Anhedonia and Anxious Arousal are Associated with Distinct Expectations about the Statistics of a Volatile Environment

Figure 1. Deriving volatility estimates from reversal learning. (A) Three-level mean-reverting hierarchical Gaussian filter (54). Level 1 (x1) represents observations of win versus no win (rewarding version) or loss versus no loss (aversive version). Level 2 (x2) represents an agent’s expectations of the outcome probabilities associated with each card deck. μ02 and m2 are the prior and equilibrium expectations, respectively, of the outcome probabilities. κ2 is a coupling parameter between levels 3 and 2, and ω2 is the rate of adjustment of reward or loss expectations (tonic volatility). Level 3 (x3) represents expectations of the changeability of the contingencies associated with each deck. μ03 and m3 are participants’ prior and equilibrium volatility estimates, respectively, and ω3 is the rate of adjustment of volatility expectations (meta-volatility). Agents’ responses (y) are determined by a softmax μ3 decision model with inverse decision temperature β. (B) Sample trial from the reward version (top) and aversive version (bottom) of a probabilistic reversal-learning paradigm. On each trial of the reward version, participants could win 50 points or lose 0 points, and on each trial of the…

Figure 2. Anxious arousal and anhedonia are associated with greater and reduced exploratory behavior, respectively, in an aversive context. Mixed-effects models with task block, task valence (gain vs. loss), and a participant-specific random intercept, controlling for paranoia, predicting (A) win-switch rates by anxious arousal and (B) U-values by anhedonia. Anxious arousal and anhedonia were measured with the 30-item adaptation of the Mood and Anxiety Symptom Questionnaire (52) and mean centered prior to analyses. Anxious arousal interacted with task valence to predict a greater win-switch rate in the loss vs. gain versions of the probabilistic reversal-learning task (t657 = 2.80, p = .005). Anhedonia interacted with task valence to predict lower U-values in the loss vs. gain versions (t657 = −2.12, p = .035). Anxious arousal and its interaction with valence were not significantly associated with lose-switch rate (ps > .519) or U-values (ps > .222). Anhedonia was not significantly related to switch behavior (ps > .166). See Figure S3 for visualizations of anxious arousal, anhedonia, and general distress with all switch rates and U-values.

Figure 3. Anhedonia and anxious arousal are associated with distinct prior beliefs about volatility. Anxious arousal and anhedonia predicting prior beliefs about volatility (μ03) derived from a 3-level hierarchical Gaussian filter fit to behavior on 2 probabilistic reversal-learning tasks. Clinical symptoms were measured with the 30-item adaptation of the Mood and Anxiety Symptom Questionnaire (52). Depicted results are from linear mixed-effects models predicting μ03 by task valence and (A) anxious arousal or (B) anhedonia, controlling for paranoia, with a random intercept for participant. Anxious arousal interacted with valence to predict a greater initial expectation of volatility in the loss vs. gain versions of the task (t219 = 2.24, p = .026), whereas anhedonia predicted a reduced initial expectation of volatility (μ03) (t219 = −2.90, p = .004), regardless of task valence (p = .987). See Figure S4 for a visualization of μ03 findings with the general distress dimension.