Mood impacts confidence through biased learning of reward likelihood

This paper demonstrates through computational modeling and two independent studies that mood disturbances bias confidence in reward-based decision-making not by directly altering self-belief, but by distorting the gradual accumulation of reward learning signals, a mechanism that is amplified in individuals with elevated hypomanic traits.

The Gist

Imagine your brain is a smart navigation app trying to figure out the best route to a destination (your goals). This app constantly learns from traffic reports (rewards) to tell you how likely you are to succeed. Usually, it updates its map based on the actual data it receives.

But what happens when your mood acts like a glitchy filter over that map?

This paper explores exactly that. The researchers wanted to understand why our confidence in our abilities can swing wildly when we are feeling down or up, even if our actual skills haven't changed. They tested two main theories:

1. The "Direct Override" Theory: Mood acts like a manual button that instantly forces the app to say, "You will succeed!" or "You will fail!" regardless of the actual traffic data.
2. The "Distorted Data" Theory: Mood doesn't change the final answer directly; instead, it twists the traffic reports as they come in. The app thinks the data is different than it really is, so it slowly builds a wrong conclusion about how likely success is.

How they tested it:
The researchers ran two studies (one in a lab and one online) where they changed people's moods and then asked them to play a game where they had to guess which choices would win rewards. After every guess, the players had to rate how confident they were.

What they found:
The "Direct Override" theory was wrong. Mood didn't just instantly flip a switch on confidence. Instead, the "Distorted Data" theory was the winner.

Here is the key discovery:

• The Slow Burn: When people's moods were manipulated, their confidence didn't change immediately. It took time. As they kept playing the game and learning, their confidence slowly drifted in the direction of their mood.
• The Ghost in the Machine: Even after the mood manipulation stopped and people felt "normal" again, the confidence they had built up during the "glitchy" mood period stuck around. It was as if the mood had secretly rewritten the history of the game in their minds.
• The Magnifying Glass: The study found that people who naturally have more unstable moods (specifically those with higher "hypomanic traits") were more sensitive to this glitch. Their navigation apps were more easily distorted by the mood filter.

The Bottom Line:
The paper concludes that when we feel a certain way, it doesn't just make us feel more or less confident. Instead, our mood actually biases how we learn from our experiences. It tweaks the way our brain processes the "evidence" of success or failure. Over time, this distorted learning creates a false sense of confidence (or lack thereof) that feels very real, even though the actual facts haven't changed.

Think of it like wearing tinted glasses while learning to drive. You aren't just feeling like a bad driver; the glasses are actually making the road signs look different, so you learn to drive differently than you would if you were wearing clear glasses.

Technical Summary

Technical Summary: Mood Impacts Confidence Through Biased Learning of Reward Likelihood

Problem Statement
Mood disturbances are characterized by significant, often non-rational fluctuations in self-efficacy and goal-directed behavior. However, the specific mechanisms linking affective states to these behavioral and cognitive shifts remain poorly understood. The central question addressed is whether mood influences confidence in reward-based decision-making through a direct, evidence-independent shift in confidence, or whether mood distorts the underlying accumulation of reward learning signals from which confidence is subsequently inferred. Furthermore, the study investigates whether these effects are amplified by trait instability in mood and self-efficacy.

Methodology
The research employed a computational modeling approach to formalize both confidence and mood within a reinforcement-learning framework. The study design consisted of two independent investigations:

1. A laboratory experiment with 35 participants.
2. A preregistered online replication with 106 participants.

In both studies, participants engaged in a reinforcement-learning task involving mood manipulation and repeated confidence ratings. The researchers utilized computational models to adjudicate between two competing accounts:

• The Direct Shift Account: Mood imposes an immediate, evidence-independent bias on confidence.
• The Biased Learning Account: Mood distorts the accumulation of reward learning signals, leading to rationally inferred but biased confidence.

Key Results
The empirical data revealed that mood systematically biased confidence ratings without altering objective accuracy. Crucially, the temporal dynamics of this bias supported the "Biased Learning" account over the "Direct Shift" account:

• The bias in confidence accumulated gradually across the learning process rather than appearing immediately following mood induction.
• This accumulation persisted even as the induced mood state itself returned toward baseline levels.

A computational model incorporating a mood-bias parameter successfully captured these dynamics. In this model, mood biases the accumulation of reward probability estimates. This single parameter accounted for:

• The observed shifts in confidence.
• The biased valuation of stimuli learned under different mood states.
• The amplification of these effects in individuals exhibiting elevated hypomanic traits.

Significance and Claims
The paper claims to provide a mechanistic account linking mood-induced distortions in confidence to disrupted goal pursuit and self-efficacy in mood disorders. By demonstrating that confidence disturbances arise from mood-biased learning signals rather than direct, evidence-independent shifts, the findings suggest that the core pathology may lie in how affective states alter the processing of reward information over time. This offers a specific computational explanation for the non-rational changes in self-efficacy observed in clinical populations, distinguishing between the immediate state of mood and its lasting impact on the learning process.