We revisit prominent learning models in which a sequence of agents make a binary decision on the basis of both a private signal and information related to past choices. We analyze the efficiency of learning in these models, measured in terms of the expected welfare. We show that, irrespective of the distribution of private signals, learning efficiency is the same whether each agent observes the entire sequence of earlier decisions or only the previous decision. In addition, we provide a simple condition on the signal distributions that is necessary and sufficient for learning efficiency. This condition fails to hold in many cases of interest. We discuss a number of extensions and variants.

We study a two-player one-arm bandit problem in discrete time, in which the risky arm can have two possible types, high and low, the decision to stop experimenting is irreversible, and players observe each other's actions but not each other's payoffs. We prove that all equilibria are in cutoff strategies and provide several qualitative results on the sequence of cutoffs. Copyright The Econometric Society 2007.