Stochastic Choice and Noisy Beliefs in Games

We elicit subjects’ beliefs over opponents’ behavior multiple times for a given game without feedback. A large majority of subjects have stochasticity in their belief reports, which we argue cannot be explained by learning or measurement error, suggesting significant noise in subjects’ unobserved “true” beliefs. Using a structural model applied to actions and beliefs data jointly, we find that such “noisy beliefs” are equally important for explaining our data as “noisy actions”—the sort of stochastic choice given fixed beliefs that is commonly assumed in empirical research. We argue that beliefs and belief-noise are driven by the payoff-salience of actions.