Formation of raiding parties for intergroup violence is mediated by social network structure

Significance

The social network structure of a small-scale society is crucial to formation of raiding parties involved in violent between-group raids. We mapped the social networks among Nyangatom men in a defined area of Ethiopia and ascertained membership in 39 intergroup raiding parties over 3 y. Although a small set of leaders initiated raids, they were not especially crucial for the composition of the raiding parties; instead, aspects of social network structure served to determine group composition and to amplify group size, once a raid was initiated. Intergroup violence, like other forms of collective action, depends on social structure and not just individual agency. This is relevant to spontaneous violent activities in settings as diverse as revolutions, gangs, and terrorist groups.

Abstract

Intergroup violence is common among humans worldwide. To assess how within-group social dynamics contribute to risky, between-group conflict, we conducted a 3-y longitudinal study of the formation of raiding parties among the Nyangatom, a group of East African nomadic pastoralists currently engaged in small-scale warfare. We also mapped the social network structure of potential male raiders. Here, we show that the initiation of raids depends on the presence of specific leaders who tend to participate in many raids, to have more friends, and to occupy more central positions in the network. However, despite the different structural position of raid leaders, raid participants are recruited from the whole population, not just from the direct friends of leaders. An individual’s decision to participate in a raid is strongly associated with the individual’s social network position in relation to other participants. Moreover, nonleaders have a larger total impact on raid participation than leaders, despite leaders’ greater connectivity. Thus, we find that leaders matter more for raid initiation than participant mobilization. Social networks may play a role in supporting risky collective action, amplify the emergence of raiding parties, and hence facilitate intergroup violence in small-scale societies.

Intergroup violence is common, worldwide, and harmful. Global annual deaths from large-scale warfare, for example, range from 0.5 to 1 million, and this does not include nonfatal physical and mental injuries (1). A diverse set of approaches has been used to study intergroup violence and warfare. Evolutionary models have credited collective violence with an important role in the development of modern human behavior (2–7), whereas cultural and ecological factors have been shown to influence small and large-scale violence (8–13). More recently, there has been increased interest in understanding the dynamics of group-based violence and the social processes that can contribute to it in the setting of insurgent and terrorist groups (14, 15); for example, online records suggests small, self-organizing groups coalesce into larger groups preceding terrorist attacks (16). Warfare has also been studied as a collective action problem—because individuals must mobilize to engage in a group activity with shared gains (e.g., deterrence, territory) and individual risks (e.g., injury, death) (17, 18).

Despite these advances, fundamental questions remain about how violent groups are formed, and the extent to which they may self-organize and emerge organically. Theoretical work suggests interindividual differences may be important for initiating and sustaining risky collective action, but empirical evidence in humans supporting this is sparse (19, 20). Research in primate behavior provides some clues regarding the emergence of violent intergroup conflict. Wild chimpanzees engage in lethal coalitionary violence against other communities (21), and a few “impact” individuals show exceptional motivation to participate in intergroup interactions (22, 23). Similarly, other primate species show interindividual variation in initiating intergroup conflict, including lemurs (24) and vervet monkeys (25). In these cases, the initiative shown by such individuals appears responsible for promoting participation by others.

To understand how violence is initiated in self-organizing groups of humans, we explore the role of social structure in collective violence in a traditional, nonstate society. Such small-scale societies offer an appealing opportunity to answer questions regarding the emergence of collective violence because they are generally free from formal institutions regulating conflict, such as are found in modern nation states (26), and there is neither conscription nor formal institutional control over violence. Unfortunately, field data on collective violence in these contexts are rare. Most studies of intergroup violence in small-scale populations have focused on the mortality rate and demographic effects of warfare, rather than the social precursors (27–29).

Although social networks are known to facilitate solutions to collective action problems (30–32) and to have a role in the emergence of both cooperation (33, 34) and violence (35), prior work on the structure of social networks (36) and their role in the emergence of violence in evolutionarily relevant populations is limited. A study among the Yanomamö examined how coparticipation in lethal intergroup violence influenced alliances later in life, finding that men who participated together in a killing were likely to live together and exchange marriage partners (37). That study provided important evidence regarding how participation in an intergroup conflict can be used strategically to advance subsequent relationships among participants; however, it did not evaluate social networks or the group composition of raiding parties.

To address how social networks influence the emergence of violence, we mapped the social ties in a nonstate society in which groups engaging in violent intergroup raiding formed organically, analyzing the role of social networks in instigating and sustaining intergroup conflict. Using observations derived from long-term ethnographic fieldwork, coupled with detailed mapping of the social network of raiding-aged men, we present data from a complete set of 39 discrete intergroup conflict events among the Nyangatom, a society of nomadic agro-pastoralists inhabiting a remote region along the border of South Sudan and Ethiopia largely outside the reach of state institutions (38, 39).

The Nyangatom

Many Nyangatom live in mobile cattle camps containing between 10 and 100 persons, and the population and number of these camps are not fixed (38). Depending on seasonal variation, camps may disband (with residents forming new camps) or they may aggregate and form larger villages. The Nyangatom also have semipermanent villages with dynamic membership, and movement between camps and villages is common. Livestock have a central place in the culture and diet of the Nyangatom and are necessary for many social exchanges, including marriage. To marry, a male is required to provide the family of the bride with bride wealth, often 30–60 cattle but sometimes as many as 100 cattle. Therefore, livestock are highly sought after, and violent conflict with other groups to obtain them is common (38–40). The Nyangatom also have a distinctive social organization involving sequential generation sets and age sets (38); most males engage in activities such as herding, socializing, and raiding with members of their age group, creating strong social bonds between members (38).

The primary type of intergroup conflict event for the Nyangatom is the raid (singular emojirimónu), in which a small group of men attempt to locate and seize livestock from other nearby ethnic groups, kill enemies they encounter when they can do so with minimal risk, and then escape unharmed. Casualties among members of raiding parties are unusual as they seek to minimize personal risk, but injuries and deaths of members of enemy groups are common, and fatalities among raiders are not unheard of. Successful raiders receive captured livestock (18, 39) and sometimes other social benefits, such as status, honorific names and scars, and public praise (9, 41). Raids generally begin with one or two individuals recruiting other participants, a process that typically takes several days. Raiding parties can also emerge when large groups of young men are congregated, such as during a ceremony. Individuals are not compelled to join a raiding party, and many young men elect not to join; and there are no formal sanctions for cowardice, desertion, or failure to participate (39).

Results

We used extensive semistructured interviews to collect information regarding intergroup conflict events that occurred between the Nyangatom and their neighbors. We comprehensively identified all 91 men residing in the study area who were of the appropriate age for raid participation (∼18–45 y) and established group composition for a complete set of all intergroup raids initiated during the study period (n = 39; Methods).

We measured a variety of attributes of potential raiders, including height, weight, kin relationships, and measures of paternal wealth (Dataset S1, Table S1). We also performed a comprehensive, sociocentric network study of this population of Nyangatom males. To measure friendship ties within this group, we used a gift task modeled on prior work with the Hadza hunter-gatherers of Tanzania (33) in which Nyangatom were asked to identify other study participants to whom they would like to give an anonymous gift.

On average, there was one raid every 4.7 wk, and raids were generally nonoverlapping in time, with raiding forays typically lasting several days. Most of the population (78 of 91) participated in at least one raid (mean participation, 2.9 raids; SD, 3.3). On average, about 7 men (SD, 3.4) participated in each raid (Fig. 1A); roughly 80% of raids in our sample were successful, resulting in an average of four head of livestock (including cattle, goats, and donkeys) being captured by each raider of a successful raid.

Fig. 1.

(A) Distribution of raid sizes over a total of 39 raids. The dashed red line shows the mean. (B) Bipartite participant–raid network. Top nodes are people, and bottom nodes are raids. Each raid has at least one leader. (C) Degree distribution (cumulative frequency) of the friendship network (red) and the average distribution of 10³ random networks with the same number of nodes and edges (green). The real distribution does not differ significantly from a random graph. (D) Participant–leader raid relation network. A leader (Bottom) is connected to an alter if the alter went on a raid with the leader.

Measured individual-level characteristics were tested for association with raid participation in bivariate models without controls—including the number of siblings, height, weight, and measures of paternal wealth (Dataset S1, Tables S2–S5). Although the number of siblings, weight, and paternal wealth were significant in bivariate models, none of these egocentric characteristics remains significant in a multivariate model (Dataset S1, Table S7). That is, we did not find evidence that these variables were independently relevant to whether a person went on a raid. However, a bootstrap analysis showed that the average weight of a leader is higher than that of a nonleader (P = 0.010), whereas neither the height (P = 0.468) nor the number of siblings (P = 0.364) is significantly different between leaders and nonleaders (SI Methods).

Fig. 1B shows a bipartite network of all 91 subjects and all 39 raids, with the five individuals identified as leaders on any raid shown in red. Leaders were clearly the most participatory, and all raids had at least one of these leaders. To complement the ethnographic data, we also used the raid participation data alone (Fig. 1B) to conduct a “minimal set analysis” to independently identify leaders (SI Methods). This analysis attempts to identify the smallest group of individuals at least one of whom participated in every raid. The procedure successfully identified all five individuals that were named as leaders on any raid by participants. These five individuals also participated in significantly more raids than expected due to chance (all P << 0.001) as determined by a procedure where we randomly reassign the identities of those who participated in each raid while keeping the distribution of raid participation fixed (SI Methods). Intriguingly, these results show that leaders can be successfully identified from raid participation data alone without prior information on the roles of each participant. Because there was no raid that did not include at least one leader, these results also suggest that leadership has an important role in the formation of intergroup violence.

The social network of raiding-age Nyangatom men is shown in Fig. 2, with those who did not participate in any raids shown in green, those who participated in at least one raid in blue, and leaders shown in red. Node size corresponds to the number of raids in which a person participated (larger indicates more raids). The mean number of incoming friendship nominations (in-degree) was 3 (SD, 2.7), and the range was 0–13.

Fig. 2.

Network of friendship ties in Nyangatom society determined using gift allocation task. Node size is proportional to raid participation (number of raids in which an individual participated). Dark gray arrows indicate reciprocal, two-way friendship ties, and light gray arrows are one-way ties. The age structure of the population is also visible, insofar as there is a rough demarcation visible here between the “northwest” and “southeast” regions of this network, with more ties within than between the two communities.

Although in-degree is associated with both wealth and number of siblings, the strongest predictor of the number of friendship nominations is leadership status. Leaders have more than twice as many friends (defined by receiving gifts) as nonleaders (5.2 vs. 2.4) and the difference is strongly significant (P = 0.01). Leaders also score significantly higher on a measure of network centrality, even when controlling for in-degree (P = 0.04) (SI Methods). This means that leaders not only have more friends but also that their friends tend to be more popular, meaning leaders also have more friends of friends as well.

We explored the ways the Nyangatom social network is similar to certain other social networks by measuring a comprehensive set of statistics (33). Although the cumulative degree distribution (Fig. 1C) does not appear to differ significantly from a random network (P = 0.76), a number of other important properties are shared with nonrandom social networks. Reciprocity (the probability that participant B names participant A as a friend, given that participant A names participant B) is significantly higher in the Nyangatom network (reciprocity, 0.37) than in a random network (P << 0.001); that is, there are significantly more two-way friendship ties in the real network than a random network. In addition, at 0.17, transitivity (the probability that two of a participant’s friends are friends with one another) is also significantly higher in the Nyangatom network than a random network with the same number of vertices and edges (P << 0.001). Finally, there is also strong homophily (the tendency of people with similar characteristics to have social ties with one another) by age group (0.88, P << 0.001) (Fig. S1) and by degree (0.10, P = 0.04). This homophily by age partly reflects the fact that the primary social interactions for males in Nyangatom society occur between members of the same age group; and our assessment of social ties also shows stronger connections within age groups than between age groups, thus supporting the validity of the gift-giving task as a measure of social connections among the Nyangatom.

Fig. S1.

Nyangatom social network. Colors denote age group membership (blue, younger age group; green, older age group). Ties within the same group are the same color as the group nodes. Ties between groups are in red. As expected, most friendship ties are within age groups.

Social network structure is highly relevant to the composition of raiding groups, and membership in raiding groups does not arise by chance. Fig. 3A shows that individuals with more social connections (i.e., higher degree) tend to go on more raids, even when we exclude the five leaders from the analysis and more popular nodes (those with higher degree) tend to go on more raids (Fig. S2). Each additional social connection is associated with an increase of 0.45 raids (SE, 0.17; P = 0.01) in the expected number of raids in which a subject participates. In fact, regression models that include in-degree, height, weight, wealth, and number of siblings show that social relationship “capital” is more strongly associated with raid participation than physical or material capital (SI Methods). Although we expect that having more social connections leads to more invitations or pressure to participate in raiding parties, it is also possible that increased raiding contributes to a greater number of social connections. Although wealth is associated with participation, the association becomes insignificant when we exclude leaders from the model (SI Methods). Our measure of network in-degree is the only variable that survives various model specifications.

Fig. 3.

(A) Number of times people joined raids as a function of social in-degree. Regression lines are shown for the full population (gray) (R² = 0.32) and excluding the leaders (black) (R² = 0.42). People who participated in no raids are green, nonleader participants are blue, and leaders are red. (B) Increase in probability of joining a raid based on geodesic social distance to leaders and to nonleader friends. Lines denote 1 SE. The large positive coefficients on first-degree connections show that direct nonleader friends are more motivating than leader friends, and both are significant. The negative coefficient on second-degree connections provides evidence against cascades beyond 1 degree in raiding-party formation. Motivation did not extend significantly to third-degree friends.

Fig. S2.

Network of friendship ties in Nyangatom society determined using a gift allocation task. Those who did not participate in any raids (nonparticipants) are shown in green, those who participated in at least one raid (participants) are shown in blue, and identified leaders are shown in red. Node size is proportional to raid participation (number of raids in which an individual participated). Dark gray arrows indicate reciprocal, two-way friendship ties, and light gray arrows are one-way ties. Dashed red lines indicate coraiding.

However, the emergence of violent collective behavior is more nuanced than leaders simply being linked by friendship ties to nonleader “followers.” We used regression analysis to evaluate the decision to join a raid, examining how this decision is associated with the total number of other people who join the raid, the number of one’s friends in particular who participate, and the number of other leaders who participate (SI Methods). In these models, we treated each individual’s decision to participate in each raid as the dependent variable, and we assessed how the presence of other potential raiders was associated with the probability that an individual would participate in a raid. To control for unobserved characteristics of individuals (e.g., their attitudes toward violence or risk, as well as other personality factors) and of raids (e.g., the distance to the raid target or the anticipated value of the raided items), we included in the model fixed effects for both individuals and raids.

Although raid size was not significantly associated with decisions to join raids, leader and friend participation was. Specifically, subjects were 6.8% (SE, 2.4%) more likely to join raids if they were directly connected by friendship to a leader in that raid. If, on the other hand, they were friends of friends with the leader (social distance 2) or friends of friends of friends (social distance 3), they were no more likely to join (Fig. 3B). This suggests that leaders may be able to mobilize their direct friendship contacts to join raids. However, further analysis yields the important observation that, if so, leaders are no more able to mobilize their friends than is anyone else in the population. Each nonleader friend who participated in a raid increased the likelihood that a person joined by 19.2% (SE, 1.4%), which is significantly higher than the boost in probability associated with leader friends participating (P < 10⁻⁵).

Although leaders appear to be less relevant than nonleaders for predicting any one decision to join a raid, recall that leaders are much better connected to the network. It is possible that leaders may have less effect per person, but a greater total effect because they are connected to more people. However, a test of this hypothesis fails. In a model where we regress total participation by a person’s friends on a person’s decision to join, their leadership status, and an interaction variable that indicates the effect of leadership on total mobilization, we find that significantly fewer people join when a leader joins than when a nonleader joins (P = 0.008) (SI Methods). In other words, the key motivating factor to join a raid once a raid is initiated is not leadership; it is friendship.

Social distance has an unusual relationship in the results for nonleader friends (Fig. 3B). After controlling for friend participation, each friend of a friend who participates in a raid actually decreases the likelihood a person will join by 1.6% (SE, 0.6%; P = 0.006). This suggests that people just outside of a person’s direct social network may actually slightly demotivate participation in raids; weak ties are apparently not useful for recruiting and may even be somewhat detrimental. This also suggests that the men indeed have different sorts of relationships with each other, even within a population of just 91 individuals. The significance of these associations survives models with various controls (SI Methods).

Finally, a model with sibling participation did not provide evidence for siblings being more or less likely than chance to raid with each other (P = 0.23) (SI Methods). Thus, it appears that kinship did not influence raiding-party composition, consistent with prior work with humans (37).

Discussion

A rich picture appears regarding the role of leadership and social network structure in the emergence of collective intergroup violence in this evolutionarily relevant population. Leaders appear to matter mechanistically, functioning as focal points or as nucleation sites for raids among the Nyangatom. Although they participate most often (the five leaders are the top five participants, by number of raids), they are not particularly good at directly mobilizing other participants. Instead, nonleaders have a critical role in amplifying the size and specifying the composition of raids once leaders initiate their formation. Although leaders may instigate raids, they have no more influence than anyone else in promoting participation. Moreover, our analysis of individual decisions to participate in raids, using fixed-effects models, shows that social network structure is key even when controlling for the “push” of individual differences in the tendency to join raids and the “pull” of differences between raiding parties that may make some parties more appealing to join than others.

Our findings are also noteworthy because we did not uncover ethnographic reports of formal sanctioning for nonparticipants. However, it is possible that the withdrawal of a friendship tie is a form of sanction (31). If so, then, the pattern of friendships may itself depend on the willingness of men, at least occasionally, to join raids together. Work among the Yanomamö suggests that coparticipation in violence may result in subsequent formation of social bonds (37), and men commonly enlist in the military with friends and are often encouraged to do so in state-sponsored wars (42). Future research should include repeated measures of network structure to ascertain the extent to which collective violence also shapes the network.

One important limitation of our study is that our network measures only provide a snapshot of the social network at one point in time, leaving open the possibility that coraiding led to the formation of the social ties we observed rather than men opting to raid with their friends. Nevertheless, based on the ethnographic evidence collected, we think friendship is a primary mechanism that contributes to coparticipation in a raid. Among East African pastoralist societies such as the Nyangatom, young men engage in many collective activities together, such as herding and ceremonies, creating opportunities to meet other members of their age group. As a result, they form very tightly bound cliques early in adolescence that are an important part of social life. Raids are risky and raiders are commonly nervous before a raid; this may be why individuals choose to raid with their friends rather than with people they are not so well acquainted with (as our findings also document, even within a relatively small population of 91 people). Rather than acting primarily as a mechanism to generate friendships with unfamiliar individuals, raids may instead act to deepen friendships or be built upon them.

Important similarities and differences emerge between our results and behavioral data on collective violence in other primates. Among wild chimpanzees engaged in group border patrols and hunting, there is little indication that kinship influences the likelihood or effectiveness of such collective action (43). Among the Nyangatom, we also found no influence of sibling relationships on raiding-party composition, suggesting alternative mechanisms for generating participation. This is also consistent with the cognitive and social complexity of humans and with prior observational work regarding the role of social ties in the emergence of both collective violence (10, 37, 42, 44) and altruism in humans (31, 32, 45).

Although we find that participation in raiding is widespread throughout the population, there is also significant individual variation. A substantial portion of the population did not participate in any raids, whereas five individuals participated in more than 10 raids and most participated in slightly less than 3 raids. The fact that the initiation of raiding parties appears to depend on leaders who function as nucleation sites for raids and who attract other participants is consistent with research showing how individual variation within a population can contribute to the resolution of collective action problems (17, 20), including in risky, intergroup violence in both humans and chimpanzees (46). Leaders may alter the costs and benefits for others—either by reducing the costs of the raid to other participants (e.g., via setting the time of the raid or by scouting) or by exerting social pressure on others to join (18, 47–49).

In sum, we find that leadership matters in initiating collective violence in this small-scale society, but that it is not an especially important factor with respect to who joins the raiding parties. However, violent group formation does not involve individuals simply copying the risky violent behaviors of other members of their group either; rather, social network structure matters in the formation of raiding parties and in the emergence of such structured violence. To the extent that Nyangatom raiding behavior mimics the general phenomenon of risky collective action, we have identified an important amplifying effect: a handful of motivated individuals, with distinctive network positions, coupled with a wider group of reinforcing individuals embedded within a network, can lead to population-level violent effects.

These results might be relevant to other informal contexts in which violence occurs, such as urban gangs (35), localized insurgencies (14, 50), revolutionary protests (10), or terrorist attacks (16). Many types of violence do not depend solely on the desires and actions of individuals or even dyads, and instead may at least partially emerge and be supported by the very social structure in which all individuals are embedded (51). These observations, finally, suggest two things with respect to the prospect of managing violence. On the positive side, attenuating the impact of a leader may prevent the original nucleation of the violence. However, on the negative side, once violence is switched on, people are likely to join from throughout the whole population, and so, once instigated, violence has a wide-reaching effect on the society.

Methods

Data were collected as part of an ongoing ethnographic study of the Nyangatom in which one of the researchers (L.G.) intermittently resided in the study area in Ethiopia between 2009 and 2012. We used semistructured interviews to collect information regarding intergroup conflict events that occurred between the Nyangatom and their neighbors, including the Turkana, Daasanach, and Suri.

We identified 91 men residing in the study area who were of the appropriate age to participate in raids (∼18–45 y). We conducted interviews with each of these individuals, collecting data on their conflict history including both successful and unsuccessful raids; raiding-party composition was validated through peer reports. The presence of a raider on a raiding party was determined by an individual’s participation in the raiding party for any portion of it; we did not measure desertion, and some individuals may have ceased their participation during the actual raid because they were afraid or for other reasons. Leadership was ascertained by cross-validated personal accounts elicited by questions about whether any person was a leader of the raid using two Nyangatom terms for leader (singular Ekarikon; singular Eketamunan).

We also performed a comprehensive, sociocentric network study of the entire population of raiding-age Nyangatom males (n = 91). To measure friendship ties within this group, we used a gift task modeled on prior work with the Hadza hunter-gatherers of Tanzania (33) in which the Nyangatom subjects were asked to identify other study participants to whom they would like to give a gift of candy. Giving a gift is an important measure of friendship in most societies (52). We chose candy as the allocation currency because of its practical ease and because Nyangatom value it. Subjects were presented with three pieces of candy and shown photo sheets containing the facial portraits of study participants to whom an anonymous allocation could be given. They were asked to indicate the three persons that they would like to receive the gift of candy and told they would not be identified as the donor. All 91 subjects (100%) participated, yielding a total of 273 social ties within this group, and distributions occurred only after all participants completed the task. We also measured a variety of attributes of the study participants including height, weight, and estimates of paternal wealth (SI Methods).

To explore associations between raid characteristics and raid participation, we evaluated linear regression models that estimated the association between an individual’s decision to join a particular raid and various raid characteristics. The basic model is as follows:

$$E\left[Y_{ir}\right]={\theta }_i+{\gamma }_r+\mathit{\beta }{\mathit{x}}_{\mathit{ir}},$$

where the dependent variable Y_ir is 1 if person i joins raid r, and 0 otherwise; x_ir is a vector of characteristics for participant i and raid r; and θ_i and γ_r are individual and raid fixed effects, respectively. We report results of the linear model for more intuitive interpretation. The results are consistent in both sign and magnitude compared with generalized linear models. See SI Methods for further description of methods.

Approval for this study was obtained from the Harvard University Committee on the Use of Human Subjects; the South Omo Zone, Southern Nations, Nationalities, and Peoples’ Region, Federal Democratic Republic of Ethiopia; and local elders. Informed consent was obtained from all participants.

SI Methods

Study Population.

The Nyangatom number between 20,000 and 30,000 and live along the Ethiopian and South Sudanese border in and adjacent to the Lower Omo Valley of Ethiopia. Research was conducted in a border area along the Nyangatom–Turkana frontier to the north of the Kibish River and west of the Kuraz Mountains.

The residence structure of the Nyangatom is dynamic. Many Nyangatom reside in mobile villages. These villages may exist for several weeks or longer before disbanding or relocating. Sometimes, members of multiple camps or villages may join together to form a larger village or larger villages may break up into smaller villages. A particular village composition usually results from considerations for providing suitable resources for livestock as well as to maximize security. There are also semipermanent villages in settled areas, but the population is highly dynamic. Individuals commonly move between these villages and the mobile villages. Young men are generally not attached to any one village for their primary residence, but instead are attached to the livestock owned by their father or other paternal relatives and change their residence based on the movements of these livestock.

The Nyangatom are organized into territorial sections (plural ngiteala). Membership provides culturally recognized rights to resources in a certain areas of Nyangatom territory. Thus, men living in a specific area generally share membership in common territorial sections. Although individuals may change their residence multiple times a year, they usually do so within a constrained area of Nyangatom territory and revisit the same areas seasonally.

Because of the complex residence dynamics of the Nyangatom in the study area, conducting an analysis of village residence was impractical. Rather, we identified individuals who resided at least seasonally in the study area and were the appropriate age to participate in raids. Men who are elders (singular ekasukout) do not participate in raids. Although we lack data on their ages, elders are estimated to be above 45 y of age. We also exclude young men below the ages of ∼18 because they usually do not participate in raids. We identified our sample (n = 91) by their membership in culturally specific age groups where men are deemed old enough to engage in activities such as cattle herding and raiding but not old enough to be elders.

These individuals are expected to be familiar with each other for several reasons. First, the dynamic residence structure ensures that individuals may have resided with each other in the same village. Second, when resources allow, multiple herds of livestock graze together in grazing areas. This usually occurs after the rainy season but can occur at other times as well. In these cases, men from a large area herd together providing collective defense. Finally, there are many ceremonial engagements in which men from throughout the area come together during which ceremonies are performed and animals are slaughtered and communally consumed. These ceremonial activities allow men from the study population to spend time together on a semiregular basis.

Male subjects had facial photographs taken for identification purposes. Each subject was assigned a unique identification number that was used to match the photograph to the subject. Photographs were compiled onto two photo sheets measuring ∼30 × 35 cm each, containing 42 and 49 photographs. Each photograph measured ∼3 × 5 cm. The photo sheets (supplemented by individual photographs) were used to allow participants to make visual identification of other study participants.

Study participants were compensated with local currency (approximately US$0.25 to US$1.00), tea, or sugar for their participation in study elements. A translator was used for the semistructured interviews, and interviews would proceed by the researcher presenting the question to the translator who would ask the question and then translate the answer back to the researcher. The researcher (L.G.) also directly asked interview questions or follow-up questions if the subject’s answers were not clear.

Because the field researcher was based at the field site for an extended period, data collection occurred throughout the duration of the study and multiple interviews were conducted with study participants. Interviews were frequently conducted with the subject alone. However, due to the open nature of the society, subjects would sometimes be joined by their friends or relatives. Subject comprehension and accuracy was validated in the field by using consistency checks within and between interviews. These involved asking participants to repeat their answers to questions to test for consistency with their previous answers.

Conflict Landscape.

The threat of conflict is a daily feature of life for those living in the study area. The Nyangatom have ongoing conflict with several of the neighboring populations. The conflicts involve the use of automatic weapons, including Kalashnikovs that were introduced in the late 1980s and are used throughout the region. Similar to other pastoralist groups in the region, the Nyangatom conduct raids that involve a small number of men who attempt to capture livestock with minimal risk. Our data focus on these small raids, ethnographically most similar to those among evolutionarily relevant groups such as mobile foragers.

Raids have a very low mortality rate because raiders seek to seize livestock using ambush and stealth and do so only when there is little risk to themselves. If they cannot find an opportunity with low risk, they will generally abandon their plans and the raid is then unsuccessful. In some cases, dehydration may cause death on raids, but no such instances occurred during the study. During the research period, no subjects were killed or wounded from their participation in a stealth raid. However, deaths for participants do occur. After the conclusion of data collection, one of leaders of this population was killed while on a raid. The raiding party attacked what appeared to be only few enemy herdsmen. In fact, there was a larger party of enemies resting nearby who quickly joined the fight resulting in the death of one member of the Nyangatom raiding party.

Raids usually have informal leadership. This usually occurs when an individual decides to initiate a raid and he may then spend several days recruiting other individuals to join him. In some cases, he may visit the village of his desired coraiders over a period of several days to convince them to go. In other cases, raids emerge after age group events in which large cohorts of men from the same age group are congregated.

During the study period, no commercial or political elements to the conflict were observed or described for the motivations of participating in conflict; thus, the circumstances under study seem appropriate as an ethnographic example of small-scale nonstate warfare.

Conflict History Data.

Since conflict is a regular feature of Nyangatom life it is discussed openly. It is common for individuals to publicly recount their participation or the participation of their peers. Conflict data were obtained through interviews with study participants about their participation in intergroup conflict events. Coparticipants were identified by the use of the facial photograph sheets. Composition was validated through peer reports with other participants, allowing confirmation of membership for each raiding party. In some cases, the researcher was unable to confirm coparticipant identity, usually because they were not members of the study population. These individuals were coded as nonsubjects, and no information was obtained about them.

Leaders were identified through interview questions with raid participants about whether any person was a leader on a particular raid the subject participated in using one of the two words for leader (singular Ekarikon; singular Eketamunan). Leadership was validated by reports from more than one raid participant not including the leader himself. This resulted in the identification of five leaders on 19 of the 39 raids. There were six additional cases where three of these five individuals were indicated as being more than a mere participant in raids in which they were not named as a “leader,” either by contributing tactical advice or selecting the location of the raid. However, because they were not identified with the locally used referents for leaders, we excluded these six cases.

We do not present specific individual contributions to raids, incidents of defection, or the outcome of the raids. Instead, we focus here on the presence or absence of any individual from the subject population in each raid.

Other Data.

We collected a variety of other data on study participants, as shown in Dataset S1, Table S1.

Anthropometry.

Body weight (in kilograms) and height (in centimeters) were measured in the field for all available subjects. Weight was measured using an electronic scale and height measured with a stadiometer. Fifteen subjects were unavailable during the anthropometric data collection and are excluded from such analyses.

Sibling relationships.

Sibling relationships were collected as part of the demographic and genealogical data collection for the study population. They were elicited through interviews with subjects, in which they were asked to identify their siblings and whether any of their siblings were among the subjects participating in the study.

Paternal wealth rankings.

Among the Nyangatom, men who are not elders seldom own more than a few head of livestock themselves. Rather, livestock are generally owned by elder male family members. Thus, measures of individual wealth are not culturally appropriate for the men in our sample. We used measures of paternal wealth to explore the relationship between the wealth of a raider’s father and raiding-party composition. Paternal wealth rankings were obtained for a subset of the study participants (n = 42). These scores were generated from a ranking task in which elder men (raters)—who were not in this sample—were asked to sort facial photographs of subjects into three piles based on the relative wealth of the father. They were initially asked to identify any men featured in the photographs that they did not know. If they could identify all of the individuals in the photographs, they were then asked to look at the photographs and determine whether they knew the father of the individual in the photograph. Six elders successfully identified all subjects and their fathers and provided the rankings of paternal wealth.

Raters were instructed that they would sort the photographs into three piles based on information about the fathers of the individual featured in the photograph. They were told that in one pile they were to place the photographs of the men who had the wealthiest fathers. They were then told that a second pile was for the men whose fathers had the least wealth. The final pile was situated between the wealthiest and the least wealthy, and raters were instructed to place the photographs of the men who were between the wealthiest and least wealthy into this pile.

Each individual was ranked six times. Each time a subject was placed in the wealthiest pile, they received a score of 3; each time they were placed in the middle pile, they received a score of 2; and each time they were placed in the least wealthy pile, they received a score of 1. The maximum any subject could receive as a wealth ranking score of 18 and the minimum of 6.

Friendship network data.

Study participants were asked to participate in a gift allocation task that was used to generate the friendship network. During this task, they were presented with three pieces of candy. Because the population has only minimal access to a market economy, novelty food items, such as candy, are valued. Study participants were presented with the two photo sheets containing the photographs of all 91 study participants. They were asked to make anonymous allocations to three individuals who they desired to receive candy and whose picture was featured on the photo sheets. They were told to do so by placing a piece of candy on the photograph of the subject. Self-allocations and multiple allocations to the same subject were not allowed. Subjects were informed that these allocations would be made at the conclusion of the study but that the identity of the donor would remain anonymous. These allocations were used to generate the friendship network data. Distribution of the candy based on the gift allocations was conducted after the completion of this study element.

We counted the number of times each person was nominated as a friend in the friendship network; this simple measure is called “in-degree.” We also used the nominations to map the full friendship network, and to calculate other social network measures (see below).

Age group membership.

Although Nyangatom have generation sets and age sets, age groups are more important in daily interaction than generation or age sets. However, age groups are not panethnic among the Nyangatom, and, in contrast to age sets, members of an age group can come from any generation. At the time of study, adult men who had not yet become ekasukout (elders) belonged to one of two age groups. Interviews with subjects collected data on the subject’s territorial section, generation set, and age group.

Using the edge-betweenness algorithm for community detection (53), we find that the vast majority of nodes (98%) are appropriately classified as belonging to their self-identified age group. Moreover, the network can be seen to be composed of two communities (Fig. S1). This provides external validation of the validity of the gift task as a measure of social ties within the Nyangatom community.

Characterizing Whether Raid Participation Is Due to Chance.

First, we identify whether raid group participation is due purely to chance using two methods: a χ² test on simulated data and a permutation test. For the χ² test, we have the null hypothesis H₀ that people are drawn uniformly, that is,

$$\text{Pr}\left(\text{drawing}\hspace{0.17em}\text{person}\hspace{0.17em}1\right)=\text{Pr}\left(\text{drawing}\hspace{0.17em}\text{person}\hspace{0.17em}2\right)=\mathrm{\dots }=\text{Pr}\left(\text{drawing}\hspace{0.17em}\text{person}\hspace{0.17em}91\right).$$

The alternate hypothesis H₁ is that not all people have the same probability of being drawn for a raid. We calculate the probability (under the null hypothesis of uniformity) of person i being drawn in raid r, where raid r has the observed number of participants n_r. We find that

$$\text{Pr}\left(\text{person}\hspace{0.17em}i\hspace{0.17em}\text{being}\hspace{0.17em}\text{drawn}\hspace{0.17em}\text{in}\hspace{0.17em}\text{raid}\hspace{0.17em}r\right)=\frac{\left(\begin{array}{c}91-1\\ n_{r-1}\end{array}\right)}{\left(\begin{array}{c}91\\ n_r\end{array}\right)}=\frac{n_r}{91},$$

and summing over all raids yields the expected number of observations,

$$E_i={\displaystyle \sum }_{r=1}^{39}\frac{n_r}{N},$$

and hence the χ² statistic,

$${\chi }^2={\displaystyle \sum }_{i=1}^{91}\frac{{\left(O_i-E_i\right)}^2}{E_i}.$$

Performing the χ² test allows us to reject the null hypothesis at P << 0.001.

We cross-validated this approach using a permutation test with synthetic data. A dataset was generated to have the same number of participants in each raid (to account for different costs and benefits associated with each) but with people having equal probability of being chosen in a raid. Here, the null hypothesis H₀ is that the two datasets come from the same distribution, and the alternate hypothesis H₁ is that the two datasets do not come from the same distribution. We pool the data, permute the observations, split into two groups of the same sizes as the original, and calculate the statistic of interest from these new data. The simulation was run 10⁶ times to obtain the null distribution of the statistic. As a natural statistic of interest, we choose the following:

$$S=\frac{{{\displaystyle \sum }}_{i=1}^{91}\left|\mathrm{\Delta }{\text{O}}_{\text{i}}\right|}{91},$$

where $\mathrm{\Delta }{\text{O}}_{\text{i}}$ is the difference of the total number of observations of person i among the two groups. Comparing with the original statistic using the observed group and the initial synthetic data, we find that P << 0.001, which is sufficient to reject the null hypothesis. Therefore, we conclude that raid group participation is not simply due to chance.

Identifying Leaders Directly from the Raid Participation Data.

We find a minimal set of raiders that account for participation in all raids. To facilitate this, we use a simple algorithm to establish an upper bound:

• 1)Calculate the number of raids in which each person participated.
• 2)Find the person who participated in the most raids.
• 3)Mark the raids in which he participated and remove those raids.
• 4)Repeat from step 1 until all raids have been removed.

We find that the five ethnographically identified leaders form precisely such a set (see Fig. 1B, which shows that every raid has at least one leader identified in this manner).

To test that this is a minimal set, we reduce the search space and enumerate all possibilities in the search for a four-person (smaller) set, M. We used a counting argument to reduce the search space. Suppose no ethnographically identified leaders are present in M. Then, the four most active allowed participants only participated a total of 26 times (7 + 7 + 6 + 6). Even if they joined completely disjoint raids, this is not sufficient to account for all 39 raids. Therefore, at least one leader must be present in M. If exactly one leader is present, the same argument shows that the maximum possible number of raids accounted for is 36. This simple argument shows that at least two ethnographically identified leaders must be present in the proposed set M. We enumerate all of the allowed possibilities and do not find such a set, so we conclude that the ethnographically identified leaders indeed form a minimal set.

This analysis also suggests that raid participation data alone would have helped us to identify the leaders that were identified via questions about leadership.

Friendship Network Structure.

Although direct comparisons of network datasets are difficult due to different contexts, different ways of ascertaining social connections, and structural differences themselves (e.g., differing network sizes, differing numbers of edges), some comparison of statistics is still informative. In the main text, we present measures of reciprocity, transitivity, and assortativity (homophily) based on degree and age, and discuss the characteristic ways that the Nyangatom social network resembles and differs from simulated random networks and other network data published previously. We find that the degree distribution of the Nyangatom social network was not significantly different from an Erdos-Renyi random network with an identical number of nodes and edges, although the other properties were different in a way similar to several modern networks; in particular, reciprocity, transitivity, and degree assortativity were significantly larger than in random networks, consistent with measures given in prior work (33). As there are multiple related definitions for these measures used in the literature, here we provide formulas for the way our calculations were performed.

“Reciprocity” (the probability that person B is nominated as a friend by person A given that person A is nominated by person B) was measured as the proportion of mutual connections. That is, given connectivity matrix A,

$$\mathrm{Reciprocity}=\frac{{{\displaystyle \sum }}_{i,j}{A}_{ij}{A}_{ij}^T}{{{\displaystyle \sum }}_{i,j}{A}_{ij}}.$$

“Transitivity” (the likelihood that two of a person’s friends are themselves friends) is calculated as a global network parameter, that is, the ratio of connected triples to the total number of possible connected triples in the graph.

To calculate assortativity (“homophily”—the probability of nodes with similar characteristics being connected), we first assign values of interest to the nodes (degree, age group). Let e_ij be the fraction of edges connecting nodes of type i and j, let $q_i^{\left(1\right)}={\displaystyle \sum _j}{e}_{ij},\hspace{0.17em}{q}_i^{\left(2\right)}={\displaystyle \sum _j}{e}_{ji}$, and let ${\sigma }^{\left(1\right)},{\sigma }^{\left(2\right)}$ be the SDs of $q_i^{\left(1\right)},\hspace{0.17em}{q}_i^{\left(2\right)},$ respectively. Then, we calculate homophily as follows:

$$\mathrm{Homophily}=\frac{{{\displaystyle \sum }}_{j,k}jk\left(e_{jk}-q_j^{\left(1\right)}{q}_k^{\left(2\right)}\right)}{{\sigma }^{\left(1\right)}{\sigma }^{\left(2\right)}}.$$

“Eigenvector centrality” assumes that the centrality of a given individual is an increasing function of the centralities of all of the individuals to whom he or she is connected. Although this is an intuitive way to think about which subjects might be better connected, it yields a practical problem—how do we simultaneously estimate the centrality of all subjects in the network? Let a_ij equal 1 if subjects i and j have a social connection and 0 if they do not. Furthermore, let x be a vector of centrality scores so that each subject’s centrality $x_j$ is proportional to the sum of the centralities of the subjects to whom they are connected: $\lambda x_i=a_{1i}{x}_1+a_{2i}{x}_2+\cdots +a_{ni}{x}_n$. This yields n equations, which can be represented as $\lambda x=A^{T}x$. The vector of centralities x can now be computed because it is an eigenvector of the eigenvalue λ. Although there are n nonzero solutions to this set of equations, in symmetric matrices, the eigenvector corresponding to the principal eigenvalue is used because it maximizes the accuracy with which the associated eigenvector can reproduce the original social network. To be sure of reaching a solution, we symmetrized all asymmetric relationships in the observed network (i.e., we assumed all friendship ties were mutual).

Modeling Individual Characteristics.

In this section, we describe methods to explore what individual characteristics are associated with raid participation, leadership, and network in-degree.

Dataset S1, Tables S2–S7, shows linear regressions that measure the association between raid participation and various individual characteristics. The basic model is the following:

$$E\left[Y_i\right]=\alpha +\mathit{\beta }{\mathit{x}}_{\mathit{i}},$$

where the dependent variable Y_i is the total number of raids in which person i participated, x_i is a vector of individual characteristics for participant i, and β is a vector of coefficients that indicate the degree of association with each characteristic. The specific independent variables are height (in centimeters), weight (in kilograms), number of siblings (paternal or maternal), and paternal wealth (only measured for n = 42 people). Models are calculated for the full population (left side of tables) and for the subset of the population who are not leaders (right side of tables). Only the significance of in-degree (net of sibling contributions, which are measured separately) remains for both the full population and the population that does not include leaders when all individual characteristics are allowed. These models suggest that social information is more important for raid participation levels than individual characteristics. As the physical egocentric variables are no longer significant when in-degree is added, there may be a path from these variables to in-degree and leadership status, which will be explored in future work. We used ordinary least squares regression to estimate these models, but count models yielded similar results.

Dataset S1, Tables S8–S12, shows a similar set of models, but the dependent variable Y_i is the in-degree of person i (net of siblings). Siblings, paternal wealth, and leadership status individually appear to be significantly associated with raid participation in both the full population and the subset of the population that does not include leaders.

Finally, Dataset S1, Table S13, shows a similar model, but the dependent variable Y_i is the eigenvector centrality of person i (net of siblings). This regression shows that even when controlling for the number of direct contacts, leaders tend to have higher centrality, suggesting that leaders not only have more friends, but their friends are more popular and they have more friends of friends as well.

Models with Raid and Social Information and Fixed Effects.

In this section, we describe methods to explore associations between raid characteristics and raid participation.

Dataset S1, Tables S14–S20, shows linear regressions that measure the association between an individual’s decision to join a particular raid and various raid characteristics. The basic model is the following:

$$E\left[Y_{ir}\right]=\alpha +{\theta }_i+{\gamma }_r+\mathit{\beta }{\mathit{x}}_{\mathit{ir}},$$

where the dependent variable Y_ir is 1 if person i joins raids r, and 0 otherwise; α is a constant (dropped if fixed effects are included); x_ir is a vector of characteristics for participant i and raid r; and θ_i and γ_r are individual and raid fixed effects, respectively.

Fixed effects are included to control for variation in stable characteristics across individuals (e.g., are some individuals inherently more likely to join raids?) and across raids (e.g., are some raids inherently more important?). This approach effectively controls for all possible stable individual and raid characteristics. For example, it ensures that personal differences that may impact the tendency of a person to engage in risky behavior or differences that may impact the importance of a raid are not driving the results. Additionally, because there are multiple (and probably correlated) observations for both raids and individuals, we adjust SEs by clustering them on both raids and individuals using multiway clustering (54).

Dataset S1, Tables S14 and S15, shows regressions of raid participation on the total number of leaders and nonleaders who joined the raid. Although Dataset S1, Table S14, shows that the number of nonleaders participating is significantly associated with the decision to join, when we control for individual fixed effects in Dataset S1, Table S15, the association ceases to be significant. Because neither the number of leaders nor the number of nonleaders survives both specifications, we turn to models based on participation by friends rather than total participation.

Dataset S1, Tables S16–S18, show regressions of raid participation on social aspects of raid composition. Dataset S1, Table S18, shows the strictest specification with both raid and individual fixed effects. As discussed in the main text, the number of first-degree leaders, first-degree friends, and second-degree friends on a raid are all significantly associated with raid participation, and these results survive multiple model specifications and strong controls for fixed individual and raid characteristics.

Dataset S1, Table S19, shows a regression of raid participation on the number of siblings also in the raid with and without individual or raid fixed effects. Only in the model without fixed effects is the number of siblings significant. For completeness, we consider the prior (full) model with siblings separated out.

Dataset S1, Table S20, shows a regression of raid participation on the social aspects of raid composition (as in Dataset S1, Table S18) with the further inclusion of the number of siblings who participated on raids. We again find that leaders of distance 1, friends of distance 1, and friends of distance 2 are significant, whereas the number of siblings on raids (net of leader and nonleader friends of distances 1, 2, and 3) is not significant. This again suggests that it is friends, not siblings, that matter for the emergence of violence.

Finally, Dataset S1, Table S21, shows a model that regresses the number of nonleader friends who join a raid on a person’s own decision to join the raid (1 = joined), their leadership status (1 = leader), and an interaction of the two. The results suggest that leaders who join raids actually mobilize significantly fewer individuals to join than nonleaders, and this is despite the fact that leaders tend to have more friends as shown in Dataset S1, Table S12.

Dataset S1, Table S22, shows a model that regresses a person’s own decision to join a raid upon the number of leaders and nonleader friends, both net of siblings, who join a raid. These results support the hypothesis that a friendship relationship is more important than a family (sibling) relationship in deciding to join raids. Again, net of siblings, the first-degree friendship effect is by far the largest determinant for joining raids.

To further test the hypothesis that it is friendship ties, and not family ties, that are important, we performed a permutation test of which raids individuals join, keeping the number per raid and the total number of raids participated in by each individual constant, and we asked what percentage of raids have any sibling pairs. We find that the observed value lies near the center of the distribution of the permuted values, suggesting that siblings do not raid in pairs more frequently than chance, in line with the regression analysis.

Network Graphs.

The friendship network was drawn with iGraph in R using the Kamada–Kawai algorithm. Node colors indicate participation status (green, nonparticipants; blue, participants; red, leaders), size is increasing with the number of raids in which a person participated (larger, more active), and arrows indicate gift-giving direction.