Bad Barrels Spoiling Good Apples in Social Dilemmas: Using Social Network Information as Escape Route
Carlos de Matos Fernandes, Andreas Flache, Jacob DijkstraBackground
In social dilemmas, one defection can trigger a cascade of further defections in a group through reciprocity in repeated interactions (Jordan et al., 2013; Axelrod, 1984). Accordingly, one bad apple can spoil the barrel, but is this idiom also valid in reverse: Can bad barrels spoil good apples? And if so, under which conditions can social networks, connecting individuals across groups, help to overcome that problem?
A problem arises in situations where groups decide to select new members partially based on information about candidates’ prior performance which is entangled with prior group performance (e.g., group grades in higher education or output of project teams in companies). In such settings, prosocial types whose behavior was “spoiled” by their group mates’ defection can get stuck in bad barrels, i.e. groups in which defection prevails. The problem is that one’s true type cannot easily be assessed based on merits alone.
In this paper, we explore by means of agent-based modelling under which conditions information via network relations, cutting across group boundaries, can help prosocial actors escaping uncooperative groups. We assume that in past dyadic interactions agents learn about others’ types (Buskens & Raub, 2013). We want to know under which structural settings such networks may buffer the effect of spoiledness during selection.
We implemented the model in NetLogo. We analyze a stochastic learning model with adaptive thresholds (Macy, 1991). Generally, if cooperation (defection) generates a positive individual outcome, thresholds decline (increase), making cooperation more (less) likely. Agents are classified as either prosocial or proself. Prosocials have initially lower thresholds than proselfs.
All agents are randomly grouped, playing iterated n-person prisoners dilemmas (INPDs) for x rounds. In a baseline condition (Rule 1), unsatisfied agents have the opportunity to leave a group after x rounds. Ungrouped agents are then sorted based on merit and matched with others with similar revealed group performance, followed by playing INPDs in new groups for another x rounds.
We compare the baseline without network information to a condition in which agents are embedded in a network with varying structural characteristics, generated via a spatial random graph algorithm (Wong, Pattison, & Robbins, 2006; Grow, Flache & Wittek, 2017). We are particularly interested in the effects of network clustering in combination with prosociality homophily because this affects the degree to which good apples can effectively signal others about their prosociality. Agents play 2-person PDs with one of their network partners, eventually generating a reputation. We pair this score with rule 1 information (Rule 2), ideally allowing privately cooperative but INPD defecting agents to signal their ‘goodness’ despite their presence in bad barrels.
Hypothesis 1. Randomly matched groups (Rule 1) generate a similar level of cooperation among proself and prosocial types. Both gradually learn to cooperate, depending on the occurrence of a self-reinforced cooperative equilibrium.
Hypothesis 2. Signalling prosocialness is more apparent when prosociality homophily is possible (Rule 2), allowing prosocials to group more easily together and thus increasing the possibility of emergence of cooperation primarily among prosocial types.