Magnitude-sensitive preference formation

Nisheeth Srivastava, Ed Vul, Paul R. SchraterNeurIPS 2014

Introduction

In classical economics, human decision-making is often modeled using an internal "utility function"—a theoretical mechanism that assigns a fixed, absolute numerical value to every object or outcome. However, psychological and neurological data consistently suggest that the human brain does not operate this way. Instead, human valuations are notoriously "labile." They fluctuate based on anchors, the set of available options, and even the initial endowment of money a person holds. Furthermore, neurons in the brain often exhibit "comparative coding," responding to whether one option is relatively better than another rather than calculating an absolute, isolated value.

This paper addresses the significant gulf between psychological reality (relative, context-dependent choices) and traditional economic modeling (absolute utility). The authors propose a novel theory of preference formation that abandons the traditional concept of utility functions in favor of a Bayesian inference model. By treating money as a magnitude label and learning preferences based on the statistical distribution of these labels across different contexts, the authors successfully replicate complex economic behaviors—like supply-demand curves and risky lottery choices—using only relative comparisons and environmental statistics.

A Bayesian Framework for Preference

The authors build upon a Bayesian preference learning framework. Instead of estimating a hidden, absolute utility function $U(x)$, the agent acts as an ideal observer learning to answer a simpler question: "Which option among the ones offered is best?"

A visual representation of the Bayesian preference learning framework, illustrating how an agent infers the best options across different contexts based on environmental statistics.

Key components of their approach include:

  • Latent Contexts: The model assumes that observations are clustered into distinct contexts (e.g., "fine dining" vs. "fast food"). Preferences are learned and applied within these specific contexts.
  • Magnitude Marginalization: To value an item, the agent marginalizes over all possible monetary labels and contexts it has previously encountered.
  • Environmental Statistics: A crucial assumption is that money labels in the real world follow a Pareto distribution (a power-law distribution where small amounts are common and large amounts are rare). This assumption is rooted in the natural scarcity of essential resources.
  • Arithmetic Ordering: The model is constrained to respect the basic rule that more money is better than less money, allowing it to transition from nominal labels to ordered magnitudes.

Validation Against Behavioral Data

Rather than introducing a new experimental dataset, the authors validate their model against several landmark behavioral studies to demonstrate its explanatory power:

  1. Pain Buy-out Study (Vlaev, 2011): The model accounts for data where subjects paid different amounts to avoid electric shocks based entirely on their initial monetary endowment.
  2. Wine Pricing Study (Plassmann, 2008): The model explains fMRI and behavioral data showing that higher price tags artificially increase the reported (and neurologically experienced) pleasantness of wine.
  3. Risky Choice Panel: The authors tested the model against a comprehensive panel of 35 different risky choice experiments involving various probabilities and expected values.
Reconstruction of the Wine Pricing Study (Plassmann, 2008) showing how price labels significantly alter reported liking compared to unlabelled options.

Key Insights and Results

By framing preferences as statistical inferences, the authors yield several profound insights into economic behavior:

  • Preferences are Inferred, Not Inherent: Preference isn't an inherent property of an object; it’s an inference based on a history of choices. We don't possess a hardwired preference for a $100 bottle of wine; rather, we infer its value based on how often we've seen similar items preferred in similar contexts.
  • Money as a Label: The model treats money not as a special "value" substance, but as a label with a natural ordering. This elegantly explains "price anchoring"—initial samples of these labels in a new context skew the Bayesian posterior, altering our perception of value.
  • The Origin of Risk Aversion: By assuming a Pareto distribution for money in the environment, the model naturally produces a concave "utility-like" curve. Risk aversion emerges not from a psychological quirk, but directly from the statistical structure of the world we live in.
  • Explaining "Anomalous" Goods: The model provides a unified explanation for Veblen goods (luxury items where demand rises with price) and Giffen goods (staples where demand rises with price) by analyzing how price changes affect the "familiarity" and "substitutability" terms in the Bayesian calculation.

In testing, the model successfully derived classical supply-demand curves without needing to manually tune utility parameters. Furthermore, the model’s predictions matched human performance across the 35 risky choice experiments. The "relative desirability" agents produced choice probabilities that were statistically indistinguishable from human data (fitting within 1 standard deviation for 34 out of 35 payoff-probability combinations). The simulation also showed that as "wealth" increases, the curvature of the inferred utility function changes, matching the empirically documented increase in relative risk aversion seen in wealthier individuals.

Limitations and Future Work

While powerful, the model has limitations. It currently assumes that the mapping of observations to latent contexts is already identified or easily clustered. In the real world, determining which "context" you are currently in can be highly ambiguous. Additionally, as the set of contexts grows, the marginalization required to calculate preferences becomes computationally expensive. The authors suggest "sparse sampling" as a biological workaround, though this remains to be fully formalized.

Finally, the authors note a "radical implication" of their theory: if you see a high-end laptop for an impossibly low price (e.g., $100), you might hesitate to buy it simply because it's an "unfamiliar price" for that context, triggering a low probability inference. This specific type of history-dependent hesitation remains an exciting avenue for future behavioral verification.

Summary updated

BibTeX
@inproceedings{srivastava2014magnitude,
  author = {Nisheeth Srivastava and Ed Vul and Paul R. Schrater},
  title = {{Magnitude-sensitive preference formation}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2014},
}

Loading related content...

Figure