## Working Papers

Strategic Experimentation with Uniform Bandit: An Experimental Study

25 Minute Presentation Video (Youtube, Iqiyi)

25 Minute Presentation Video (Youtube, Iqiyi)

Abstract

We study how people perform risky experimentation to generate information when they can also learn from each other. We develop and experimentally test a modified version of the Keller et al. (2005) two-armed bandit model. Our modified model predicts that the information generated by a group of players is no more than that generated by a single player in any perfect Bayes equilibrium. To implement this model in the lab, we design a novel dynamic information structure that can trivialize the posterior calculation for any sequence of signal realizations. We find that 1) when experimenting alone, the median subject generates almost exactly the same amount as the theoretical prediction, that 2) when experimenting with others, the median subject tends to generate more information than when alone, which is against the theoretical prediction, and that 3) the subjects only react to the posterior belief and do not condition their actions on what other players’ past actions, thus the folk theorem breaks down completely in our infinitely repeated environment.

Abstract

This experiment focuses on testing Bayesian persuasion (Kamenica and Gentzkow, 2011) through minimal design. We adopt an experimental design in which the Sender chooses a partition of the state space instead of an information structure. This experimental design makes Bayesian persuasion highly interpretable and, more importantly, eliminates the burden of Bayesian updating for the subjects. We find that 1) the overall behavior of the Senders is qualitatively optimal in the sense that they set the posterior probability of the weaker signal near zero, but 2) they quantitatively do not best respond to the Receivers in the sense that the stronger signals are systematically lower than what the Receivers require, resulting in a persistently high rejection rate of the stronger signal. Moreover, the uncertainty about the requirement of the Receivers is the key impeding factor for the Senders to persuade in that 3) once we replace the Receivers with a robot that plays a known strategy, most Senders learn to play the optimal strategy. This suggests that Bayesian Persuasion, a supposedly difficult problem, is easy to learn once all other sources of difficulties that are not essential to the key strategic element are lifted, although strategic uncertainty from the Receiver side can impede learning.

## Works in Progress

Dynamic Assignment with Limited Commitment

(with Renkun Yang)

(with Renkun Yang)

Abstract

We study the optimal dynamic mechanism design when transfers are not allowed and the principal cannot commit to future allocations. Specifically, the principal (she) decides in each period whether to allocate a good to the agent (he), whose private value evolves over time. The efficiency-maximizing principal bears a cost of allocation that is not internalized by the agent. The optimal mechanism in the two-period model has two main features. First, the principal elicits truthful report from the low-type agent in the first period by promising him one unit in the second period. Second, to fulfill this promise the principal strategically garbles (without observing) the agent's initial report. As the time horizon expands, the efficiency loss from both the noisy communication and the distortion of future allocation is backloaded and vanishes in the infinite horizon limit.

Gender Manipulation

(with Xiaomin Bian)

(with Xiaomin Bian)

Abstract

Gender can be a choice variable in some environments. For example, a person can choose to use a gendered username and/or avatar in online shopping platforms, and a charity organization can choose whether to send a male or female solicitor to collect money from households. Such choices can influence how one is treated by others, which introduces some room for strategies. We use an experiment to first examine whether the recipient’s gendered avatar can influence the proposer’s decision in the dictator game. Then we allow the recipient to choose his/her avatar and examine how the receiver chooses the avatars and how the proposer reacts to it. We find that the proposer does not give more to the recipient with a female avatar but the recipient still chooses a female avatar more often.