Zoning Out? Inferring Latent Engagement States During Perceptual Decision-Making in Humans

Introduction: Psychology experiments typically assume participants remain focused on the task, without explicitly accounting for periods of disengagement. However, attention naturally fluctuates, and standard experimental designs may not capture these dynamics. Here, we use a generalized linear model–hidden Markov model (GLM-HMM) to infer latent engagement states directly from human choice behaviour. The present study is part of a larger project on engagement dynamics and aimed to: 1) optimize the statistical pipeline for subsequent large-scale analyses, and 2) assess the robustness and reproducibility of the GLM-HMM framework by replicating findings from Ashwood et al. (2022). Methods: We analysed choice behaviour from 27 participants performing a two-alternative motion discrimination task. A GLM-HMM was used to infer latent engagement states from trial-by-trial responses. Model parameters were optimized in a pilot dataset (n = 3) to evaluate convergence, prior settings, and model...Show moreIntroduction: Psychology experiments typically assume participants remain focused on the task, without explicitly accounting for periods of disengagement. However, attention naturally fluctuates, and standard experimental designs may not capture these dynamics. Here, we use a generalized linear model–hidden Markov model (GLM-HMM) to infer latent engagement states directly from human choice behaviour. The present study is part of a larger project on engagement dynamics and aimed to: 1) optimize the statistical pipeline for subsequent large-scale analyses, and 2) assess the robustness and reproducibility of the GLM-HMM framework by replicating findings from Ashwood et al. (2022). Methods: We analysed choice behaviour from 27 participants performing a two-alternative motion discrimination task. A GLM-HMM was used to infer latent engagement states from trial-by-trial responses. Model parameters were optimized in a pilot dataset (n = 3) to evaluate convergence, prior settings, and model stability. The final analysis replicated Ashwood et al. (2022) using the refined pipeline. Model performance was assessed via cross-validation, and state-dependent effects on behavioural parameters, reaction times, and accuracy were evaluated using linear and generalized linear mixed-effects models Results: Cross-validation indicated that a two-state GLM-HMM explained choice behaviour better than a single-state model (ΔLL = 0.012 bits/trial). Sensory evidence weights were higher in state 1 than state 2 (t(25) = 2.63, p = .015, d = 0.52), whereas use of the win–stay, lose–switch strategy was greater in state 2 (t(25) = −4.73, p < .001, d = 0.93). Reaction times were slightly longer in the less engaged state (β = 0.015, p < .001), and accuracy tended to be lower. The probability of engagement declined across trials (β = −0.00060, p < .001). Discussion: The findings provide preliminary evidence that the GLM-HMM can identify latent fluctuations in task engagement from behavioural data. Two distinct states reflected varying reliance on task-relevant and task-irrelevant strategies. Our results closely replicate those of Ashwood et al. (2022), underscoring the robustness and reproducibility of the framework. By refining the statistical pipeline, this study establishes a foundation for large-scale analyses aimed at understanding how engagement fluctuates during perceptual decision-makingShow less

Leiden University Student Repository

Documents

In Collections

Zoning Out? Inferring Latent Engagement States During Perceptual Decision-Making in Humans