Partner choice and cooperation in social dilemmas can increase resource inequality

The study investigates how partner choice affects cooperation and resource distribution in social dilemmas when individuals differ in endowment (resources available to contribute) and productivity (returns generated by contributions). While partner choice often promotes cooperation by allowing cooperators to assort and avoid defectors in equal societies, real-world contexts feature inequality in resources and productivity, which may influence partner attractiveness and access to cooperative opportunities. The authors experimentally test whether partner choice in an unequal setting leads to (a) inequality-averse selection that could reduce disparities, (b) homophily-driven assortative pairing by similar types, or (c) a population-wide preference for highly endowed, highly productive partners (HH) that segregates the advantaged from others, potentially amplifying inequality. The purpose is to assess consequences of partner choice for cooperation, segregation, and the evolution of resource inequality within groups.

Prior research shows partner choice enhances cooperation by enabling assortative interactions among cooperators and motivating defectors to switch (e.g., computational and experimental evidence). Most past work assumes equal ability to reciprocate, though some studies consider heterogeneity in productivity/endowments. Individuals use cues like wealth/endowment and productivity to select partners, affecting cooperation opportunities. Homophily—preference for similar others—can shape network structures and segregation, supported by theoretical models and empirical observations in social networks and neighborhoods. Inequality in networks influences cooperation patterns, with evidence that people preferentially cooperate with wealthier partners to maintain ties, potentially increasing system-level inequality. The current work extends this literature by experimentally examining partner choice under ex-ante inequality in endowment and productivity, testing whether partner choice fosters segregation and exacerbates resource disparities despite potential cooperation benefits.

Design: 2 (pairing condition: partner choice vs assigned partner) × 4 (type: HH, HL, LH, LL) between-subjects experiment. Sample: n = 336 across 42 groups of 8; 21 groups per condition (n = 168 each). Within each group, two participants per type: HH (high endowment 75 units, high productivity 1.7), HL (75, 1.3), LH (25, 1.7), LL (25, 1.3). Recruitment: 32 groups via Leiden University (n = 256) and 10 groups via Prolific (n = 80); identical procedures and incentives. Ethics approved; pre-registered (AsPredicted #53435). Task: Participants played 24 rounds of a two-person public goods game. In each round, paired participants simultaneously decided how many units of their endowment to contribute. Each contributed unit was multiplied by the contributor’s productivity factor; the resulting total was split equally between pair members. Full mutual cooperation maximized joint outcomes; unilateral defection maximized individual gain. Pairing manipulation: Assigned partner condition—pseudo-random algorithm ensured each participant interacted six times with each partner type (uniform exposure across types). Partner choice condition—at the start of each round, participants ranked partner types (1–4). Matching algorithm paired participants with their highest-ranked feasible partner type based on mutual compatibility; if not feasible, moved to next ranked type; random tie-breaking when necessary. Participants learned partner type before contribution decisions. Measures: Primary outcomes included (i) segregation/assortment (frequency of same-type vs mixed-type pairs; stability measured as average consecutive rounds together), (ii) partner preferences (popularity, rejection rates, avoidance dynamics, and their change over time), (iii) cooperation (relative cooperation: contributions as percentage of own endowment; changes over time; dependence on partner type and pairing stability; effect of being matched with first-choice type), and (iv) accumulated resources (units earned at the end of the game; distributional inequality via shares and Gini index). Analyses: Multilevel (logistic) models using lme4 in R with random intercepts for participants nested within groups (for repeated measures), or for groups (for accumulated resources at end). Two-tailed tests; Bonferroni corrections where multiple contrasts applied. Exploratory checks assessed differences between participant pools (Leiden vs Prolific). Incentives: Units converted at 2600 units = €1; additional incentivized tasks included expectation accuracy and a social value orientation slider measure.

• Segregation under partner choice: Partner choice increased same-type pairings compared to assigned partners (MLLM, z = −7.56, b_condition = −1.68, p < 0.001, 95% CI [−2.18, −1.19]). Same-type pairs were more stable (MLM, t(292) = 12.14, b_similar = 2.77, p < 0.001, 95% CI [2.32, 3.23]). HH-HH pairings were most prevalent (73% of HH rounds), and LL-LL pairings were also frequent (66.9% of LL rounds).
• Partner preferences: HH types were most popular (65.1% of all first choices were HH; MLLM, z = 9.76, b_type = 5.56, p < 0.001, 95% CI [4.25, 6.86]). LL types were most rejected (could not be paired with first choice in 75% of rounds; MLLM, z = 6.34, b_type = 1.85, p < 0.001, 95% CI [1.18, 2.53]) and actively avoided (after LL partner, higher likelihood to prefer a different type; z = 5.67, b_previouspartner = 0.72, p < 0.001, 95% CI [0.43, 1.01]).
• Dynamics over time: Non-HH participants’ preference for HH decreased over rounds (MLLM, z = −9.24, b_round = −0.07, p < 0.001), while HH preferences did not change (z = 0.15, b_round = 0.002, p = 0.879). Homophily increased over time for HL (z = 2.57, b_HL×round = 0.06, p = 0.010), LH (z = 1.98, b_LH×round = 0.04, p = 0.048), and LL (z = 3.19, b_LL×round = 0.07, p = 0.001).
• Cooperation: In the assigned partner condition, relative cooperation declined over time (MLM, t(7726) = −4.38, b_condition*round = −0.33, p < 0.001), whereas no significant time trend was found in the partner choice condition (t(7726) = −1.54, b_round = −0.08, p = 0.125), noting exploratory pool differences. Participants adjusted preferences after encountering a less cooperative partner (MLM, z = 3.67, b_contribution = 0.37, p < 0.001). Stable pairings were associated with higher cooperation (MLM, t(389) = 6.03, b_stability = 2.46, p < 0.001). Cooperation was higher toward HH partners (M = 70.75%, SE = 5.61) and lower toward LL partners (M = 50.83%, SE = 5.55). Being paired with a first-choice type increased cooperation (MLM, z = 13.28, b_ranking = 14.11, p < 0.001).
• Resource accumulation and inequality: HH participants earned substantially more under partner choice relative to other types (MLM, z = 49.09, b_type = 1345.93, p < 0.001) and more than HH in the assigned condition (z = 7.70, b_condition = 326.95, p < 0.001). LL participants earned less relative to others under partner choice (z = −41.05, b_type = −1125.44, p < 0.001) and less than LL in the assigned condition (z = −3.34, b_condition = −141.71, p = 0.005). Total population earnings did not differ by condition (z = −1.17, b_condition = −36.04, p = 1.00). Distribution skew increased markedly with partner choice: Gini rose from 0.24 (assigned) to 0.44 (partner choice); HH received 51.8% of generated earnings (vs 33.6% in assigned), LL received 9.0% (vs 20.2% in assigned).

Findings support the hypothesis that when individuals differ in endowment and productivity, partner choice produces segregation and amplifies pre-existing resource inequalities. A strong population-wide preference for highly endowed, highly productive partners led HH participants to almost exclusively pair with each other, while LL participants were frequently rejected and clustered together not by choice but due to lack of demand. This assortment stabilized cooperation overall relative to assigned partnerships but concentrated cooperative gains within advantaged pairs. Over time, partner choice fostered emergent homophily among non-HH types, likely driven by rejection avoidance and feasibility constraints. The results align with network studies showing preferential cooperation with wealthier partners and consequent increases in system-level inequality. In this anonymous setting where only general characteristics (endowment/productivity) were visible, structural attributes overshadowed behavioral reputation in partner selection, further fueling segregation and unequal resource accumulation. The study thus nuances the view of partner choice as universally beneficial for cooperation: in unequal contexts, it can be a mechanism of cumulative advantage that stabilizes within-group cooperation while widening disparities across types.

The study demonstrates that in unequal environments, partner choice—despite stabilizing cooperation—drives segregation and magnifies resource inequality. Advantaged (HH) participants preferentially pair and cooperate with each other, capturing a disproportionate share of generated resources, while disadvantaged (LL) participants cluster together, cooperate less, and fall further behind. Contributions include (i) causal evidence that partner choice under ex-ante inequality increases within-population resource disparities, (ii) identification of dynamic preference shifts and emergent homophily under feasibility constraints, and (iii) quantification of cooperation and distributional outcomes (e.g., increased Gini). Future research should examine how providing behavioral reputation alongside structural attributes affects partner selection and whether observed patterns persist with larger stakes, different network sizes/structures, or mechanisms that facilitate inclusion of disadvantaged actors (e.g., subsidies, matching schemes, visibility of cooperative behavior).

• External validity: The experiment used artificial types with fixed endowment/productivity and anonymity; real-world settings may include richer reputational information and dynamic attributes.
• Incentives and stakes: While incentivized, payoff magnitudes were modest; generalization to higher-stakes contexts is inferred from meta-analytic patterns but untested here.
• Sample composition: Two participant pools (Leiden University and Prolific) showed exploratory differences (e.g., time trends in cooperation under partner choice), which may limit generalizability.
• Measurement scope: Partner choice was based on type-level information; behavior-based partner selection might yield different outcomes.
• Statistical/modeling: Some model assumptions were not fully met, though multilevel models are robust; multiple testing corrections applied, but residual risks remain.
• Pre-registration deviations: A planned response-time exclusion was not implemented due to lack of time limits; however, no participants were excluded, and robustness checks excluding failed attention checks did not alter results.