Confidence predicts speed-accuracy tradeoff for subsequent decisions

Kobe Desender, Annika Boldt, Tom Verguts, Tobias H. Donner

TL;DR

This paper reveals that our confidence in a past decision directly influences how we approach the next decision, making us more cautious (slower but more accurate) after feeling uncertain. This effect is mediated by a specific brain signal, the "error positivity" (Pe) component of EEG, suggesting our brains dynamically adjust decision-making strategies based on internal confidence estimates, even without external feedback.

What Problem Does This Paper Address?

When we make decisions, especially under uncertainty and without immediate feedback, how do we adjust our strategy for future choices? Traditional models often assume fixed decision parameters. However, it's intuitive that if we feel unsure about a past decision, we might want to be more careful next time. This paper investigates whether and how our subjective feeling of confidence in a decision influences the speed-accuracy tradeoff in subsequent decisions, particularly in the absence of external reinforcement. They hypothesize that low confidence leads to a more conservative decision strategy, characterized by higher decision bounds, thus prioritizing accuracy over speed.

Methods

The researchers employed a multi-pronged approach:

  • Drift Diffusion Model (DDM): This is a standard computational model for decision-making that describes how evidence is accumulated over time until it reaches one of two boundaries, representing a choice. They used a hierarchical Bayesian version of the DDM to fit behavioral data from participants across three experiments. This allowed them to estimate key parameters like decision bounds and drift rates.
  • Behavioral Experiments: Three distinct perceptual choice tasks were used. Participants made decisions, reported their confidence, and in some cases, their choices were followed by a brief period of continued sensory evidence or a blank screen.
  • EEG Recording: In one experiment, electroencephalography (EEG) was recorded to capture brain activity. They focused on specific event-related potential (ERP) components known to be associated with confidence and error processing: the Error-Related Negativity (ERN) and the Error Positivity (Pe).
  • Regression Analysis: They used regression models to link behavioral parameters (like decision bounds) to confidence ratings and EEG signals (Pe and ERN amplitudes).

Datasets

The study utilized data from three distinct experiments conducted by the authors. No new benchmark datasets were introduced.

  • Experiment 1: A dot motion discrimination task.
  • Experiment 2: A visual task involving judging which of two fields contained more dots.
  • Experiment 3: A visual color categorization task.

Key Insights and Ideas

  • Confidence as a "Next-Trial" Policy Setter: The most striking finding is that confidence in a previous trial directly impacts the decision-making strategy on the current trial. This suggests a dynamic, trial-by-trial adjustment of cognitive control.
  • The "Cautious" Response to Uncertainty: When participants reported low confidence (especially after perceived errors), their decision bounds on the next trial were significantly increased. This means they required more evidence before committing to a decision, leading to slower but more accurate choices.
  • Neural Basis of Confidence-Driven Adjustment: The study identified the "error positivity" (Pe) component of EEG as a key neural correlate. Higher Pe amplitudes, which are linked to higher confidence, linearly predicted increased decision bounds on the subsequent trial. This suggests the brain uses internal confidence signals to proactively adjust behavior.
  • Independence from Objective Accuracy: The effect of confidence on decision bounds was observed even when controlling for the objective accuracy of the previous trial, highlighting the role of subjective confidence rather than just error detection.
  • Generalizability Across Tasks: The observed effect of confidence on decision bounds was consistent across the three different perceptual tasks, suggesting a fundamental mechanism of cognitive control.

Key Results

  • Increased Decision Bounds: Low confidence and perceived errors on trial n led to significantly increased decision bounds on trial n+1. For instance, perceived errors led to a larger increase in decision bound compared to low confidence.
  • No Significant Effect on Drift Rate: While decision bounds were consistently modulated by confidence, the effect on the drift rate (the speed of evidence accumulation) was smaller and less consistent across experiments.
  • Pe Predicts Decision Bound: Regression analysis showed a significant positive relationship between Pe amplitude and subsequent decision bound (p < 0.001).
  • ERN Not a Predictor: In contrast, the ERN, another confidence-related signal, did not significantly predict subsequent decision bound.
  • Linear Scaling: The relationship between Pe amplitude and subsequent decision bound was approximately linear.

Limitations and Future Work

  • Individual Differences: The use of a hierarchical Bayesian DDM, while beneficial for handling unbalanced data, constrains strong conclusions about individual differences. Future work could collect more extensive data per participant to explore these variations.
  • Source of Confidence Variation: The paper notes that confidence can be low due to insufficient evidence quality (low drift rate) or insufficient evidence accumulation (low decision bound). Future research could disentangle the strategic adjustments driven by these different sources of uncertainty.
  • Drift Rate Modulation: The observed modulation of drift rate by confidence was less consistent and smaller than the effect on decision bounds. Further investigation is warranted to understand its role and potential mechanisms (e.g., attentional focus).
  • Neural Generators of Pe: While the Pe was identified as a key neural marker, its precise neural generators remain unknown and potentially widespread. Further research could pinpoint these sources.
  • Ecological Validity: The study's findings are particularly relevant for real-world decision-making where external feedback is often absent, but the specific mechanisms of how internal confidence signals translate to adaptive policies in various contexts could be further explored.

Summary updated

Abstract

When external feedback about decision outcomes is lacking, agents need to adapt their decision policies based on an internal estimate of the correctness of their choices (i.e., decision confidence). We hypothesized that agents use confidence to continuously update the tradeoff between the speed and accuracy of their decisions: When confidence is low in one decision, the agent needs more evidence before committing to a choice in the next decision, leading to slower but more accurate decisions. We tested this hypothesis by fitting a bounded accumulation decision model to behavioral data from three different perceptual choice tasks. Decision bounds indeed depended on the reported confidence on the previous trial, independent of objective accuracy. This increase in decision bound was predicted by a centro-parietal EEG component sensitive to confidence. We conclude that internally computed neural signals of confidence predict the ongoing adjustment of decision policies.
BibTeX
@inproceedings{desender2019confidence,
  author = {Kobe Desender and Annika Boldt and Tom Verguts and Tobias H. Donner},
  title = {{Confidence predicts speed-accuracy tradeoff for subsequent decisions}},
  booktitle = {eLife},
  year = {2019},
  doi = {10.7554/elife.43499},
}
Figure