Abstract
In repeated-game applications, where both the collusive and noncollusive outcomes can be supported as equilibria, it is crucial for researchers to understand when each equilibria is more likely to result. Controlled experiments have provided a selection criterion for the two-player repeated prisoner's dilemma. The theoretical selection criterion uses game primitives to measure the set of beliefs such that an agent prefers to defect in all rounds relative to cooperating conditionally. The set of such beliefs is large when uncertainty with respect to what others will do--strategic uncertainty--is large, and the prediction is that a noncooperative outcome would emerge. In this paper, we experimentally test this model of selection and its underlying mechanism of strategic uncertainty by manipulating the total number of players. Our results affirm the model as a tool for predicting when tacit collusion is likely/unlikely to be successful. Extending the analysis, we corroborate the findings and mechanism of the model, in alternative settings, and for the decisions of non-human AI agents.