Evidence of direct and indirect reciprocity in network-structured economic games

The study investigates how social standing and reputation structure cooperation, exploitation, and punishment in human communities. While kin selection and direct reciprocity explain some cooperation, they fall short of accounting for widespread cooperation among non-kin and strangers. Indirect reciprocity—where individuals condition behavior on others’ reputations—has been theorized to stabilize large-scale cooperation by incentivizing prosocial behavior and norm enforcement. The authors test predictions about positive and negative, direct and indirect reciprocity in real-world social networks using network-structured economic games and peer-rated reputations in four rural Colombian communities. The central questions are: (1) Do individuals condition cooperative, exploitative, and punitive behaviors on their dyadic perceptions of others’ standing? (2) Do individuals’ behaviors shape how they are perceived by others? (3) Do generalized, community-wide reputations predict being targets of cooperation, exploitation, and punishment beyond dyadic relationships?

Background research establishes that human cooperation extends beyond kinship ties and is influenced by reciprocity, status, and reputation. Direct reciprocity can maintain cooperation in constrained settings but is limited in larger groups and transient interactions. Models of indirect reciprocity link observable behavior to reputation, creating incentives to cooperate and to punish defectors. Theoretical work distinguishes positive indirect reciprocity (helping those in good standing) and negative indirect reciprocity (targeting those in bad standing), with recent models proposing that targeted exploitation of individuals with selfish reputations can catalyze cooperation emergence. Empirical evidence across societies shows roles for punishment, partner choice, and reputation in cooperation, yet few studies simultaneously examine cooperation, exploitation, and punishment networks with both dyadic perceptions and generalized reputations, particularly in non-WEIRD contexts. This study addresses that gap.

Setting and participants: The study was conducted in four rural Colombian communities: coastal (n = 186 egos; 220 alters), lowland (n = 154; 178), highland (n = 45; 53), and altiplano (n = 111; 160). Across the broader project, participants average 39.8 years (sd = 17.9), 58% female, 44% Afrocolombian, 11% Emberá, and 45% Mestizo. Communities vary in subsistence ecology (fishing, horticulture, wage labor, agriculture, floriculture) and socioeconomic austerity.

Design and procedures: Respondents completed three RICH network-structured economic games and provided dyadic peer ratings of generosity and selfishness using a photo roster of adult community members. Games were recipient-identity conditioned and decider-confidential.

• RICH allocation (giving) game: Each respondent received an endowment (20,000–30,000 COP depending on site) in 1,000-peso coins to keep (self-photo) or allocate to others (other photos).
• RICH exploitation (taking) game: A coin was pre-allocated to each photo (500 or 1,000 COP depending on site). Respondents could anonymously remove coins from others to benefit themselves.
• RICH costly reduction (punishment) game: Respondents received 15,000 COP in 1,000-peso coins and could purchase punishment tokens that reduced a target’s payout with a 4× multiplier (4,000 COP reduction per token).
• Dyadic peer ratings: Participants placed tokens on photographs indicating which community members were especially generous (green) or selfish (purple). No fixed limits on tokens; typical use was 7–9 per category.

Data volume: Nind = 496 individuals; Nobs = 53,876 ratings/transfers. Descriptive network statistics (density, reciprocity, transitivity, average distance) were reported for each community and layer.

Ethics: Approved by the Department of Human Behavior, Ecology, and Culture (MPI-EVA). Verbal informed consent obtained; participants paid in cash.

Analytical strategy: A multiplex Bayesian Social Relations Model (SRM) jointly modeled binary ties G_{j,k,m} (placing a coin/token by j on k in layer m) with Bernoulli-logistic regression. The linear predictor included: θ_{j,k,m} = γ_m + α_{j,m} (sender) + β_{k,m} (receiver) + δ_{j,k,m} (dyad) + …

Generalized reciprocity was estimated via multivariate normal structures for sender and receiver effects across layers, with weakly informative priors: unit normal raw effects, Exponential(2.5) priors on SDs, LKJ(2.5) priors on correlation Cholesky factors. Dyadic reciprocity was modeled by concatenating dyad effects across layers with constraints ensuring symmetry and equal variances across paired directions (j→k, k→j), using priors to enforce block-structured correlation matrices. The approach estimated within-layer and cross-layer correlations (e.g., if giving to k predicts not exploiting k). Models were coded in Stan (rstan), diagnostics (trace plots, R-hat, ESS) indicated good convergence.

Dyadic structure and direct reciprocity:

• Within-layer dyadic reciprocity was positive: • Giving reciprocated: correlations ~0.61–0.72 across sites (e.g., 0.61 [0.52, 0.68] coastal; 0.70 [0.63, 0.77] lowland; 0.63 [0.48, 0.77] highland; 0.72 [0.65, 0.79] altiplano). • Exploitation reciprocated: ~0.21–0.53 (e.g., coastal 0.21 [0.15, 0.27]; lowland 0.53 [0.48, 0.57]; highland 0.53 [0.37, 0.70]; altiplano 0.36 [0.29, 0.43]). • Punishment reciprocated in three sites (e.g., coastal 0.40 [0.21, 0.58]; lowland 0.46 [0.25, 0.65]; altiplano 0.56 [0.36, 0.74]); not reliable in highland (0.01 [−0.48, 0.48]).
• Mutual perceptions reciprocate: Dyadic reciprocity for generosity ratings (e.g., ~0.30–0.57) and for selfishness ratings (~0.59–0.66).

P1: Behavior conditioned on dyadic perceptions (direct positive/negative reciprocity):

• Giving: Individuals gave more to those they personally rated as generous (e.g., 0.75–0.85 across sites) and less to those they rated as selfish (negative correlations; e.g., coastal −0.33 [−0.42, −0.23]; lowland −0.14 [−0.22, −0.06]; highland −0.42 [−0.63, −0.21]; altiplano −0.03 [−0.13, 0.07], not reliable).
• Exploitation: Individuals were less likely to exploit those they rated as generous (e.g., −0.53 to −0.81 across sites) and more likely to exploit those they rated as selfish (e.g., 0.24–0.49).
• Punishment: Individuals were less likely to punish those they rated as generous (e.g., coastal −0.28 [−0.39, −0.16]; lowland −0.51 [−0.64, −0.37]; highland −0.12 [−0.47, 0.27]; altiplano −0.34 [−0.49, −0.18]) and more likely to punish those they rated as selfish (e.g., coastal 0.73 [0.64, 0.82]; lowland 0.77 [0.67, 0.85]; highland 0.16 [−0.25, 0.52]; altiplano 0.82 [0.72, 0.90]).

P2: Behavior predicts perceived standing (dyadic effects of j’s behavior on k’s ratings of j):

• Being given to increased rating j as generous (e.g., 0.63–0.72 across sites).
• Being exploited by j decreased rating j as generous (e.g., −0.36 to −0.64 across sites). Punishment by j tended to decrease ratings of j as generous (significant in lowland: −0.26 [−0.41, −0.12]; weaker/uncertain elsewhere).
• Giving to k reduced k’s rating of j as selfish at lowland and highland (e.g., −0.18 [−0.27, −0.08]; −0.39 [−0.58, −0.19]); not reliable at coastal/altiplano. Exploiting or punishing k tended to increase k’s rating of j as selfish in three of four sites.

P3: Generalized (community-wide) reputations predict behavior beyond dyads:

• Individuals with generous reputations received more giving (e.g., 0.50–0.66 across sites) and were less likely to be exploited (e.g., −0.21 to −0.52 across sites).
• Generous reputation did not consistently reduce being punished; in the coastal site, generous-reputed individuals were more likely to be punished (0.23 [0.05, 0.41]).
• Individuals with selfish reputations were more likely to be punished (e.g., 0.57–0.84 across three sites; highland estimate imprecise), but were not generally more exploited (except highland: 0.36 [0.11, 0.59]).

Additional patterns:

• Profiles of punishers varied by community: in lowland, punishers were more generous and less exploitative; in coastal and altiplano, punishers tended to be more exploitative and less generous.
• Mixed evidence for negative indirect reciprocity at community level: while selfish reputation predicted punishment, it did not robustly predict exploitation; exploitation by individuals did not consistently attract more punishment at the generalized level.

Network descriptives (illustrative): For example, exploitation layers were dense (density ~0.41–0.59) with high transitivity; punishment layers were sparse (density ~0.011–0.031).

Findings support theories that both direct and indirect reciprocity shape cooperative networks. At the dyadic level, individuals cooperated with those they perceived as generous and targeted those they perceived as selfish for exploitation and punishment, demonstrating standing-conditional behavior consistent with direct reciprocity framed by interpersonal sentiments. Behavior also fed back into standing: giving increased being perceived as generous; exploiting/punishing increased being perceived as selfish, supporting ecological validity of the games as proxies for real-world behavior. At the generalized level, community-wide reputations predicted who received allocations and who was punished, indicating that aggregated social perceptions guide interpersonal actions beyond specific dyads. Variation across communities in the profile of punishers suggests local socioecological and normative contexts (e.g., need, moral norms) modulate the linkage between prosociality and norm enforcement. Evidence for hypothesized negative indirect reciprocity was mixed: selfish reputations attracted punishment but not systematic exploitation, and those who exploited many others were not more likely to be punished at the generalized level—possibly reflecting the conditional and covert nature of exploitation. Overall, the results indicate that positive and negative reciprocity, together with costly punishment, help sustain community-wide cooperation networks.

The study contributes an integrated, multiplex approach combining network-structured economic games with dyadic and generalized reputation measures, analyzed via a Bayesian Social Relations Model. It demonstrates that people condition cooperation, exploitation, and punishment on perceptions of others’ standing, and that generalized reputations for generosity attract allocations while selfish reputations attract punishment. These findings provide empirical support for theories of direct and indirect reciprocity in ecologically valid, non-WEIRD settings. Future research should incorporate experimental manipulations within RICH game frameworks, collect longitudinal data to track dynamic reciprocity processes, and expand reputation measures (e.g., multiple status dimensions, additional interpersonal sentiments) to refine understanding of how socio-cognitive features shape multiplex cooperation networks.

The design is correlational; causal mechanisms cannot be definitively established. No experimental treatments were implemented to isolate specific drivers of behavior. Punishment was rare in the highland site, reducing precision of related estimates. The study focused on two reputation dimensions (generosity, selfishness); other reputational and status constructs were not measured. Cross-sectional snapshots may miss temporal dynamics; longitudinal follow-up is needed. Demographic covariates (e.g., age, gender, wealth) were not incorporated into models here, which may leave confounding unaddressed.