Hierarchical Bayesian Reinforcement Learning Models of Loss-Avoidance under Threat: Orthogonalised Go/No-Go Task Insights into Anxiety and its Treatment
Year: 2026
Type: dissertation
Abstract: This thesis is both a methods development thesis and a hypothesis testing one. To investigate the computational mechanisms of avoidance and anxiety, I have developed and implemented state of the art-hierarchical Bayesian reinforcement learning models to single studies, mega-analysis across studies, and in between-subjects RCT designs. Maladaptive avoidance is a hallmark of anxiety disorders, it is a main factor in their impact on daily life functioning, their maintenance, and the comorbidity with depression. This thesis specifically explores the computational mechanisms of anxiety and avoidance. Throughout this thesis I use an orthogonalised version of the go/no-go task with a threat-of-shock component. In the methodological innovation chapter, I describe in detail the new RL methods I have developed and implemented. In the first empirical chapter, I sought to replicate previous research that found that individuals with anxiety disorders have a heightened reliance on Pavlovian avoidance bias when compared to healthy controls, and that this is particularly the case in the face of acute stress as induced by threat-of-shock. I did not fully replicate these findings. I found individuals with anxiety disorder exhibited a Pavlovian bias whereas controls exhibited a Pavlovian approach bias; threat was associated to an increase of both Pavlovian biases. In the second empirical chapter, I modified my model to aggregate the data from three di`erent studies into a mega-analysis; the increased sample size a`orded to see that individuals with anxiety disorders have a decreased learning rate under threat, I also saw that threat increased both Pavlovian approach and avoidance parameters, as well as action bias; while it decreased learning rates. In the last empirical chapter, I used an experimental, between-subjects RCT design to study the computational mechanisms of subchronic escitalopram; I found no computational e`ects related to treatment nor any other manipulation except for threat-of-shock being associated to Pavlovian approach bias and a decrease in punishment learning rates. The general discussion discusses these findings, considers its implications, limitations, and possible future directions.
Author: Mena Poblete, Carlos
Language: English
Subfield: Experimental and Cognitive Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7223499161