Social Support and Network Formation in a Small-Scale Horticulturalist Population
Human evolutionary research has historically conceptualised social support as a purely dyadic phenomenon (e.g., see Refs. 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16). That is, given some trait pertaining to two persons i and j — e.g., their genetic relatedness, history of helping each other, physical proximity, or difference in wealth — does i help j? Both elegant and tractable, this dyad-centric view of social support evokes classic theoretical models of cooperation as a “Prisoner’s Dilemma” within a void consisting only of ego (i) and alter (j)17. Yet it also belies the fact that aid relationships (i.e., who helps who) constitute complex networks of supportive social bonds that emanate throughout entire human communities.Members of such networks may, in principle, unilaterally help whomever they wish. And their decisions to help — or to not help — specific others comprise a dynamic, supra-dyadic relational context that shapes one’s plausible set of aid targets at the micro level18,19,20,21,22. Put simply, in social support networks, aid is targeted and interdependent across dyads such that the patterning of cooperation among multiple alters jointly affects whom any one network member helps. This sociocentric (i.e., whole network) view of social support is distinct from the perspective taken by evolutionary graph theorists who study the emergence of cooperation on network structure and other spatial substrates (e.g., square grids) that may be fixed or dynamic (e.g., see Refs. 23,24,25). And it is distinct from the perspective taken by analysts of egocentric (i.e., personal) networks who study how the arrangement of intimate relationships exclusively between one’s closest contacts (e.g., the extent to which one’s friends are also friends) eases access to help (e.g., see Martí, Bolíbar, and Lozares26).Differences between the dyad-centric and the sociocentric perspectives on social support are not merely cosmetic. Indeed, the dyad-centric stance of human evolutionary research has led to a situation wherein the relational context of helping behaviour is underexplored. And this has, in turn, impaired understanding of the relative importance of fundamental evolutionary mechanisms to the structuring of cooperative relationships in human communities.Specifically, human evolutionary research on helping behaviour generally takes the theories of kin selection and reciprocal altruism as lodestars. In so doing, sociometric data from subsistence societies across the globe have been used to investigate whether consanguinity (i.e., genetic kinship) and reciprocity govern aid unconditionally and in relation to multiple social and demographic factors. These include affinity (i.e., marriage-based kinship), physical proximity, relative need, homophily (e.g., based on age and gender), social closeness, friendship, religiosity, reputation, conflict, status, and anthropometric measurements such as size, height, and strength. And, on balance, evidence1,2,3,4,5,6,7,8,9,10,13,14,16,27,28,29,30,31,32,33 suggests that helping family and responding in kind when helped are the primary mechanisms by which humans informally distribute resources vital to day-to-day survival (e.g., advice, information, food, money, durables, and physical assistance).However, despite laudable exceptions2,7,15,28,29,30,31,32,33,34 and perhaps due to the influence of methodological trends in the wider behavioural ecology literature on relationships between animals (see Refs. 35,36,37), human evolutionary studies of real helping behaviour have typically relied on non-network methods — namely, monadic regression, dyadic regression, and permutation tests (e.g., see Refs. 1,2,3,5,6,8,9,10,11,12,13,14,16,27). Respectively, these techniques treat the supra-dyadic structure of social support networks as ignorable, reducible to dyads, or a nuisance to be corrected for38. Yet, sociocentric research by sociologists39,40,41,42,43,44,45,46,47,48,49 firmly establishes that humans create and maintain relationships in accordance with factors intrinsic to the supra-dyadic arrangement of network structure itself (e.g., processes of degree-reinforcement and group formation involving at least three persons). And this sociological research makes clear that network-structure-related dynamics can operate simultaneously and independently of non-network factors (e.g., age and kinship).Ultimately, reliance on methods that disregard complex interdependences between aid obscures the extent to which helping family and responding in kind when helped outrank the dynamics of the cooperative system within which decisions to assist specific individuals take place. This uncertainty represents a substantial gap in our scientific understanding of altruism. Accordingly, here I tackle a major point of interest in evolutionary anthropology and human behavioural ecology50 specifically through the lens of the sociology of social networks18,21,51, asking:RQ: How important is helping family and responding in kind when helped relative to supra-dyadic, network-structure-related constraints on the provision of aid?The Current StudyTo answer my research question, I use Koster’s27 recently-released cross-sectional data on genetic relatedness and the habitual provision of tangible aid (e.g., firewood, food, valuable items, and/or physical assistance). Re-analysed here due to their exceptional detail and measurement quality in addition to their broad relevance to the scientific community (see Methods), these data were collected in 2013 and concern a complete population. Specifically, they cover all 108 adult (18+) residents (11,556 ordered dyads) of the 32 households of Arang Dak — a remote village of 279 indigenous Mayangna and Miskito swidden (i.e., “slash-and-burn”) horticulturalists. Arang Dak sits on the Lakus River in Nicaragua’s Bosawás Biosphere Reserve, a neotropical forest in the Department of Jinotega.In total, the tangible aid network that I analyse — i.e., x(t2013)— consists of 1,485 asymmetric aid relationships between the adult residents of Arang Dak. Of the 1,485 aid relationships, 1,422 are verified by the source and the recipient of help. That is, xij(t2013) = 1 if villager i reported in 2013 that they give tangible aid to villager j at least once per month and villager j reported in 2013 that they receive tangible aid from villager i at least once per month. Still, note that Koster’s27 data document self-reported resource flows as opposed to observed transfers. Named sources and targets of aid are based on the village roster — not freely recalled from memory. See Methods for a summary of the data and details on the measurement of the network and kinship.Modelling StrategyTo analyse tangible aid in relation to supra-dyadic network structure (Fig. 1), I use generative network models following Redhead and von Rueden32 and von Rueden et al.33, amongst other human evolutionary scientists2,7,15,28,29,30,31,32,33,34. Specially, I rely on Stochastic Actor-Oriented Models (SAOMs) which are used for observational (i.e., non-causal) analyses of the temporal evolution of networks.Put simply, SAOMs are akin to multinomial logistic regression. More formally, SAOMs are simulations of individual network members’ choices between outgoing relationships with different rewards and costs. These simulations are calibrated or “tuned” to the observed network data. That is, conditional on x (i.e., the observed states of a dynamic network), SAOMs simulate network evolution between successive observations or “snapshots” of the network at (M) discrete time points — i.e., (xleft({t}_{m}right)to xleft({t}_{m+1}right)) — as a continuous-time, Markovian process of repeated, asynchronous, and sequential tie changes. The Markovian process is defined on the space of all possible directed graphs for a set of N = {1, …, n} network members40,42,44,52,53,54,55.SAOMs decompose change between successive network observations into its smallest possible unit. Specifically, “change” means creating one outgoing tie if it does not exist, dropping one outgoing tie if it does, or doing nothing (i.e., maintaining the status quo network). More formally, during a SAOM simulation, focal actors i (ego) myopically modify just one of their outgoing relationships with some alter j in the set of network members N (i.e., j ∈ N, j ≠ i). The change made by i is the change that maximises a utility or “evaluation” function. In this respect, the evaluation function captures the “attractiveness”44 of tie changes — where “attraction” means “…something like ‘sending a tie to [an actor j] with a higher probability if all other circumstances are equal.’” (Snijders and Lomi56, p. 5).The evaluation function itself is a weighted sum of parameter estimates (widehat{beta }) and their associated covariates k (i.e., SAOM “effects”44) plus a Gumbel-distributed variable used to capture random influences55. The simulated tie changes or “ministeps”44 made by i shift the network between adjacent (unobserved) states. These states differ, at most, by the presence/absence of a single tie40,42. And the probabilities of the ministeps — a large number of which are required to bring one observation of the network to the next (i.e., (xleft({t}_{m}right)to xleft({t}_{m+1}right))) — are given by a multinomial logit which uses the evaluation function as the linear predictor.Each covariate k used to specify the evaluation function summarises some structural (i.e., purely network-related) feature or non-structural feature of i’s immediate (i.e., local) network — e.g., the sum of the in-degrees of i’s alters, the number of reciprocated dyads that i is embedded in, or i’s number of outgoing ties weighted by genetic relatedness. These features correspond to theoretical mechanisms of interest (e.g., preferential attachment, reciprocal altruism, or kin selection) and generally take the form of unstandardised sums.SAOM parameter estimates (widehat{beta }) (log odds ratios) summarise the association between the covariates and the simulated tie changes or “ministeps”. Specifically, should a focal actor i have the opportunity to make a ministep in departure from some current (i.e., status-quo) network state x in transit to a new network state x±ij — i.e., the adjacent network defined by i’s addition/subtraction of the tie xij to/from x — ({widehat{beta }}_{k}) is the log odds of choosing between two different versions of x±ij in relation to some covariate k. For example, ({widehat{beta }}_{{rm{Reciprocity}}}=1.7) would indicate that the log odds of i creating and maintaining the supportive relation xij is, conditional on the other covariates k, larger by 1.7 when xij reciprocates a tie (i.e., xji) compared to when xij does not reciprocate a tie (i.e., reciprocated ties are more “attractive”). In contrast, ({widehat{beta }}_{{rm{Reciprocity}}}=-1.7) would indicate that the log odds of xij is, conditional on the other effects, smaller by −1.7 when xij reciprocates a tie compared to when xij does not reciprocate a tie (i.e., reciprocated ties are less “attractive”).Given the longitudinal nature of the model, the gain in the evaluation function for a ministep is determined by the difference Δ in the value of the statistic s for a covariate k — i.e., Δk,ij(x, x±ij) = sk,i(x±ij) − sk,i(x) — incurred through the addition/subtraction of xij to/from x (see Block et al.42 and Ripley et al.44 on “change statistics”). Accordingly, ({widehat{beta }}_{{rm{Reciprocity}}}=1.7), for example, is the value that xij positively contributes to the evaluation function when xij increases the network statistic sk,i(x) underlying the Reciprocity effect by the value of one (i.e., ΔReciprocity,ij (x, x±ij) = sReciprocity,i(x±ij) − sReciprocity,i (x) = 1 − 0 = 1).The probabilities of network members being selected for a ministep is governed by a separate “rate” function. And the baseline rate parameter λ is a kind of intercept for the amount of network change between successive observations of the analysed network. Larger baseline rates indicate that, on average, more simulated tie changes were made to bring one observation of the network to the next (i.e., (xleft({t}_{m}right)to xleft({t}_{m+1}right))).However, as the data from Nicaragua are from a single point in time (i.e., 2013), I use the cross-sectional or stationary Stochastic Actor-Oriented Model (cf. von Rueden et al.33). Accordingly, Arang Dak’s tangible aid network is assumed to be in “short-term dynamic equilibrium.” As Snijders and Steglich40 (p. 265) discuss in detail, “this ‘short-term equilibrium’ specification of the SAOM is achieved by requiring that the observed network is both the centre and the starting value of a longitudinal network evolution process in which the number of change opportunities per actor [i.e., λ] is fixed to some high (but not too high) value.”Practically speaking, this means that the cross-sectionally observed network is used as the beginning and the target state for a SAOM simulation — i.e., (xleft({t}_{2013}right)to xleft({t}_{2013}right)) — during which actors are allowed to make, on average, very many changes (i.e., λ) to their portfolio of outgoing ties. These simulated tie changes produce a distribution of synthetic networks with properties that are, on average, similar to those of the cross-sectionally observed network in a converged SAOM — where the target properties correspond to the researcher-chosen SAOM effects k. Put simply, “[cross-sectional] SAOMs assume that the network structure, although changing, is in a stochastically stable state.” (Krause, Huisman, and Snijders57, p. 36–37). Thus, the estimated parameters (widehat{beta }) continue to summarise the rules by which ministeps unfold. However, the asynchronous, sequential, simulated tie changes, in a sense, “cancel out” and thus hold the network in “short-term dynamic equilibrium”40,42. Formally, the cross-sectional SAOM is defined as a stationary distribution of a Markov Chain with transition probabilities given by the multinomial logit used to model change between adjacent network states40,42.The rate parameter λ is fixed at 36 for my analysis. The value of 36 is the maximum observed out-degree in the source-recipient-verified tangible aid network x(t2013). Accordingly, under λ = 36, all members of the tangible aid network have, on average, at least one opportunity to modify their entire portfolio of outgoing ties during the simulations. Nevertheless, to ensure the robustness of my results, I also fit a second set of models for which λ was fixed to 108 (i.e., thrice the maximum out-degree).Model SpecificationTo assess the importance of kinship and reciprocity to hypothetical decisions to help others (i.e., ministeps), I use four archetypal specifications of the SAOM’s evaluation function. These model specifications feature nested sets of covariates (i.e., the SAOM “effects”44). And, using language found in prior evolutionary studies3,5, I refer to these archetypal specifications as the “Conventional Model” (Model 1) of aid, the “Extended Model” (Model 2) of aid, the “Networked Aid Model (Limited)” (Model 3), and the “Networked Aid Model (Comprehensive)” (Model 4).The first specification (i.e., Model 1) comes from Hackman et al.3 and Kasper and Borgerhoff Mulder5 who respectively label it the “Human Behavioural Ecology” and “Conventional” model. This specification is comprised of just four dyadic covariates — one each for consanguinity (i.e., Wright’s coefficient of genetic relatedness), affinity (i.e., Wright’s coefficient of genetic relatedness between i’s spouse s and his/her blood relative j), the receipt of aid, and geographic distance. The first three covariates are used to test long-standing predictions of helping in order to reap indirect and direct fitness benefits in line with the theories of kin selection and reciprocal altruism (see Refs. 1,5,27,58,59 for primers). And the fourth covariate is used to adjust for tolerated scrounging — i.e., what Jaeggi and Gurven4 (p. 2) define as aid resulting from one’s inability to monopolise resources due to costs imposed by the resource-poor — where a covariate for distance operationalises pressure to help imposed by those who are spatially close4.The second specification (i.e., Model 2) reflects Kasper and Borgerhoff Mulder’s5 and Thomas et al.’s9 extensions to the conventional model (see also Page et al.16). Specifically, and following important work by Allen-Arave, Gurven, and Hill1, Hooper et al.14, and Nolin7, it is distinguished by nuanced tests of kin selection and reciprocal altruism via interactions between: (i) consanguinity and the receipt of aid; (ii) consanguinity and relative need; and (iii) consanguinity and geographic distance. Furthermore, Kasper and Borgerhoff Mulder’s5 and Thomas et al.’s9 extended model includes covariates for the non-network-related attributes of individuals (e.g., gender, wealth, and physical size), thus adjusting for homophily, trait-based popularity, trait-based activity, and local context (e.g., results from a gift-giving game9 or, in the present case, infidelity and discrimination based on skin-tone27).The third specification (i.e., Model 4) is my revision of the second. It is geared to make the relational context of aid explicit. This is done using nine covariates that account for the breadth of sociologists’ contemporary understanding of supra-dyadic interdependence between positive-valence (i.e., not based on disliking or aggression), asymmetric social relationships39,40,41,42,43,44,45,46,47,48,49. In keeping with the nature of the SAOM, each of these covariates summarises some structural feature of a villager’s immediate (i.e., local) network (e.g., the number of transitive triads that she is embedded in). Accordingly, each structural covariate is used to capture a form of self-organisation — i.e., network formation driven by an individual’s selection of alters in response to network structure itself (Lusher et al.49, p. 10–11 and 23–27).Specifically, the covariates added in the third specification reflect predictions derived from three fundamental sociological theories of the emergence of non-romantic relationships. The first is structural balance theory which posits that individuals create and maintain ties that move groups of three people from an intransitive to a transitive state (i.e., transitive closure), the latter of which is understood to be more psychologically harmonious or “balanced” (see Refs. 39,43,47,48,60,61,62 for primers). The second is Simmelian tie theory which posits that, once formed, individuals will maintain relationships embedded in maximally-cohesive groups of three people such that 3-cliques (i.e., fully-reciprocated triads) are resistant to dissolution (see Refs. 43,48,63 for primers). The third is social exchange theory (as it relates to structured reciprocity) which posits that individuals will unilaterally give benefits to others in response to benefits received such that indirect reciprocity (i.e., returns to generosity) and generalised reciprocity (i.e. paying-it-forward) in groups of three people encourage cyclic closure — i.e., the simplest form of chain-generalised exchange (see Refs. 19,20,43 for primers). Furthermore, the third specification reflects the broad prediction that individuals vary in their propensity to send and receive relationships based on their structural position alone (e.g., popularity-biased attachment) leading to dispersion in the distribution of in-degrees and out-degrees (see Refs. 39,44,49 for primers) — especially for ties with an inherent cost to their maintenance39,42.Last, I consider a fourth specification (i.e., Model 3) that uses a subset of the nine network-structure-related covariates included in Model 4. This limited set of structural effects typifies the specifications used in prior human evolutionary studies of empirical help that present generative models of entire networks2,7,15,28,29,30,31,32,33,34. Specifically, the fourth specification features just three network-structure-related covariates to account for structural balance theory, self-reinforcing in-degree (i.e., popularity-bias), and the interplay between in-degree and out-degree.Descriptive statistics for the relevant attributes of the 108 residents of Arang Dak appear in Table 1. Formulae used to calculate the network statistics sk,i(x) underlying each effect k used to specify my SAOMs, alongside verbal descriptions to aid reader interpretation, appear in Online-Only Table 1. See Methods for additional rationale behind the third specification.Table 1 Descriptive statistics for the monadic and dyadic attributes of the residents of Arang Dak.Full size tableModel ComparisonCompared to prior human evolutionary research on social support networks, I take two novel approaches to gauging the importance of kinship and reciprocity to help. First, I use a technique41 specifically designed to measure the relative importance of individual effects in SAOMs (see Methods). And second, I evaluate each specification’s ability to produce synthetic graphs with topologies representative of the structure of the analysed tangible aid network64.Judging model specifications using topological properties reflects one of the core purposes of methods such as the SAOM and the Exponential Random Graph Model (ERGM) — i.e., to explain the emergence of global network structure (see Refs. 40,42,46,47,49 also Refs. 18,48), not simply the state of individual dyads (i.e., is aid given or not?). Admittedly, explaining global network structure is not a stated primary aim of dyadic-centric or sociocentric studies of help by human evolutionary scientists, including those wherein authors rely on SAOMs or ERGMs2,7,15,28,29,30,31,32,33,34. Still, topological reproduction is an important, strong test of the relative quality of the four archetypal specifications as each encodes the set of rules presumed to govern network members’ decisions about whom to help.To clarify, recall that here it is assumed, a priori, that network members can, in principle, cooperate with whomever they wish, that their cooperative decisions are intertwined across multiple scales, and that their micro-level decisions ultimately give rise to macro-level patterns of supportive social bonds (see Refs. 18,19,20,21,22). The macro-level patterns generated by SAOMs and ERGMs can differ dramatically based on specification40,46,47,49,64,65. Thus, the empirical relevance of a candidate model rests with its ability to produce synthetic graphs similar to the observed structure40,42,46,47,48,49,64. Ultimately, divergence between the real and simulated graphs suggests that a candidate specification is suspect as it does not describe how some network of interest could have formed. More
