Learning not to fear: Metacognitive contributors to safety computations
For humans, the inability to identify safety is a hallmark of anxiety, which is linked to poor health (Felger, 2018) and psychological outcomes (Mennin et al., 2005). Existing computational accounts fail to explain why individuals suffering from anxiety have difficulty identifying safety. This lack of understanding may be due, in part, to existing focus on threat processes with an assumption that safety is an inverse of threat. This perspective presumes that deficits in safety recognition are a result of threat overestimation, but fails to consider circumstances in which threat is accurately estimated, but safety is unrecognized. Safety may instead reflect independent computations related to the self that mediate threat estimates (Tashjian et al., 2021).
The proposed research partitions the complexity of safety recognition into two main evaluative components: an external focus on threat and an internal focus on self-competence. Empirical investigations of safety processing are lacking, and for the few that exist, metacognition (the internal assessment of one’s ability to achieve a desired outcome) is not evaluated. However, metacognition is likely crucial given poor metacognition can impair perceptual judgment and error integration, which may ultimately influence safety estimations (Rigoni et al., 2013).
Identifying internal contributors to safety underestimation has significant potential to improve treatment efficacy for anxiety disorders. Nearly 50% of clinically anxious individuals do not fully respond to treatment focused on external threat responsivity (Ginsburg et al., 2018). Thus, safety learning may provide an alternative pathway for reducing anxiety in those resistant to threat-based treatment. The current project aims to develop a computational account of safety learning to increase precision in the treatment of anxiety. Findings will shed light on the component and interactive processes by which threat perception and metacognition contribute to anxiety.
This proposal will facilitate development of a new task to define the computational underpinnings of internal and external safety evaluations. Behavioral data will be collected from 250 adult participants (ages 18-40) using Prolific. Computational modeling will be used to provide a mechanistic account of how individuals engage in safety learning and decision making. Self-reported anxiety, intolerance of uncertainty, and anxiety-related control will be measured as constructs hypothesized to have relevance for individual differences in safety processing.
Learning phase: First participants learn about their own strength, which changes each round according to how much food they have consumed (metacognitive learning). Confidence and performance estimates will be measured to assess metacognitive accuracy. Participants also learn about the wolves strength, which changes each round due to perceptual features (external learning). Learning rate curves will be compared to test the hypothesis that learning for the self has a greater functionality (increased overall learning), but a slower learning onset.
Decision phase: Participants will also complete a series of decisions about how much protection they need to combat the wolves. In order to obtain protection, participants must complete a cognitive task that varies in difficulty based on their avatar’s strength. Psychophysical curves will be compared to examine decision making when self contingencies change but threat remains constant versus when threat contingencies change and the self remains constant. This comparison tests the hypothesis that participants become more risk averse when external threats increase versus when self strength decreases, suggesting heightened attending to external factors over internal factors when evaluating safety. More anxious participants are hypothesized to choose suboptimal protection when the cognitive task is required due to underestimations of competence.
Affective spillover: Lastly, participants will perform an affective spillover task where they do not receive an aversive outcome for encountering a predator. Reaction time responses to predator images will be measured as an index of threat salience.
A sample size of N=250 is sufficient to achieve 90% power to detect a small effect of d=0.20 with a two-tailed alpha=0.05. K-fold cross validation will be used to estimate out-of-sample accuracy. Participants will be excluded from analyses if they do not respond within the time limit on more than 20% of all trials. Prior work suggests an estimated 5% participants will be excluded based on quality checks (n=12). Payment will not be affected by quality assurance.
Study methods and hypotheses have been preregistered on Open Science Framework (OSF): https://osf.io/zfsk2
Open and Reproducible Science
I actively participate in open science via preregistration, open data, open materials, and pre-prints. The Safety Task, computational pipelines, and deidentified data will be shared on OSF.
The task takes approximately 2 hours to complete (100 trials for each phase, 50 trials for spillover task, questionnaires). Participants will be paid USD$9.50 per hour based on Prolific suggested funding (USD$19.00 total). Participants with high performance will be placed in a lottery for an extra $100 bonus, which will be provided to 3 participants. The bonus lottery improves attention and engagement with the task.
Based on prior piloting for novel task development with Prolific, this proposal estimates a need for 90 pilot participants to allow for 30 pilot participants over 3 task iterations. After task development is complete, the preregistered sample size of 250 plus an extra 12 as a quality buffer will be collected.
Including service and VAT, pilot funds needed are USD$2280.00, full sample funds (262 participants) are USD$6,637.33, and bonus funds are USD$400. In sum, this proposal application totals USD$9,317.33.