TL;DR
This paper proposes that our "reference points" for decision-making aren't static but are dynamically learned through Bayesian structure learning. They demonstrate that abrupt changes in outcomes lead to new reference points, altering risk preferences, while gradual changes lead to continuous updates of a single reference point. This offers a novel computational framework for understanding how context shapes our choices under uncertainty.
What Problem Does This Paper Address?
Decision-making under risk is heavily influenced by reference points – the subjective benchmarks against which outcomes are evaluated as gains or losses. While current theories often assume these reference points are based on expectations, a critical question remains: where do these expectations come from, and how do they adapt? Do we rely on a single, global expectation, or do we maintain multiple, context-specific ones? This paper tackles this by investigating how people form and update these reference points, particularly when faced with changing environmental statistics. They aim to bridge the gap between economic models of reference dependence and psychological theories of learning and context.
Methods
The authors propose a Bayesian structure learning model to explain how reference points are formed. This model assumes that agents infer latent "contexts" or "structures" within the environment, with each context having its own expected outcome, serving as a reference point. When the statistical properties of outcomes change, the model can either update the existing reference point (for gradual changes) or infer a new context and thus a new reference point (for abrupt changes).
They tested this model through two experiments involving a binary choice task where participants chose between a certain lottery and a risky lottery with the same expected value. Experiment 1 involved randomized blocks of trials with different expected value distributions (A and B), while Experiment 2 presented these distributions in a monotonically increasing order. They used mixed-effects logistic regression to analyze participants' choices and compare the fit of their full structure learning model (allowing multiple reference points) against a restricted model (enforcing a single reference point).
Datasets
The study used data collected from human participants recruited via Amazon Mechanical Turk (MTurk). No new benchmarks or datasets were introduced; the focus was on analyzing behavioral data from controlled experimental tasks.
Key Insights and Ideas
- Structure Learning as the Source of Reference Points: The core idea is that our brains actively infer underlying structures in the environment to form reference points, rather than passively updating a single expectation.
- Abrupt vs. Gradual Change: Dramatic shifts in outcome distributions trigger the formation of new reference points, leading to non-monotonic changes in risk preferences. In contrast, gradual changes lead to the continuous updating of a single reference point, resulting in monotonic shifts in preferences.
- Context-Dependent Risk Preferences: The model explains how the same objective outcome can be perceived differently (as a gain or loss) depending on the inferred context, leading to context-specific risk attitudes.
- Non-Monotonicity in Risk Preferences: A key prediction, confirmed experimentally, is that risk preferences can change non-monotonically with changes in the environment, a phenomenon not easily explained by simpler models.
Key Results
- Experiment 1 (Randomized Blocks): Participants showed non-monotonic changes in their probability of choosing the risky option for distribution A as the mean of distribution B varied. This suggests they formed separate reference points for A and B trials when the distributions were sufficiently distinct.
- Experiment 2 (Monotonic Blocks): When the mean of distribution B changed gradually, participants' choices showed monotonic changes, consistent with a single, continuously updated reference point.
- Model Fit: The full structure learning model (allowing multiple reference points) provided a significantly better fit to the data than the restricted model (forcing a single reference point), particularly for Experiment 1.
- Individual Differences: The study accounted for individual differences in risk preferences (positive vs. negative value-dependent risk preferences), showing consistent patterns across these groups.
Limitations and Future Work
- Generalizability of Utility Function: While the study used a quadratic utility function for simplicity, they note that the same structural inferences could be achieved with other functions, like Prospect Theory's, suggesting the core finding is about reference point formation, not the specific utility mapping.
- Neural Correlates: The authors suggest that future research could use neuroimaging (fMRI) to find neural correlates of reference points and investigate how structure learning influences them.
- Broader Applications: They propose that the same logic of structure learning could apply to other economic variables like probabilities and delays, and could be extended to multi-attribute decisions.
- Rational Inattention: The work draws inspiration from rational inattention theories, suggesting a path for integrating cognitive constraints into economic models.
Summary updated