Reputations for treatment of outgroup members can prevent the emergence of political segregation in cooperative networks

The study examines how different types of reputation systems influence cooperation and the evolution of social ties when salient social identities (here, political party) are known. The central research questions are: (1) How does visibility of political identity and the way reputations are tracked (undifferentiated, parochial, or distinct intra/intergroup) affect early cooperation, especially toward outgroup members? (2) How do these early cooperation patterns shape subsequent tie formation and dissolution, leading to different levels of political segregation in networks? The context is rising affective polarization in the U.S., where cross-party ties are rarer and distrust is high. The authors hypothesize that when political identities are visible, reputations that do not distinguish by target (undifferentiated) or that only reflect treatment of ingroup members (parochial) will foster ingroup favoritism and political segregation, whereas reputations that separately track treatment of ingroup and outgroup members (intra/intergroup) could mitigate segregation by incentivizing cooperative behavior toward outgroup members.

Prior work shows reputations underpin cooperation and partner choice in dynamic networks, fostering assortative mixing of cooperators and cluster formation. However, social identity theory and evidence on ingroup favoritism indicate people tend to cooperate more with ingroup than outgroup members, especially in political contexts. Undifferentiated reputations may send ambiguous signals in polarized settings, encouraging reliance on category-based trust and ingroup bias. Parochial reputations (tracking only ingroup-directed behavior) reflect real-world information flows that are often ingroup-centric, potentially strengthening incentives to prioritize ingroup cooperation and ignore outgroup treatment. Conversely, research on unbounded indirect reciprocity indicates that reputation concerns increase cooperation even toward outgroups when behavior is visible, suggesting that distinct intra- and intergroup reputation scores could motivate cross-boundary cooperation and reduce segregation. The authors position these perspectives as competing predictions for the intra/intergroup reputation system.

Design: A large web-based experiment embedded 1073 U.S. participants (Prolific) into 40 dynamic networks (initial average size ≈26.8; Erdős–Rényi with density 0.167, ~4.5 ties per node). Each tie represented a dyadic iterated Prisoner’s Dilemma (PD). Sessions ran 18 rounds (unknown to participants). Each participant began with 1000 monetary units (MUs; 1000 MU = $1). Game and incentives: In each round and for each alter, participants independently chose how many MUs (0–50, in increments of 10) to send. Sent MUs were deducted from sender and doubled for the receiver. The payoff structure satisfied PD inequalities (T=100>R=50>P=0>S=-50; and 2R=100>T+S=50). After decisions, participants observed how much they received from each alter. Network dynamics: Every fourth round, participants could sever one existing tie and, if they did, propose one new tie to a non-neighbor (proposal required alter’s acceptance). Tie deletions were unilateral; additions required approval. If a participant ended a tie-update with zero ties, they were isolated from further interaction and moved to a post-study questionnaire. Experimental conditions (between-network): Four conditions varied political identity visibility and reputation information (reputations based on average giving over prior three rounds): - Control: Political identities hidden; single undifferentiated reputation (average giving to all alters). - Undifferentiated reputations: Political identities visible; single reputation not distinguishing ingroup vs outgroup treatment. - Parochial reputations: Political identities visible; reputation based only on how the alter treated their own ingroup members. - Intra/Intergroup reputations: Political identities visible; two reputation scores per alter, one for treatment of their ingroup and one for treatment of their outgroup. Reputations were objective image scores (average giving), not standing-based norms. Participants and assignment: To balance party representation, recruitment targeted self-identified liberals vs conservatives/moderates on Prolific; within sessions, participants reported political leaning on a 6-point scale (1–3 Republican; 4–6 Democrat). In visible-identity conditions, nodes were color-coded (red/blue). Comprehension checks required ≥4/5 correct to proceed; 20–28 qualified participants per session. Measures: Cooperation (MUs sent), direct reciprocity (MUs received from alter in prior round), tie deletion/selection decisions, acceptance of tie requests, network clustering (transitivity, igraph), and political segregation (fewer cross-party ties than expected given density; netseg freeman measure). Statistical analyses: To focus on pre-equilibrium effects, cooperation was modeled for rounds 1–8 with linear mixed models (alters nested in participants; participants nested in networks; AR(1) for serial correlation). Inference used permutation of outcomes within network-rounds (1000 permutations). Direct reciprocity was included where applicable. Tie-deletion choice among alters and new-tie selection among candidates used conditional logistic regression; nonparametric inference used permutations when feasible; otherwise, sandwich SEs. Acceptance of tie requests was modeled with predictors including proposer’s reputation and same-party indicators. Network-level clustering and segregation were analyzed via OLS across 40 networks. Bootstrapping (90% resamples, 1000 reps) provided CIs for some margins; details and sensitivity analyses appear in SI. Payments: $2 (quiz), $1 (start), and $1 per 1000 MUs earned.

- Early cooperation (rounds 1–8): Overall early cooperation was highest in Control and Intra/Intergroup, lowest in Parochial, with Undifferentiated intermediate (Fig. S2). Ingroup favoritism in early cooperation varied significantly by condition (χ²=188.52, DF=3, p<.001; comparing Models 1 vs 2 in Table 3). Parochial showed the largest ingroup favoritism, Undifferentiated the second largest; Intra/Intergroup had smaller ingroup favoritism comparable to Control when accounting for direct reciprocity (Fig. 1B; Table 3, Model 3). - Cooperation convergence: By the end, cooperation was high across all conditions and ingroup-outgroup differences diminished (Fig. 1A). - Severing ties: Participants in Intra/Intergroup were least likely to drop alters across tie-update phases (Fig. S4). When selecting which alter to drop, less cooperative alters were more likely to be dropped (β≈-0.048, p<.001). Political homophily affected dropping mainly in Parochial: participants were significantly less likely to drop ingroup members relative to Control (β=-0.914, p=0.002). Effects in Undifferentiated (β=-0.58, p=0.053) and Intra/Intergroup (β=-0.54, p=0.09) were not significantly different from Control (Fig. 2A; Table S2). In Parochial, an outgroup alter would need to give 14.4 more MUs than an ingroup alter to have equal drop probability. - Tie formation proposals: Across identity-visible conditions, participants more often proposed ties to ingroup members, strongest in Parochial (β=1.445, p<.001), then Undifferentiated (β=1.214, p<.001), and weakest yet still significant in Intra/Intergroup (β=0.958, p<.001) (Fig. 2B; Table S3). In Parochial, an outgroup candidate needed ~23-point higher ingroup reputation to offset homophily; in Undifferentiated, ~18.4 points. - Tie acceptance: No significant same-party advantage in acceptance decisions (β=0.075, p=0.459); acceptance primarily tracked proposer’s reputation (β=0.084, p<.001) (Table S4). - Network structure: Clustering increased over time in all conditions with no significant endline differences (Table 4, clustering model). Political segregation increased significantly in Undifferentiated and Parochial relative to Control, but not in Intra/Intergroup (Table 4; Fig. 3B, Fig. 4A): • Parochial effect on segregation: 0.430 (p<.001) • Undifferentiated effect on segregation: 0.388 (p<.001) • Intra/Intergroup effect on segregation: 0.148 (p=0.092, ns) Thus, merely revealing political identity increased segregation only when reputations were undifferentiated or parochial; distinct intra/intergroup reputations prevented emergent segregation. - Sample sizes: 1073 participants; 40 networks; numerous decision instances (e.g., 67,774 cooperation decisions overall; 26,053 participant-round-alters in Model 3).

The findings show that when political identities are visible, reputation systems shape not only cooperation but also who forms and dissolves ties across party lines. Undifferentiated and parochial reputations produce strong early ingroup favoritism in cooperation and partner choice, leading to politically segregated networks. In contrast, providing separate reputations for treatment of ingroup and outgroup members sustains high early cooperation, reduces ingroup favoritism to control-like levels, and prevents political segregation. These results align with unbounded indirect reciprocity: opportunities to build a positive reputation for treating outgroup members well increase cross-boundary cooperation and selection, even in polarized contexts. Contrary to social identity theory predictions emphasizing ingroup advantage and potential penalization for aiding outgroups, participants did not shun ingroup members who cooperated with outgroup members; instead, prosocial reputations toward outgroups could attract ties. The study highlights design principles for reputation systems in polarized environments: making outgroup-directed behavior legible and valued can counteract identity-based sorting while maintaining cooperation.

The study demonstrates that in dynamic cooperative networks with visible political identities, reputation systems strongly influence whether networks segregate by party or integrate based on cooperative behavior. Parochial and undifferentiated reputations foster early ingroup favoritism and lead to politically segregated networks. In contrast, intra/intergroup reputation systems—tracking treatment of both ingroup and outgroup members—sustain high cooperation and prevent political segregation, yielding integration levels comparable to contexts where political identities are hidden. These insights inform the design of platforms and institutions that aim to preserve cooperation and cross-group ties in polarized settings. Future research should test generalization to domains with symbolic or zero-sum stakes (e.g., social media), incorporate subjective reputation formation, assess downstream attitude change (intergroup contact effects), and explore institutional interventions complementing relationship-level mechanisms.

- Context and incentives: The environment was conducive to mutual monetary gain; results may differ in domains emphasizing symbolic rewards or perceived zero-sum political outcomes (e.g., social media). - Reputation type: Reputations were objective image scores (average giving) and did not encode standing-based norms or subjective evaluations. Real-world reputations may be filtered by partisan perceptions. - Scope and generalizability: Political identity is one social identity; applicability to other identities or more acrimonious contexts may be limited. Effects might attenuate under stronger sectarian norms. - Network design: Only four tie-update phases and one tie change per phase constrain strategic network formation; broader strategic behaviors were limited by design. - Measurement: The study did not measure post-interaction attitudes toward the outgroup, thoughts, or institutional-level processes central to polarization. - Dependence and inference: Network and temporal dependence complicate standard errors; nonparametric permutation and bootstrapping were used, but some estimates lack conventional CIs. - Sample and attrition: Online Prolific sample; some dropouts and isolates occurred, though distributions were analyzed and not evidently biasing conditions.