People, across a wide range of personal and professional domains, need to accurately detect whether the state of the world has changed. Previous research has documented a systematic pattern of over- and under-reaction to signals of change due to system neglect, the tendency to overweight the signals and underweight the system producing the signals. We investigate whether experience, and hence the potential to learn, improves people’s ability to detect change. Participants in our study made probabilistic judgments across 20 trials, each consisting of 10 periods, all in a single system that crossed three levels of diagnosticity (a measure of the informativeness of the signal) with four levels of transition probability (a measure of the stability of the environment). We found that the system-neglect pattern was only modestly attenuated by experience. Although average performance did not increase with experience overall, the degree of learning varied substantially across the 12 systems we investigated, with participants showing significant improvement in some high diagnosticity conditions and none in others. We examine this variation in learning through the lens of a simple linear adjustment heuristic, which we term the “δ-ϵ” model. We show that some systems produce consistent feedback in the sense that the best δ and ϵ responses for one trial also do well on other trials. We show that learning is related to the consistency of feedback, as well as a participant’s “scope for learning” how close their initial judgments are to optimal behavior.

The underpricing of initial public offerings (IPO) is a well-documented fact of empirical equity market research. Theories explain this underpricing with market imperfections. We study three empirically relevant IPO mechanisms under almost perfect market conditions in the laboratory: a stylized book building approach, a closed book auction, and an open book auction. We report underpricing in each of these IPO mechanisms. Uncertainty about the aftermarket behavior may partly explain IPO excess returns but underpricing persists even in the repeated setting where uncertainty is negligible and despite the equilibrium adjustment dynamics, that we observe in the data. The data reveal a market-wide impact of investors’ reluctance to sell in the aftermarket at a price below the offering price. We conclude that a behavioural bias similar to the disposition effect fosters IPO underpricing in our setting.

In many market environments, for example in investment banking, salesforce management and others, workers and supervisors work closely as a team. Workers are paid a fixed salary and supervisors determine any raises, which are typically dependent on how well the organization does. In such scenarios, a supervisor who constantly offers suggestions can create a problem—typically a worker cannot ignore his supervisor's advice, yet if such advice is wrong and is followed, it will only decrease firm profits. We conduct a laboratory experiment to address a question critical for such settings—does the relationship between advisor and worker interfere with the learning abilities of the worker? The answer is a resounding no. In fact, subjects who have a supervisor advising them and whose advice is costly to ignore actually learn better than those with an advisor whose advice can be ignored. An even more striking result is that advisees as well as advisors in both these conditions learn better than subjects with no advisors. Our result can be attributed to the presence of advice and has direct relevance to learning in many environments.

Learning models predict that the relative speed at which players in a game adjust their behavior has a critical influence on long term behavior. In an ultimatum game, the prediction is that proposers learn not to make small offers faster than responders learn not to reject them. We experimentally test whether relative speed of learning has the predicted effect, by manipulating the amount of experience accumulated by proposers and responders. The experiment allows the predicted learning by responders to be observed, for the first time.

This paper reports an experiment designed to detect the influence of strategic uncertainty on behavior in order statistic coordination games, which arise when a player's best response is an order statistic of the cohort's action combination. Unlike previous experiments using order statistic coordination games, the new experiment holds the payoff function constant and only changes cohort size and order statistic.

The Individual Evolutionary Learning (IEL) model explains human subjects’ behavior in a wide range of repeated games which have unique Nash equilibria. Using a variation of ‘better response’ strategies, IEL agents quickly learn to play Nash equilibrium strategies and their dynamic behavior is like that of humans subjects. In this paper we study whether IEL can also explain behavior in games with gains from coordination. We focus on the simplest such game: the 2 person repeated Battle of Sexes game. In laboratory experiments, two patterns of behavior often emerge: players either converge rapidly to one of the stage game Nash equilibria and stay there or learn to coordinate their actions and alternate between the two Nash equilibria every other round. We show that IEL explains this behavior if the human subjects are truly in the dark and do not know or believe they know their opponent’s payoffs. To explain the behavior when agents are not in the dark, we need to modify the basic IEL model and allow some agents to begin with a good idea about how to play. We show that if the proportion of inspired agents with good ideas is chosen judiciously, the behavior of IEL agents looks remarkably similar to that of human subjects in laboratory experiments.

In many search environments, searchers are learning about the distribution of offers in the market. I conduct an experiment exploring a broad class of search problems with learning about the distribution of payoffs. My results support the prediction that learning results in declining reservation values, providing evidence that learning may be an explanation for recall. Theory predicts a “one step” reservation value strategy, but many subjects instead choose to set a high reservation value in order to learn about the distribution before adjusting based on their observations. Under-searching in search experiments may stem from a reinforcement heuristic and lack of negative feedback after using sub-optimal strategies.

We present an experiment designed to separate the two commonplace explanations for behavior in ultimatum games—subjects’ concern for fairness versus the failure of subgame perfection as an equilibrium refinement. We employ a tournament structure of the bargaining interaction to eliminate the potential for fairness to influence behavior. Comparing the results of the tournament game with two control treatments affords us a clean test of subgame perfection as well as a measure fairness-induced play. We find after 10 iterations of play that about half of all non-subgame-perfect demands are due to fairness, and the rest to imperfect learning. However, as suggested by models of learning, we also confirm that the ultimatum game presents an especially difficult environment for learning subgame perfection.

We report an experiment on a decision task by Samuelson and Bazerman (1985). Subjects submit a bid for an item with an unknown value. A winner's curse phenomenon arises when subjects bid too high and make losses. Learning direction theory can account for this. However, other influences on behaviour can also be identified. We introduce impulse balance theory to make quantitative predictions on the basis of learning direction theory. We also look at monotonic ladder processes. It is shown that for this kind of Markov chains the impulse balance point is connected to the mode of the stationary distribution.

This paper studies how subjects in a three-person sequential step-level public good game learn to punish free riders more over time. Our current work makes several additions to the literature on other regarding behavior. First, our experiment provides evidence that subjects care about the actions that lead to an outcome as well as the outcome itself, replicating the results of A. Falk, E. Fehr and U. Fischbacher (Economic Inquiry, in press), J. Brandts and C. Sola (Games and Economic Behavior, 36(2), 138-157, 2001.) and J.H. Kagel and K. Wolfe (Working paper, Ohio State University, 1999). Second, our experiment provides one of the first tests of the newer theories of reciprocity by A. Falk and U. Fischbacher (Working paper, University of Zurich, 2000) and G. Charness and M. Rabin (Quarterly Journal of Economics, in press) that take a psychological games approach. We find that these theories fail to explain the experimental data. Finally, we examine the mechanism by which subjects learn to punish free-riding more ofter over time.

We augment a standard bilateral gift exchange game so employees can send messages at the same time as choosing an effort level. Employee effort (controlling for wages) is unaffected by allowing messages, but wages dramatically increase. Messages affect wages because employees give managers advice to set higher wages, usually explaining that this will result in higher effort. This advice prompts managers to try higher wages, helping them learn that raising wages increases their payoffs. In a follow-up experiment, we directly provide managers with additional information about the relationship between wages and effort. This too causes wages to increase, but to a lesser extent than allowing messages. Our results highlight the critical role of learning in generating gains from positive gift exchange.

Bargaining and dilemma games have developed in experimental economics as fairly separate literatures. More than a few analysts are now persuaded that the patterns of behavior in these games are closely related, and considerable effort is being put into a search for models that bridge the gap between the two types of games. I focus on a handful of models that, when taken together, outline the conceptual issues, and provide a sense of the progress that has already been made.

The lack of a behavioral isomorphism between theoretically equivalent auction institutions is a robust finding in experimental economics. Using a near-continuous time environment and graphically adjustable bid functions, we are able to provide subjects with extensive feedback in multiple auction formats. We find that (1) First Price and Dutch Clock auctions are behaviorally isomorphic and (2) Second Price and English Clock auctions are behaviorally isomorphic. We further replicate the established result (1) that prices in Dutch Clock auctions exceed those of English Clock auctions and (2) that prices in First Price auctions exceed those of Second Price auctions. The latter pattern is often attributed to risk aversion which changes the equilibrium bidding strategy for First Price and Dutch Clock auctions. Because we observe each participant’s bid function directly, we find evidence suggesting a different explanation, namely that bidders are best responding to the distribution of observed prices.

We use a limited information environment to assess the role of confusion in the repeated voluntary contributions game. A comparison with play in a standard version of the game suggests, that the common claim that decision errors due to confused subjects biases estimates of cooperation upwards, is not necessarily correct. Furthermore, we find that simple learning cannot generate the kind of contribution dynamics commonly attributed to the existence of conditional cooperators. We conclude that cooperative behavior and its decay observed in public goods games is not a pure artefact of confusion and learning.

We analyze a class of coordination games in which the Kth player to submit an entry wins a contest. These games have an infinite number of symmetric equilibria and the set of equilibria does not change with K. We run experiments with 15 participants and with K = 3, 7, and 11. Our experiments show that the value of K affects initial submissions and convergence to equilibrium. When K is small relative to the number of participants, our experiments show that repeated play converges to or near zero. When K is large, an equilibrium is often not reached as a result of repeated play. We seek explanations to these patterns in hierarchical thinking and direction learning.

Is the assumption that people automatically know their own preferences innocuous? We present an experiment studying the limits of preference discovery. If tastes must be learned through experience, preferences for some goods may never be learned because it is costly to try new things, and thus non-learned preferences may cause welfare loss. We conduct an online experiment in which finite-lived participants have an induced utility function over fictitious goods about whose marginal utilities they have initial guesses. Subjects learn most, but not all, of their preferences eventually. Choice reversals occur, but primarily in early rounds. Subjects slow their sampling of new goods over time, supporting our conjecture that incomplete learning can persist. Incomplete learning is more common for goods that are rare, have low initial value guesses, or appear in choice sets alongside goods that appear attractive. It is also more common for people with lower incomes or shorter lifetimes. More noise in initial value guesses has opposite effects for low-value and high-value goods because it affects the perceived likelihood that the good is worth trying. Over time, subjects develop a pessimistic bias in beliefs about goods’ values, since optimistic errors are more likely to be corrected. Overall, our results show that if people need to learn their preferences through consumption experience, that learning process will cause choice reversals, and even when a person has completed sampling the goods she is willing to try, she may continue to lose welfare because of suboptimal choices that arise from non-learned preferences.

We design a novel experiment to study how subjects update their beliefs about the beliefs of others. Three players receive sequential signals about an unknown state of the world. Player 1 reports her beliefs about the state; Player 2 simultaneously reports her beliefs about the beliefs of Player 1; Player 3 simultaneously reports her beliefs about the beliefs of Player 2. We say that beliefs exhibit higher-order learning if the beliefs of Player k about the beliefs of Player $k-1$ become more accurate as more signals are observed. We find that some of the predicted dynamics of higher-order beliefs are reflected in the data; in particular, higher-order beliefs are updated more slowly with private than public information. However, higher-order learning fails even after a large number of signals is observed. We argue that this result is driven by base-rate neglect, heterogeneity in updating processes, and subjects’ failure to correctly take learning rules of others into account.

This paper estimates depth of reasoning in an Iterative Best Response model using data from Weber (2003) ten-period repeated guessing game with no feedback. Different mixture models are estimated and the type (Level-0, Level-1, etc) of each player is determined in every round using the Expectation Maximization algorithm. The matrices showing the number of individuals transitioning among levels is computed in each case. It is found that most players either remain in the same level or advance to the next two levels they were in the previous period. The lowest levels (Level-0 and Level-1) have a higher probability of transitioning to a higher level than Level-2 or Level-3. Thus, we can conclude that subjects, through repetition of the task, quickly become more sophisticated strategic thinkers as defined by higher levels. However, in some specifications the highest levels have a relatively large probability of switching to a lower level in the next period. In general, depth of reasoning increases monotonically in small steps as individuals are subjected to the same task repeatedly.