Drazen Prelec’s research while affiliated with Massachusetts Institute of Technology and other places

Prelec, Seung and McCoy (2017) proposed a crowd wisdom model where ideal Bayesian observers receive discrete i.i.d. signals si ∈ {s1,...,sn} conditional on an unknown ‘state,’ i.e., a possible world aj ∈ {a1,...,am}. Signals and worlds are presumed drawn from a distribution p(si,aj) known to observers, but unknown to the analyst. The paper asserted via an example but without formal proof that if ideal observers specify a belief matrix as conditional distributions p(aj|si), and a meta-knowledge matrix (beliefs about other observers’ signals) as p(si|sk) = ∑j p(si|aj)p(aj|sk), then the analyst can, in the large sample limit, derive p(si,aj) and identify the actual signal-generating world. Here we provide a proof based on computing the signal prior as the stationary distribution of the meta-knowledge matrix.

According to the previous literature, only a few papers found better accuracy than a chance to detect dishonesty, even when more information and verbal cues (VCs) improve precision in detecting dishonesty. A new classification of dishonesty profiles has recently been published, allowing us to study if this low success rate happens for all people or if some people have higher predictive ability. This paper aims to examine if (dis)honest people can detect better/worse (un)ethical behavior of others. With this in mind, we designed one experiment using videos from one of the most popular TV shows in the UK where contestants make a (dis)honesty decision upon gaining or sharing a certain amount of money. Our participants from an online MTurk sample ( N = 1,582) had to determine under different conditions whether the contestants would act in an (dis)honest way. Three significant results emerged from these two experiments. First, accuracy in detecting (dis)honesty is not different than chance, but submaximizers (compared to maximizers) and radical dishonest people (compare to non-radicals) are better at detecting honesty, while there is no difference in detecting dishonesty. Second, more information and VCs improve precision in detecting dishonesty, but honesty is better detected using only non-verbal cues (NVCs). Finally, a preconceived honesty bias improves specificity (honesty detection accuracy) and worsens sensitivity (dishonesty detection accuracy).

Research Summary Managers regularly deal with dynamic tasks, where decisions impact immediate payoffs as well as long‐term capabilities. Research shows that people do poorly in dynamic tasks, but the underlying mechanisms are unclear. These may range from unsystematic problem‐solving to rational learning in complex environments. In a series of experiments, we tease apart alternative explanations, showing that poor performance is due to behavioral difficulties. Remarkably, we find that people do poorly even if provided with complete information about the payoff function, thus, eliminating any need for learning. They unsystematically search among possible solutions and end up with inefficient heuristics. The results show that differences in thinking through a dynamic problem may lead to substantial variation in performance, even if common sources of complexity and ambiguity are excluded. Managerial Summary Why do people, including managers, have difficulty managing systems where taking action today impacts future outcomes? Difficulty of learning in a complex environment has been proposed as the key challenge. Using experiments, we show that people find such tasks difficult even when all relevant information is provided to them and there is nothing to learn. Using trial and error most participants learn satisfactory, but inferior, heuristics. Those who systematically think through tradeoffs over time significantly outperform others even in a simple task, suggesting such thinking adds value in realistic managerial settings as well.

Experimental studies of dishonesty usually rely on population-level analyses, which compare the distribution of claimed rewards in an unsupervised, self-administered lottery (e.g., tossing a coin) with the expected lottery statistics (e.g., 50/50 chance of winning). Here, we provide a paradigm that measures dishonesty at the individual level and identifies new dishonesty profiles with specific theoretical interpretations. We found that among dishonest participants, (a) some did not bother implementing the lottery at all, (b) some implemented but lied about the lottery outcome, and (c) some violated instructions by repeating the lottery multiple times until obtaining an outcome they felt was acceptable. These results held both in the lab and with online participants. In Experiment 1 ( N = 178), the lottery was a coin toss, which permitted only a binary honest/dishonest response; Experiment 2 ( N = 172) employed a six-sided-die roll, which permitted gradations in dishonesty. We replicated some previous results and also provide a new, richer classification of dishonest behavior.

by SIAM. We consider the problem of eliciting truthful responses to a survey question when the respondents share a common prior that the survey planner is agnostic about. The planner would therefore like to have a “universal” mechanism, which would induce honest answers for all possible priors. If the planner also requires a locality condition that ensures that the mechanism payoffs are determined by the respondents’ posterior probabilities of the true state of nature, we prove that, under additional smoothness and sensitivity conditions, the payoff in the truth-telling equilibrium must be a logarithmic function of those posterior probabilities. Moreover, the respondents are necessarily ranked according to those probabilities. Finally, we discuss implementation issues.

Рассматривается проблема выявления правдивых ответов на вопросы некоторого исследования в случае, когда респонденты имеют общее априорное распределение, не интересующее составителя опроса. В такой постановке составителю опроса желательно иметь универсальное правило, стимулирующее респондентов отвечать правдиво при любом априорном распределении. Если дополнительно выполняются условие локальности (которое гарантирует, что платежные функции правил определяются апостериорными вероятностями фактического состояния системы) и условие достаточной гладкости, мы доказываем, что равновесная платежная функция в случае правдивых ответов респондентов является логарифмической функцией апостериорных вероятностей. Более того, респонденты должны быть упорядочены в соответствии с этими вероятностями. В заключение обсуждаются вопросы применения полученных результатов.

We consider the problem of eliciting truthful responses to a survey question when the respondents share a common prior that the survey planner is agnostic about. The planner would therefore like to have a "universal” mechanism, which would induce honest answers for all possible priors. If the planner also requires a locality condition that ensures that the mechanism payoffs are determined by the respondents' posterior probabilities of the true state of nature, we prove that, under additional smoothness and sensitivity conditions, the payoff in the truth-telling equilibrium must be a logarithmic function of those posterior probabilities. Moreover, the respondents are necessarily ranked according to those probabilities. Finally, we discuss implementation issues.