UC Berkeley

This dissertation explores economic implications of misinferring from others' behavior. The first two chapters study misinference in models of social learning. They explore in turn two distinct inferential errors: (1) taste projection---the tendency for people to overestimate how similar others' tastes are to their own, and (2) redundancy neglect---people fail to realize that those acting before them also infer from the behavior of predecessors. The final chapter draws out the implications of taste projection in auctions.

More specifically, within social-learning environments, Chapter 1 explores the implications of "taste projection": agents overestimate how common is their own taste. For instance, investors with varied risk preferences learn which prospect minimizes risk and which maximizes expected return. Inference requires agents to assess if surprising action frequencies are likely provoked by uncommon tastes or contrary private information. Taste projectors miscalculate these odds. In settings where rational agents correctly learn their optimal choice, projection stops some or all from ever learning the right choice. Long-run beliefs and behavior are determined by a player's taste, the degree of all players' biases, and the nature of uncertainty. First, when each thinks her taste is most common, society comes to believe a single action is best for all, irrespective of whether this is true. Second, when the bias is weaker, social beliefs and behavior perpetually cycle---history never provides a clear message about the optimal choice. Third, when quality is highly uncertain, popularity due to taste is systematically over-attributed to quality. Finally, this form of biased learning can exacerbate and perpetuate a false-consensus effect: if people neglect differences in perceptions when learning about the distribution of tastes from others' choices, then a small initial bias eventually leads all types to think their own taste is most common. For contrast, I also characterize rational learning among Bayesian agents with taste-dependent beliefs over the distribution of preferences.

Across a range of social-learning settings, Chapter 2 follows Eyster and Rabin (2010) in studying the implications of agents who neglect the redundancy in information when learning from others. Players naively think each predecessor's action reflects solely that person's private information. We explore new implications that arise in environments richer than ER's canonical binary-state setting. Whereas in both classical learning models and ER society will with positive probability come to believe the true state, we characterize a set of states that agents will always come to disbelieve even when true. Typically when the truth lies in this set, society will "unlearn": an early generation learns the truth, but society's beliefs move away and converge to wrong beliefs. Society only remains confident in those hypotheses such that the behavior observed when people are fully confident in the hypothesis most closely resembles the behavior we'd see by privately informed agents if that hypothesis were true. We provide specific implications of these principles. First, in cases where options such as restaurants or stocks have independent quality, people form polarized beliefs---they come to believe that the best option is the best it could be and that all lesser options are the worst they could be. Second, in an investment setting, polarized perceptions lead investors to allocate all their wealth to a single prospect, generating a welfare loss through under-diversification. Third, agents generally overestimate the extent of private information in the economy.

Chapter 3 explores how taste projection affects bidding in auctions. We consider auctions for a good with both private- and common-value elements. We model projection by assuming bidders with higher private values perceive a distribution of valuations that first-order stochastically dominates the perception of those with lesser private values. Those with low private values perceive a distribution shifted to the left whereas those with high private values perceive one shifted to the right. We draw out the implications of this assumption in first- and second-price sealed-bid auctions and English auctions. When the good has only private value, projection leads players to misperceive the extent of competition. This induces overbidding, on average, in first-price auctions, but has no effect in second-price or English auctions. If the good also has some common-value component, players draw inference about others' signals from their equilibrium bids. No matter the auction format, projection leads to distorted inference that reduces efficiency. The probability the player with the highest value receives the good is decreasing in the extent of projection.