Risk taking

The Balloon Analogue Risk Task (BART) is widely-used to measure risk propensity in theoretical, clinical, and applied research. In the task, people choose either to pump a balloon to increase its value at the risk of the balloon bursting and losing all value, or to bank the current value of the balloon. Risk propensity is most commonly measured as the average number of pumps on trials for which the balloon does not burst. Burst trials are excluded because they necessarily underestimate the number of pumps people intended to make. However, their exclusion discards relevant information about people's risk propensity. A better measure of risk propensity uses the statistical method of censoring to incorporate all of the trials. We develop a new Bayesian method, based on censoring, for measuring both risk propensity and behavioral consistency in the BART. Through applications to previous data we demonstrate how the method can be extended to consider the correlation of risk propensity with external measures, and to compare differences in risk propensity between groups. We provide implementations of all of these methods in R, MATLAB, and the GUI-based statistical software JASP.

People are often faced with repeated risky decisions that involve uncertainty. In sequential risk-taking tasks, like the Balloon Analogue Risk Task (BART), the underlying decision process is poorly understood. Accurate depiction of the task requires modeling the mental representation of the object or the problem in the task environment and the cognitive processes operating on these representations to produce responses. Dual-process theory proposes that human cognition involves two main families of processes, often referred to as System 1 (fast and automatic) and System 2 (slow and conscious). We crossed models of the BART with different assumptions about the interaction between the two systems (serial vs. parallel), and built a pool of computational dual-process models with varying representations and process configurations. This model framework was designed to explain both choice and response time data in the BART. A model comparison study examined the statistical properties of the models (i.e., parameter recovery, model recovery, and predictive accuracy). The best-performing subset of models was then evaluated by fitting them to data collected in three experiments whose manipulations were intended to engage the two systems in different ways. Results showed that the model assuming simultaneously activated two systems, a drift rate influenced by experience, and an evaluation process based on frequency representation, outperformed the others. Findings shed light on how modeling multiple processes and representations can benefit our understanding of sequential risk-taking behavior.

When people make financial decisions, they need not only think about their current financial situation, but also about changes in future wealth. This work investigates people's beliefs about their future wealth and how these beliefs impact financial decisions. Using a joint experimental and computational cognitive modeling approach, we show that people's future beliefs serve as reference points when making investment decisions. These results are further supported by data from a large-scale cross-sectional survey (n = 4,606) showing that people's beliefs about the future value of their assets are related to investment decisions between risky (i.e., stock market index) and safe (i.e., bond earning a fixed amount per year) options. In both the experiments and survey, we hypothesize that outcomes that are nominally stated as sure gains can become coded as losses due to belief-based reference points. This pattern leads to an increase in riskier choices across positive outcomes for individuals with optimistic beliefs about their future wealth.

Reinforcement learning is hypothesized to play a strong role in the development and maintenance of impulsivity-related (or externalizing) disorders across the lifespan. For children with the hyperactive-impulsive presentation of Attention-Deficit/Hyperactivity Disorder (ADHD-HI), blunted sensitivity to reward is thought to produce subjective states of irritability and negative affect, which evoke excessive approach behaviors that function to down-regulate negative affect. Blunted sensitivity to punishment is thought to produce impaired extinction/reversal learning of previously reinforced action-outcome contingencies, in addition to a general lack of risk aversion. Combined, blunted reward and punishment sensitivity can potentiate aggressive and antisocial behaviors, setting a developmental course for Oppositional Defiant Disorder, Conduct Disorder, Substance Use Disorders, and Antisocial Personality Disorder. In the current study, we examine the relationships between (1) cognitive mechanisms underlying reinforcement learning on a passive avoidance paradigm, (2) neural responses to feedback collected with task-based fMRI, and (3) trait measures of externalizing and internalizing psychopathology, within a sample consisting of control adolescents (n=66), and those diagnosed with externalizing (n=134) and/or internalizing (n=74) disorders (comorbid n=67). We use a computational model of passive avoidance learning behavior, and a multidimensional item response theory model of self-report responses to develop a joint, generative model spanning all levels of observed data. Generative modeling allowed us to estimate person-specific parameters from behavioral data that correspond to key theoretical mechanisms (i.e. reward and punishment sensitivity), and then infer relationships between these and latent externalizing and internalizing traits while accounting for measurement error that would otherwise attenuate such individual difference correlations. We found a negative association between trait impulsivity and reward learning rate, punishment learning rate, and reward frequency sensitivity. We also identified interactions between externalizing and internalizing traits that can potentiate or protect against certain forms of impulsive and anxious behaviors and symptoms.

There are two theoretical approaches accounting for how people preferentially choose lotteries in the classical gamble task. Following the rationality postulate, a person chooses according to the utility associated with each lottery. For example, prospect theory (PT, Kahneman & Tversky, 1979) proposes that at a final decision stage, a decision maker calculates utilities for each gamble by combining the gamble’s attributes . Alternatively, following the bounded rationality postulate, a rule-based heuristic may be used to evaluate each gamble’s attributes one-at-a-time (e.g. Priority Heuristic, Brandstätter, et al, 2006). This approach implies serial information processing and that a decision is made when the critical difference between compared attributes is evaluated. To validate the core assumptions of these approaches we employed a parametric version of the Systems Factorial Technology (SFT), which can be used to diagnose whether processes are organized in serial or parallel mental architectures, whether a stopping rule is exhaustive or self-terminating, and whether the processes are interdependent. Using the joint analysis of preferential choice response time distributions, we compared stochastic versions of several decision-making models: serial, parallel, parallel interactive, mixture, and the full-parameter models. The results indicated differences in how participants processed gambles’ attributes. Some participants adopted either serial or parallel processing, while some relied on their trial-to-trial mixture. Some of the model fits, matched to those of the statistical full model, speaking strongly in favor of both the Take-the-Best and Weighted-Additive models. In general, these findings invite reconsiderations for heuristic-based approaches to decision making and on boundedly rational decision models.