Affective context shape confidence and motivated task behavior irrespective of objective performance

Background:

Humans are constantly confronted with novel challenges, in which it is unclear whether one’s own action will yield a desired outcome. To navigate these challenges, learning from past experience may inform the subjective probability to achieve the desired outcome, while subjective value of the outcome and potential consequences determine subsequent effort investment. Past studies have investigated how individuals weigh fixed effort expenditures against possible rewards, whereas less research has been conducted on the effect of changing action-outcome contingencies and affective contexts (i.e. gain or loss) on subsequent effort investments. Here, we attempt to fill this gap by employing a novel experimental task, investigating how subjective and computational measures of uncertainty as well as affective contexts shape effort investment in a constantly changing environment. We hypothesize that the potential of losing points, exerts influence on effort exertion and subjective uncertainty, while leaving model-derived estimates of uncertainty unaffected.

Methods:

Participants (n=94) performed a two-armed bandit learning task and earned money based on their performance. Participants were randomly assigned to one of two paradigm versions, offering different affective contexts: in the loss-version, participants (n=50) invested effort on each trial using a hand-dynamometer and gained or lost points in proportion to that effort depending on whether the preceding bandit-choice was correct or not, while in the no-loss-version (n=44) incorrect choices yielded zero points. On a subset of trials, participants rated their confidence (subjective uncertainty) in correctly predicting the next outcome. Computational facets of uncertainty were captured using the Hierarchical Gaussian Filter (HGF). To assess the effects of affective context on behavior and task performance, we employed hierarchical linear models (HLM) and unpaired two-sided t-tests. Potential systematic differences in uncertainty representations across contexts were examined by fitting HGF variants to combined and separated datasets, with the best-fitting model identified via Bayesian Model Selection (BMS). Finally, HGF-derived uncertainty estimates were used as predictors of trial-wise behavior.

Results:

Task performance did not differ between conditions, indicating that affective context had no significant effect on learning (p=0.35). Across both contexts, trial-wise confidence positively predicted subsequent effort investment (p<0.001). However, participants in the loss condition invested significantly less effort overall (p<0.001), while unexpectedly confidence ratings did not differ between conditions (p=0.58). HLM analyses further revealed significant interactions between affective context and trial outcome: following incorrect choices, participants in the loss condition showed both higher switch probability (p<0.001) and lower confidence ratings (p=0.035). As predicted, BMS identified the HGF fitted to the combined dataset as the best-fitting model of participants’ choice behavior (exceedance probability=1). HGF-derived uncertainty estimates significantly predicted both effort investment (p<0.001) and confidence (p<0.001).

Discussion:

Together, our results support the notion that affective contexts shape both effort investment and confidence, particularly after punishment, while leaving computational estimates of uncertainty unaffected. This dissociation underlines how affective context drives confidence and motivated behavior independently of objective performance, with particular relevance for conditions characterized by heightened negative affect, such as depression and anxiety.

References

No references