Anxious individuals have difficulty learning the causal statistics of aversive environments

Abstract

Statistical regularities in the causal structure of the environment enable us to predict the probable outcomes of our actions. Environments differ in the extent to which action-outcome contingencies are stable or volatile. Difficulty in being able to use this information to optimally update outcome predictions might contribute to the decision-making difficulties seen in anxiety. We tested this using an aversive learning task manipulating environmental volatility. Human participants low in trait anxiety matched updating of their outcome predictions to the volatility of the current environment, as predicted by a Bayesian model. Individuals with high trait anxiety showed less ability to adjust updating of outcome expectancies between stable and volatile environments. This was linked to reduced sensitivity of the pupil dilatory response to volatility, potentially indicative of altered norepinephrinergic responsivity to changes in this aspect of environmental information.

Figures

Figure 1. Task Structure

a) Example trial. Participants had to choose one of two shaped Gabor patches. Each shape contained a two digit number which indicated the magnitude of electrical shock that might be received should that shape be chosen. Following option presentation, onset of a response cue indicated that participants could make their choice. After response, a variable interval was followed by outcome delivery. The shape associated with the electrical shock for that trial was displayed in the centre of the screen for 2s. If the participant had chosen this shape, an electrical shock of the indicated magnitude was delivered at the onset of the outcome period. b) Outcome probabilities across the course of the task. The task comprised two blocks. In the stable block (shaded), one shape (for example, the circle) had a 75% probability of resulting in an electrical shock being delivered, if it was chosen; the other shape (e.g. a square) has a 25% probability of resulting in shock delivery. In the volatile block (unshaded) the probability that choice of a given shape would result in shock delivery switched every 20 trials between 80% and 20%. Participants were randomly assigned to complete the task with the stable block first (as shown), or with the volatile block first.

Figure 2. Estimates of Participants’ Learning Rates

a) Participants’ choices during the stable and volatile blocks of the aversive learning task were fitted with a Rescorla Wagner learning model in which learning rate was allowed to vary. Estimates of individual participants’ learning rates are displayed (circles) separately for the stable and volatile blocks for the two task schedules (Schedule 1= stable task block first, n=15, Schedule 2 = volatile task block first, n=15). A logarithmic scale is used. Black lines display mean (+−SEM) of participant learning rates, grey dotted lines link the learning rates in volatile and stable blocks for each participant. Participants showed higher learning rates in the volatile versus stable blocks regardless of the order in which they were completed, F(1,28)=16.3, p<0.001. b) The relative log learning rate for the volatile versus the stable blocks (i.e. log (LR in volatile block) – log (LR in stable block)) was negatively correlated with participant trait anxiety, r(28)=−0.42, p= 0.02. The black dotted line indicates the degree to which the model of an optimal Bayesian learner (as described by Behrens and colleagues) adjusted its learning rate. As can be seen, low trait anxious participants altered their learning rates to a similar degree to the Bayesian Learner, with high trait anxious participants showing a reduced adaptation of learning rate between the volatile and stable blocks of the task. Error bars represent the standard deviation of the estimated parameters from the behavioral model for each subject.

Figure 3. Post Outcome Pupil Dilation Tracks Both Environmental Volatility and Outcome Surprise

Time courses for the effect of trial-wise estimates of (a) volatility and (b) surprise on pupil dilation following presentation of the outcome. The graphs show the mean across participants (n=28) of the beta weights obtained by regressing post outcome pupil dilation against trial-wise estimates of environmental volatility and outcome surprise. Post outcome pupil dilation was greater for trials where environmental volatility was high, F(1,26)=9.8, p=0.004, and the outcome was surprising, F(1,26)=9.2, p=0.005. Asterisks indicate 1s time bins in which the effect of volatility or surprise on pupil dilation post outcome differed significantly from zero (bonferonni corrected for multiple comparisons, ps corrected <.05 the effect of trial-wise volatility was longer lasting and had a later onset than that outcome surprise. degree to which an individual pupil tracked as mean beta weight across second post-outcome period predicted change in learning rate between volatile stable blocks>r(26)=0.37, p=0.05. (d) The degree to which an individual’s pupil tracked surprise predicted extent of surprise-related choice reaction time slowing on the subsequent trial, r(26)=0.44, p=0.02. Shaded regions in panels a and b represent the standard error of the mean. Error bars in panels c and d represent the standard deviations of the regression coefficients (beta weights) from the pupil analysis and the parameter estimates from the behavioral model for each subject.

Figure 4. The Relationship Between Trait Anxiety and Post Outcome Pupil Dilation as a function of trial-wise estimates of Volatility and Surprise

a) The degree to which participants’ pupil dilation, post outcome, tracked environmental volatility was negatively related to trait anxiety, r(26)=−0.51, p=0.005. b) Using a median split on participants’ trait anxiety scores, low anxious participants (n=15) showed a clear pupil response to environmental volatility whereas high anxious participants (n=13) did not (asterisks indicate 1s time bins in which Bonferonni corrected t-tests differed between the groups at p<.05 corrected, 2-tailed.) c) Pupil response to outcome surprise was not related to individual differences in trait anxiety (p=0.4), d) this is illustrated using a median split on trait anxiety. Error bars in panels a and c represent the standard deviations of the regression coefficients (beta weights) from the pupil analysis for each subject. Shaded regions in panels b and d represent the standard error of the mean.