Introduction

Defined as the “intentional use of physical force or power threatened or actual, against oneself, another person, or a group or community” (Krug et al., 2002, p. 5), violence has numerous adverse consequences, ranging from destruction to death. The economic impact of the consequences of violence amounted to an estimated 13.3% of the global gross domestic product (GDP) in 2015 (OECD, 2016). The edited volume of Suder (2004) in the aftermath of the 9/11 terrorist attacks in the United States (US) was arguably the beginning of research into the effects of violence on international business (IB). In an increasingly interconnected world with armed conflicts spilling over into neighboring states, international criminal networks, and terrorism, no country is safe from violence and its effects (Imbusch & Veit, 2011). However, violence presents a particular scourge for countries trying to overcome economic development-related challenges.

The typical concerns of IB scholars, namely the specificities of individual countries (Kibria et al., 2020; Moore, 2021) and firms (Dai et al., 2023; Oetzel & Getz, 2012), have been dealt with in scholarship seeking to understand violence in IB. However, the intractability of violence makes for extreme complexity, and evidence suggests violence can deter (Li & Vashchilko, 2010; Witte et al., 2017) or accelerate (Skovoroda et al., 2019) inward foreign direct investment (IFDI), and that the relationship also functions in the reverse (Pinto & Zhu, 2022). This leaves practitioners, scholars, and policymakers in a bind: what is certain and thus actionable in a violent country?

In this paper, we use uncertainty as our point of departure. We examine violence in IB using Knight’s (1921) distinction between risk as something that is measurable and can be assigned a probability, and uncertainty as what is unmeasurable; what Phan and Wood (2020: 429) referred to as “unknown-knowns.” Most IB research is about the risks associated with violence (e.g., Busse & Hefeker, 2007; Oh & Oetzel, 2017) and scholars are often concerned with how that risk can be managed (e.g., Dai, 2009; Driffield et al., 2013; Oetzel & Oh, 2014). We do not work in that tradition but instead contribute to the smaller body of work theorizing the uncertainty associated with violence (e.g., Cornwell et al., 2023; Hiatt & Sine, 2014). Consequently, we build on the finding of Oh and Oetzel (2017) that sometimes neither experience with conflict in general nor country-specific experience provides learning advantages to multinational enterprises (MNEs). When that happens, we suggest violence is characterized by Knightian uncertainty and cannot be mitigated in the way most IB research suggests.

Systematic research on violence, conflict, and peace emerged in political science research after World War II. However, from the outset, the field has been interdisciplinary, with important contributions stemming from international relations (e.g., Midlarsky et al., 1980), sociology (e.g., Collins, 1974), and economics (e.g., North et al., 2009). This work has been done amidst an ongoing struggle for conceptual clarity, given the multidimensional nature of violence. For instance, the influential work of Galtung (1969) differentiated between violence dimensions, such as personal and structural, latent and actual, and physical and psychological. It is however evident that violent events are often characterized by many of these dimensions at the same time. Half a century later, Balcells and Stanton (2021) demonstrated that the macro-/micro-level divide (across or within a country or conflict-afflicted area) often does not hold. Further, Valentino (2014) showed that clarity is lacking about the motives or conditions under which violence is more likely. However, there is scholarly consensus that violence is “primarily, if not exclusively” orchestrated by powerful actors for specific purposes.

The topic of violence against civilians has received extensive attention from scholars studying peace and violence, as it has proven particularly hard to understand. Violence against civilians involves one-sided, deliberate violent acts perpetrated against unarmed non-combatants (Raleigh et al., 2010). It can originate internationally or be driven by individuals, and every permutation in between (Balcells & Stanton, 2021). Violence can be perpetrated by insurgents or incumbents (even governments during contested periods), while groups of different actors like rebels, militias, and government troops can work together or stand in conflict against each other to perpetrate violence against civilians (Raleigh, 2012). Welsh (2023) found the interaction between territorial control and competition (between different militias or between militias and government agents) introduces drastic changes in both the locations of more versus less intense violence against civilians and the timing of such incidents.

To be effective in these locations, businesses need to understand contested territories, who are powerful versus power-hungry actors, and the most recent configuration of shifting allegiances. Substantial managerial effort in conflict-torn contexts is directed at doing that (Cornwell et al., 2023). However, MNEs are disadvantaged by the inherent intractability of violence and by the fact that they are foreign. Cornwell et al. (2023) highlighted the costs and difficulties of collecting information about violent contexts, the reliance of MNEs on local employees for validating such information, and also that MNEs leave conflict-ridden locations if they fail to obtain greater knowledge of local conditions.

As DeGhetto et al. (2020, p. 7) explained, “executives do not want to endanger themselves, their expatriates, their local employees, or their firms’ assets.” This is consistent with Knight (1921: 3.VII.44), who argued that under conditions of uncertainty, “the decisions of responsible business managers, for the most part, have little similarity with conclusions reached by exhaustive analysis and accurate measurement.” Therefore, we suggest that Knightian uncertain violence deters investment in countries grappling with such violence.

Accordingly, in this paper, we test three hypotheses about Knightian uncertain violence. We suggest that (1) volatility in the number of violent events against civilians, (2) the number of civilian fatalities, and (3) their interaction, have a negative impact on IFDI flows. Using a range of measures and analytic approaches, we examine 48 African countries from 1997 to 2021 and find support for all our hypotheses. Intriguingly, in our post hoc analysis, we also find that exporting has the opposite effect, where Knightian uncertain violent events are associated with greater levels of exports, a finding we further explore.

Our work introduces to IB an overview of the literature on violence from various disciplines and a more in-depth engagement with IB research on the same topic. In particular, through our consideration of Knightian uncertainty in IB, we help clarify the apparent anomaly of why MNEs sometimes learn to operate in violent contexts and why violence occasionally deters IFDI. Our findings thus add important additional nuances to the long-standing body of work about the promise, albeit also caveats, of IFDI-assisted development (Narula & Driffield, 2012; Narula & Pineli, 2019). We reflect on how governments can respond when confronted with the challenges of high levels of violence against civilians, high levels of civilian fatalities, and substantial volatility in violence in their countries, and suggest numerous avenues for future research.

The remainder of the paper is organized as follows. The next section contains the literature review and hypotheses development. After that, we describe the research setting and the methodology. Then, we present the study’s results, followed by a discussion thereof and policy recommendations. Finally, the conclusion is presented.

Literature review and hypotheses development

In this section, we provide an overview of violence and peace research, before discussing how IB scholars have approached the topic of violence. We discuss the notion of Knightian uncertainty, using it to develop our hypotheses.

Better understanding of violence

In this paper, we focus on violence against civilians, which involves one-sided, deliberate violent acts perpetrated against unarmed non-combatants (Raleigh et al., 2010). It is characterized by asymmetry, as the victim is unarmed and not in a position to counter the violence, and the perpetrator is deemed the only actor using violence. Civilians can be violently targeted across various types of violent events, making this type of violent incident particularly insidious. In addition, more than 50 years of research on violence has provided few useful conceptual categories, instead highlighting the complexities associated with seeking to categorize or predict violence.

Research on political violence and conflict has always been interdisciplinary, with contributions from international relations (e.g., Midlarsky et al., 1980), sociology (e.g., Collins, 1974), and economics (e.g., North et al., 2009). The current consensus is that competition between powerful and/or power-hungry parties for control over populations and/or territories is at the heart of politically motivated violent events (Raleigh, 2012; Valentino, 2014; Welsh, 2023). Valentino (2014, p. 91) suggested that political violence is “primarily, if not exclusively” orchestrated by powerful actors for specific purposes, but indicated that there is not yet clarity about the motives or conditions under which violence is more likely. Civilians are affected by political violence: Welsh (2023) used geolocation mapping over time to find that higher levels of civilian targeting result from an interaction between territorial control and competition at the subnational level between different militias or between militias and government agencies.

As violence research has advanced, it has become increasingly clear that political-institutional and socio-economic violence is deeply intertwined (Fox & Hoelscher, 2012; Imbusch et al., 2011). To give an illustration: In July 2021, South Africa’s ex-president Zuma was jailed for contempt of court. This triggered rioting across South Africa that caused billions in damages and killed more than 350 people, most in just 3 days (Africa et al., 2021). However, although some commentators explained the rioting as political, others regarded it as essentially socio-economic in natureFootnote 1.

The work on social violence has evolved separately, at the same time as work on politically motivated violence, and was also interdisciplinary from the outset, involving especially psychological and sociological perspectives (Wolfgang & Ferracuti, 1967/2001). Social violence refers “to a broader manifestation of grievances, criminal behaviors, and interpersonal violence in society,” such as crime, homicides, and interpersonal and self-directed violence (MacClinchy & Raleigh, 2016, p. 20). An important trigger for social violence is poverty (Braithwaite et al., 2016), with studies on droughts (Miguel et al., 2004), income (Braithwaite et al., 2014), and food prices (Fielding & Shortland, 2010; Smith, 2014) all finding a link to violence.

Regardless of whether violence is a response to political or social issues, violence can trigger a conflict trap. A conflict trap emerges when violent conflict perpetuates itself because it negatively changes the society in which it occurs (Hegre et al., 2017). For example, violence reduces educational opportunities (Kibris, 2015), including via higher teacher absenteeism and turnover (Jarillo et al., 2016), a greater propensity for children to leave school early (Rodríguez & Sánchez, 2012), and the destruction of physical infrastructure (Barrera & Ibañez, 2004). Conflict traps happen distressingly often. Studying 103 countries that experienced a civil war between 1945 and 2009, Walter (2011) found that only 44 countries escaped returning to civil war. Since 2003, every civil war has been the extension of a previous civil war.

In short, as much as scholars of peace and violence are making progress in teasing out what triggers violence and when violence is likely to occur, the intractability of violence in the empirical world is mirrored by its intractability in attempts to gain a scholarly understanding of the phenomenon. It is known that both political and social motives can trigger violence against civilians, and general explanations have been identified (e.g., the search for power or grievances of hungry people). However, there is not yet a robust scholarly understanding of when and why unarmed civilians going about their business become victims of violent attacks.

Violence in international business research

Given that violence is hard to theorize, it is unsurprising that matters get even more complicated when adding another multidimensional consideration, namely business across borders. To get a sense of the state of current knowledge, we did a systematic search of the seven leading IB journals—Journal of International Business Studies, Journal of World Business, Global Strategy Journal, Journal of International Management, Management International Review, International Business Review, and Journal of International Business Policy—from 2001 onwards. “Violence” as a search term in the title or abstract only yielded eight results, while broadening the search to include “conflict” increased the number to 40 relevant papers (see Appendix 1). Half of the papers appeared from 2019 onwards, suggesting that the field has become of increasing interest to IB scholars in recent years. The papers in Appendix 1 provide a useful indication of the state of current knowledge in the field, but we cite more broadly in our discussion of themes. Already in 2004, the Journal of Peace Research published a paper on how violence affects investment location decisions when Fielding (2004) showed that capital flight both resulted from and predicted violent incidents. In IB, Dai (2009) examined MNE strategy during interstate warfare, yet the work was not published in a mainstream IB journal and remains little cited.

Interest in understanding the effects of violence on IB was triggered by the 9/11 terrorist attacks in the US (Czinkota et al., 2010; Suder, 2004), an event with dimensions of both manageable risk and non-manageable uncertainty (Liesch et al., 2006; Phan & Wood, 2020). However, it took a while for empirical work on terrorism to start appearing. Terrorism is regarded as “the use of violence or the threat of violence to attain political or ideological goals and the willingness to attack noncombatants” (Oh & Oetzel, 2011, p. 660). Early work focused on the psychological impact of terrorism on employees (Bader & Berg, 2013; Bader & Schuster, 2015), with later authors investigating firm-level (Abrahms et al., 2019, 2023; Liu et al., 2022) and more macro outcomes (Dimitrova et al., 2022; Jiménez & Lupton, 2021; Osgood & Simonelli, 2020).

As is already evident from the work on terrorism, IB researchers have mainly considered violence with political roots. Only a few IB papers focus on violence driven by socio-economic factors, with Ramos and Ashby (2013, 2017) making a crucial contribution with their work on the effect of organized crime in Mexico on IB. Unmanaged migration (Reade et al., 2019) is another paper in that vein. Most extant IB research on violence examines political conflict, namely the “use of force towards a political end that is perpetrated to advance the position of a person or group defined by their political position in society” (MacClinchy & Raleigh, 2016, p. 20). Scholars have focused on how IB is affected by military dyadic conflicts (Li et al., 2020; Li & Vashchilko, 2010) or different types of war (Chen, 2017; Oh & Oetzel, 2017).

An important exception to the emphasis on the political roots of violence is found in papers that conduct an in-depth examination of a single country (or group of countries). While the political dimension is not absent in those papers, there is general recognition of the multiple co-existing factors that together create a violent context (e.g., Alaydi et al., 2021; Barnard & Luiz, 2018; Luiz et al., 2019; Parente et al., 2019). Consistent with the findings of violence scholars (Fox & Hoelscher, 2012), researchers find that elements like social dissent and protest, poverty, and fragile institutions function together in response to (often misguided) political decisions to create a context that is hard for businesses to navigate.

The diverse types and mechanisms of violence mutually affect the complexities of not only operating in different firms and industries but also across varying cultures and institutions. This does not allow for easy categorization. We hone in on a single dimension: the Knightian distinction between risk and uncertainty.

Knightian risk and uncertainty in international business research

Knight (1921) argued for the need to differentiate between risk and uncertainty. Risk is measurable and can be managed with routines, standard operating procedures, and other such expectations of a (relatively) predictable future. In contrast, uncertainty is not measurable. Instead, under conditions of uncertainty—and lacking some probabilistic measure to guide decision-making—opinions, estimates, and convictions form the basis of decisions.

Perhaps because uncertainty involves seeking to understand “unknown-knowns” (Phan & Wood, 2020, p. 429), it has not received much scholarly attention in management research. Scholars have used strong statements to describe it: “events so massive that they defy imagination” (Phan & Wood, 2020, p. 429) and “a hall of mirrors” (Alvarez & Porac, 2020, p. 735). As more work on the topic emerges (e.g., Feduzi et al., 2022; Mousavi & Gigerenzer, 2014; Packard & Clark, 2020; Rindova & Courtney, 2020), strong emphasis is placed on knowledge and the use of decision-making heuristics. This is consistent with Knight (1921: III.VIII.41), who proposed the “most thoroughgoing methods of dealing with uncertainty; i.e., by securing better knowledge of and control over the future.”

In IB research, the distinction between uncertainty and risk has been mentioned, rather than theorized (e.g., Buckley, 2018; Forsgren, 2016; Vahlne & Johanson, 2017; see Young et al., 2018 for an exception). In the discussion of violence, the emphasis tends to be on risk in the environment, especially political (e.g.,Henisz et al, 2010; Shan et al., 2018), and on a range of firm capabilities to deal with such risk (Albino-Pimentel et al., 2021; Buckley et al., 2016; Lu et al., 2018; Stevens & Newenham-Kahindi, 2017).

This emphasis is understandable, as management scholars are primarily interested in issues that can be managed. Nevertheless, in seeking to understand emerging MNEs’ apparent greater tolerance for risk, Buckley et al. (2018) differentiated between controllable and “noncontrollable” risks, such as political instability. Noncontrollable risk is conceptually close to Knightian uncertainty and suggests that sometimes events defy managerial intervention. In IB literature on violence, several observations have suggested the presence of uncertainty, rather than risk.

For example, in their paper on “dodging bullets,” Witte et al., (2017, p. 866) proposed that the “discontinuous risk of infrequent and episodic events is closer to pure uncertainty than continuous, Knightian risk of predictable events.” Oh and Oetzel (2017) did not use a Knightian lens, but uncertainty (rather than risk) seems the more likely explanation for the finding that neither experience with conflict in general, nor country-specific experience appear to provide benefits for MNEs when conflicts occur between organized non-governmental armed groups. Our paper contributes to this smaller body of work.

Hypotheses development

To develop our argument, we focused on violence against civilians. Although MNEs, to some extent, benefit from being outsiders in politically violent countries (Dai et al., 2013, p. 572), the benefit disappears when violence is directed at civilians. This is because violence against civilians is asymmetric and there is no benefit to seeking to remain outside of a conflict. Given the concerns of managers about MNE employees in violent locations (Cornwell et al., 2023; DeGhetto et al., 2020), managers are likely to be particularly concerned about violence directed at non-aggressors.

Violence literature provides ample evidence of the uncertainty associated with violence against civilians. Welsh (2023) argued that the groups perpetrating such violence seek to display dominance and punish defectors and that violence is greater in locations where groups control territory or face competition, but less when both occur, because of violent groups’ need to ensure civilian support. As for the composition of perpetrator groups, Raleigh (2012, p. 463) mentioned a bewildering array of parties:

Both rebel/insurgents and governments/incumbents use VAC [violence against civilians] during periods of major unrest (for example, civil war), and rebels are by far the most violent. Yet, VAC patterns are complicated by the coexistence of multiple agents within civil wars, including government troops, multiple rebels, and militia groups. While rebels may pose a more direct danger to civilians, governments often use informal militia groups to conduct VAC in both controlled and frontline contested zones during civil wars (for example, Janjaweed in Sudan; Mayi-Mayi in DR-Congo; FESCI in Cote D’Ivoire). Governments also rely on VAC in non-war contexts: a high proportion of violence against civilians takes place outside of civil wars, in states experiencing elections, internal pressures, or government repression.

Hence, there is extensive uncertainty in the drivers, locations, and timing of violence against civilians. The intuition that such uncertainty will affect business is supported by Hiatt and Sine’s (2014) description of how a small mechanic business in Medellín, Colombia, was affected during the worst era of drug-related conflict. Drug-related battles would erupt all over the city, but there was no telling when and in which parts of the city the battles would happen. The erratic eruption of violence induced uncertainty and customers did not bring in their cars for servicing at the appointed time. The times when customers felt safe to travel were not necessarily times when the mechanic’s employees felt safe to do so; they also experienced uncertainty about when and where violence would erupt. Consequently, when customers showed up, the mechanic’s business was often understaffed.

At the MNE level, Cornwell et al. (2023) showed that in a conflict-ridden context, firm-level factors are most salient—for instance, MNEs’ experience with violent contexts or their relational connections to local networks and public officials through which they can obtain information. This is followed by host-country factors—for example, the constellations and attitudes of actors or the geographic scope, intensity, and duration of conflicts, with only a minimal role played by the international arena. For businesses to manage such complexity, they need to understand contested territories, who are powerful versus power-hungry, and the most recent configuration of shifting allegiances. In a violent context, the world of a manager shrinks; one manager commented that bombings 50 miles away “may have been the other side of the world” (Cornwell et al., 2023, p. 12).

Disadvantaged not only by the inherent intractability of violence but further by foreignness, managers constantly assessed factors like transportation infrastructure, such as “blocked roads from demonstrations, bombing and fighter presence at critical infrastructure” (Cornwell et al., 2023, p. 13); the flexibility or not of operations, such as reserves of inventory and geographical dispersion of locations; and the tone used to refer to the MNE in the local media. Knight (1921: III.VIII.1) explained this type of situation as follows:

It is impossible to form [assess] a group of instances because the situation dealt with is to a high degree unique. The best example of uncertainty is in the exercise of judgment or the formation of those opinions as to the future course of events, which opinions (and not scientific knowledge) guide most of our conduct.

The need for such local and personalized (rather than contextualized and institutional) information in uncertain contexts has long been known (Mascarenhas, 1982) and is consistent with Knightian views of uncertainty. Knight (1921: III.X.2) underlined the need for relationships and personal verification to increase “the possibility of knowing the accuracy of other men’s [sic] knowledge”. As hard as it may be to make sense of a violent environment, it is even harder when it is highly changeable. Given investors’ almost axiomatic preference for certainty (Barnard & Luiz, 2018; Canh et al., 2020; Chari & Banalieva, 2015), we expect that not only violence per se, but specifically changes in the violence (e.g., violent attacks against civilians that seem to be abating before suddenly increasing again, or even a rapid but inexplicable reduction in high levels of violence) will be of concern. As explained by Cornwell et al., (2023, p. 17), “a slight change in conflict-related events can quickly alter the situation.”

The uncertainty associated with civilian deaths poses a particular challenge, as death is irreversible. Knight (1921: III.VII.44) spoke of “responsible business managers.” To this end, DeGhetto et al. (2020) developed a measure for what they termed safety risk and found that high levels of safety risk deter investment by MNEs. They concluded that “multinational corporations [MNCs] are concerned about safety risk because it creates uncertainty and endangers the physical security of MNCs’ personnel and assets” (DeGhetto et al., 2020, p. 7). MNE employees’ security is imperative not only in terms of their welfare but also from a strategic perspective. In a conflict-ridden location, “employees are paramount—an MNE can be endowed with information access, reconfiguration capabilities, and stock flexibility, but if employees are unwilling to remain, such endowments become redundant” (Cornwell et al., 2023, p. 16). Therefore, in terms of IFDI, we hypothesize:

Hypothesis 1

The greater the volatility in the number of violent events against civilians, the greater the negative impact on IFDI flows.

Hypothesis 2

The greater the number of civilian deaths, the greater the negative impact on IFDI flows.

Finally, we argue that there is an interaction effect, that the most uncertain contexts are those with both high levels of civilian fatalities and high volatility in violence. To explain our reasoning, consider an MNE with an existing subsidiary in a location where a relatively peaceful period was suddenly disrupted by violent attacks, or where peace has seemingly descended on a long-violent location. As explained by Steen et al. (2006, p. 309), international trade and investment networks are “complex systems and complex systems carry the property of dissipative structures where non-linear processes have the potential to produce unpredictable future outcomes.” Similarly, Cornwell et al. (2023) highlighted the interrelationships between different factors and the extent to which even small changes can alter the overall situation.

Given that violence produces unpredictable results, sending employees to such a location for even routine activities—e.g., introducing new initiatives, providing maintenance of specialized machinery, or reviewing operations—will likely be delayed until the MNE has greater certainty about what is going on. If this was also a location with high levels of civilian fatalities, the MNE would have to decide whether to deploy employees in a location characterized by highly consequential and irreversible events, namely civilian fatalities. This additional consideration is likely to substantially magnify the concerns of the MNE.

In this paper, we are not concerned with the choices of MNEs already in a location; we are looking at de novo FDI inflows. However, the example is useful to highlight how the confidence of investors, stemming from the fact that they possess some useful knowledge about a location, is likely to be eroded by conditions swinging from relative peace to numerous incidents of violent conflict and back to greater peace again. If investors are, at the same time, also confronted with events they cannot mitigate—civilian deaths—they will likely judge a location as simply not amenable to investment. This leads to our final hypothesis:

Hypothesis 3

The interaction between the volatility in violence against civilians and the number of civilian deaths has an even greater negative impact on IFDI flows.

In short, locations will be judged as uncertain to the extent that violent conditions appear to be constantly reversing course—whether getting better or worse—and where a substantial number of civilian fatalities occur. Such a deep level of uncertainty will deter IFDI.

Research setting

This study is set in Africa, spanning 25 years from 1997 to 2021. Comprising 54 countries (of which 48 are included in this study) with considerable institutional variance (Fon et al., 2021), Africa is a continent of stark contrasts in terms of overall economic development.Footnote 2 The 2021 average African GDP per capita was US $1645, with Burundi being the lowest (US$237) and countries like Gabon, Botswana, and oil-rich Equatorial Guinea all over US $7300 (with some island economies’ averages even higher). The average GDP growth in Africa was 4.1% in 2021, with lows of – 10.8% (South Sudan) and – 3.5% (Congo), and highs of 7.5% (Kenya), 8.7% (Eritrea), 10.9% (Rwanda), and 11.4% (Botswana). Seventy percent of Africans have completed primary school, varying from 27% in South Sudan, 41% in Chad and Equatorial Guinea to—the island economies aside – 100% in Kenya and more than 90% in Ghana along with most of Southern Africa. Internet usage in Africa is at 30%, with only 1% of Eritreans, but 70% of South Africans having access to the Internet.

We use Africa as our research setting for the following reasons. First, while not the most violent continent—that dubious recognition goes to Latin America (Imbusch et al., 2011)—Africa is characterized by a colonial history (Lange & Dawson, 2009) and widespread poverty (Miguel, 2007), both known contributors to violence. Second, attracting FDI inflows matters for African countries because they have scarce capital and high unemployment rates (Marandu et al., 2019). Moreover, recent research has found that Africa has the highest IFDI growth rate globally (Lu et al., 2018). Thus, the complex African terrain offers ample opportunity to advance our understanding of IB (Barnard et al., 2023).

Table 1 shows the trends of violence against civilians experienced by the countries in our sample. Evidence from the Armed Conflict Location and Event Data Project (ACLED) database suggests that African countries can be clustered broadly into five groups: (1) improving violence, (2) worsening violence, (3) remitting-relapsing improvement in violence, (4) remitting-relapsing worsening of violence, and (5) ongoing turbulence in terms of levels of violence.

Table 1 Trends in attacks on civilians in Africa (1997–2021)
Full size table

A small number of countries (7) have experienced a steady or only occasionally relapsing reduction in attacks on civilians. Countries such as Angola, Eritrea, and Sierra Leone have succeeded in lowering their levels of violence against civilians. For example, Angola emerged from a 27-year civil war in 2002 (Musacchio et al., 2012). Almost the same number of countries (eight) experienced ongoing turbulence in violence with no discernible trend. Examples are countries like Egypt and Zimbabwe with its hyperinflation crisis (Hanke & Kwok, 2009).

Consistent with the literature on conflict traps, 13 countries experienced a steady increase in attacks on civilians, such as Algeria, Cameroon, Morocco, and Nigeria. For example, Nigeria has been increasingly plagued by political violence driven by, among others, Boko Haram since the 2000s (Obadare, 2022), while in Cameroon, hundreds of thousands of civilians have been displaced due to conflict between government and armed forces since 2017 (Orock, 2022).

However, most countries (20) have experienced a remitting-relapsing trend of increasing violence against civilians. Underlining how hard it is to overcome violence, these countries underwent periods during which attacks decreased. However, these periods were followed by a renewed and intensified trend of attacks on civilians. The South African example highlights that ending violence seldom happens smoothly. Although the end of apartheid brought an end to violence, the lack of service delivery by the ruling party soon resulted in a new wave of social unrest and violence (Luiz & Barnard, 2022). A similar trend is evident in Tunisia after the Arab Spring (El-Haddad, 2020).

The trends do not capture the intensity of violence. Therefore, Fig. 1 illustrates the overall fatalities of civilians during 1997–2021, normalized by population as of 2010. It is important to note that like violent events against civilians, fatalities vary over time and the total number of fatalities from 1997 to 2021 is represented here. Nonetheless, the evidence is useful for putting the trends into perspective.

Fig. 1
figure 1

Source ACLED database

Total fatalities between 1997 and 2021 (per 100,000 people).

Full size image

For instance, Table 1 shows that violence against civilians in Botswana follows a remitting-relapsing worsening trend, and in Morocco a steadily worsening trend. However, these two countries have the lowest intensity of violence on the continent, with Botswana and Morocco having a fatality rate of five and seven deaths per 100,000 people, respectively. Similarly, countries with very high levels of violence include Eritrea, Rwanda, Angola, and Libya, although these countries have demonstrated improvements in the levels of violence against civilians experienced over the past quarter century. The country with the highest level of fatalities is Eritrea, with 24,241 deaths per 100,000 people over 25 years. This represents around a full percent of the population killed per year (more during more violent periods). The evidence points to extremely high levels of violence during violent periods.

Methodology

The question of how to measure uncertainty has long challenged scholars and there is still no consensus thereon. Nonetheless, existing literature finds very different empirical outcomes when proxy measures tap into uncertainty, rather than risk (Nishimura & Ozaki, 2004, 2007). Similar differences in empirical outcomes can be expected for risk versus uncertainty from the perspective of MNEs. When violence presents a risk, it is possible to develop strategies to manage it, but that is not the case when violence is uncertain. Therefore, we deemed it crucial to seek to measure the uncertainty in violence.

For statistical analysis to be meaningful, it is necessary to identify occurrences with which systematic patterns can be associated after the fact but, conceptually, it is important that they did not appear systematic or predictable at the time the events occurred. The most common approach for measuring uncertainty has been via several variance measures (e.g., Cascaldi-Garcia et al., 2023; Goodell et al., 2021). It is imperative to note that variance can only be used as a measure of uncertainty under certain circumstances. Volatility in fields like finance and economics [“the dispersion of short-term shocks around a long-term mean,” as per Aït-Sahalia et al. (2021, p. 1)] does not necessarily trigger uncertainty (i.e., the difficulty of forecasting a distribution). Volatility in a field like finance is often used to signal risk, which is understood as having an upside and a downside (i.e., higher risk, but higher return).

When volatility measures are used to measure Knightian uncertainty, the emphasis is on the extent to which “disagreement, disconformity, and discord” (Dibiasi & Iselin, 2021, p. 2115) can surface. Thus, we suggest that volatility in the context of violence functions differently from how it functions in a field like finance. Considering how the conflict trap functions (Hegre & Nygård, 2015) and that global trade and investment networks are complex systems characterized by dissipative structures and non-linearity in processes and outcomes (Steen et al., 2006), volatility in violence could trigger unforeseen consequences (i.e., it has uncertain potential outcomes).

Our choice of the changeability of levels of violence against civilians and civilian deaths is consistent with the literature on violence (Valentino, 2014; Welsh, 2023) that emphasizes the numerous, interwoven motives affecting both. These measures also capture the disconformity and discord associated with Knightian uncertainty measures. From the perspective of MNEs, there is generally an expectation (or at least a hope) that violent conditions—if present—will improve and that there will be no fatalities (DeGhetto et al., 2020). Knightian uncertainty occurs to the extent that such a hope proves to be misplaced.

Data

To empirically test our hypotheses on the relationship between violence and FDI inflows, we used data on 48 African countries from 1997 to 2021.Footnote 3 Our data on violence were sourced from the ACLED database (Raleigh et al., 2010; Ahmed et al. 2011; Armed Conflict Location & Event Data Project 2019; Sahoo and Dash 2022), a database created by Prof. Clionadh Raleigh from the University of Sussex. The events-based dataset initially focused on African countries but has since expanded to include all regions in the world. Real-time data on violence are collected by researchers from several local and regional sources, such as media publications, ReliefWeb, Factiva, government and non-governmental organization reports, and partner organizations, including humanitarian agencies. Once the data have been collected, multiple rounds of reviews are conducted by the ACLED data reviewers to ensure accuracy and reliability before weekly publication.

The ACLED database not only contains comprehensive information on the number of violent events in each country but also provides fine-grained data on the type of violence, location, time and duration of violence, number of fatalities as well as the actors involved in these events. The five main event types of violence categorized by ACLED are battles (e.g., armed clashes), demonstrations (e.g., excessive force against protestors), riots (e.g., mob violence), explosions/remote violence (e.g., missile attacks), and violence against civilians. This categorization provides a useful foundation for making a distinction between risk and uncertainty.

We regard the probability of all event types other than violence against civilians as generally more knowable and therefore less uncertain. This includes interactions between rebel or militia groups, rebel factions, governments, or government splinter groups, as they all involve actors who decide to participate in an event that proves to be violent. For the same reason, following scholars who see the 9/11 terrorist attacks on the US as having dimensions of both manageable risk and non-manageable uncertainty (Liesch et al., 2006; Phan & Wood, 2020), we also see events as instances of risk, rather than uncertainty, if civilians were collateral damage. For example, there is a risk of violence associated with civilians protesting the non-delivery of essential services, even if at the time they did not expect the event to be violent. The ACLED coding uses a hierarchical coding structure that captures “violence against civilians” separately from civilian targeting as part of other types of violence, thus allowing this distinction.

Violence against civilians is generally uncertain because it involves unilaterally and asymmetrically targeting unarmed non-combatants. We see sexual attacks, violent attacks, and kidnappings/forced disappearances as uncertain events—that is, violence that happened simply because of being at the wrong place at the wrong time. Thus, we believe violence against civilians is a stringent measure of uncertainty. The ACLED codebook,Footnote 4 and for the thinking behind it, Raleigh et al. (2010) provide more information about the coding.

Data on IFDI flows and country-level indicators were sourced from the World Bank’s World Development Indicators (WDI) and Worldwide Governance Indicators (WGI) as well as from United Nations (UN) data.

Empirical model

Due to unbalanced panel data, we used the Arellano–Bond linear dynamic panel-data estimator (xtabond command in Stata) to test our hypotheses (Arellano & Bond, 1991). The Arellano–Bond approach uses the generalized methods of movement (GMM) estimator and was ideal for our data, as it accounts for potential unobserved country-specific heterogeneity (Alessandri & Seth, 2014). Since there are potential endogeneity concerns due to reverse causality in the violence-IFDI, fatalities-IFDI, and violence-fatalities relationships, our Arellano–Bond dynamic estimator included three lags of the dependent variable as instrumental variables. According to Arellano and Bond (1991), the efficiency of the model increases when more lags of the dependent variable are used as instruments. Additionally, since panel data, especially at the country level, have unobserved heterogeneity and predetermined regressors, this model also allows for control of unobserved panel-level effects.

The Arellano–Bond estimator is preferred because it relies on minimal assumptions and has consistent estimates for unbalanced panel datasets (Moral-Benito et al., 2019; Samant et al., 2023).

$${\text{IFDI}}_{it}\,=\,{\beta }_{0}+{\beta }_{1 }{\text{IFDI}}_{i, t-1}+ {\beta }_{1 }{\text{IFDI}}_{i, t-2}+ {\beta }_{1 }{\text{IFDI}}_{i, t-3}+{\beta }_{4-6 }{\text{Independent} \text{Variables}}_{i, t-1}+ {\beta }_{7-15 }{\text{Controls}}_{i, t-1}+ {\theta }_{i}+{\varepsilon }_{it}$$

where \({\text{IFDI}}_{it}\) is the key dependent variable and \({\beta }_{1 }{\text{IFDI}}_{i, t-1}\) is the 1-year lag of the dependent variable that enters the model as a control. \({\beta }_{1 }{\text{IFDI}}_{i, t-2}\) and \({\beta }_{1 }{\text{IFDI}}_{i, t-3}\) are the 2- and 3-year lags of the dependent variable; \(i\) refers to the country and \(t\) refers to the year in the country-year panel; \({\theta }_{i}\) refers to the country-specific effects; and \({\varepsilon }_{it}\) are disturbances that are distributed across countries.

Variables and measures

Inward foreign direct investment Our dependent variable was measured as the 5-year moving average of all IFDI flows by country year. We recognized that this is a highly variable measure and used IFDI stock in our robustness tests. However, IFDI flows better reflect the extent to which a country can attract new FDI. Given our interest in the ability to attract FDI against the backdrop of a (more or less) violent society, this variability is not simply “noise,” but contains useful information.

Volatility in violence against civilians To test H1, we created a variable to capture the volatility of violence against civilians. This was done in four parts. First, we created a moving average of 5 years; second, we created a moving standard deviation for 5 years; and third, we squared the standard deviation. Lastly, this variable was lagged by 1 year. This is a frequently used measure of volatility (Clougherty & Zhang, 2021).

Fatality To test for H2 regarding fatalities, we constructed a 5-year moving average for total fatalities of civilians by country year with a 1-year lag.

Interaction term We created an interaction term between volatility in violence against civilians and fatality to test H3.

Controls In addition to the key dependent and independent variables, we included several country-level control variables that can influence IFDI. To control for the overall level of violence in the country, we created the total events except the violence against civilians variable, which is the sum of all events without those related to violence against civilians. Similar to the volatility in violence against civilians variable, we used a 5-year moving average with a 1-year lag.

Based on prior IB literature (Dunning, 1993), we controlled for the four key motivations behind IFDI. First, we controlled for efficiency-motivated IFDI using GDP per capita as a measure to capture the average standard of living and purchasing power of consumers in a country (Kingsley & Graham, 2017; Li & Vashchilko, 2010). Second, as has been used in prior research, we measured and controlled for market-seeking IFDI using total population as a proxy variable (Kingsley & Graham, 2017; Li & Vashchilko, 2010). Third, strategic asset-seeking motivated IFDI was controlled by using the count of papers published in scientific and technical journals, which captures the overall level of knowledge in a country (Asongu, 2014). Fourth, we controlled for resource-seeking IFDI using mineral rent as a percentage of GDP (Bokpin et al., 2015), since IFDI in many African countries is driven by resource-seeking motives.

Thereafter, we controlled for host-country characteristics. We included GDP growth as a control variable because the country’s growth level influences its ability to attract IFDI (Chowdhury & Mavrotas, 2006; Iamsiraroj & Doucouliagos, 2015), as well as the Gini index to control for income inequality. This index has been widely used to measure income inequality in the African context (Odhiambo, 2022) and the variable ranges from 0 to 1, with 0 indicating perfect income equality and 1 indicating complete income disparity. Thus, a higher score on the Gini coefficient index is an indicator of higher income inequality in a country (Lerman & Yitzhaki, 1984).

To control for institutional factors, we also included six controls from WGI. These included: (1) voice and accountability, which captures the extent to which the citizens of a country can select their government and is a proxy for the presence of democracy; (2) regulatory quality, which measures the government’s ability to formulate and implement regulations in the country; (3) government effectiveness, which captures the quality of public and civil services and their independence from political pressures; (4) rule of law, which captures the quality of contract enforcement, property rights, police, and courts; (5) control of corruption, which captures the extent to which public power is exercised for private gain; and (6) political stability to capture the overall stability of the home-country government. The corruption control variable includes petty and grand forms of corruption (Skovoroda et al., 2019). All six WGI variables were rescaled to run from 0 to 1, with 1 indicating better perceptions for the given country (Kaufmann et al., 2007). These indicators reflect the quality of institutions in the countries.

Due to the impact of financial openness on the level of FDI inflows in a country (Braithwaite et al., 2014), we created a financial openness variable using the Chinn-Ito index, which measures the extent of openness in capital account transactions (Chinn & Ito, 2008; Ito, 2006). Moreover, we controlled for trade openness, measured as the sum of exports and imports as a share of GDP, because of the strong relationship between IFDI and trade openness (Filippaios et al., 2019; Li & Vashchilko, 2010). To address potential problems with endogeneity, all control variables were lagged by 1 year (Oh & Oetzel, 2011).

Results

Descriptive analysis

Table 2 presents the correlations of the variables in our empirical analysis, including the means and standard deviations of the variables. Our descriptive statistics showed that the mean of violence against civilians in total events was 47.34. Although total fatalities had a mean of 570.59, the standard deviation was 3223.57, with a minimum of 0 and a maximum of 73,811, showing considerable dispersion across countries and years in the number of fatalities.

Table 2 Descriptive statistics and correlation matrix
Full size table

Most variables had low correlations and no correlation was greater than 0.5. In addition, we examined variance inflation factors (VIFs). Our VIFs were much lower than the threshold value of 10, ruling out concerns around multicollinearity (Cohen et al., 2003).

Regression results

Table 3 presents our results from testing Hypotheses 1, 2, and 3. Model 1 outlines the results of ordinary least squares (OLS) pooled regressions, controlling for time and country dummies, as our baseline estimation; and Model 2 conveys the results with fixed-effects (FE) estimation. As shown by prior research, these two models provide the upper and lower bounds for our GMM coefficients (García-Manjón & Romero-Merino, 2012; Grilli & Murtinu, 2014; Heid et al., 2012). The significance levels of coefficients in Models 1 and 2 are similar to those seen when we apply the Arellano–Bond estimator, which we discuss in greater detail below.

Table 3 Arellano–Bond linear dynamic panel-data-Hypotheses 1–3 (dependent variable—IFDI)
Full size table

Models 3–6 present results from our Arellano–Bond linear dynamic panel data, with Model 3 as the base model that has IFDI as the dependent variable and only the control variables. Model 4 introduces volatility in violence against civilians, the independent variable used to test H1. We hypothesized that the volatility in the overall levels of violence against civilians would have a negative impact on our dependent variable, IFDI. Our results from Model 4 confirm this hypothesis, as the estimated coefficient for this independent variable was negative and highly significant (β = − 0.54, se = 0.12, p = 0.000). We found that, on average, for a unit increase in the volatility of violence against civilians, there was a decline of 0.54 in IFDI. Thus, H1 was supported.

Model 5 introduces fatalities as another independent variable to test H2. This variable was negative and highly significant (β = − 0.47, se = 0.05, p = 0.000), suggesting that, on average, an increase in fatalities by one unit resulted in a decline of 0.47 in IFDI for the countries in our sample. Thus, we found support for H2. Model 6 introduces the interaction term between volatility and fatalities. The coefficient was negative and highly significant (β = − 0.61, se = 0.17, p = 0.000), showing that greater volatility in violence against civilians combined with higher fatalities had a negative impact on overall IFDI flows. Therefore, Model 6 offers support for H3. Together, these results strongly suggest that volatility in violence has a negative impact on FDI inflows. Moreover, this negative impact was greater when higher levels of civilian fatalities were involved. Overall, our results for the key independent variables were consistent across OLS, FE, and Arellano–Bond estimators.

The coefficients for the control variables were also consistent across Models 1–6. The coefficient for the variable total events except violence against civilians was negative, but not significant, which suggests that decreases in IFDI flows are explained primarily by violence against civilians compared to other types of violence.

Regarding our controls for IFDI motives, the coefficients for scientific and technical journals were not significant in any of the four models. This finding was consistent with prior research regarding the limited strategic asset-seeking IFDI in Africa (Gunessee & Hu, 2021). The other controls for the motives for IFDI (GDP per capita, total population, and mineral rent as a percentage of GDP) were positive and significant. The positive coefficients for mineral rent as a percentage of GDP were in line with prior research that found investments in Africa to often be motivated by resource-seeking motives (Anwar et al., 2022). The coefficient for our proxy for market-seeking IFDI, total population, was highly positive and significant in Model 3 but had a decline in significance as we introduced our key variables to the models. This suggests that market seeking in our context is complicated by violence-related considerations.

Additionally, the coefficient for GDP growth was positive and highly significant, suggesting that countries with higher growth rates have an increase in IFDI flows. The coefficient for the Gini index was negative and highly significant, which shows that IFDI is lower when there is greater income inequality in the host country. The coefficients of the six WGI variables (i.e., voice and accountability, regulatory quality, government effectiveness, rule of law, control of corruption, and political stability) were positive and significant, suggesting that when the country has better institutions, there is greater FDI inflow. Among the six, control of corruption had the highest level of significance.

The coefficient for financial openness was positive and highly significant, suggesting that the more financially open a country is, the greater the FDI inflows. The coefficient for trade openness was highly significant but consistently negative. This suggests that trade is a substitute for IFDI, a finding consistent with prior research (Pontes, 2007). We return to the implications of this finding later.

Robustness Tests

In addition to the empirical models outlined in the previous section, we estimated several robustness tests, which confirmed our findings related to our three hypotheses. We discuss our robustness tests based on our use of alternate dependent and independent variables, controls, and alternate estimators.

Alternate dependent variables The results presented in Table 3 use FDI inflows as the dependent variable. The IFDI variable was created using a 5-year moving average of IFDI flows. First, as part of our robustness tests, we re-estimated the empirical models using a 3-year moving average, instead of a 5-year moving average. Our results did not change, as the impact of violence on IFDI remained unchanged. Second, replacing moving averages, we used IFDI flows to the country in a given year as a dependent variable. The direction and significance of the coefficients for our key independent variables remained unchanged. Third, we used IFDI stock instead of FDI inflow to test our hypotheses, yet the results and impact on violence against civilians and fatalities did not vary.

In our Arellano–Bond linear dynamic panel-data estimation models, presented in Table 3, we used a 1-year lag of the dependent variable that entered the model as a control. As part of our robustness tests, we replaced the 1-year lags of the dependent variable with the 3- and 5-year lags of the variable. The results with the different lags remained consistent with our initial findings, as reported in Table 3.

Alternate independent variables Our data on violence came from the ACLED database. As part of our robustness tests, we collected violence data from the Uppsala Conflict Data Program (Chen, 2017; Davies et al., 2022; Moore, 2021). We ran our models using data from this alternate source and the results for our key variables related to violence remained consistent with the original findings.

For our ACLED data, in addition to the models estimated with 1-year lags for the independent variables, we conducted robustness tests with 3- and 5-year lags. The Akaike information criterion (AIC) measure (Akaike, 1974) was used to determine the ideal time lag for our models. The AIC is a measure to compare the relative quality of models based on goodness of fit (Piscitello & Thakur-Wernz, 2023). Using the estat ic command in Stata to test AIC, we found that AIC is smaller for the 1-year lag model, compared to the 3- and 5-year lag models, suggesting the former is a better model. Therefore, we only report findings for our 1-year lag model. Moreover, we tested for curvilinear relationships for our key independent variables, but the results were not significant, and hence are not reported in our paper.

Alternate controls To test the effect of alternate control variables, we substituted scientific and technical journals (included to measure strategic asset-seeking) with total patent applications by the country for a given year. We used patent data from the World Intellectual Property Organization for our measure. Furthermore, we replaced our GDP growth variable with total GDP in our models, and our results’ strength and direction did not change. Furthermore, we replaced control of corruption, one of the six variables to proxy home-country institutions, with a count of bribery incidents data from WDI. Additionally, we substituted the rule of law variable from WGI with the strength of the legal rights index from WDI. Our results did not change with these alternate measures for our control variables and hence are not reported in this paper.

In our robustness tests, we also included regional trade agreements as controls, as this also impacts the level of trade openness in the country. Specifically, we controlled for the following regional agreements using binary variables, which were 1 if the country was part of the agreement and 0 otherwise. There are instances where the country is part of two or more regional agreements. Based on recent research (Getachew et al., 2023), we controlled for the following regional trade agreements—Arab Maghreb Union, Central African Economic and Monetary Community, Community of Sahel-Saharan States, Common Market for Eastern and Southern Africa, Eastern African Community, Economic Community of West African States, Southern African Customs Union, Southern African Development Community, and West African Economic and Monetary Union. We did not control for the African Continental Free Trade Area (AfCFTA) in our analysis because this trade agreement was established in 2018, with phased implementation taking place between 5 and 10 years from January 1, 2021 (ElGanainy et al., 2023). Therefore, as our data ended in 2021, we did not expect AfCFTA to impact trade openness in the countries in our sample. None of the controls for the regional trade agreements in Africa were significant, which is in line with findings from prior research that show these trade agreements are not very effective in impacting trade and IFDI in the participating countries (Candau et al., 2019).

Alternate methods We also supplemented our Arellano–Bond linear dynamic panel-data estimation models with an instrumental variables regression model, since there is potential for endogeneity due to reverse causality in the violence and FDI inflows relationship. We utilized the two-stage least-squares generalization (G2SLS) of panel data estimators (xtivreg command in Stata). There are two implementations for this estimator—G2SLS from Balestra and Varadharajan-Krishnakumar (1987) and EC2SLS from Baltagi (2013)—and we estimated our models with both.

As we had three potentially endogenous variables—violence, conflict, and the interaction between the two—we used three different instrumental variables based on prior research on conflict and FDI. Using data from the United Nations Population Division, our first instrumental variable, the male–female ratio, was captured by the total number of males per 100 females (Desa, 2022). According to prior research, uneven sex ratios (i.e., unequal number of males to females) in a country’s population have an impact on the level of violence in the country (Archer, 2022; Diamond-Smith & Rudolph, 2018; Hesketh & Xing, 2006). Thus, the male–female ratio satisfied the relevance condition. However, there was no direct impact of gender imbalance on IFDI. Hence, this variable also satisfied the exogeneity condition. Our second instrumental variable was flood, a categorical variable that is 1 if the country had floods in a given year and 0 otherwise. Data for this variable were obtained from Dartmouth Flood Observatory’s Global Active Archive of Large Flood Events (Miao, 2019). Research has found a strong correlation between the occurrence of flooding in a country and excess fatalities (Jonkman & Kelman, 2005; Petrucci, 2022), including in the African context (Di Baldassarrev et al., 2010). However, occurrences of floods do not directly affect IFDI flows into a country, thus this variable meets the relevance and exogeneity conditions needed for strong instrumental variables. The last instrumental variable was ethnic, which is a count of the total number of ethnic groups in a country. The data for this variable were collected from the Central Intelligence Agency’s World Factbook. We used a count of ethnic groups as an instrumental variable because this impacts the overall violence in the country as well as the number of fatalities from this violence (Berkley, 2001; Daley, 2006). We did not expect the number of ethnic groups in a country to impact its IFDI flows.

Our initial results were upheld in our robustness tests with alternate methods. Table 4 presents the results of our robustness test using the G2SLS estimator with the Balestra and Varadharajan-Krishnakumar (1987) model. Moreover, Table 4 presents the tests conducted to confirm the validity of our instruments, namely the Angrist–Pischke test for excluded variables. Appendix 2 presents the first stage results of our G2SLS, which show that the likelihood of violence and fatalities increases when there are a greater number of males relative to females in a country, when there are floods in a given year in a country, and when there are more ethnic groups in a focal country.

Table 4 Robustness test—G2SLS—Hypotheses 1–3 (dependent variable—IFDI)
Full size table

Post hoc analysis–exports

One of the surprising findings in our empirical analysis is the highly significant but negative coefficient for the trade openness control variable, measured as the sum of exports and imports as a share of GDP. We therefore, as part of our post hoc analysis, re-estimated our regression models with exports as the dependent variable instead of IFDI. We tested all three hypotheses with our key independent variables, namely volatility in violence against civilians, fatalities, and volatility in violence against civilians × fatalities. The results of our post hoc analysis are presented in Table 5. The coefficients for volatility in violence against civilians, fatalities, and volatility in violence against civilians × fatalities were all positive and highly significant. In contrast to IFDI, which decreases with higher levels of violence against civilians and higher fatalities, exports increase with higher violence and fatalities.

Table 5 Post hoc analysis––Hypotheses 1–3 (dependent variable—exports)
Full size table

We also conducted robustness tests using trade openness as our dependent variable but the results remained consistent with those presented in Table 5. When there is substantial uncertainty associated with violence, IFDI decreases, but exporting by local firms increases. We deal with this potentially substitutive effect in the discussion and policy recommendations.

Discussion

Our empirical evidence demonstrates that certain types of violence should not be seen as knowable and manageable risk, but instead as largely unknowable uncertainty. It is well known that uncertainty in general deters IFDI (Canh et al., 2020; Choi et al., 2021), and this is also true when uncertainty is associated with violence. Violence against civilians deters IFDI, especially if it involves fatalities and if there is substantial variability in the levels of violence over time. We argue that all of these reflect the functioning of uncertainty.

Before considering the policy implications of these findings, we need to acknowledge a central limitation to our work: Our conceptual model and empirical analysis are at the country level, as we could not obtain reliable FDI data at the sub-national level. Thus, we provide evidence at the aggregate, but we could not consider how widely violence was distributed. This limitation matters because violence is not evenly distributed; there is considerable variance in violence not only between, but also within countries (Michalopoulos & Papaioannou, 2020).

It is testimony to the damage that violence wreaks that we can observe the negative consequences of Knightian uncertain violence on IFDI at the national level, even though violence typically happens in specific parts of a country. However, the fact that not all regions are equally violent has important implications for policymakers.

In their book Violence and Social Orders, Douglas North and co-authors argue that the fundamental difference between developed, wealthy countries with “open access orders” and the many limited access order countries (by their estimation where 85% of the world population lives) lies in how violence is managed: In all open access societies, “states possess a monopoly on the legitimate use of violence” (North et al., 2009, p. 22). In limited access orders, the state “exercises limited control over violence” (North et al., 2009, p. 46).

For example, members of African ethnic groups that were artificially separated by colonial-era borders often engage in violence against their national government (and also, albeit to a lesser extent, against civilians) because they feel a greater affinity with co-ethnic individuals on the other side of the national border (Michalopoulos & Papaioannou, 2016). The limits of its control over violence are likely very clear to a government dealing with such disaffected ethnic groupings. At the same time, there are likely parts of the country where the authority of the government is better established. MNEs feel more affected when they are in closer proximity to violent incidents (Cornwell et al., 2023; Dai et al., 2023) and it therefore seems that policy implications are likely to differ across sub-national regions. IFDI will likely prove very hard to attract in areas prone to Knightian uncertain violence, but less so in less violent regions. Establishing the extent to which this is indeed the case is an important area for future research.

With this important caveat, and acknowledging that our empirical work does not provide evidence of how violence can be reduced, we also believe that violence cannot be accepted as a given, no matter how intractable it may appear. Violent attacks on and killings of civilians are not justifiable. In considering how policymakers in countries grappling with Knightian uncertain violence can remain internationally and economically connected, we first synthesize insights about how violence can be reduced, and the role of MNEs in the process.

Reducing violence

Reducing violence is not the same as suppressing violence. The suppression of violence is not only expensive—requiring ten or more counterinsurgents per thousand residents (Goode, 2010; Quinlivan, 1995)—but as our evidence shows, relapses do more harm than the relatively peaceful periods between them do good. Empirical evidence suggests that strongman actions are likely to exacerbate, rather than reduce violence, as can be seen in the cases of Egypt (Fielding & Shortland, 2010), Thailand (Croissant, 2005), and Tunisia (Boukhars, 2017).

Instead, multiple role players need to work together to reduce violence. MNEs can be important partners in this process, and MNEs that help build peace can also benefit from doing so (Moore, 2021; Oetzel & Miklian, 2017). However, the Sierra Leonean experience suggested that IFDI can also “exacerbate tensions and grievances that are at the heart of conflict and fragility” (Ganson & M’cleod, 2019, p. 620). This implies that business reform must include “addressing drivers of conflict and fragility” (Luiz et al., 2019, p. 232), and a systemic approach is needed (Luiz et al., 2019; Van Tulder & Keen, 2018). Cross-sectoral partners from business, government, and civil society need to jointly engage in the complex task of working to effect change (Van Tulder & Keen, 2018). Unfortunately, we were unable to capture nuances in the FDI-related decision-making process at the firm level and could not consider the relationship between the various players, including MNEs, local government, various politically and socially aligned groups, and elite members of society. Future research should seek to provide evidence about these mechanisms at a more granular level.

Although the capabilities MNEs need to operate in fragile societies are “atypical” (Luiz et al., 2019, p. 217), the domains within which action needs to be taken are familiar. Much additional research is needed to confirm how these mechanisms function in the context of IB, but the work of Menocal (2011) and others indicate that the processes of promoting peace and supporting development are in many ways complementary, with three mechanisms repeatedly mentioned, namely reducing poverty, greater educational attainment and improved governance (Bormann et al., 2019; Hegre & Nygård, 2015; Mustasilta, 2019).

Poverty is an important trigger of violence (Braithwaite et al., 2014, 2016; Fielding & Shortland, 2010; Miguel et al., 2004; Smith, 2014), and addressing poverty can help create a less violent society. Organizations like MNEs seldom occupy a neutral position in discussions on such matters, as they tend to be “viewed as central causes and/or solutions to societal issues” (Ballesteros & Magelssen, 2022, p. 1519). Thus, some scholars see IB as a “stabilizing force” (Moore, 2021, p. 457) and MNEs as potentially “the most important institution in determining whether countries advance toward peace, remain in a limbo of conflict, or revert back to war” (Forrer & Katsos, 2015, p. 446). But MNEs can also increase poverty (Brandl et al., 2022) or reactivate the triggers of violence (Ganson & M’cleod, 2019). It is important to be mindful of these tensions in assessing the (actual or potential) developmental contribution of MNEs.

For example, the significance of the Gini coefficient index in our work underlines the importance of focusing not only on economic growth per se but on ensuring that benefits extend beyond the local elites. The work on MNEs and inequality suggests that it is important to consider not just inequality in outcomes, but also opportunities (Narula & Van der Straaten, 2021). Like societies, MNEs are disadvantaged by inequality (Krammer et al., 2023), and governments and MNEs need to engage to enable mutually advantageous outcomes (Rygh, 2021; Van der Straaten et al., 2023). To facilitate such engagement, we need a better understanding of the relationship between the motives for IFDI and inequality and, we suggest, also violence. It is a limitation of our paper that we do not have fine-grained country-level data that breaks down IFDI based on the motivations of the MNEs investing in the host country. While we controlled for resource-, market-, efficiency-, and asset-seeking motivations for IFDI in our empirical analysis, we echo Narula and Van der Straaten (2021) on the need for further research that examines these relationships according to the different IFDI motivations.

Regarding education, MNEs’ need for skills has long been known and they need not only highly developed scientific expertise but also a stock of solid foundational skills (Narula & Dunning, 2000, 2010). Therefore, partnerships to improve education are important for both MNEs seeking skilled employees as well as for the reduction of violence. An obvious focus for MNEs would be to play a role in restoring (or expanding) the physical infrastructure for education as well as financing educational interventions. Partnerships to deal with issues such as student and teacher absenteeism and improving teacher quality are also avenues MNEs can explore to help a country escape a conflict trap.

A critical mechanism to reduce violence is improved governance (Hegre & Nygård, 2015), and our empirical evidence provides clear support for the importance of governance elements like the rule of law, control of corruption, and political stability. It is important to note that scholars repeatedly underline the importance of not just developing formal institutions, but specifically their clear enactment. For example, Mustasilta (2019) points to the value of including traditional leaders to improve governance; Bormann et al. (2019) show the importance of actually sharing power in formal institutional power-sharing regimes, and Moore (2021) documents the importance of transparency and accountability in post-conflict environments.

We suggest that improved governance does not only serve to reduce violence but is in fact critical to enabling IFDI to benefit countries grappling with Knightian uncertain violence. In the next section, we explain the critical importance of governance where countries grappling with violence seek IFDI-assisted development.

Knightian uncertain violence and IFDI-assisted development policies

The role of IFDI in supporting development is foundational in IB research, dating back to when Dunning (1958/1998) examined how US investment helped the industrial recovery of the post-war United Kingdom. Subsequently, large bodies of work have examined related topics including spillovers from IFDI (e.g., Crespo & Fontoura, 2007; Kokko, 1996), IFDI-assisted development (e.g., Lall & Narula, 2013) and even more positively, IFDI-led development (e.g., Herzer & Klasen, 2008). Attracting IFDI has become a crucial policy objective for many governments, with the hope that IFDI can provide an infusion of capital, spur job creation, and trigger learning. Our work provides evidence of how difficult it is for a country grappling with Knightian uncertain violence to attract IFDI.

The preconditions and caveats that can prohibit IFDI from fulfilling its developmental potential have been repeatedly pointed out (Narula & Driffield, 2012; Narula & Dunning, 2000, 2010). An important caveat has been about the “right kind” of IFDI (Narula & Dunning, 2000, p. 150) and the consensus is that it is the host country’s task to ensure it provides an environment within which both investors and the host country can benefit (Narula & Pineli, 2019), with recent policy advice targeted at helping governments achieve greater benefits from IFDI (Sauvant, 2021). When Narula and Dunning (2000) argued for the right kind of IFDI, they had in mind IFDI that would induce technological spillovers and beneficial externalities in the host country. For countries experiencing Knightian uncertain violence, an important additional criterion for positive IFDI is IFDI that does not exacerbate violence.

Where close-to monopoly rents and weak governance co-exist, there is a risk of IFDI worsening violence in a host country (Henisz et al., 2010; Pinto & Zhu, 2022). For example, Wegenast and Schneider (2017) show that where foreign natural resource-seeking firms have limited certainty about their property rights, they seek to quickly extract as many resources as possible, triggering state repression greater than where the state-controlled resources. Ganson et al. (2023) argued that capital-intensive and heavily regulated sectors (e.g., infrastructure development and large-scale agriculture) all tend to be oligopolistic, with a high risk of contested relationships between stakeholders. Pinto and Zhu (2022) also reinterpret the well-known literature on the technology gap between local and foreign firms. Should that gap be so large that domestic firms cannot compete in their home market, reduced competition, increased market concentration, and higher monopoly rents can be expected. Therefore, because “higher rent creation increases the size of the spoils and thus the expected returns to appropriating those rents” (Pinto & Zhu, 2022, p. 1015), there is a risk of increased violence.

Certain large-scale projects, for example, infrastructure development, are required for economic growth and the expansion of opportunities. It is therefore important to underline that it is not the possibility of monopoly rents per se that results in violence. Instead, violence is triggered by the simultaneous presence of monopoly-type rents and weak governance (Ganson et al., 2023; Pinto & Zhu, 2022).

There are various reasons why weak governance increases IFDI-related violence. This includes dependence on the inflows of revenues associated with IFDI (Pinto & Zhu, 2022; Richani, 2005), susceptibility to the demands of foreign firms (Wegenast & Schneider, 2017), and even simple government ineptitude, e.g., ignoring community concerns (Nuhu, 2023). The importance of governance quality has long been recognized in IB (Slangen & Van Tulder, 2009) and we suggest that strengthening the governance of IFDI is critical to both reduce violence and realize the benefits of IFDI. This underlines the importance of initiatives like Investment Facilitation for Development (Berger et al., 2022), and the need to explicitly consider the challenge of violence in those discussions.

Our work is limited in that we cannot differentiate between reinvested earnings and de novo FDI. This matters because reinvested earnings on average account for just more than half of global FDI, and although the proportion for Africa is lower than elsewhere, it is nonetheless substantial (UNCTAD, 2020). Moreover, Barry (2018, p. 270) points out that although “Multinationals look for sustained peace when pursuing new ventures,” MNEs that are already in a location tend to remain there. The motivations for reinvested earnings are different than for de novo entry (Lundan, 2006; Salorio & Brewer, 1998), and likely also the developmental impact of such investment in violent locations. Unfortunately, data limitations prevented us from addressing the question of whether the lower IFDI in locations grappling with Knightian uncertain violence reflects or not the choices of existing ventures in these locations. This underlines the importance of studying the outcomes of MNEs’ presence in especially violent locations over time.

Exports under conditions of Knightian uncertain violence

To understand why uncertainty so strongly deters IFDI, it is useful to go back to Johanson and Vahlne’s (1977) foundational work on internationalization. Johanson and Vahlne (1977) framed internationalization (in the title of their classic work) as a process of “knowledge development and increasing foreign market commitment.” Given that Knight (1921) equated uncertainty with unknowability, it follows that if knowledge cannot be developed—because violence-induced uncertainty results in unknowability—the level of commitment to the foreign location is unlikely to be increased.

Our evidence demonstrates that exports increase as IFDI decreases. Much as Knightian uncertainty can be expected to trigger a preference for lower-commitment modes of internationalization, this is a puzzling finding, and future research should seek to better understand what the mechanisms are explaining this pattern. In their work on the Mexican war on drugs, Gorrín et al. (2023) found that violence depressed overall economic growth as well as exports from the country. Consistent with evidence that large IFDI projects in capital-intensive industries are targeted (Ganson et al., 2023; Pinto & Zhu, 2022), evidence from both the Philippines (Crost & Felter, 2020) and Mexico (Herrera & Martinez-Alvarez, 2022) suggest that the perpetrators of violence targeted large domestic firms in similar capital-intensive industries to fund their operations.

This raises the intriguing possibility that the rise in exports could be driven by small enterprises from the home country, especially since research on emerging economies has shown that exporting by firms from these countries helps them improve their capabilities (Abubakar et al., 2019; Piscitello & Thakur-Wernz, 2023). It has been previously documented that exporting by emerging market small enterprises represents a way to escape home country challenges like corruption (de Oliveira et al., 2021; Pindado Tapia et al., 2023; Qi et al., 2020; Wang & Ma, 2018; Wu & Deng, 2020). Exporting as a way to escape violence would be consistent with the findings of that body of scholarship, and future scholarship needs to explore that possibility.

Moreover, evidence is emerging that these small enterprises develop resilience in often-unexpected ways. Larsen and Witte (2023) showed that exporters often originate in the informal economy where they are less constrained by rules and regulations, and although they do not specifically consider exporters, Branzei and Abdelnour (2010) found small, informal enterprises to be especially resilient to terrorist attacks. They argued that these enterprises often flourish under extreme adversity given “positive psychology predictions that disruption, shock, and trauma may encourage (re)engagement in enterprise activities” (Branzei & Abdelnour, 2010, p. 806). How both smallness and informality play out for exporters from violent contexts are important avenues for future research.

The location of trade partners is another topic deserving of future research. Trade with neighboring countries contributes most to reducing conflict, probably because it is in the interest of neighbors to encourage peace or at least not stoke violence (Calì & Mulabdic, 2017). This is positive; UNCTAD (2024, p. 35) highlights that small firms “often internationalize by expanding in neighboring countries and within home regions.” Seeking to support the growth of greater numbers of small exporters into often-poor neighboring countries may seem like an inefficient policy option, but as Del Prete et al., (2023, p. 2) comment in their study of firm performance during the Second Libyan Civil War, “the characteristics that allow firms to survive in a conflict are possibly different from those favoring growth in times of peace.”

An important characteristic that helped flower producers in Kenya retain export markets when electoral violence broke out was direct relationships with buyers in export markets (Ksoll et al., 2023). The importance of direct, personal ties in providing trusted knowledge is comparable to the way that the knowledge and views of local MNE employees in conflict-ridden areas were decisive in MNEs’ decision to remain in or exit a country (Cornwell et al., 2023). Personal relationships with trade partners can provide a way for foreigners to obtain direct, unmediated, and, importantly, trusted knowledge of the violent foreign environment. Hence, another important area for future research is whether and how exporters’ ability to provide knowledge to foreign trade partners could mitigate the negative effects of violence.

Exporting is known to offer comparable benefits to IFDI (Atkin et al., 2017; Freixanet & Federo, 2023; Thakur-Wernz & Bosse, 2023 in their review of learning by exporting). As policymakers seek to reduce violence and encourage development, we want to suggest that both IFDI and exporting offer the benefit of economic international connectedness.

Conclusion

We reviewed the wide range of literature dealing with violence and IB studies of violence, highlighting the complexities of how violence happens. Interest in the topic of violence is increasing and our review of the field provides a useful foundation for future scholars. IB literature has tended to focus on the risks associated with violence, with findings generally underlining the possibility, albeit challenging task, of learning to manage such risk (e.g., Dai, 2009; Driffield et al., 2013; Oetzel & Oh, 2014). Nevertheless, MNEs do not always learn from their presence in violent locations (Oh & Oetzel, 2017).

Our paper offers the Knightian distinction between risk and uncertainty as an explanation for that apparent anomaly. Given the challenges of measuring uncertainty, we believe that we used a robust measure of Knightian uncertain violence. Nonetheless, we need to acknowledge that our use of the ACLED data on violence is associated with all the limitations of using secondary data. In particular, although we had information on the targets of violence, we had little visibility into other potentially relevant characteristics of violent events, such as the underlying motives for these events.

Our analysis of 25 years of African data shows lower IFDI when violence is directed at civilians, civilians are killed, and violence levels are constantly changing. We believe focusing on Africa is useful for examining our research question. However, as indicated by Amankwah-Amoah and Debrah (2017), negative perceptions and stereotypes about Africa abound. It may be that violent events are interpreted differently when they occur in Africa, compared to other regions that may not suffer from a similar level of prejudice. To the extent that the response to violent conflict is heightened by such prejudice, our work may either overstate the effects of violence on FDI or, as argued by Barnard (2020), allow us to uncover mechanisms that are also at work elsewhere, albeit in a less pronounced way.

There are many other regions globally experiencing violence, such as South and Central America or parts of Asia like the Philippines. We recommend future studies that examine our hypotheses in the context of other violent regions. The question of how violence-induced uncertainty is understood in countries that do not often experience intense violence like extensive civilian killings also needs further examination.

Nonetheless, and given the numerous complexities of operating across borders, our work suggests that the conceptually consequential but empirically hard-to-model distinction between risk and uncertainty may be more important than previously recognized. We hope that scholars and policymakers will more fully consider the distinction in making sense of an increasingly turbulent geopolitical world.