Social networks and cooperation in hunter-gatherers

Abstract

Social networks show striking structural regularities, and both theory and evidence suggest that networks may have facilitated the development of large-scale cooperation in humans. Here, we characterize the social networks of the Hadza, a population of hunter-gatherers in Tanzania. We show that Hadza networks have important properties also seen in modernized social networks, including a skewed degree distribution, degree assortativity, transitivity, reciprocity, geographic decay and homophily. We demonstrate that Hadza camps exhibit high between-group and low within-group variation in public goods game donations. Network ties are also more likely between people who give the same amount, and the similarity in cooperative behaviour extends up to two degrees of separation. Social distance appears to be as important as genetic relatedness and physical proximity in explaining assortativity in cooperation. Our results suggest that certain elements of social network structure may have been present at an early point in human history. Also, early humans may have formed ties with both kin and non-kin, based in part on their tendency to cooperate. Social networks may thus have contributed to the emergence of cooperation.

Figure 1

(a) Cumulative in-degree distributions for the campmate and gift networks are significantly different from random networks with the same number of nodes and edges (Kolmogorov-Smirnov test, p<10⁻¹⁵) and have fatter tails; the random distributions are shown in gray, separately for campmate and gift networks. The gift networks within each camp (ordered by size of camp from smallest, yellow, to largest, blue) show similar distributions of in-degree. (b) Estimates based on dyadic models of social ties (see SI) show that a 1SD change in similarity in characteristics between two people significantly increase the likelihood of a social tie (homophily). Horizontal lines indicate 95% confidence intervals. For the campmate networks, sex is not included because all ties are same sex; homophily for height is not shown due to scale (the estimate is 801%, 95% C.I. 549%–1148%); and homophily for cooperation is shown in Figure 2c. (c) Graphs of the campmate networks show that cooperators tend to be connected to cooperators and cluster together (see also Figure 2b). Node colour and size indicates donation, shape indicates sex. Arrows point from an ego (the naming person) to an alter (the named person). Arrow colours indicate whether the ego and alter are related genetically, affinally (by marriage), or not at all (friendship).

Figure 2

Donations in the public goods game are associated with social network characteristics. A comparison of variance in observed donations with variance in 1000 simulations where donations were randomly shuffled between all individuals in the population (a) shows that between-group variance in cooperation is significantly higher than expected, and within-group variance is significantly lower than expected, at the camp level. An analysis of cooperative behaviour across all camps (b) shows that correlation in cooperation extends to one degree of separation in the campmate networks and two degrees (to one’s friend’s friends) in the gift networks. Moreover, there is anti-correlation at three degrees of separation in the campmate network, suggesting polarization between cooperators and non-cooperators. This correlation cannot be explained by cooperators being more likely to form or attract social ties (c). Instead, subjects with similar levels of giving are significantly more likely to be connected at the dyadic level (c). Finally, several measures of proximity are independently associated with similarity in donations, but social proximity (the inverse of the degree of separation between two people in the network) appears to be just as important as genetic proximity (relatedness) and physical proximity (residence in the same camp) in a multivariate test (d). (Gift networks are defined only within camps and so are not presented for “camp” and “geographic” proximity in 1d.) Vertical lines indicate 95% confidence intervals and stars indicate estimates with p<0.05. See SI for details of the models.