When does group chat promote cooperation in shared resource governance?

The paper investigates when and how group communication—implemented as text-based group chat—promotes cooperation in shared resource (commons) dilemmas. Prior work shows communication generally improves cooperative outcomes, yet the mechanisms and contextual conditions remain unclear. The authors pose a comparative research question across four experimental games representing different commons settings: Which features of chat content (sentiment and topics) are associated with cooperative behavior, and how do these associations vary by institutional rules, biophysical conditions, and uncertainty? The study’s purpose is to leverage computational text analysis to move beyond treating communication as a binary treatment and instead quantify the content and valence of chats to explain variations in cooperation, thereby informing theory and experimental design in commons governance.

A broad literature links cooperation in commons to trust, group structure, norms, rules, monitoring, and uncertainty. Communication reliably enhances cooperation by clarifying coordination problems, enabling detection/punishment of free-riding, fostering commitments, and building common knowledge and trust through repeated interactions and homophily. Social-psychological perspectives emphasize group identity, though effects are mixed and sensitive to framing. Institutional perspectives highlight how rules and biophysical/material contexts shape incentives, perceptions, and norms, with uncertainty exerting heterogeneous and context-dependent effects. Despite recognition that communication matters, the content of chat has been under-analyzed. Existing single-game studies suggest that strategy-focused communication promotes cooperation, external information can crowd out benefits of communication, and even negative maintenance statements (criticism/disapproval) can improve cooperation. The authors argue for a multi-game, systematic, computational analysis of chat to compare patterns across contexts.

Design and data: The study aggregates round-level data from four laboratory/online behavioral experiments with group text chat: Foraging Game (FOR), Irrigation Game (IRR), Groundwater Game (GG), and Port of Mars (POM). Across 143 groups and approximately 1470 round-level observations, the dataset includes group chat logs and game performance/decision variables. Basic corpus stats: FOR (123 communication rounds; 4470 chats; 27,003 words; 41 groups), IRR (879 rounds; 13,675 chats; 66,437 words; 44 groups), GG (249 rounds; 1419 chats; 7743 words; 25 groups), POM (219 rounds; 4095 chats; 26,595 words; 33 groups). Chats occur each round (timing varies per game), providing synchronized, group-wide text communication.

Dependent variable (group-level cooperation): Defined comparably per game on a 0–1 scale. FOR: total tokens collected per round divided by the theoretical social optimum (1233). IRR: group earnings per round divided by the round-specific social optimum given infrastructure. GG: proportion choosing the low-water-use crop (0 if all choose high-use; 1 if all choose low-use). POM: reinvestment sufficient to avoid system health decline (0 if none; 1 if sufficient reinvestment).

Modeling strategy: Hierarchical linear regression models at the round level, with rounds nested in groups and groups nested in game types. Primary specification includes sentiment (main independent variable) and covariates; random intercepts (and in some split analyses random slopes) accommodate clustering. Models progressively add fixed and random effects. LR tests support multilevel modeling.

Sentiment analysis: VADER (rule-based, −1 to 1) is the primary sentiment measure; polarity scores also considered. Robustness checks employ Bing and NRC lexicons. A dictionary-based approach is used to efficiently handle short texts; domain terms were not expected to invert conventional sentiment semantics.

Topic modeling (STM): Structural Topic Model (R stm) with K=5 per game using spectral initialization. Documents are defined as chat sessions between rounds. Preprocessing includes stop-word removal, stemming, and rare-word filtering. Topic labeling combines model-based metrics (FREX, representativeness, topic quality) and document reading. Topic prevalence (topic proportion per document) is regressed on cooperation and controls using estimateEffect, with uncertainty via the “Global” option. Covariates across models include stock level (infrastructure/water/system health), number of chats, communication inequality (Gini of chats), gender composition, resource variability, limited visibility, and game/round controls as appropriate.

Controls: Stock (remaining resource or system health), variability (1 if variable dynamics apply), limited visibility (1 if information is limited), number of chats (proxy for engagement/understanding), Gini-chats (communication inequality), male percentage, and fixed/random effects for nesting structure.

Sentiment and cooperation (full sample): In multilevel models (N=1470), sentiment (VADER) is generally positively correlated with cooperation in fully specified models but not statistically significant; naïve models showed mixed signs. LR tests confirm the need for multilevel modeling. Communication inequality (Gini) tends to be negatively associated with cooperation in several specifications. The number of chats is negatively associated with cooperation in multiple models (interpreted as low-performing groups communicating more), per narrative results. Higher remaining resource stock correlates with lower cooperation (players exploit more when abundant). Resource variability is positively associated with cooperation; limited visibility is negatively associated. Gender shows little impact overall.

Topic prevalence and cooperation: STM-derived topics (K=5 per game) capture recurring themes: coordination, strategizing, sensemaking/evaluation, and off-topic/socializing.

• Foraging (FOR): Coordination on where to harvest is positively associated with cooperation (Estimate≈1.085, SE=0.623, p≈0.084). Off-topic/socializing Topic 5 is negatively associated (Estimate≈−1.868, SE=0.591, p=0.002). Coordination on when to harvest is positive but not significant.
• Irrigation (IRR): Off-topic/socializing is negatively associated with cooperation (Estimate≈−0.111, SE=0.037, p=0.003). Other coordination/negotiation/evaluation topics show no significant associations.
• Groundwater (GG): Crop choice affirmation is negatively associated (Estimate≈−0.268, SE=0.055, p<0.001), suggesting more affirmation talk in low-cooperation rounds. Long-term goal discussion is positively associated (Estimate≈0.131, SE=0.069, p≈0.057). Other topics are not significant.
• Port of Mars (POM): No topic shows a significant association with cooperation; effects are small and imprecise.

Overall: Similar conversational patterns emerge across games (coordination, strategizing, socializing), but which topics correlate with cooperation is context-dependent. Off-topic/socializing correlates with lower cooperation in multiple games. Coordination is beneficial in FOR but not uniformly across games; addressing long-term uncertainty matters in GG.

Findings indicate that the effects of chat content on cooperation are contingent on the task environment’s rules, biophysical structure, and uncertainty. In FOR, clear spatial/temporal structure and visibility make concrete coordination (e.g., where to harvest) especially effective, while off-topic chat undermines cooperation. In IRR, structural power asymmetries (upstream vs downstream) and limited chat channels complicate coordination; emotionally charged exchanges around perceived inequity can hinder cooperative alignment, with off-topic/socializing negatively related to cooperation. In GG, simple coordination is insufficient; groups that articulate long-term goals under hydrological and horizon uncertainty cooperate more, whereas frequent crop-choice affirmation appears in struggling groups with enforcement issues and waning trust. In POM, pervasive unpredictability and disturbances weaken the link between discussion topics and cooperative outcomes; standard coordinating or sensemaking talk shows little detectable impact.

Null aggregate effects for sentiment underscore that positive or negative tone alone does not reliably predict cooperation; impacts likely depend on alignment with group identity/functions and context. The computational approach reveals cross-game similarities but also critical differences in which topics matter, highlighting that communication’s effectiveness depends on institutional and biophysical features and the nature of uncertainty.

The study introduces a multi-game, computational text analysis of group chat to explain cooperation in shared resource dilemmas, combining sentiment analysis and STM. It shows that while groups commonly engage in strategizing, coordination, and socializing, the cooperative relevance of these topics varies by game context. Coordination discussions promote cooperation in the Foraging game, long-term goal discussions promote cooperation in the Groundwater game, and off-topic/socializing correlates with lower cooperation in Foraging and Irrigation; no topics significantly explain cooperation in Port of Mars. Contributions include a unified cross-game dataset and method that move beyond treating communication as a treatment to quantifying content and prevalence. Future work should integrate human-coded functional categories with supervised learning, develop domain-specific sentiment/semantics, and explore multi-dimensional cooperation metrics. Extending analyses to real-world settings should consider institutional complexity and the role of mis- and disinformation.

Methodological: Dictionary-based sentiment may miss functional nuances (e.g., criticism that constructively promotes cooperation); STM may fail to capture influential but rare or short-lived topics. Statistical power for some game-topic associations may be limited. Construct validity: Cooperation metrics differ by game, potentially shaping observed relationships. External validity: Laboratory/game contexts differ from real-world commons, where legal/normative constraints and mis/disinformation may alter communication effects. Context dependence: Results hinge on specific rules, biophysical conditions, and uncertainties of each game; findings may not generalize to other designs without adaptation.