Hi! I’m a Ph.D. Candidate in Economics at Cornell University, with an interest in games, information, and applied theory. My work focuses on practical applications of information design. My committee consists of David Easley, Tommaso Denti, and Larry Blume.
I consider a model of Bayesian persuasion with observational learning. A sender designs a test to encourage adoption of a new product, which maybe either good or bad. The test and its outcome are viewed by an opinion leader, who chooses whether or not to adopt the sender’s product. If the opinion leader chooses to adopt, the test and its outcome are next viewed by an audience who also makes an adopt/reject decision. Both opinion leader and audience strictly prefer to adopt the product when it is good, and reject it when it is bad. The cost of adoption is heterogeneous for each individual, and each individual’s adoption cost is their own private information. The sender’s goal is to maximize the probability of audience adoption. Compared to a benchmark in which the audience views the sender’s test directly, I show that the presence of the opinion leader weakly increases the informativeness of the sender’s test. However, it may lower the probability that the audience views the sender’s test, and so the overall effect on audience welfare is ambiguous. I extend this model to include two forms of collusion. Under information-sharing, the opinion leader informs the sender of her private type. In this case the audience weakly prefers the opinion leader to be present, and they may or may not prefer the opinion leader to share her type with the sender. Under cooperation, the opinion leader adopts the sender’s product regardless of the test result. In this case, the audience receives the same welfare as in the benchmark where the sender communicates with them directly.
The Bayesian persuasion framework assumes a sender can choose any random map from the state to any outcome space; I present a model which questions the importance of this assumption. A biased researcher chooses how many subjects to enroll in a trial; each subject improves with some probability when the treatment is good and complementary probability when it is bad. Under pre-registration, the researcher commits to a sample size, while under sequential sampling he observes each subject’s condition before deciding whether to continue or end the trial. I show that as the information contained in a given subject outcome vanishes, under sequential sampling the sender can obtain his first-best Bayesian persuasion equilibrium outcome, but under pre-registration preliminary results suggest the sender cannot do much better than full revelation.
“Hear-No-Evil” Equilibrium (Work in Progress)
A Bayes correlated equilibrium of an incomplete information game (G,S) is any outcome that an omniscient mediator could induce with an appropriate information design policy. Bergemann and Morris (2016) show that the mediator can restrict attention to messages which are incentive-compatible action recommendations, and also that set of BCE is shrinking in the baseline level of information S. However, even though the mediator’s message may be incentive-compatible for an agent to follow after hearing it, it is possible that the agent would prefer not to hear it in the first place. In this paper I define a refinement of Bayes correlated equilibrium which I refer to as hear-no-evil Bayes correlated equilibrium, in which the mediator’s messages must also be incentive-compatible for agents to hear; I show that the set of HNEBCE of (G,S) is shrinking in S if and only if this hear-no-evil condition has no bite.