The Effects of Shocks on Cooperative Social Networks: A Comparative Analysis of Experimental and Simulated Behavior in Small Networks

Co-authors: Keith Burghardt, Luba Levin-Banchik, and Zeev Maoz

Status: Under review

PDF of Manuscript

Abstract: Many networks of cooperation, such as international alliance and trade networks, change in time as agents form, maintain, or drop ties with each other to maximize their utilities. Previous studies suggest that utility can be maximized due to direct ties, interactions between agents’ neighbors, and ties with the same neighbors across different networks. Furthermore, costs associated with maintaining ties, such as tariffs in trade networks, can vary in time, sometimes in abrupt ways. What is less understood is the mechanisms governing the evolution of these connections, in particular, how changes in tie costs impact agent decisions to form, maintain, or end ties? We address these mechanisms via a human experiment in which subjects can form ties with each other to increase their individual utility. After several rounds, some or all subjects were shocked: the cost of forming ties dramatically changes. We compare the experiment to agent-based simulations in which (exploitative) agents act to maximize their utility or make some random (explorative) moves. The model matches experiment data much better when agents are allowed to explore. Moreover, explorative agents can create significantly higher utility than myopic utility maximizing agents. Disagreement between the simulation and experiment suggests that human subjects cooperate more than simulated agents. This results in a paradoxical outcome. Specifically, experimental subjects’ utilities significantly outperform those of sim- ulated agents when tie-costs are low. However, under high tie-costs, excessive cooperative behavior by human subjects hurts their utility. Finally, we find that the effects of shocks are not localized to shocked agents; agents’ behavior is affected by shocked neighbors. We discuss the implications of these results for real-world networks.