Search

In this study, we introduce BEAST-Net, an interpretable model of human decision making under uncertainty, that fuses the foundational principles of BEAST, a behavioral model grounded in psychological theory, with the capabilities of machine learning techniques. Our strategy involves mathematically formalizing BEAST as a differentiable function and parameterizing it via a neural network. This approach facilitates the learning of model parameters from data and optimizes it through backpropagation. BEAST-Net scales to larger datasets and adapts to new data more efficiently, while preserving the psychological interpretability of the original model. Evaluations of BEAST-Net on the most extensive publicly accessible human choice task datasets demonstrate its superior performance over several baselines, including the original BEAST model. Importantly, BEAST-Net provides interpretable explanations for choice behavior, leading to the extraction of novel psychological insights from the data. This research demonstrates the potential of machine learning techniques to enhance the scalability and adaptability of models rooted in psychological theory, without compromising their interpretability or insight generation capabilities.

Experts and employees in many domains make multiple similar but independent decisions in sequence. Often, the serial position of the case in the sequence influences the decision. Explanations for these serial position effects focus on the role of decision-makers' fatigue, but these effects emerge also when fatigue is unlikely. Here, we suggest that serial position effects can emerge due to decision-makers' motivation to be or appear consistent. For example, to avoid having inconsistencies revealed, decisions may become more favorable toward the side that is more likely to put a decision under scrutiny. As a context, we focus on the legal domain in which many high-stakes decisions are made in sequence and in which there are clear institutional processes of decision scrutiny. We analyze two field data sets: 386,109 U.S. immigration judges' decisions on asylum requests and 20,796 jury decisions in 18th century London criminal court. We distinguish between five mechanisms that can drive serial position effects and examine their predictions in these settings. We find that consistent with motivation-based explanations of serial position effects, but inconsistent with fatigue-based explanations, decisions become more lenient as a function of serial position, and the effect persists over breaks. We further find, as is predicted by motivational accounts, that the leniency effect is stronger among more experienced decision-makers. By elucidating the different drivers of serial position effects, our investigation clarifies why they are common, when they are expected, and how to reduce them. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Research on small repeated decisions from experience suggests that people often behave as if they underweight rare events and choose the options that are frequently better. In a pandemic, this tendency implies complacency and reckless behavior. Furthermore, behavioral contagion exacerbates this problem. In two pre-registered experiments (Ntotal= 312), we validate these predictions and highlight a potential solution. Groups of participants played a repeated game in one of two versions. In the basic version, people clearly preferred the dangerous reckless behavior that was better most of the time over the safer responsible behavior. In the augmented version, we gave participants an additional alternative abstracting the use of an application that frequently saves time but can sometimes have high costs. This alternative was stochastically dominated by the responsible choice option and was thus normatively irrelevant to the decision participants made. Nevertheless, most participants chose the new (“irrelevant”) alternative, providing the first clear demonstration of underweighting of rare events in fully described social games. We discuss public policies that can make the responsible use of health applications better most of the time, thus helping them get traction despite being voluntary. In one field demonstration of this idea amid the COVID-19 pandemic, usage rates of a contact tracing application among nursing home employees more than tripled when using the app also started saving them a little time each day, and the high usage rates sustained over at least four weeks.

External enforcement policies aimed to reduce violations differ on two key components: the probability of inspection and the severity of the punishment. Different lines of research offer different insights regarding the relative importance of each component. In four studies, students and Prolific crowdsourcing participants (N(total) = 816) repeatedly faced temptations to commit violations under two enforcement policies. Controlling for expected value, we found that a policy combining a high probability of inspection with a low severity of fines (HILS) was more effective than an economically equivalent policy that combined a low probability of inspection with a high severity of fines (LIHS). The advantage of prioritizing inspection frequency over punishment severity (HILS over LIHS) was greater for participants who, in the absence of enforcement, started out with a higher violation rate. Consistent with studies of decisions from experience, frequent enforcement with small fines was more effective than rare severe fines even when we announced the severity of the fine in advance to boost deterrence. In addition, in line with the phenomenon of underweighting of rare events, the effect was stronger when the probability of inspection was rarer (as in most real-life inspection probabilities) and was eliminated under moderate inspection probabilities. We thus recommend that policymakers looking to effectively reduce recurring violations among noncriminal populations should consider increasing inspection rates rather than punishment severity.

Choice prediction competitions suggest that popular models of choice, including prospect theory, have low predictive accuracy. Peterson et al. show the key problem lies in assuming each alternative is evaluated in isolation, independently of the context. This observation demonstrates how a focus on predictions can promote understanding of cognitive processes.

External enforcement policies aimed to reduce violations differ on two key components: the probability of inspection and the severity of punishments. Different lines of research offer competing predictions regarding the relative importance of each component. In three incentive compatible studies, students and Prolific crowdsourcing participants (Ntotal=430) repeatedly faced temptations to commit violations under two enforcement policies. Controlling for expected value, the results indicated that a policy combining High probability of Inspection with Low Severity of fine (HILS) was more effective than a policy combining Low probability of Inspection with High Severity of fine (LIHS). Consistent with the prediction of Decisions from Experience research, this finding held even when the severity of the fine was stated in advance to boost deterrence. In addition, the advantage of HILS over LIHS was greater as participants’ baseline rate of violation (without enforcement) was higher, implying that HILS is more effective among frequent offenders.

This paper argues that two of the common methods used in behavioral and social sciences to reduce the chances that models overfit the available data, namely heavy reliance on benchmark models and rigorous parameter estimation techniques, can slow the advancement of these sciences. An examination of classical decision research highlights how applying these methods shaped the field but have also led to limited success. As an alternative, the paper proposes a prediction-oriented approach to the development of behavioral models. Evaluating and comparing models based on their predictive power inherently guards against overfitting and also facilitates accumulation of knowledge. The paper reviews research employing the prediction-oriented approach in behavioral decision research and demonstrates that, in contrast to a common misconception, the focus on predictions can also facilitate better understanding of the underlying processes.

The COVID-19 pandemic poses a major challenge to policy makers on how to encourage compliance to social distancing and personal protection rules. This paper compares the effectiveness of two policies that aim to increase the frequency of responsible health behavior using smartphone-tracking applications. The first involves enhanced alert capabilities, which remove social externalities and protect the users from others’ reckless behavior. The second adds a rule enforcement mechanism that reduces the users’ benefit from reckless behavior. Both strategies should be effective if agents are expected-value maximizers, risk averse, and behave in accordance with cumulative prospect theory (Tversky and Kahneman, 1992) or in accordance with the Cognitive Hierarchy model (Camerer et al., 2004). A multi-player trust-game experiment was designed to compare the effectiveness of the two policies. The results reveal a substantial advantage to the enforcement application, even one with occasional misses. The enhanced-alert strategy was completely ineffective. The findings align with the small samples hypothesis, suggesting that decision makers tend to select the options that lead to the best payoff in a small sample of similar past experiences. In the current context, the tendency to rely on a small sample appears to be more consequential than other deviations from rational choice.

©American Psychological Association, 2020. This paper is not the copy of record and may not exactly replicate the authoritative document published in the APA journal. Please do not copy or cite without author's permission. The final article is available, upon publication, at: https://doi.apa.org/doi/10.1037/bul0000220 To what extent are research results influenced by subjective decisions that scientists make as they design studies? Fifteen research teams independently designed studies to answer five original research questions related to moral judgments, negotiations, and implicit cognition. Participants from 2 separate large samples (total N 15,000) were then randomly assigned to complete 1 version of each study. Effect sizes varied dramatically across different sets of materials designed to test the same hypothesis: Materials from different teams rendered statistically significant effects in opposite directions for 4 of 5 hypotheses, with the narrowest range in estimates being d = 0.37 to 0.26. Meta-analysis and a Bayesian perspective on the results revealed overall support for 2 hypotheses and a lack of support for 3 hypotheses. Overall, practically none of the variability in effect sizes was attributable to the skill of the research team in designing materials, whereas considerable variability was attributable to the hypothesis being tested. In a forecasting survey, predictions of other scientists were significantly correlated with study results, both across and within hypotheses. Crowdsourced testing of research hypotheses helps reveal the true consistency of empirical support for a scientific claim.

Experimental and field studies of repeated decisions document high sensitivity to the probability of regret; experience leads people to underweight rare risks and select the actions that provide the best outcomes in most cases. We clarify the implications of this pattern to policies designed to combat the spread of coronavirus. Our analysis questions the value of policies that encourage adherence to physical distancing guidelines without continuous enforcement. Yet, it highlights the potential of policies that encourage the use of voluntary health code applications, which estimate the risk a person is infectious. To be effective, these applications should gently enforce physical distancing, and the policies should ensure that the common experience from using the applications is better than the common experience from not using them.

Contains fulltext : 228242.pdf (Publisher’s version ) (Open Access) Drawing on the concept of a gale of creative destruction in a capitalistic economy, we argue that initiatives to assess the robustness of findings in the organizational literature should aim to simultaneously test competing ideas operating in the same theoretical space. In other words, replication efforts should seek not just to support or question the original findings, but also to replace them with revised, stronger theories with greater explanatory power. Achieving this will typically require adding new measures, conditions, and subject populations to research designs, in order to carry out conceptual tests of multiple theories in addition to directly replicating the original findings. To illustrate the value of the creative destruction approach for theory pruning in organizational scholarship, we describe recent replication initiatives re-examining culture and work morality, working parents’ reasoning about day care options, and gender discrimination in hiring decisions. Significance statement It is becoming increasingly clear that many, if not most, published research findings across scientific fields are not readily replicable when the same method is repeated. Although extremely valuable, failed replications risk leaving a theoretical void - reducing confidence the original theoretical prediction is true, but not replacing it with positive evidence in favor of an alternative theory. We introduce the creative destruction approach to replication, which combines theory pruning methods from the field of management with emerging best practices from the open science movement, with the aim of making replications as generative as possible. In effect, we advocate for a Replication 2.0 movement in which the goal shifts from checking on the reliability of past findings to actively engaging in competitive theory testing and theory building. Scientific transparency statement The materials, code, and data for this article are posted publicly on the Open Science Framework, with links provided in the article. 19 p.

To help predict choice behavior, behavioral economics research tries to identify robust deviations from rational choice, and explain them by assuming distinct biases. Our analysis questions the value of this convention and proposes an alternative. First, we demonstrate that six known deviations from rationality that emerge when people gain experience are not robust; they can be reversed by small changes in the incentive structure. For example, published research shows that experience triggers underweighting of rare outcomes, and our study shows that experience can also trigger oversensitivity to rare outcomes. Then, we show that it is not necessary to assume situation specific biases. Simple models, assuming reliance on small samples of memories, capture all 12 contradicting deviations, and provide useful ex-ante predictions. The practical implications of our analysis include the clarification of the conditions that trigger rational choice. One example involves individual choice tasks in which the payoff maximizing option also maximizes the probability of success. The paper’s methodological contributions include the elucidation of conditions that elicit aggregation gain (Grunfeld & Griliches (1960)). When people use a wide set of rules that imply reliance on small samples of similar past experiences, it is easy to develop a usefully specified model of the aggregate behavior, even when it is difficult to correctly specify the factors that impact individual decisions. We believe that part of the popularity of situation-specific explanations, to the deviations that we capture with a single model, reflects the incorrect belief that aggregation is always bad.

Analyses of human learning reveal a discrepancy between the long- and the short-term effects of outcomes on subsequent choice. The long-term effect is simple: favorable outcomes increase the choice rate of an alternative whereas unfavorable outcomes decrease it. The short-term effects are more complex. Favorable outcomes can decrease the choice rate of the best option. This pattern violates the positive recency assumption that underlies the popular models of learning. The current research tries to clarify the implications of these results. Analysis of wide sets of learning experiments shows that rare positive outcomes have a wavy recency effect. The probability of risky choice after a successful outcome from risk-taking at trial t is initially (at t + 1) relatively high, falls to a minimum at t + 2, then increases for about 15 trials, and then decreases again. Rare negative outcomes trigger a wavy reaction when the feedback is complete, but not under partial feedback. The difference between the effects of rare positive and rare negative outcomes and between full and partial feedback settings can be described as a reflection of an interaction of an effort to discover patterns with two other features of human learning: surprise-triggers-change and the hot stove effect. A similarity-based descriptive model is shown to capture well all these interacting phenomena. In addition, the model outperforms the leading models in capturing the outcomes of data used in the 2010 Technion Prediction Tournament.

Experts in many domains, from education and finances to sports and entertainment, make multiple similar but independent decisions in sequence. Does the mere serial position of a case in a sequence influence the decision? We focus on the legal domain in which many high-stakes decisions are made in a sequence and justice considerations imply irrelevant factors like serial position of a case must not impact decision making. We identify four mechanisms that can drive serial position effects and note each can have different predictions, hence existence and direction of such effects are empirical questions. Indeed, most empirical evidence from non-legal settings suggests decisions become more favorable with serial position, but a previous study of sequences of parole hearings finds the opposite effect. We analyze two field datasets, 386,109 US immigration judges’ decisions on asylum requests and 20,796 jury decisions in 18th century London criminal court, and find decisions become more favorable (lenient) the later a decision is made in a sequence of decisions. To complement the analysis, we run three controlled experiments with laypeople and find similar results. We conclude that while serial position effects may be context-specific, from the point of view of the individual involved, it is often best to be last.

Searching for and acting upon perceived patterns of regularity is a fundamental learning process critical for adapting to changes in the environment. Yet in more artificial, static settings, in which patterns do not exist, this mechanism could interfere with choice maximization and manifest as unexplained choice variability in later trials. Recently however, Ashby et al. (2017) found that choice variability in later trials of a repeated choice setting is correlated with levels of diversification in policy tasks, in which patterns can never be exploited. They concluded that in repeated choice tasks, choice-variability in later trials is unlikely the result of following perceived patterns. Here, we demonstrate that correlations between choice variability and policy diversification can actually be the result of pattern seeking, rather than serving as evidence against it. We review evidence for the robustness of pattern seeking mechanisms in repeated choices and explain how such mechanisms could in fact create the results observed by Ashby et al. To examine our interpretation for their results, we conducted a sequential dependencies analysis of their data and find evidence that many participants behaved as if they believed trials are inter-dependent, even though they were explicitly instructed that the environment is stationary. The results of a new experiment in which sequential patterns are directly manipulated support our interpretation: Experiencing patterns affected both choice variability in later trials and policy diversification. Finally, we argue that decisions from experience tasks are a valid tool to examine the generation of preferences via fundamental learning processes.

Experimental studies of choice behavior document distinct, and sometimes contradictory, deviations from maximization. For example, people tend to overweight rare events in 1-shot decisions under risk, and to exhibit the opposite bias when they rely on past experience. The common explanations of these results assume that the contradicting anomalies reflect situation-specific processes that involve the weighting of subjective values and the use of simple heuristics. The current article analyzes 14 choice anomalies that have been described by different models, including the Allais, St. Petersburg, and Ellsberg paradoxes, and the reflection effect. Next, it uses a choice prediction competition methodology to clarify the interaction between the different anomalies. It focuses on decisions under risk (known payoff distributions) and under ambiguity (unknown probabilities), with and without feedback concerning the outcomes of past choices. The results demonstrate that it is not necessary to assume situation-specific processes. The distinct anomalies can be captured by assuming high sensitivity to the expected return and 4 additional tendencies: pessimism, bias toward equal weighting, sensitivity to payoff sign, and an effort to minimize the probability of immediate regret. Importantly, feedback increases sensitivity to probability of regret. Simple abstractions of these assumptions, variants of the model Best Estimate and Sampling Tools (BEAST), allow surprisingly accurate ex ante predictions of behavior. Unlike the popular models, BEAST does not assume subjective weighting functions or cognitive shortcuts. Rather, it assumes the use of sampling tools and reliance on small samples, in addition to the estimation of the expected values. (PsycINFO Database Record

We introduce a synergetic approach incorporating psychological theories and data science in service of predicting human behavior. Our method harnesses psychological theories to extract rigorous features to a data science algorithm. We demonstrate that this approach can be extremely powerful in a fundamental human choice setting. In particular, a random forest algorithm that makes use of psychological features that we derive, dubbed psychological forest, leads to prediction that significantly outperforms best practices in a choice prediction competition. Our results also suggest that this integrative approach is vital for data science tools to perform reasonably well on the data. Finally, we discuss how social scientists can learn from using this approach and conclude that integrating social and data science practices is a highly fruitful path for future research of human behavior.

Many behavioral phenomena, including underweighting of rare events and probability matching, can be the product of a tendency to rely on small samples of experiences. Why would small samples be used, and which experiences are likely to be included in these samples? Previous studies suggest that a cognitively efficient reliance on the most recent experiences can be very effective. We explore a very different and more cognitively demanding process explaining the tendency to rely on small samples: exploitation of environmental regularities. The first part of our study shows that across wide classes of dynamic binary choice environments, focusing only on experiences that followed the same sequence of outcomes preceding the current task is more effective than focusing on the most recent experiences. The second part of our study examines the psychological significance of these sequence-based rules. It shows that these tractable rules reproduce well-known indications of sensitivity to sequences and predict a nontrivial wavy recency effect of rare events. Analysis of published data supports this wavy recency prediction, but suggests an even wavier effect than these sequence-based rules predict. This pattern, and the main behavioral phenomena documented in basic decisions from experience and probability learning tasks, can be captured with a similarity-based model assuming that people follow sequences of outcomes most of the time but sometimes respond to trends. We conclude with theoretical notes on similarity-based learning.