Bayesian spatial modelling of Ebola outbreaks in Democratic Republic of Congo through the INLA‐SPDE approach

Ebola virus (EBV) disease is a globally acknowledged public health emergency, endemic in the west and equatorial Africa. To understand the epidemiology especially the dynamic pattern of EBV disease, we analyse the EBV case notification data for confirmed cases and reported deaths of the ongoing outbreak in the Democratic Republic of Congo (DRC) between 2018 and 2019, and examined the impact of reported violence on the spread of the virus. Using fully Bayesian geo‐statistical analysis through stochastic partial differential equations (SPDE) allows us to quantify the spatial patterns at every point of the spatial domain. Parameter estimation was based on the integrated nested Laplace approximation (INLA). Our findings revealed a positive association between violent events in the affected areas and the reported EBV cases (posterior mean = 0.024, 95% CI: 0.005, 0.045) and deaths (posterior mean = 0.022, 95% CI: 0.005, 0.041). Translating to an increase of 2.4% and 2.2% in the relative risks of EBV cases and deaths associated with a unit increase in violent events (one additional Ebola case is associated with an average of 45 violent events). We also observed clusters of EBV cases and deaths spread to neighbouring locations in similar manners. Findings from the study are therefore useful for hot spot identification, location‐specific disease surveillance and intervention.


| INTRODUC TI ON
The 2018-2019 human Ebola virus outbreak in several sub-Saharan African countries is especially unprecedented. The largest outbreak, which challenged the health systems of Sierra Leone, Liberia and Guinea, was recorded between 2014 and 2016 (WHO Ebola Response Team, 2016). More than 28,000 cases and 11,000 confirmed dead, mostly from the populations of these three countries, were recorded (Subissi et al., 2018;WHO Ebola Response Team, 2016). Thus, the general populace, especially the healthcare workers in the different West African countries, are exposed to greater risk. A recent study indicated that the short-term (3 and 6 weeks) probability of international spread outside the African continent is small but not negligible and that any further spread in more African countries could enhance the likelihood of global dissemination (Gomes et al., 2014;Pigott et al., 2014).
The outbreak that began in August 2018 is mainly concentrated in three provinces located in the North/South Eastern region of the Democratic Republic of Congo (DRC). As of 15 December 2019, the World Health Organization (WHO) had recorded about 3,348 EBV cases and 2,213 deaths (World Health Organization, 2019b). The outbreak is concentrated in cities and towns housing major domestic and international airports, and it is the tenth and largest in DRC since the first was reported in 1976 (Dyer, 2018;Maxmen, 2018;Wannier et al., 2019). The epidemiology of EBV disease is complex because of the pattern of spread, which is favoured by densely populated areas. The rate at which the outbreak is evolving cannot be overemphasized, and the geographical extension of its spread is widening which, in part, is due to violence targeted at health workers (Wannier et al., 2019).
There is extensive literature on the relationship between conflict and infectious diseases (Adegboye & Danelle, 2014;Gayer et al., 2007;Qadri et al., 2017;Sharara & Kanj, 2014;Stone-Brown, 2013). Armed conflicts provide fertile ground for the emergence and spread of infectious diseases and lead to the destruction of resources and health infrastructure, displacement of large parts of the population to crowded refugee camps and thus accelerating infections (Adegboye & Danelle, 2014). The 2018 EBV outbreak is reported to be occurring in a long-standing conflict zone, and it is suspected that there is a relationship between the number of violent events and the rate of transmission. The transmission rates between the conflict zones suggest that violent events contribute to the increased transmission of the disease (Wannier et al., 2019). During the World Health Assembly at Geneva, Switzerland, the DRC's minister for health, Oly Ilunga Kalenga, told reporters that the country's government is unable to contain the spread of the disease due to the increase in violent events against healthcare facilities. Kalenga stated that 'the real emergency we face right now is security' (Bibbo, 2019).
In this study, we took a One Health approach by mapping the spatial distribution of the 2018 EBV outbreak and conflict events in DRC, adopting a Bayesian kriging approach through the stochastic partial differential equations (SPDE), and parameter estimation was based on the integrated nested Laplace approximation (INLA) proposed by Rue et al.,( 2009). INLA is becoming a popular approach for complex models as applicable in different fields (Schrödle & Held, 2011;Selle et al., 2019;Ugarte et al., 2014) and with SPDE for Spatio-temporal geo-statistical data analysis (Abd Naeeim et al., 2020;Baquero & Machado, 2018;Blangiardo & Cameletti, 2015;Cameletti et al., 2013;Mayfield et al., 2018). For continuous spatial variables, which are only measured at some finite set of specific points, like the case of the Ebola outbreak, it might be useful to predict their values at some unobserved locations within the geographical unit. This approach might play an important role in health risk management in order to understand and identify areas where the risk of exceeding potentially harmful thresholds is higher. Therefore, the present study aimed at quantifying the extent of the 2018-2019 EBV outbreak in DRC and to predict its occurrence at continuous spatial locations. In addition to measuring spatial dependence, we aimed at examining the association between violent events and the spread of EBV in the affected regions. The analysis was performed for reported cases and repeated for recorded deaths.
Our findings will contribute to expanding the present understanding of the transmission dynamics underlying the outbreak that may be useful in the different frontiers of its upsurge, the possibility of its containment and the design of control strategies in the future.

| DRC Ebola outbreak and violent event data
Weekly counts of Ebola cases report at specific locations in the health zones (HZ) ( Figure S1 (Raleigh et al., 2010) and linked them to the locations where the EBV cases were observed based on proximity. ACLED provide real-time data that capture armed and non-armed conflict events, especially in developing countries on the African and Asian continents. It is also widely used for analysis source on political violence and protest around the world (Raleigh et al., 2010).

| Descriptive analysis
We began our analysis with descriptive summaries presented as counts and percentages, and data visualization on maps. Using a One Health approach, we estimated the local indicators of spatial associations (LISA) (Anselin, 1995) using bivariate local Moran's I statistics to assess the correlation of EBV cases and violent events with reference to the spatial location (spatial autocorrelation at each specific health zones) (Saffary et al., 2020). For the spatial weights, we used Queen-style contiguity 1st-order nearest neighbour (i.e. two districts are neighbouring if they share common borders or a point).

| INLA-SPDE
Spatial data can be continuous or areal. The intrinsic conditional autoregressive (CAR) model (Besag et al., 1991) is often used for modelling areal data; it summarizes disease burden of a geographical entity based on its county or regional division. However, it is not flexible enough to capture the complex localized spatial structure likely to be present in the residual spatial autocorrelation (Lee et al., 2014). In the present application, we modelled the observed cases and deaths from EBV at specific locations through the SPDE approach with INLA, adjusting for the explanatory variable, a number of violent events at specific locations. Specifically, we let y i be the reported cases of (or deaths from) EBV in the study locations. The response variables were considered to have come from a Poisson distribution such that. so that where subscript i denotes locations, µ i is the mean of the observed cases/deaths, E(y i ), y i is the observed cases/deaths in each location i, β 0 is the model intercept, β 1 is the linear parameter quantifying the effect of the violent events, and u(s i ) is the spatial random component for the point-referenced data that capture the study area's spatial variability.
The SPDE approach involves representing the continuously indexed spatial process (Gaussian Field (GF)) with a Matérn covariance function as a discretely indexed spatial random process. Thus, a basis function representation provides the link between the GF and Gaussian Markov random field (GMRF) making it easier to use fast computational numerical methods such as the (INLA) as implemented in the R-INLA package (Selle et al., 2019).
The Matérn covariance function is given by.
where ||s i-s j || is the Euclidean distance between any two locations s i and s j, K λ is a modified Bessel function of the second kind and order λ > 0 measures the degree of smoothness of the process, σ 2 is the marginal variance, and k is a scaling parameter related to the range r, which is the distance at which the spatial correlation becomes almost null. The empirical definition is r = The link between SPDE and the Matérn parameters is given by.
The solution to the SPDE represented by the stationary and isotropic Matérn GF (s) is approximated through a basis function representation defined on a triangulation, which divides the spatial domain into a set of non-intersecting triangles, of the domain D: where G is the total number of vertices of the triangulation, ϖ g Is a set of basis functions and ϖ g are zero-mean Gaussian distributed weights. To ascertain a Markov structure, the basis function is chosen to have local support and to be piecewise linear in each triangle so that ϖ g =1 at vertex g and 0 at other vertices (Blangiardo & Cameletti, 2015). In practice, the mesh definition ( Figure S2) is a trade-off between the accuracy of the GMRF representation and computational costs that depend on the number of vertices used in the triangulation.

| Hierarchical models
We considered four hierarchical models and made a comparison Carlo (MCMC) method (Blangiardo & Cameletti, 2015). We employed default priors for GF (Lindgren & Rue, 2015) and penalized priors (Simpson et al., 2017) for the precision parameters' hyperpriors. The marginal posterior distributions for the parameters are presented, including their mean and 95% credible intervals. For details, we refer the reader to Rue et al., (2009) andBlangiardo andCameletti (2015).

| Exploratory data analysis
Of the 3,351 case notification data from DRC for the period April 2019-December 2019, 3,233 were confirmed EBV cases. Table 1 presents the descriptive summaries for the data. During this period, 2099 deaths from EBV were confirmed, yielding a fatality rate of 64.9%. More females (56.3%) were infected with EBV than males, and about 28.1% of the infected were less than 18 years of age.
There were a total of 1878 violent events recorded in the three   For both confirmed EBV cases and deaths, the inclusion of spatial effect (Models 3 and 4) improved the fit compared with the other two models (Models 1 and 2). However, the difference in DIC values for the third and fourth models is less than ten and so none can be adjourned to be better than the other, and normally, the model with fewer parameters should be embraced in such situation. In order to evaluate the effects of the conflict variable (violent events), we present the results of the last model as contained in Table 3 and Figure 3.  of predicted EBV cases and deaths. The maps show that predicted EBV deaths and cases are higher and concentrated in the bordering health zones between Ituri and North Kivu, respectively, after accounting for the number of violent events (Figure 3a,c). The spatial variations of the standard deviations are similar to the posterior mean for both cases and deaths. They range between 0 to 25 for EBV deaths and 0 to 30 for EBV cases. The uncertainty spread out from the same bordering health zones (Figure 3b,3d).

| D ISCUSS I ON
Geo-statistical analysis of infectious disease burden is an essential approach in disease prevention and control strategies because it provides a graphical explanation of the relationship between disease burden, human population and geographical place of residence (Adegboye et al., 2018;Manto, 2005). The analysis of the spatial pat- Our results confirmed that the 2018 Ebola disease outbreak varied spatially with higher concentrations at specific locations and spreads to neighbouring locations. Also, as expected, the pattern of the spatial spread of confirmed EBV cases is similar to that of recorded deaths.
Findings from a recent study (Richardson et al., 2016) (Bausch & Schwarz, 2014). All these take place in the face of a weak and dysfunctional healthcare system (Tomori, 2014).
We found a significant positive association between the number of violent events and EBV cases and deaths. This finding is especially intuitive for Ebola disease since violence leads to the destruction other mitigating factors, such as environmental factors and socioeconomic factors, could be considered in future studies.

| CON CLUS ION
The use of the SPDE approach based on continuous spatial random field to study the spatial distributions of the EBV outbreak in parts of DRC has allowed us to quantify the spread at every point of the spatial domain. Thus, allowing the proper view of the impact at every location, including places where data were not collected. This approach is more advantageous than adopting a method that sums the disease occurrence based on geographical areas. We have expanded on the dynamics of the ongoing EBV outbreak in DRC by predicting its occurrence at continuous spatial locations and, in particular, examined the effects of the conflicts on EBV outbreak in the region.
Findings from the study are therefore useful for hot spot identification, location-specific disease surveillance and intervention, control strategies, health planning and funding allocation for curtailing the disease. Proper disease surveillance and response systems need to be strengthened in the affected areas and indeed in all parts of the country to enhance early detection and control thus, curtailing further spread whether now or in the future.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

E TH I C A L A PPROVA L
This study was based on publicly available data and did not require ethical approval nor informed consent.

DATA AVA I L A B I L I T Y S TAT E M E N T
This study is based on two publicly available data sources, the Ebola disease outbreak news published by the World Health Organization and the Armed Conflict Location & Event Data Project (ACLED).