Experimental Economics

The empirical observation of the common ratio effect (CRE) is often interpreted as evidence of an underlying common ratio preference (CRP). However, prior research has demonstrated that, in the presence of noise, expected utility can generate a CRE in standard paired choice tasks. We expand on that research to describe how the existence or absence of a CRE may reveal little about whether there exists an underlying CRP. We then propose an alternative approach to test for the existence of a CRP using paired valuation tasks that is robust to heterogeneity and noise. We implement this approach in an online experiment with 900 participants, and we find no evidence of a systematic CRP. To reconcile our findings with existing evidence, we present the same participants with standard paired choice tasks, and we demonstrate how appropriately chosen experimental parameters can generate a CRE even in our population that has no systematic CRP.

There is ample evidence of a gender gap in confidence in male-typed domains, but we know little about how this gap affects others' views about men and women. On one hand, if women indicate their performance is worse than equally-performing men in self-evaluations—and if others, such as employers, expect and account for this gap when reviewing these self-evaluations—the impact of the gender gap in confidence may be mitigated. On the other hand, the impact of gender gap in confidence may be exacerbated if others fail to account for it. In a series of experiments involving more than 5,000 participants, we find robust evidence for the latter. Even when the gender gap in confidence is expected, individuals fail to account for it. As a result, the gender gap in confidence causes individuals to believe that there is a gender gap in performance even when there is not. These results persist across a variety of conditions, interventions, and types of individuals. Ultimately, we find that—while awareness of the gender gap in confidence is not sufficient to prevent its negative consequences—interventions that facilitate Bayesian updating may be.

In many types of decisions, individuals can influence the probabilities of good or bad outcomes by their actions, but there is still a role for chance in determining final outcomes. If punishment and rewards are conditioned on such random outcomes, this violates a property of optimal incentives. It has been posited since ancient times that humans do assign punishments and rewards based on factors outside of actors’ control, a tendency called “moral luck.” This paper provides new evidence on the prevalence and robustness of moral luck, and on a key open question of whether moral luck is a preference or a bias. The results are from controlled experiments that can cleanly identify moral luck, but also involve real, consequential moral choices that are a matter of life and death for a third party (a mouse). We find moral luck in punishment, and show that this is at least partly due to a bias. Our findings support a causal chain in which random outcomes lead to biased judgments and incentivized beliefs about the nature of the actor, even though they contain zero information, and this in turn causes punishments to vary with outcomes. We also show that the bias is strong enough to remain in the face of an intervention that encourages deliberation. The bias is prevalent, but not universal, it is unrelated to most demographics, and is present regardless of high or low cognitive ability or education. We also find evidence that actors exhibit internalized moral luck in how they evaluate themselves based on outcomes.

This paper theoretically and empirically analyzes how identity (one’s sense of self)affects consumption or investment decisions. We first present a model where identity distorts individuals’ beliefs about uncertain outcomes and imposes psychic costs on identity-incongruent actions. Then, using two large field experiments on soccer betting in the UK and Kenya, we experimentally varied material incentives for betting on matches where soccer fans are neutral or favor one of the teams playing. Finally, we combine the model with respondents’ portfolio allocations across different matches to disentangle biases in investment decisions due to over optimistic beliefs and those due to the psychic costs from identity-incongruent choices. We find that, on average, respondents misallocate 10% of their investment budget because of their over-optimistic beliefs and misallocate an additional 15% of their budget to avoid psychic costs. Our experimental findings also suggest that the impact of debiasing information campaigns that target biased beliefs is limited in contexts with identity concerns than otherwise.

We compare preferences for resolution of uncertainty when the uncertainty is resolved about a probability rather than a value. In various existing frameworks–e.g., Kreps and Porteus (1978)–, preferences over gradual versus one-shot resolution do not depend on whether values or probabilities define the main object of uncertainty. Yet, in our experiment, a large majority of subjects preferred to resolve uncertain values gradually but uncertain probabilities all at once–both with uncertainty defined over gains and losses. Interestingly, we find this discrepancy to be history-dependent: it fades away when subjects learned that the best outcomes were no longer possible. Our analyses suggest that in our experiments subjects put larger decision weights on uncertain values–in particular, high values–relative to uncertain probabilities. We discuss the deviations from the expected utility framework that are needed to explain our data.

Using data from a field experiment on exercise, we analyze the relationship between imperfect memory and people’s awareness of their limited self-control. We find that people overestimate past gym attendance, and that larger overestimation of past attendance is associated with (i) more overestimation of future attendance, (ii) a lower willingness to pay to motivate higher future gym attendance, and (iii) a smaller gap between goal and forecasted attendance. We organize these facts with a structural model of quasi-hyperbolic discounting and naivete, estimating that people with more biased memories are more naive about their time inconsistency, but not more time-inconsistent.

Individual forecasts of economic variables show widespread overreaction to recent news, but laboratory experiments on belief updating typically find underinference from new signals. We provide new experimental evidence to connect these two seemingly inconsistent phenomena. Building on a classic experimental paradigm, we study how people make inferences and revise forecasts in the same information environment. Participants underreact to signals when inferring about underlying states, but overreact to signals when revising forecasts about future outcomes. This gap in belief updating is largely driven by the use of different simplifying heuristics for the two tasks. Additional treatments suggest that the choice of heuristics is affected by the similarity between cues in the information environment and the belief-updating question: when forming a posterior belief, participants are more likely to rely on cues that appear similar to the variable elicited by the question.

Policymakers routinely make high-stakes decisions of which programs to fund. Assessing the value of a program is difficult and may be affected by bounded rationality. In an experiment with policymakers in the U.S. government, we find that respondents’ valuations of programs are inelastic with respect to the program’s impact. A complementary experiment among a representative sample of the general public reveals even more pronounced inelasticity in a population less familiar with making program funding decisions. We design and test two portable decision aids, one which presents two alternative programs side-by-side rather than in isolation and another which translates total program cost into an annual cost per person impacted. The decision aids increase elasticity by 0.20 on a base of 0.33 among policymakers and by 0.21 on a base of 0.21 among the general public. We provide evidence that cognitive noise—noisy assessments of complex inputs—is a mechanism that can help explain the observed inelasticity of program valuation with respect to impact.

We investigate how increasing the complexity of the message space can reduce misrepresentation in strategic communication. We develop a theoretical model that extends the standard cheap talk approach by i) allowing for communication about both a payoff relevant state and non-payoff relevant attributes which are correlated with the state, and ii) supposing that agents are boundedly rational in understanding the relationship between the states and their attributes. We show that although babbling is the only equilibrium for perfectly rational agents, boundedly rational agents induce an equilibrium that features informative messages. We adopt a novel experimental design to test our predictions and explore mechanisms that drive changes in the informativeness of communication. We find that increasing the number of messages that could be sent, while keeping the number of messages sent fixed, can significantly reduce misrepresentation, particularly when the receiver can anticipate what questions the messages are in response to. We find that the informativeness of communication is dependent both on the complexity level and CRT of agents. Our results shed light on why principals in mechanisms may not use a direct mechanisms, but instead use an indirect mechanism which elicits multi-dimensional information even when the additional information seems redundant, as in tax reporting or insurance claims.

We document a new source of discrimination in hiring that arises through sequential spillover effects. Employers in an incentivized resume rating experiment evaluate a sequence of hypothetical candidates with randomly assigned characteristics. Candidates are rated worse when following a white man than when following women or minorities. Exploring the mechanisms, we find that employers directly favor white men when resumes are high quality, but when evaluating low quality white men, they instead indirectly favor them through the spillover effect. Thus, our findings highlight the powerful effects of implicit bias, and how it may manifest even when direct favoritism is constrained.

In absence of policy interventions like affirmative action, employers’ biased beliefs about underrepresented groups may not correct on their own due to less hiring of and subsequent learning about those underrepresented. This paper explores whether temporary affirmative action can improve representation even after the policy is lifted in settings where employers hold biased beliefs about performance. I experimentally elicit employer hiring decisions and beliefs about potential employee performance in two between-subject experimental treatments: a control treatment without affirmative action and a temporary affirmative action treatment. While beliefs and hiring are biased against women in the control condition, I find in the temporary affirmative action treatment that representation improves even after affirmative action is lifted. Exposure substantially increases the likelihood that women are hired even beyond the policy instance. This increase is partially driven by employers gradually learning that their beliefs about women’s performance are biased downward where employers who are most likely to discriminate against women show the greatest reduction in gender bias in beliefs. The results shed new light on using a temporary affirmative action policy to fundamentally break a cycle of underrepresentation by correcting biased beliefs.

High school graduates in Germany who lack parents with college experience are40 percentage points less likely to attend college than those with college-educated par-ents, despite the fact that in Germany college is free. This study provides evidence that parental influence explains a significant portion of this socio-economic gap through at least two channels: one, parental pressure and two, the intergenerational transmission of beliefs and preferences. To understand parental influence, I conduct a field experiment with 1,195 students and 819 parents in Germany. Importantly, I experimentally make students’ stated college plans visible to parents. In the first finding, visibility to parents doubles the socio-economic gap in college plans to 27 percentage points. This is mainly driven by a large increase in college plans among students with college-educated parents. To disentangle mechanisms, I collect detailed survey data on students’ and parents’ subjective expectations for various career tracks and estimate a structural model of career choice under uncertainty. Model simulations indicate that 40% ofthe socio-economic gap in college plans is explained by parental pressure and 44% by students internalizing family-specific beliefs.

Can machine learning (ML) help people predict behavior in high-stakes prisoner’s dilemmas? We show participants videos from the TV show Golden Balls, in which contestants communicate before choosing whether to cooperate or defect. Participants show limited ability to predict contestants’ behavior; simple ML algorithms provide significantly more accurate predictions. We then flag contestants that ML predicts are highly likely to cooperate or defect. We find that flagging the extreme predictions helps participants improve their predictions. In addition, 54% of participants choose to fully delegate their predictions to the algorithm, improving their predictions as a result.

The application of reference-dependent models is often complicated by the modeler's uncertainty regarding the reference point (referent) that agents adopt. We develop a powerful and minimally parametric approach to testing whether decisions could be rationalized by a general reference-dependent model with a specific referent. Our approach builds from the observation that, when both payoffs and the true referent are randomly varied, a marginal increase in all payoffs will have an equivalent effect as a marginal decrease in the referent. The observation that this equivalence holds at all payoff/referent combinations, when applied to decisions over properly constructed gambles, allows us to generate our test through modifications to existing tools for rejecting single-index representations. We assess the performance of this test in a simulation study and find that it is highly diagnostic even in the comparatively small sample sizes that are common in experimental economics. We then utilize this approach in an online experiment in which  we randomly vary the salience of both goal-based and expectations-based referents. In this experiment, we confirm the common assumption that salient goals could serve as reference points. Illustrating the importance of salience, we reject that either reference point is adopted when it is not salient. Perhaps surprisingly, we reject the adoption of expectations as a reference point even when they are salient.

We study paternalistic preferences in two large-scale, incentivized experiments with participants sampled from the general population in the United States. Participants, acting as third-party spectators, decide whether to intervene to prevent another individual, the stakeholder, from making a mistake. We find causal evidence for the nature of the intervention being of great importance for the willingness to intervene: only about a third of the spectators intervene by restricting the stakeholder’s choice set, while a large majority intervene by providing information. In contrast, the stakeholder’s personal responsibility for the mistake does not have a substantial effect on the spectators’ willingness to intervene. We introduce a simple framework that allows us to classify most participants as either welfarists, libertarian paternalists, or authoritarians. Our results shed light on attitudes toward paternalistic policies and why libertarian paternalism has gained strong support in recent years.

We study in the laboratory the behavior of children and adolescents (ages 7 to 16) in two repeated coordination games, the stag hunt and battle of the sexes. Coordinating on the efficient and fair long run outcome (EFO) requires participants to share intentions and beliefs. This exercise is arguably complex in the battle of the sexes, as it requires taking turns between the two static Nash equilibria, hence coordinating the strategies. By contrast, in the stag hunt it only requires repeating the action that leads to the Pareto efficient outcome, hence coordinating the actions. We obtain four main findings. First, for both games, we show a significant and remarkably stable increase in the ability to coordinate on the EFO with age. Second, the majority of participants in all ages adhere to one of a small number of relatively simple strategies. Third, jointly profitable outcomes are more prevalent in the stag hunt than in the battle of the sexes. Last, behavior improves between the first and second supergame. This evidence suggests that we gradually learn how to share intentions and beliefs, an ability that we train rapidly and export to new interactions, but that is limited by game complexity.

Many decisions – such as what educational or career path to pursue – are dynamic in nature, with individuals receiving feedback at one point in time and making decisions later. Using a controlled experiment, with two sessions one week apart, we analyze the dynamic effects of feedback on beliefs about own performance and decision-making across two different domains (verbal skills and math). We find significant gender gaps in beliefs and choices before feedback: men are more optimistic about their performance and more willing to compete than women in both domains, but the gaps are significantly larger in math. Feedback significantly shifts individuals’ beliefs and choices. Despite this, we see substantial persistence of gender gaps over time. This is particularly true among the set of individuals who receive negative feedback. We find that, holding fixed performance and decisions before feedback, women update their beliefs and choices more negatively than men do after bad news. Our results highlight the challenges involved in overcoming gender gaps in dynamic settings.

We report experimental evidence that people have difficulty effectively engaging in statistical discrimination, leading to lower accuracy gains from discriminating than a rational model would predict. As a result, discrimination can be significantly reduced without lowering accuracy, simply by improving the way people use information. We show that this inefficiency stems from subjects putting excess weight on their subjective judgements while simultaneously applying crude contrast-driven group-level biases. A series of treatment interventions give us insight into the psychological drivers of these errors and guidance on policies likely to be effective at removing them.

Declining engagement over time is a central obstacle to achieving long-term goals, especially in education. A common route to addressing the fall-off among students through the course of a semester is to schedule assignments and tests throughout the term. In this field experiment, we study how the timing of a potentially rewarding but cognitively taxing assignment, taking notes and posting them, affects students’ academic behavior and performance. We find that assigning such tasks to low-performing students in the middle of the term, compared to early or late in the semester, improves their performance more, e. g., in terms of higher attendance, homework grades, and exam grades. We argue that, rather than “early intervention,” possibly with the goal of habit formation in studying, or “crunch time intervention,” engagement interventions are most effective if they target the time when students start to fall off because of accumulating frictions and complications in their semester schedule.

The influence of behavioral biases on aggregate outcomes like prices and allocations depends in part on self-selection: whether rational people opt more strongly into aggregate interactions than biased individuals. We conduct a series of betting market, auction and committee experiments using 15 classic cognitive bias tasks. We document that some cognitive errors are strongly reduced through self-selection, while others are not affected at all. A large part of this variation is explained by the quality of people’s meta-cognition. In some cognitive tasks, confidence and performance are strongly positively correlated, while for others this link is absent or even negative.