During recent EVD outbreaks in 2018 to 2021, we quantified the risk of conflict event exposure and found an increased risk at a national level by a similar magnitude in the DRC and Guinea, at 1.88 and 1.98 times, respectively. The uncertainty in the IRR values for Guinea throughout all the analyses was much wider than the DRC. In the distribution of the exposures and events data fit to the model, there was overlap with areas of high reported conflict and EVD case frequency. Several of the regions/territories analysed had higher IRR values than the national risk: Mbandaka, Mwenga, Goma, Butembo (all in the DRC) and N’Zérékoré (in Guinea), which ranged from 2.06 to 4.26.

We evaluated if certain types of conflict may be more impactful than others, to contribute to our understanding of how conflict can change disease risk. The conflict sub-event types with the greatest impacts in terms of EVD risk were armed clashes in Guinea and shelling/artillery/missile attack in the DRC. However, both armed clashes and shelling/artillery/missile attacks had the largest uncertainty. Violent demonstrations had the strongest and most significant impact in Guinea, while in the DRC, peaceful protests, excessive force against protesters, looting/property destruction, protests with interventions and violent demonstrations had the largest, most significant impact, with little uncertainty.

The cornerstone of EVD control is through identifying cases and reducing contact with infected bodily fluids (and vaccination) [8,10]. Conflict can impact these control measures in multiple ways. In previous EVD outbreaks, weak health systems (due to access, supplies, safety of staff and patients and fear), complex socio-cultural and political environments (creating mistrust in governments and NGOs) and ineffective messaging and community engagement have all created problems in EVD response [20,38]. Additionally, there have been difficulties in administering EVD vaccines due to damaged transportation and infrastructure, although continued efforts aim to reach those most at risk [11,16,39]. Conflict and EVD can create a positive feedback loop, and it has been previously observed that conflict increases more during EVD outbreaks compared to controls [18,19]. A suggested explanation is that EVD outbreaks may further destabilise fragile regions and fuel conflicts. For example, mistrust increased in the DRC during the 2018-2019 EVD outbreak due to the government’s decision to ban populations in the affected provinces from participating in the presidential election [18,19].

A potential explanation for the large number of protest-related impacts on EVD cases found here, may be due to these events being more disruptive to a greater number of people. Protests rely on large congregations of people, increasing contact among individuals and therefore fueling EVD transmission. Furthermore, protests and demonstrations may be more likely to impact transportation, education and healthcare, due to the widespread disruption they can cause in dense urban areas [40,41]. Disruption of healthcare and transportation may be particularly damaging in controlling EVD, as it could complicate contact tracing and delay testing, diagnosis and treatment. Further evidence for this explanation may be that within these very impactful sub-events, looting/property destruction was also found highly impactful and important, which is often the product of protests and demonstrations and results in damage to healthcare infrastructure and supplies.

During 2018 to 2021, battles were the furthest away from major settlements in both countries, compared to protests and riots. Proximity may impact the effect of the conflict on the population, both in terms of the number of people affected and the magnitude of the effect. Proximity may also add further evidence to the above hypothesis that protests and demonstrations, although potentially less violent, may be more impactful than battles and armed clashes. Battles (in terms of the ACLED’s definition [29]) are more likely to be fought on pre-defined battle “grounds”, which are less likely to be close to settlements. Alternatively, these settlements, including their healthcare and testing facilities are quickly displaced and abandoned following the onset of conflict due to the immediate threat to life [42]. The absence of testing facilities may result in cases not being identified and reported, therefore making battles appear less impactful to EVD transmission.

A potential explanation for why remote violence/explosions and battles were very impactful but often with higher uncertainty is that these types of conflict occurred less frequently, but when they did, they had a very large impact on EVD cases. The fewer events therefore resulted in the large uncertainty and more data (if available) may help to understand if the larger impact of these conflict types is consistent through time. Additionally, peoples’ behaviour and risks must be considered during more “extreme” forms of violence such as battles and explosions, compared to protests and demonstrations. Violent conflict may have a very large initial impact (accounting for the high IRR values), but then the risk stabilises in the area as civilians take precautions (e.g., displacement) to protect their life [43].

Alternatively, there is the potential for more violent forms of conflict (such as battles and explosions) to decrease the risk of EVD. People may be more likely to stay at home during these events or be subject to curfews and not mix in public spaces, decreasing their exposure to the pathogen [44]. Furthermore, severe conflict may be more fatal (of both civilians and combatants), reducing the susceptible pool for EVD cases. More work is needed to understand how to quantity conflict severity for scientific research purposes, which could be used to identify dose-response relationships with disease, while still considering that conflicts impact individuals differently in terms of disruption and severity.

When extending the exposure periods, the effect of conflict on EVD cases was almost eliminated by 10 weeks in the DRC, while in Guinea, the risk of EVD cases decreased to 4 weeks and then increased by 10 weeks. However, it appears that extending the lag periods may have been diluting the effect of conflict on EVD cases, as changing the lag periods to a single week made all the lag periods for both countries more significant and impactful. These results track with previous research which found synchronicity and correlation between the reporting of EVD cases and conflict in the DRC and suggest that the effect of conflict may be long-lasting [11,19].

The results show that conflict is still highly influential multiple months following the event, and in some instances, conflict had a greater and more significant effect on EVD with time. A potential explanation may be due to the disruption caused by the conflict event having a cumulative effect. Previous findings suggest that conflict lengthens the EVD treatment process, e.g., longer times to detection, isolation, treatment and vaccination, creating a lasting effect and prolonging outbreaks [18]. Therefore, time would be needed for EVD cases to increase and be reported, after elevated transmission.

There are several limitations to the datasets used here, the first being the assumption that if EVD or conflict occurs, it is correctly identified (e.g., confirmed and suspected EVD cases were included) and reported on the correct day with no delay, where the transmission event/symptom onset occurred. These assumptions are unlikely, as EVD cases may be missed if cases never seek formal medical assistance or are in hard-to-reach areas. Furthermore, the categorisation of conflict events is subjective (category definitions were included in Additional file 1) and several categories are ambiguous. Data availability was limited and sub-setting the data (by sub-national unit or conflict sub-event type) led to small sample sizes in some instances, resulting in wide uncertainty, statistical insignificance or the models not being able to fit, particularly in Guinea. However, generally the findings were replicated in both countries, in terms of sub-event type, the effects of lag and geographic distribution of risk.

The methodology accounts for confounders by fixing them across the administrative unit. It is not as effective across multiple administrative units. Sub-national units are different geographical sizes, with different population sizes and risk factors. Therefore, comparing different regions/territories can create misleading results. Additionally, the use of vaccination during these more recent outbreaks [45,46], makes comparison to previous work challenging. While data on Ebola vaccination are available, the use of ring vaccination makes it difficult to incorporate vaccination data into the model used here (as a binary outcome) [38,46]. For example, the sub-national region with the largest IRR here was Équateur, part of the DRC which has received relatively little vaccination, compared to the eastern provinces [38].