.00	4.69	0.030 a
Water contamination	0	0	40	20.00	9.34	0.002
Sewage management	0	0	20	10.00	3.66	0.056
International travel/trade a	12	26.67	17	8.50	9.94	0.002 a
Intranational trade/travel	15	33.33	7	3.50	46.44	<0.001
Socioeconomic change	3	6.67	5	2.50	0.92	0.339
Cultural/religious beliefs or practices	18	40.00	3	1.50	64.66	<0.001
Public health infrastructure b	15	33.33	15	7.50	20.47	<0.001 b
Antibiotics	0	0	8	4.00	0.81	0.382
Vaccination breakdown	0	0	10	5.00	1.24	0.265
Medical procedures	26	57.78	15	7.50	63.08	<0.001
War/conflict a	5	11.11	5	2.50	4.93	0.026 a
Weather conditions	1	2.22	13	6.50	0.58	0.446
Poverty a	7	15.56	8	4.00	6.64	0.001 a

Rows in bold: drivers that differed significantly at α = 0.05; Bold black rows: drivers found more frequently in filovirus outbreaks; Bold blue rows: drivers found more frequently in background outbreaks. Only drivers scored in ≥3% of all outbreaks included in study are shown. See Tables S6–S30 in the Supplemental Appendix Table for analyses investigating the influence of sample bias covariates. Results for drivers significant at p < 0.001 were generally robust to the influence of these factors.

^a Analyses of these drivers were not usually statistically significant at α = 0.05 in multivariate models, including sample bias covariates (see Tables S8, S12, S17, and S22 in the Supplemental Appendix Table).

^b One of the models, including sample bias covariates, showed a p value of only 0.077, correlations for this driver were statistically significant at α = 0.05 in all other models (see Table S16 in the Supplemental Appendix Table).

In MCA analysis, filovirus outbreaks diverged from completely random samples of zoonotic outbreaks primarily along the first principal dimension of variation (Fig. 3 and Fig. S1 in the Supplemental Mateirals), but primarily along the second principal dimension when compared to outbreaks of other viral hemorrhagic fevers (Fig. S2 in the Supplemental Materials).

FIG. 3.

Multiple correspondence analyses of zoonotic outbreak drivers and the drivers that load on the first two dimensions of multivariate driver space. Definitions of abbreviations are as follows: AntiB, antibiotics; CultRelB, cultural/religious beliefs or practices; FoodC, food contamination; HealthInf, public health infrastructure; HPD, human population density; HumAnCon, human-animal contact; InLSProd, industrial livestock production; InterTT, International travel/trade; IntraTT, intranational trade/travel; LocalLSProd, local livestock production; MedP, medical procedures; Poverty, miscellaneous stressors related to poverty; SewageM, sewage management; SocChange, broadscale socioeconomic change (e.g., change in government regime); VaccBD, vaccination breakdown; War, war/armed conflict; WLhunt, wildlife hunting, processing, and consumption. The notations “_1” indicates that a driver was typically scored for outbreaks in that part of the graph, whereas “_0” indicates a given driver was not typically observed. Isolines indicate high, and potentially overlapping, point density. See Supplemental Dataset S1 for a list of outbreaks and their observed drivers.

The variables that loaded strongly on these axes included poverty, intra-national trade and travel, cultural beliefs, and wildlife hunting/consumption (Fig. 3 and Figs. S1 and S2 in the Supplemental Materials), all of which were more frequently found in filovirus outbreaks than background outbreaks (Table 2). Although overall they differed significantly (Table S4 in the Supplemental Materials), several outbreaks of nonfiloviral hemorrhagic fevers clustered with filoviral outbreaks in MCA analyses (Fig. S2 in the Supplemental Materials).

In analyses using all background outbreaks, nine drivers were found more often in filovirus outbreaks compared with background examples and three drivers were found more often in background outbreaks compared with filovirus outbreaks (Table 2). Results for many drivers, including wildlife hunting and consumption (Table S7 in the Supplemental Appendix Table), cultural and religious beliefs (Table S8 in the Supplemental Appendix Table), human-animal contact (Table S10 in the Supplemental Appendix Table), international trade and travel (Table S13 in the Supplemental Appendix Table), and breakdowns or application of incorrect medical procedures (Table S17 in the Supplemental Appendix Table), were robust to the inclusion of sample-bias covariates.

Inadequate public health infrastructure (Tables S17 and S27 in the Supplemental Appendix Table) was significant in 10 of 11 models (with a p value of 0.07 in the 11th). Local livestock production (Table S15 in the Supplemental Appendix Table), sewage management (Table S19 in the Supplemental Appendix Table), and water contamination (Table S23 in the Supplemental Appendix Table) were significantly more likely to be reported drivers in background outbreaks than filovirus outbreaks. Results for three additional drivers, indicated by an asterisk in Table 2, were significant in bivariate analyses but were not robust to the influence of sample-bias covariates.

When filovirus outbreaks were compared with outbreaks of other viruses, seven drivers were found more often in filovirus outbreaks and three were found more often in viral background outbreaks (Table S2 in the Supplemental Materials). These differences were generally not significant in analyses, including sample-bias covariates. However, three drivers, cultural and religious beliefs (Table S33 in the Supplemental Appendix Table), human-animal contact (Tables S37 and S50 in the Supplemental Appendix Table), and intranational trade and travel (Table S39 in the Supplemental Appendix Table), were consistently more common in filovirus outbreaks across all models.

Finally, when compared to outbreaks of other viral hemorrhagic fevers, only four drivers differed significantly (Table S5 in the Supplemental Materials), with weather conditions and local livestock production being less and cultural and religious beliefs and wildlife hunting being more common in filovirus outbreaks. Results for three of these drivers proved highly robust to influence of sample bias covariates (Tables S55, S56, and S60 in the Supplemental Appendix Table). Results for wildlife hunting were not statistically significant when more than three covariates were included in models (Table S61 in the Supplemental Appendix Table). However, this result was also robust to the influence of any individual covariate, perhaps indicating type II error in models with five or six covariates.

Discussion

Filovirus versus background outbreaks

Alexander (2015) and Jacob et al. (2020) review Ebolavirus outbreak drivers, and Kuhn (2008) reviews filovirus outbreaks in general. In addition to these narrative syntheses, there are many accounts of the drivers of individual filovirus outbreaks (e.g., main table in Supplemental Dataset S1). Our study goes a step further in that it quantifies the relative contribution of various drivers to filovirus outbreaks compared to other virus outbreaks and with zoonotic outbreaks in general. Our results support the hypothesis that the driver profile of filovirus outbreaks is distinct from that of most zoonotic pathogens. MCA analyses revealed a pattern where filovirus outbreaks diverge from outbreaks of other zoonotic pathogens along the first principal dimension of variation in multivariate “driver space” (Fig. 3).

Statistical analyses also showed that differences were highly significant, with a p value of 0.004 or less in all omnibus tests comparing Ebola and Marburg to background outbreaks (Table 1). Our background data included outbreaks of 38 diseases from 79 countries, compared to 2 diseases (of 5 viral species) and 17 countries for filovirus outbreaks. The background outbreaks were also globally distributed, with a slight bias toward European and North American countries (Stephens et al. 2021). We included three filovirus outbreaks from Europe and North America, but for the most part, filovirus outbreaks included in our study occurred in Africa. This raised the possibility that many of the patterns we documented could have been driven solely by factors that varied among countries and regions, such as differences in surveillance capabilities, diagnostic resources, or background levels of disease diversity.

However, contingency table analyses and analyses of individual drivers were conducted, including covariates that have been used to represent variation in such country level attributes in previous studies (Smith et al. 2014, Stephens et al. 2021). Omnibus tests were statistically significant regardless of what covariates were included and whether filoviruses were compared to both bacterial and viral (Table 1) or only other viral (Table S2 in the Supplemental Materials) background outbreaks.

We investigated the specific drivers that contributed to this pattern with additional analyses of each driver included in an omnibus test. Regardless of the sample bias covariates included in models, wildlife hunting and consumption (Table S7 in the Supplemental Appendix Table), cultural and religious beliefs (Table S8 in the Supplemental Appendix Table), trade and travel within national boundaries (Table S14 in the Supplemental Appendix Table), and breakdowns or application of incorrect medical procedures (Table S16 in the Supplemental Appendix Table) were reported significantly more often in filovirus than background outbreaks. Inadequate or degraded public health infrastructure was also significant in 10 out of 11 models (Tables S17 and S27, with a in p values of 0.07 in the 11th in the Supplemental Appendix Table) suggesting that the one model in which it was not significant represented type II error.

Results for human-animal contact (Table S10 in the Supplemental Appendix Table) were also robust, and in many cases represented a spillover event related to wildlife hunting or consumption. Consumption of contaminated meat was documented as a contributing factor in 43% of outbreaks where human-animal contact was found to be important.

Results for three factors were also highly robust in the opposite direction, with local livestock production (Table S15 in the Supplemental Appendix Table), sewage management (Table S18 in the Supplemental Appendix Table), and water contamination (Table S27 in the Supplemental Appendix Table) significantly more likely to be reported in background outbreaks compared with filovirus outbreaks. We speculate that this result was related to differences in the primary transmission modes of filoviruses vs background pathogens. Stephens et al. (2021) reported that environmental transmission is by far the most common mode of transmission used by zoonotic pathogens they considered. That study, based on the same background outbreaks as this study, found that 87% of the 300 outbreaks considered were caused by pathogens that can be environmentally transmitted.

In contrast, environmental transmission appears to play a limited role in filovirus transmission (Kuhn 2008, Jacob et al. 2020) with a single example of environmental transmission of Ebolavirus in a laboratory setting (Jaax et al. 1995). In general, filoviruses have not been found to persist long in the environment and direct transmission is by far the predominant transmission mode (Alimonti et al. 2014, Bibby et al. 2015).

When we compared outbreaks of filoviruses to those of other viral pathogens, rather than both viral and bacterial pathogens, filovirus outbreaks were still significantly different from background outbreaks in omnibus tests (Table S2 in the Supplemental Materials). However, fewer and slightly different factors were found to distinguish filovirus outbreaks from those of other viruses (Table S3 in the Supplemental Materials and Tables S31–S54 in the Supplemental Appendix Table). Human-animal contact (Tables S37 and S50 in the Supplemental Appendix Table), cultural and religious beliefs (Table S33 in the Supplemental Appendix Table), and travel within national boundaries (Table S39 in the Supplemental Appendix Table) were still found more frequently in filovirus outbreaks regardless of the sample bias covariates included in models.

Some additional drivers showed mixed results. Wildlife hunting/consumption (Table S47 in the Supplemental Appendix Table, more frequent in filovirus outbreaks) and local livestock production (Table S51 in the Supplemental Appendix Table, more frequent in background outbreaks) differed significantly (α = 0.05) in models with two predictors regardless of which sample bias covariate was included, but were rarely significant in models with three or more predictors (Tables S31 and S40 in the Supplemental Appendix Table). Results were similar for changes in vector abundance (Table S48 vs. Table S34 in the Supplemental Appendix Table) and food contamination (Table S49 vs. Table S36 in the Supplemental Appendix Table), which were found less frequently in filovirus outbreaks, perhaps reflecting a greater importance of environmental transmission among background viruses.

Finally, although they still differed significantly (Table S4 in the Supplemental Materials), the overall driver profile of viral hemorrhagic fevers was more similar to that of filoviruses than other groups of pathogens we considered. Fewer drivers distinguished the two sets of outbreaks (Table S5 in the Supplemental Materials), and in MCA analyses they diverged along the second rather than first principal dimension of multivariate driver space (Fig. S2 in the Supplemental Materials). In this analysis, the first dimension explained 21% of variation in the independent variables and, the second, along which the two sets of outbreak diverged, explained 18% of variation. A number of hemorrhagic fever outbreaks also clustered with filoviral outbreaks even along the second principal orthogonal dimension of driver variation (Fig. S2 in the Supplemental Materials).

However, there was no obvious pattern to the specific outbreaks that clustered with filoviral outbreaks. Outbreaks of Rift valley fever, Lassa Fever, and Crimean Congo hemorrhagic fever, and from both within and outside Africa, were represented. Overall, it appears that outbreaks of other viral hemorrhagic fevers are slightly more likely to have a profile similar to that of an African filovirus outbreak than a zoonotic disease chosen completely at random, although on the whole, purely socioeconomic factors still tend to be less important even for these diseases than for filoviruses.

Common drivers of Filovirus outbreaks

Here, we discuss putative drivers reported in at least five filovirus outbreaks (i.e., >10% of the 45 outbreaks we considered), and how they typically manifested in Ebola and Marburg outbreaks. Note that the drivers we associated with each outbreak should not be considered to have always been documented by a peer-reviewed study, since we also reviewed gray literature such as ProMED e-mails. We also only scored drivers as contributing to an outbreak if specifically reported by at least one source. For example, in the outbreak sources we reviewed, poverty and economic factors were rarely directly discussed, so we scored it somewhat infrequently. However, we speculate that it may contribute at least indirectly to other factors such as inadequate health infrastructure and wildlife hunting that we more often scored as contributing to outbreaks.

Human-animal contact

Ebolavirus persistence in human populations for more than a year following a spillover event seems to be rare (CDC 2021c), and has never been documented in Marburg. The majority of past filovirus outbreaks (excluding laboratory accidents) probably started with human-animal contact (Kuhn 2008), although the importance of human-to-human transmission (from chronically infected cases) has increased in recent years (Keita et al. 2021). However, even in some outbreaks likely to have been zoonotic in origin, no spillover event has been documented, so we did not score human-animal contact. For example, an extensive retrospective contact tracing study of patients in the first recorded Ugandan Ebolavirus outbreak (Francesconi et al. 2003) identified three index cases with no identifiable human source of infection.

Even after conducting in-depth interviews with survivors and other contacts of patients, potential animal sources of infection were not identified. We only scored cases as positive for human-animal contact when at least one reference provides evidence of a spillover source. We also treated the spread of an outbreak to a new country as a separate outbreak event, and such events were generally the result of human-to-human transmission (e.g., Gatherer 2014, Shuaib et al. 2014). Beyond these examples, human-to-human transmission may cause more outbreaks than is generally appreciated. In at least one case, a strain was maintained by transmission in human populations for more than 2 years (Bell 2016, Kaner and Schaack 2016), and recent evidence suggests that a single person may have carried and transmitted Zaire ebolavirus after up to 5 years (McCrae 2021). It remains to be seen how isolated these events were.

When spillover occurs, human-animal contact is often related to wildlife hunting and consumption, particularly of bats and primates (e.g., Georges et al. 1999, WHO 2005, Kuhn 2008, Leroy et al. 2009). Index cases have also been linked to contact with the corpses of infected animals without the consumption of infected meat (e.g., Le Guenno et al. 1995, Georges et al. 1999).

Some index cases have been associated with high bat population densities, and prolonged contact either within buildings or caves where bats roost or during periods of naturally high seasonal abundance (WHO 1978b, Leroy et al. 2009, Amman et al. 2012). Although the evidence for it is tenuous, one outbreak may have started with a family caring for a sick kitten found in a dumpster (Rangamani et al. 2012). The kitten, which showed similar disease symptoms to affected family members, was tested and found negative for a number of common bacterial infections, but it was not tested for Ebola.

In all, documented human-animal contact appears to have directly contributed to at least 37 (82.2%) filovirus outbreaks we included (Table 2). Until the 2014–2016 regional Ebola epidemic, in which international travel and human transmission spread the disease to multiple countries after an initial zoonotic origin in Guinea (Kaner and Schaack 2016), filovirus outbreaks initiated by human-to-human transmission are rare in our data.

Wildlife hunting and food contamination

Wildlife hunting often contributes to infections through consumption of contaminated meat (e.g., Georges et al. 1999, WHO 2005, Kuhn 2008, Leroy et al. 2009), either by the hunter themselves or buyers/handlers/consumers. Wildlife hunting was documented as contributing to 16 of 45 (35.5%) of filovirus outbreaks, and in 13 of these instances consumption of contaminated meat was also documented (i.e., food contamination was also a driver). The importance of wildlife consumption in filovirus outbreaks is striking, given its rarity in background outbreaks (Table 1 and Table S1 in the Supplemental Materials), where it was only implicated in a single outbreak of monkeypox in 2007 in the Republic of Congo (MacNeil et al. 2009).

Oral shedding of infectious Marburg virus by fruit bats has been documented in laboratory settings (Amman et al. 2015), which implies that contamination of fruit or other foods consumed in an area where bats are feeding or roosting could play a role in transmission. However, this has not yet been documented as contributing to a spillover event.

International travel and intranational travel

For any outbreak that spreads to more than one country through human-to-human transmission (e.g., Gatherer 2014, Shuaib et al. 2014), international travel is a contributing factor by definition. In any outbreak within a country that spreads to more than one locality, intranational travel (i.e., travel within national boundaries) may have caused secondary transmission among localities. However, because repeated spillover from a natural widespread reservoir is also possible in such cases, we only included intranational travel as an outbreak driver when movement of infected individuals between localities was specifically mentioned by at least one source (e.g., Gatherer 2014). When we noted either form of travel, it generally referred to travel of infected individuals.

However, in at least one outbreak, travel by members of the international press to affected areas in large numbers disrupted response efforts (Heymann et al. 1999). International trade and travel were documented as a contributing factor in 12 of 45 (26.7%) filovirus outbreaks, while intranational trade and travel were documented in 15 (33.3%) of these outbreaks (Table 1).

Cultural and religious beliefs or practices

We used this to refer to outbreaks starting, being spread, or exacerbated (e.g., in which a driver contributes to diminished effectiveness of response efforts) by either a cultural practice or belief that facilitates disease transmission. In the context of Ebola, funerals and funeral rites are a common source of transmission events (Kuhn 2008). For example, initial cases of Ebolavirus disease in the 2014 Sierra Leone outbreak were mourners at the funeral of a well-known herbalist and faith healer (Koroma and Lv 2015). This initial burst of transmission may have led to as many as 365 deaths. Resistance to containment efforts that conflict with traditional funeral practices, such as cremation of bodies, is also common (Manguvo and Mafuvadze 2015).

Use of traditional healers frequently contributes to outbreaks by delaying disease detection by medical authorities, and because healers themselves and their families may become infected (Georges et al. 1999, Manguvo and Mafuvadze 2015). Containment efforts can also be delayed by mistaken beliefs, such as that the disease is caused by witchcraft rather than by a pathogen (WHO 2012). Finally, in some cases response efforts were hampered by mistrust of individuals from outside local communities, such as health care workers from agencies like the WHO and CDC, or fear of stigmatization within local communities, which led some families to conceal symptoms or hide the corpses of victims from health care workers (Kerstiëns and Matthys 1999). Manguvo and Mafuvadze (2015) present a review of how religions and traditional belief systems can influence Ebolavirus transmission and Ebola virus disease outbreaks. Cultural factors were documented as contributing to 18 (40%) filovirus outbreaks.

Medical procedures

These are cases where either incorrect procedures were followed or there was some failure in the execution of procedures. In early outbreaks, it was common for physicians to know little about the disease that they were treating, and so to use procedures that proved ineffective. For example, in one of the earliest outbreaks Ebola was mistaken for typhoid, yellow fever, and Marburg virus (Bres 1978), and treatments designed for Marburg were ineffective (WHO 1978a). Even more recently, Ebola infections have sometimes been initially misdiagnosed. For example, the index case in the 2014 Nigeria Ebola outbreak was initially treated for malaria (Shuaib et al. 2014). The early course of Ebola infection presents symptoms similar to yellow fever, typhoid, malaria, and other diseases which are extremely common in sub-Saharan Africa (Beeching et al. 2014).

When new species or strains appear, they can also present unusual symptoms. For example, among 26 laboratory confirmed cases treated in a hospital in the first outbreak of Bundibugyo ebolavirus, only 7 showed any hemorrhagic symptoms (Roddy et al. 2012). Patients infected by the strain of Sudan ebolavirus that caused the 2012 outbreak in Uganda also presented diminished or absent bleeding compared to typical late-stage Ebola patients (ProMED 2012a), causing confusion among health care workers. Despite awareness of an outbreak being underway, it was not diagnosed as Ebola until at least 20 patients were infected and 14 had died (ProMED 2012b). Finally, failures of medical procedures also sometimes lead to infections among personnel attempting to follow established procedures (Heymann et al. 1999, Bell 2016). Medical mistakes were documented as a driver of 26 (57.78%) filovirus outbreaks.

Health care facilities/infrastructure

These are cases where a lack of resources, such as too few health care workers, lack of supplies, or facilities that were in poor condition, contributed to an outbreak. These were often instances where capacity was diminished compared to what would be considered normal for an area. For example, after years of conflict, many health care professionals had left Liberia, leading to severe shortfalls in treatment capabilities at the beginning of the 2014 outbreak (Varpilah et al. 2011, Stanturf et al. 2015).

Some areas are sufficiently impoverished that procedures intended to stretch limited resources, such as needle sharing among patients, are commonplace (WHO 1978a). In this category, we also included instances where no surveillance and reporting networks similar to those available in other Ebola countries were present, which made it difficult for authorities to recognize the need for and organize a response. This was most common in countries that had never previously recognized Ebola outbreaks (e.g., Lamunu et al. 2004). Public health infrastructure was documented as a contributing factor in 15 (33.3%) of filovirus outbreaks.

Human population density and urbanization

To date, most filovirus outbreaks have occurred in relatively remote areas. However, urbanization may have played an important role in 2014–2016 Ebola regional epidemic. An outbreak that began Guinea in December 2013 spread to Conakry, the capitol of Guinea, by February 1, 2014 (Kaner and Schaack 2016). The outbreak subsequently spread to urban areas in Liberia, Sierra Leone, and Nigeria (Kaner and Schaack 2016, CDC 2019). In most cases, affected areas of high population density were also urban areas. However, in one case, an entire district in Uganda where an outbreak occurred was described as having a high population density due to concentrations of displaced individuals living in precarious conditions (MacDonald 2000). Human population density was reported as a driver in five filovirus outbreaks, and in four of these outbreaks, relevant areas of high population density were also urban.

War/conflict

There were two main ways in which armed conflicts tended to contribute to outbreaks. In some cases, the medical infrastructure of an area where an outbreak occurred was severely deteriorated due to wars in previous years from which a region had not yet fully recovered (e.g., Stanturf et al. 2015). In other cases, it was difficult for health care workers and other responders to operate in an area where an outbreak was occurring due to active ongoing conflicts (e.g., Maurice 2000). War and/or armed conflict were directly mentioned by at least one source as contributing to five (11.1%) outbreaks.

Miscellaneous stressors related to poverty

Lack of medical resources was often related to poverty in an area. For example, in some cases, hospitals resorted to needle sharing among patients because very few were available (WHO 1978a). Impoverished medical facilities also lacked basic safety supplies such as masks, gowns, and gloves in some cases (e.g., Heymann et al. 1999). In at least one case, a caregiver from an impoverished area attempted treatment strategies that were inexpensive before bringing her ward to qualified health care workers (ProMED 2011). Note that we only score this for a given outbreak if at least one source mentioned it specifically. Among the sources that we reviewed, poverty was only documented as directly contributing to seven outbreaks. However, we speculate that it may have contributed to many more, at least indirectly.

Socioeconomic factors and future work

Overall, perhaps the single most striking result of our study is that socioeconomic factors have played a more prominent role in filovirus outbreaks compared with outbreaks of other zoonotic pathogens (Table 2, Table S2 and Table S4 [in the Supplemental Materials], and Fig. 3). This suggests a critical need to incorporate more detailed data on such factors into statistical models of filovirus spillover or outbreak dynamics. For example, models of filovirus spillover risk have mostly focused on environmental factors such as temperature and rainfall (Schmidt et al. 2017), host diversity (Olivero et al. 2017a), and deforestation (Olivero et al. 2017b).

The only socioeconomic factor that has been directly included in models of spillover or outbreak risk that we are aware of is human population density (Walsh and Haseeb 2015, Olivero et al. 2017b, Schmidt et al. 2017). While this is correlated with many socioeconomic factors (Smith et al. 2007, Ohlan 2015, Walsh and Haseeb 2015, Nieves et al. 2017), more direct data on economic activity and patterns of land use could potentially increase model accuracy. Socioeconomic data could also allow direct testing of many of the hypotheses that our results suggest, for example, that inadequate medical infrastructure is a risk factor for filovirus outbreaks (Table S16 in the Supplemental Appendix Table).

Our results also suggest that economic and social marginalization may be of particular importance, especially when considering filovirus outbreak risk. Variation in economic activity is correlated with a host of socioeconomic factors (e.g., Fuchs 2004, Rothstein and Holmberg 2011, Korotayev et al. 2018, Hill et al. 2019). Economic marginalization, particularly, could increase the chances that a spillover event develops into a full-blown outbreak by reducing the quality and accessibility of health care (Williamson and Fast 1998, Shi and Stevens 2005, Peters et al. 2008), and resources available for detection and containment of spillover events (Rouquet et al. 2005, Singh et al. 2017).

Economic marginalization may also affect spillover risk by increasing reliance on wildlife (De Merode et al. 2004, Davies and Brown 2008, Merson et al. 2019). We note that most areas where filovirus outbreaks have occurred have relatively low per capita GDP (Fig. S3 in the Supplemental Materials) when compared to all countries globally. In contrast, the few outbreaks that have occurred in high-income countries were contained relatively quickly (Kissling et al. 1970, Slenczka 1999, Chevalier et al. 2014). Further, a study in Sudan demonstrated that Ebolavirus transmission rates were roughly doubled in impoverished areas compared to more affluent ones during the 2014 outbreak (Fallah et al. 2015).

These results suggest that more accurate statistical models of filovirus outbreak risk could be developed by including covariates that capture spatial variation in social factors. For example, there are several variables that can be used to quantify economic activity (Rutstein and Johnson 2004, Nordhaus 2006, Gibson et al. 2021) and could thereby identify areas that are absolutely or relatively impoverished. To our knowledge, which covariates best correlate with outbreak risk has not been investigated in the context of filoviruses. The influence of poverty and other socioeconomic factors such as armed conflict (Eck 2012), land use (Gutman et al. 2008, Li et al. 2017), and access to medical resources (Ray and Ebener 2008) represents an essential area for future studies attempting to accurately model filovirus outbreak risk. Regarding the management of outbreaks, our results highlight the importance of cultural and religious factors, which have been discussed by Alexander et al. (2015) and Manguvo and Mafuvadz (2015).

Abramowitz (2017) describes how anthropologists played a key role in the response to the 2014 regional Ebola epidemic and could contribute to the management of future epidemics. Our results also point to a critical need for greater investment in medical resources and training of personnel in countries commonly affected by filoviruses. We found that lack of medical resources, problems related to medical procedures (misdiagnosis and use of incorrect procedures or errors in the application of correct procedures by health care workers), or both were reported to have contributed to 30 of the 45 filovirus outbreaks we considered. Such investments would also obviously have benefits that go well beyond managing filovirus outbreaks.

Availability of Data and Materials

All data used for analyses are included in Supplemental Datasets S1 and S2.

Authors' Contributions

P.R.S., N.G., A.M.S., and J.P.S. designed outbreak driver schema, compiled, and scored background outbreaks. P.R.S. and M.S. compiled and scored filovirus outbreaks. P.R.S. and S.F. compiled data on sample-bias covariates. N.G. and K.F.N. compiled additional outbreak data included in the Supplemental Data. P.R.S. and M.S. conducted statistical analyses. P.R.S. created the first article draft, and all authors contributed to writing the final article.

Author Disclosure Statement

The authors declare that they have no competing interests.

Funding Information

This work was supported by NIH R01Al156866 “Spillover of ebola and other filoviruses at ecological boundaries” and by the UGA Center for the Ecology of Infectious Diseases.

Supplementary Material

Supplemental Data

Supplemental Dataset S1

Supplemental Dataset S2

Supplemental Appendix Table