Post-Outcome Valence, Not Emotion Prediction Errors, as a Primary Predictor of Behavior

What can feelings tell us about human behavior that we cannot know from directly observable states alone? From a reinforcement learning perspective, adaptive behavior is calibrated based on deviations from predicted states, called prediction errors (PEs). This perspective has typically focused on outcomes in the environment (e.g., monetary or food rewards, actions, etc.), relatively de-emphasizing agents’ internal feeling states. Heffner and colleagues (2021) complemented this focus by additionally probing deviations from predicted feelings (i.e., emotion PEs). In four experiments using two-dimensional valence-arousal grids, participants reported how they expected to feel about their predicted outcomes and how they subsequently felt about the actual outcomes. In this context, “outcomes” refers to monetary offers in social settings (i.e., variations of the Ultimatum Game). Measuring outcome PEs alongside emotion PEs made it possible to use both PE types as joint predictors of choice behavior (i.e., reactions to offers). When accounting for outcome PEs as well as predicted emotion, emotion PEs remained significantly predictive of choice behavior. Moreover, the association between emotion PEs and choice behavior was altered in participants at risk for depression. Based on these associations, the authors claimed that emotion PEs not only explain variance in choice behavior beyond outcome PEs but even guide socially adaptive behavior. Here, we revisit the role of self-reported emotions in behavior. Our follow-up analyses revealed post-outcome valence—how good or bad participants felt after an outcome—to be a better predictor of behavior, and thus a stronger candidate than emotion PEs for “guiding” it.