TL;DR
This paper introduces a Bayesian sequential sampling model that explains how people integrate information about options when making decisions, considering both what they know (expectations) and what they don't know (uncertainty). The model predicts, and experiments confirm, that attention and confidence boost choices for better-than-average options but hinder choices for worse-than-average ones, offering a nuanced view of decision-making under uncertainty.
What Problem Does This Paper Address?
When faced with multiple choices, we rarely have complete information about each option. We might spend more time looking at some options (attention) or feel more certain about their value (confidence). Previous research has shown that more attention and higher confidence often lead to choosing an option. However, this raises a question: should we always prefer options we know more about? Intuitively, no. The value we expect an option to have should critically influence how additional information affects our preferences. This paper aims to formalize this intuition and test its implications, bridging the gap between simple evidence accumulation models and the complex reality of human decision-making where prior expectations and uncertainty play a crucial role.
Methods
The core of this research is a novel Bayesian sequential sampling model. This model extends traditional approaches by incorporating "expectation-dependent updating" and "precision-weighted updating." * Expectation-dependent updating: The model assumes that how an option's estimated value is updated by new information depends on whether that information aligns with prior expectations. If an option is expected to be good and the new information confirms this, the estimate increases. If it's expected to be good but the information suggests otherwise, the estimate decreases. * Precision-weighted updating: This refers to how confidence influences the integration of new information. Higher confidence means new samples (information) are given more weight in updating the estimated value.
The model uses Bayesian inference to update an agent's belief about an item's true value based on noisy samples. Attention and confidence are modeled as factors that influence the precision (inverse variance) of these samples. The authors then tested the model's predictions using two experiments where they manipulated item presentation duration (as a proxy for attention) and collected confidence ratings.
Datasets
The study utilized consumer item datasets for their experiments. Participants were first familiarized with items, then rated them individually, and finally made choices between pairs of items. The specific datasets themselves are not described as novel benchmarks but rather as stimuli for the experimental paradigm.
Key Insights and Ideas
- Context matters: Our preferences aren't just about the inherent value of an option, but also about our expectations of what values are typical in a given context.
- Attention and confidence are double-edged swords: They don't just increase the likelihood of choosing an item; their effect is reference-dependent. They boost choices for options that are better than expected but decrease choices for options that are worse than expected.
- Bayesian updating as a flexible strategy: The model suggests that humans use Bayesian principles to dynamically adjust their decision-making strategies based on available information and their confidence in it.
- Metacognition is key: The ability to assess one's own knowledge (confidence) is crucial for adaptive decision-making, allowing for more efficient information processing.
Key Results
The model's predictions were largely confirmed by the experimental data:
- Attention: Increased presentation duration (attention) led to higher choice probability for items with values above the average, but lower probability for items below the average. This effect was more pronounced when the overall value of the choice set was high.
- Confidence: Similar to attention, higher confidence in an item's value also increased its choice probability when it was better than average and decreased it when it was worse than average.
- Speed and Consistency: Participants made choices faster and more consistently when their overall confidence in the items' values was higher.
- Model Fit: The Bayesian updating model provided a better account of the observed behavioral patterns than alternative models that did not incorporate expectation-dependent or precision-weighted updating.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Parameter Sensitivity: While the core qualitative predictions were robust, some quantitative predictions (like response times) were more sensitive to parameter choices.
- Experimental Manipulation of Attention: The study manipulated presentation duration rather than allowing free gaze. Future work should compare these approaches to understand the role of active information search.
- Biased Priors: The model's predictions can be influenced by the initial prior expectations. Future research could explore how these priors are formed and updated in different contexts.
- Beyond Simple Choice: The current model focuses on binary choices. Extending it to more complex decision scenarios with multiple alternatives and different types of information is a key next step.
- Neural Correlates: The authors suggest their model can inform predictions about neural activity related to information regulation in decision-making, encouraging future neuroscientific investigations.
- Metacognitive Mismatches: The study touched upon over/underconfidence, suggesting future work to explore how systematic biases in confidence affect decision-making speed and consistency.
Summary updated