Prosocial preferences can escalate intergroup conflicts by countering selfish motivations to leave

The study investigates the micro-foundations of stay-or-leave decisions in intergroup conflict. The central questions are when and why potential defenders choose to stay and fight versus leave and avoid risk. The authors hypothesize that (a) the personal cost of leaving shapes leaving propensity, with lower costs increasing leaving; (b) individual risk aversion and loss aversion increase leaving since staying entails risk of defeat and potential losses; and (c) prosocial preferences and solidarity with in-group members decrease leaving and increase willingness to stay and contribute to defense. Understanding these mechanisms is important because stay-or-leave choices influence conflict intensity, group success, and overall social welfare, potentially escalating or de-escalating conflicts depending on participation decisions.

Archival and experimental literatures indicate that conflict often triggers migration and desertion, especially in low-income contexts, and that group norms of cooperation and solidarity are associated with lower desertion. Research on intergroup contests (attacker-defender games) documents how individuals contribute to aggression or defense and how coordination failures waste resources. Theories of risk and loss aversion (prospect theory) suggest heightened sensitivity to potential losses in conflict, increasing incentives to avoid risky participation. Social value orientation (SVO) and in-group identification literatures show that prosocial preferences and solidarity promote cooperation and collective action, including parochial altruism in intergroup settings. Together, prior work suggests competing motivations: self-regarding cost-benefit calculations that favor leaving and other-regarding concerns that favor staying to support the group.

The authors developed the Intergroup Attacker-Defender Contest with Exit (IADC-E), a stylized game with two three-person groups (attackers vs. defenders). Each round, participants receive 20 Experimental Money Units (EMU). Attackers and any defenders who stay decide how much to invest (0–20 EMU) in their group conflict pool. Investments are sunk. If attackers’ total contributions exceed defenders’ total, attackers win and seize the non-invested EMU of defenders who stayed; defenders who stayed then earn zero. If defenders’ total equals or exceeds attackers’, all players keep their remaining EMU. Defenders can alternatively leave, paying a fixed exit cost L, thereby evading attack and earning 20−L EMU. Decisions are simultaneous, and all players learn how many defenders left before investing. If all defenders leave, no conflict occurs, defenders earn 20−L, and attackers keep endowments. Study 1 (n=122 after exclusions; online, Qualtrics, Prolific) was a one-shot implementation. Participants first read instructions, completed comprehension checks, and as defenders reported willingness to pay to leave, stay/leave choices for each cost L∈{0,…,20}, and planned defense contributions conditional on 0–2 others leaving. They then made attacker contributions conditional on number of staying defenders. Incentives: random role assignment to 3v3 groups; a randomly drawn L implemented; pay based on EMU (conversion specified). Predictions compared to mixed-strategy Nash equilibrium (IADC without exit) benchmarks and prior empirical earnings. Study 2 (n=240; lab, oTree) used fixed groups across four 10-round blocks with varying exit costs: L=5, 7, 10 EMU, and a baseline block with no exit option. Order was counterbalanced. Before gameplay, participants completed SVO slider (prosocial vs selfish) and a risk preference task (iterative multiple price list with 50/50 lottery vs sure amount). Comprehension checks preceded play. Each round: defenders chose to leave or stay, then those who stayed and all attackers chose investments. After each block, participants rated in-group solidarity (composite of closeness, bonding, commitment, solidarity). Expectations about others’ decisions were elicited in blocks (predicted number leaving when exit possible; predicted group contributions when exit not possible). Payments were based on averages of randomly selected rounds and included incentives for belief accuracy. Study 3 (n=240; lab, oTree) fixed L=5 but manipulated asymmetric leaving opportunities across four 10-round blocks: (i) no defender could leave, (ii) exactly one defender could leave, (iii) exactly two defenders could leave, (iv) all three defenders could leave. Which specific individuals could (not) leave remained constant across relevant blocks to minimize reciprocity confounds. No belief elicitation in Study 3. Analyses: Multilevel (logistic) and linear mixed-effects models (lme4 in R) with random intercepts for participants nested in groups accounted for repeated measures and interdependence. Key predictors included leaving cost, number of defenders staying/leaving, previous round success/failure, individual risk preference, and SVO angle. Additional maximum-likelihood modeling in Study 1 estimated a power-function discounting parameter for the stay option. All tests were two-tailed; assumptions checked; no multiple-comparisons correction as no multiple contrasts within models were tested. Preregistrations for Studies 2 and 3 were filed on AsPredicted.

- Economic incentives and risk preferences: - Lower leaving costs increased the likelihood that defenders left (Study 2 MLLM: z=3.18, b_leaving cost 5 vs 7=0.50, p=0.001; z=−4.03, b_7 vs 10=−0.58, p<0.001). - Defenders anticipated this pattern (beliefs: MLM t(6958)=9.57, b_5 vs 7=0.20, p<0.001; t(6958)=−25.51, b_7 vs 10=−0.53, p<0.001). - More risk-averse individuals were more likely to leave (MLLM z=−2.70, b_risk taking=−0.05, p=0.007). - Prior defeat increased subsequent leaving (MLLM z=−8.72, b_previous success=−1.13, p<0.001). - Conflict intensity and outcomes: - Fewer defenders leaving increased group-level conflict contributions for both attackers and defenders (attackers: MLM t(1104.76)=7.69, b_num defenders=3.13, p<0.001; defenders: MLM t(1110.36)=12.20, b_num defenders=4.34, p<0.001). - When leaving was possible, defenders contributed less (MLM t(1113.09)=−7.58, b_leave possible=−4.69, p<0.001), and defenders were less successful compared to no-exit blocks (MLLM z=−3.57, b_leave possible=−0.71, p<0.001). With no exit, defenders won 77% of rounds but 38% of EMU were wasted. - Individually, leaving dominated staying in earnings across cost levels (MLM t(3416.64)=−33.67, b_stay=−6.03, p<0.001). Mixed profiles of some staying and some leaving reduced earnings, revealing a coordination problem (MLM t(3310.09)=14.12, b_dummy coordination=2.60, p<0.001). - Social factors and prosociality: - Defenders were less likely to leave when others in their group had stayed previously (MLLM z=−8.72, b_previous others stayed=−1.08, p<0.001). - Higher prosocial SVO predicted lower leaving (Study 2: MLLM z=−3.14, b_SVO angle=−0.03, p=0.002) and higher in-group solidarity among those who stayed (Study 2: MLM t(3505)=12.76, b_stay=0.36, p<0.001; replicated in Study 3: t(2307)=10.91, b_stay=0.41, p<0.001). - Asymmetric leaving opportunities (Study 3, L=5): - When only one or two defenders could leave, those with the option left far less often than when all could leave (leaving rates: 71.58% when all could leave vs 38.75%/39.00% under asymmetry; MLLM z=−16.07, b_asymmetric leave=−1.94, p<0.001). Those who could leave under asymmetry earned less than when all could leave (MLM t(3556.78)=−6.89, b_dummy could leave=−1.47, p<0.001). - More defenders present increased defense success under asymmetry (MLLM z=4.58, b_num defenders=0.52, p<0.001). - Defenders unable to leave contributed more than in no-exit blocks (MLM t(2358.98)=6.59, b_dummy could leave=1.07, p<0.001) and more than voluntary stayers who could leave (MLM t(1341.53)=−5.72, b_stayed could leave=−1.44, p<0.001), earning less when some could leave (MLM t(39)=−12.49, b_dummy could leave=−5.45, p<0.001). - Among voluntary stayers under asymmetry, 51.43% invested more than the leaving cost (clear sacrifice), 12.82% invested exactly the leaving cost, and 35.74% invested less than the leaving cost (free-riding attempts). In 69.85% of the latter cases, they earned more than the leaving cost. - Prosocial SVO predicted staying (MLLM z=−3.16, b_SVO angle=−0.06, p=0.002) and higher conflict contributions under asymmetry (MLM t(101.34)=2.76, b_SVO angle=0.06, p=0.007). Prosocial defenders earned less than selfish defenders overall (MLLM t(118)=−2.67, b_SVO angle=−0.06, p=0.009). - Study 1 calibration: - In a one-shot IADC-E, some defenders chose to stay at cost levels where leaving was economically superior based on game-theoretic benchmarks (NE expected earnings ≈11.4 EMU; past studies’ weighted average ≈6.13 EMU; expected staying earnings benchmark 12.7 EMU noted), indicating non-selfish motivations to stay.

Findings demonstrate that individual economic incentives (lower exit costs, higher risk aversion, learning from defeat) increase leaving, while social considerations (prosocial preferences, solidarity, observing others stay) reduce leaving and promote defense participation. This dual-motivation framework explains when defenders fight or flee. Allowing exit reduces conflict expenditures and increases individual earnings but can lower defense success rates; forbidding exit increases defense success at the cost of greater resource waste, highlighting a social dilemma and coordination problem. Under asymmetric exit abilities, many who could leave nonetheless stayed—often at personal cost—helping those unable to leave. However, a sizable minority stayed strategically to free-ride on compelled contributors. Thus, staying cannot be taken as unequivocal evidence of solidarity; motivations are mixed. Broader implications include that prosociality, while essential for cooperation and public goods, can inadvertently escalate intergroup conflict by increasing participation and resource expenditure. The results inform theories of parochial altruism, coalitionary conflict, and cost–benefit decision-making under risk and social preferences.

The paper identifies the causal determinants of defenders’ stay-or-leave decisions in intergroup conflict. Economic incentives and risk drive leaving, whereas prosocial preferences and solidarity reduce leaving and increase defense participation. Allowing exit generally improves individual and aggregate welfare by lowering conflict waste but reduces defense success; restricting exit does the opposite. Under asymmetric exit abilities, prosocial actors stay and contribute at personal cost, improving group survival but increasing conflict intensity, while some actors stay to free-ride strategically. The work advances microfoundations of conflict participation and highlights a paradox: the same social concerns that sustain cooperation can escalate intergroup conflict. Future research should examine attackers’ ability to exit (and signaling abstention), effects of information ambiguity and sequential decisions, the role of non-economic and reputational exit costs, and the influence of moral norms (e.g., heroism, duty) and credible commitment signals on conflict dynamics.

- External validity: Laboratory groups were anonymous and randomly formed, likely underestimating real-world in-group identification and solidarity; generalization to migration/desertion in real conflicts requires caution. - Exit modeling: Leaving carried only monetary costs; real-world exits involve multifaceted costs (emotional, social capital, reputational) not captured here. - Attackers’ exit option: Attackers could not explicitly leave (though they could invest zero). Allowing an active abstention option could alter norms and coordination on attacking versus abstaining. - Information structure: Participants knew exactly how many defenders left and decisions were simultaneous. Real conflicts often involve misinformation, ambiguity, and sequential decisions, which may change behaviors. - Strategic heterogeneity: Mixed motivations (prosocial vs free-riding) among stayers complicate interpretation of staying as solidarity; measuring and distinguishing motives in real time remains challenging. - Statistical scope: Although multilevel models were used and assumptions assessed, multiple testing corrections were not applied as no multiple contrasts within models were tested; replication across contexts would strengthen robustness.