Modeling cognitive reappraisal using model-based RL - Revised

This study utilizes a modified version of the sequential decision-making task originally registered (https://osf.io/2etby/overview). The primary adjustment was necessitated by initial pilot data (N=20) indicating that the revaluation effect in the original design was too weak and inconsistent to reliably measure individual differences, as participants frequently failed to maintain continuity between phases. To address this, we implemented specific structural changes in the current version (v3). We strengthened the incentives by increasing the magnitude of the revaluation shift, with outcomes now spanning ±20 points instead of the previous smaller range, to create a more salient value conflict. We also enforced attention by adding a dedicated "Attention Check" phase and interstitial "Preview" screens before ratings to prevent rapid clicking and ensure processing. Furthermore, we show total points to explicitly signal that performance has direct consequences for the bonus payment. Finally, we changed the revaluation phase from a passive instruction screen to an active "Click-to-Reveal" process with a mandatory 10-second reflection period to support model-based planning over habitual responding. A subsequent pilot (N=15) using this updated structure confirmed significant revaluation effects, validating the design changes. However, an asymmetry was observed where accuracy in the Feather condition was significantly lower than in the Triangle condition during the initial revaluation phase. To robustly test our individual difference hypotheses despite this variance, we are proceeding with a full-scale collection of N=150. We aim to detect a medium effect size of r = 0.3 regarding the correlation between revaluation performance and ERQ/Anxiety scores. To achieve 95% power, accounting for potential exclusions and interaction effects in the Linear Mixed Model, a sample of 150 participants is required, which exceeds the base requirement of N=112 for 90% power. The study employs a web-based sequential decision-making paradigm designed to operationalize reappraisal computationally via model-based reinforcement learning. The task centers on a two-step state-space map where participants navigate from an Initial state (Circle) to one of two Middle states (Feather or Triangle), and finally to one of four Terminal states (Square, Lamp, Amphora, or Cloud). Transitions to the Middle states are deterministic, while transitions to the Terminal states are probabilistic, featuring a 75% common and 25% rare transition structure. The session begins with a Structure Learning phase consisting of 32 forced-choice trials to build a cognitive map of the transitions, followed by a Structure Quiz that enforces 100% accuracy to proceed. Participants then advance to the Value Learning phase, comprising 20 trials where reward contingencies are introduced: the Feather yields +5 points and the Triangle yields -5 points. To ensure engagement, a real-time display of total points tracks the score throughout this phase. Following this, a strict Attention Check explicitly warns participants that performance affects their bonus payment, after which they complete Rating 1, providing a baseline subjective valuation of the Middle states using a Visual Analogue Scale (VAS). Subsequently, participants enter the first "Structure Update" phase (Revaluation 1), which assesses model-based revaluation. This is an active learning process where participants must click on Terminal states to reveal a major value shift: the Square (T1) and Lamp (T2) drop to -20 points, while the Amphora (T3) and Cloud (T4) rise to +20 points. This creates a direct value conflict where the previously rewarding Feather pathway is now punishing, and the previously punishing Triangle pathway is now rewarding. After revealing these values and undergoing a mandatory 10-second Reflection Period to plan their strategy, participants complete Rating 2. This second rating assesses the extent to which they have updated their subjective valuation of the Middle states before making any new choices. Immediately following this, they complete Query 1, a series of 8 binary decisions between navigating a Middle state or accepting a fixed offer ranging from -20 to +20. The task concludes with a second revaluation phase (Revaluation 2) to test flexibility. Participants again actively reveal updated values for the Terminal states: while the Square (T1) remains at -20 points and the Cloud (T4) remains at +20 points, the Lamp (T2) and Amphora (T3) shift to 0 points (neutral). This nuanced change requires a fine-tuning of the expected values. Following another mandatory reflection period, participants complete Rating 3 to capture their final subjective valuations. This is followed by Query 2, a final set of decision probes to assess behavioral adaptation to this second shift. Upon completing the behavioral task, participants answer three questionnaires: the Emotion Regulation Questionnaire (ERQ) to measure reappraisal strategies, a Trait Anxiety Inventory assessing worry, and a Physiological Anxiety Checklist assessing physical symptoms. We hypothesize that effective reappraisers, characterized by high ERQ scores, will calculate the optimal Expected Value (EV) of each state by planning their next move to maximize reward generally across revaluation queries. Specifically, our first hypothesis (H1) posits that individual differences in ERQ Reappraisal scores will positively correlate with revaluation accuracy, defined as choice behavior that reflects the updated EV. Our second hypothesis (H2) posits that individual differences in Trait Anxiety, specifically worry, will negatively correlate with revaluation accuracy. We predict that high-worry individuals will too easily reevaluate positive states to negative due to an over-sensitivity to threat and will take longer to reevaluate negative states to positive. Both tests are directional and corrected for multiple comparisons with an alpha level of 0.025. Based on our power analysis for a medium effect size of approximately 0.3, our sample of 150 participants provides greater than 95% power to detect these effects. Finally, we expect to find again the effect of better initial revaluation of triangle (negative->positive) than feather (positive->negative), which was significant using a linear-mixed model in the pilot data, and at n=150 is powered at 0.99. you can find the data from our most recent pilot at: https://github.com/adipizanti/Modeling-cognitive-reappraisal-using-model-based-RL