Felipe A. Araujo, PhD

We study the effect of on-the-job experience on base-rate neglect, which is a common bias in assessing conditional probabilities. We do so by carrying out experiments with medical professionals, who are routinely exposed to conditional-probability problems in the form of diagnostic tests, and non-medical professionals, who are not. As such, medical workers with more years of experience will have had more exposure to base-rate type problems than non-medical workers with similar years of experience. We estimate the effect of on-the-job experience by comparing the answers of more or less experienced professionals in both the medical and non-medical domains. Though the incidence of the bias is high for both groups and all levels of experience, we find that more experienced medical workers (a) have lower rates of perfect base-rate neglect (i.e., completely ignoring the base rates), (b) provide more accurate posterior estimates, and (c) adjust their estimates more in response to changes in the base rates. We observe no such difference for non-medical workers. We conduct a number of robustness checks and consider possible mechanism, such as education, job or survey attrition, selectivity into medical professions, and experience with false positives. Our results suggests that on-the-job experience mitigates, but does not eliminate, base-rate neglect.

We demonstrate pitfalls when extrapolating behavioral findings from static to dynamic settings. We focus on regret theory and the design of "regret lotteries", which have been advocated as more effective than standard incentives. Findings from one-shot settings have been used to promote regret as a tool to boost incentives in recurrent decisions across a variety of settings. Using theory and experiments, we replicate regret lotteries as the superior one-shot incentive; however, for repeated decisions we show the comparative static is entirely reversed. The results highlight the issues extrapolating from static to dynamic settings, particularly for behavioral policy design.

We study the effects of exposure to unawareness on risk preferences using a novel experimental task. The task has solutions that are difficult to find, but easy to verify and so exposes subjects to unawareness in a natural way. We find that increased exposure to unawareness alone does not affect risk taking. The role of context, however, is shown to be important. For treatments inducing higher unawareness, subjects are more risk averse when the risk elicitation task is framed in the same context as the unawareness-inducing task versus framed in a neutral way; we observe no such differences for the control treatment. We propose a novel decision theoretic model that guides the interpretation of the experimental findings. Our results could inform the decision and game-theoretic literatures, as most models of unawareness assume risk preferences are orthogonal to varying awareness.

We propose an extension of the classical regret theory model (Loomes & Sugden, 1982; henceforth LS) incorporating the notion of a reference point. As in LS, the model can account for a number of documented deviations from expected utility theory. In addition, we show that our model is consistent with a class of behaviors known as omission bias, for example, a reluctance to exchange lottery tickets, and generates predictions that are consistent with recent empirical evidence on the common-ratio effect with correlated outcomes. The model also provides a novel interpretation for risk aversion in small stakes, equiprobable gambles. The predictive power of the theory, as well as its relative shortcomings and advantages, is examined and compared to that of other extensions of regret theory and three alternative reference-dependent models.

We examine a common value dynamic matching environment where adverse selection accrues slowly over time. Theoretical best responses are therefore time varying, and the prior experimental literature suggests that sequential environments might lead to greater understanding of adverse selection in this dynamic setting. However, while a sophisticated minority in our experiment do condition on time and are close to a best response, the majority use a stationary response, even after extended experience. In an environment with persistent uncertainty, our results indicate that sequentiality is insufficient for the large majority of participants to recognize the effects of adverse selection.

Real-effort experiments are frequently used when examining a response to incentives. For a real-effort task to be well suited for such an exercise its measurable output must be sufficiently elastic over the incentives considered. The popular slider task in Gill and Prowse (Am Econ Rev 102(1):469–503, 2012) has been characterized as satisfying this requirement, and the task is increasingly used to investigate the response to incentives. However, a between-subject examination of the slider task's response to incentives has not been conducted. We provide such an examination with three different piece-rate incentives: half a cent, two cents, and eight cents per slider completed. We find only a small increase in performance: despite a 1500 % increase in the incentives, output only increases by 5 %. With such an inelastic response we caution that for typical experimental sample sizes and incentives the slider task is unlikely to demonstrate a meaningful and statistically significant performance response.