Loss aversion is one of the most prominent concepts in the study of decision-making under risk. Traditionally, previous research has investigated loss aversion in contexts where participants are aware of the payoff distributions of risky options between which they have to choose (i.e., decisions from description). But to what extent does loss aversion also emerge in situations in which the payoff distributions are initially unknown, and only learned from experiential sampling—that is, in decisions from experience? To investigate this question, we first conducted a meta-analysis, synthesizing and re-analyzing all existing datasets (total n > 430 per condition) that allow for a direct comparison of decisions from description and decisions from experience based on mixed gambles (i.e., where the options can lead to either a gain or a loss). Analyzing choices with cumulative prospect theory through a Bayesian multi-level modeling approach revealed substantial differences between learning modes, demonstrating that loss aversion is more pronounced in decisions made from experience compared to decisions made from description. Second, and building on these findings, we ran a pre-registered large-scale study to address potential caveats of the meta-analysis (e.g., few mixed choice problems, small samples). Study materials and raw data from the meta-analysis and own study will be shared openly during the publication process. Our results corroborate the novel and robust description-experience gap. In further analyses, we illuminate potential contributors to the gap in loss aversion between description and experiential paradigms.

During decision-making, people with schizophrenia (PwS) often exhibit biases associated with rigid beliefs. Here we explored these biases in the context of morality evaluations. We tested hypotheses that PwS would seek less information, hold stronger initial impressions, or less flexibly update beliefs when inferring a potentially “bad” person’s morality. PwS (n=45) and controls (n=45) completed two tasks that involved periodically rating their beliefs. The first involved receiving preliminary information about accused criminals and choosing how to explore optional evidence. The second entailed predicting choices of people inflicting pain on others. For the latter, a Hierarchical Gaussian Filtering model yielded belief volatility (ω) and decision noise (β) parameters. On task one, initial ratings of guilt from PwS did not differ from controls, but did indicate greater certainty (F(1,88)=8.33, p=.005, η2=.09). PwS also explored less evidence (F(1,88)=12.91, p<.001, η2=.13). On task two, PwS had comparable morality ratings, but their initial ratings were more extreme (F(1,88)=13.53, p<.001, η2=.13). PwS also had lower decision noise (inverse temperature β; F(1,88)=7.26, p=.008, η2=.08), despite no differences in belief volatility (p=0.52). PwS felt more certain about initial impressions of a morally questionable person’s “badness” and explored less information. Modeling suggests normative flexibility in belief updating, but less consistent use of beliefs to predict decisions of others (i.e., more random predictions) in PwS.

I present two projects on the link between decisions about exploration vs exploitation and symptoms of depression and anxiety. The first project focusses on decision making when risks are involved, using a gamified multi-armed bandit task and examining psychiatric traits as well as task-related affect. The second project focusses on the reliability and validity of common computational modelling approaches to exploratory decision making. I show flaws in the current modelling approaches and make suggestions for improving the reliability and validity of model parameters.

Increases in risky behaviors such as substance abuse are common in schizophrenia (SZ) and lead to worsening symptoms and outcomes, but behavioral measures from risk-taking tasks such as the Balloon Analogue Risk Task (BART) find mixed results. However, reversal learning paradigms display consistent behavioral alterations across different stages of SZ. To account for this, we introduced a version of the BART with a reversal of risk contingencies and collected subclinical questionnaires for schizotypal traits (SPQ) and psychotic-like experiences (PLEs; CAPS, PDI) to investigate trait-relevant behavioral alterations. We found no behavioral correlation between adjusted pumps on the BART and schizotypal traits or PLEs. We then used three of the leading BART computational models (EWMV, Four Parameter, STL) across pre-reversal and post-reversal data separately and performed model comparisons to find which best describes participant behavior. Using the winning model (EWMV) we found that pre-reversal SPQ, PDI, and CAPS scores were all positively associated with the model’s updating exponent, meaning individuals with higher schizotypal traits and PLEs updated initial risk beliefs more quickly at the beginning of the task. Post-reversal, PDI and SPQ scores were negatively associated with inverse temperature, meaning that when risk contingencies in the task changed, individuals with higher PDI/SPQ scores displayed more stochastic behavior. Lastly, a negative correlation was seen between SPQ and loss aversion post-reversal. In summary, these exploratory results provide evidence that changing contingencies is beneficial to risk-taking assessments in subclinical SZ research due to computationally tractable trait-level alterations in risk-taking mechanisms.