Dynamic and Stochastic Rational Behavior

Nail Kashaev, Victor H. Aguiar, Martin Plávala, Charles GauthierarXiv 2023 · DOI · arXiv

TL;DR

This paper introduces the Dynamic Random Utility Model (DRUM), a new framework for understanding consumer choice behavior that accounts for preferences changing over time. DRUM offers a more flexible and empirically relevant alternative to traditional static models, particularly when analyzing panel data, and is shown to better explain real-world behavior where static rationality often fails.

What Problem Does This Paper Address?

The paper tackles a fundamental challenge in economics: how to model and test rational decision-making when preferences aren't fixed. The standard model, based on utility maximization with stable preferences (Samuelson-Afriat framework), has faced increasing empirical scrutiny due to evidence of preference instability over time. While Random Utility Models (RUM) offer a way to incorporate stochasticity, they typically analyze choices at a single point in time or across a population without explicitly modeling dynamic preference changes. This leaves a gap in understanding how rational choices evolve over time, especially when individuals face sequences of decisions. DRUM aims to bridge this gap by allowing for preferences to vary dynamically while still maintaining a notion of rationality.

Methods

The core of the paper is the development of a revealed preference characterization for DRUM. This involves:

  • Dynamic Random Utility Model (DRUM): The central model where decision-makers randomly draw a utility function in each period and maximize it subject to a menu. This allows for unrestricted time correlation and cross-section heterogeneity in preferences.
  • Revealed Preference Characterization: The authors derive testable conditions for DRUM using panel data (choices and menus over time). This involves extending static RUM characterizations to a dynamic setting.
  • Link to Static Models: A key methodological contribution is showing how DRUM can be represented as a finite mixture of deterministic behaviors, which in turn are products of static rationalizable behaviors. This connection allows them to translate existing RUM characterizations into the dynamic DRUM framework.
  • Axiomatic Characterization: The paper develops axiomatic characterizations of DRUM using linear inequalities, including dynamic triangle conditions and D-monotonicity, which provide behavioral restrictions on choice probabilities over time.
  • Generalization of Weyl-Minkowski Theorem: The authors introduce a generalization of this theorem to the economic context, which is crucial for deriving recursive characterizations of DRUM.

Datasets

The paper uses an experimental dataset collected from Amazon Mechanical Turk for its empirical application. This dataset consists of choices made by 2135 decision-makers (DMs) over binary comparisons of lotteries.

Key Insights and Ideas

  • Preferences Aren't Static: The most significant insight is that real-world decision-making often involves dynamic preference shifts, which static models fail to capture. DRUM provides a framework to analyze this dynamic aspect of rationality.
  • Panel Data is Crucial: Analyzing choices over time (panel data) is essential for uncovering the richer implications of DRUM compared to static RUM, which primarily relies on cross-sectional data.
  • DRUM is More Empirically Bitey: DRUM imposes stronger behavioral restrictions than static RUM, especially when analyzing longitudinal data, leading to more informative tests of rational behavior.
  • Bridging the Gap: DRUM effectively bridges the gap between the strictness of the Samuelson-Afriat framework (constant preferences) and the limitations of static RUM (ignoring time dynamics).
  • Computational Tractability: Despite its complexity, the paper demonstrates that DRUM can be characterized and tested using mathematical tools like Kronecker products and generalizations of the Weyl-Minkowski theorem, leading to computational gains.

Key Results

  • Empirical Application: In their application to binary lottery choices, the authors find that while the static utility maximization model fails to explain population behavior (8% of DMs are inconsistent), DRUM cannot be rejected as an explanation for the data.
  • DRUM vs. Static Models: The paper shows that consistency with static RUM for marginal and conditional distributions is not sufficient for DRUM consistency. DRUM imposes additional restrictions on the joint distribution of choices over time.
  • D-Monotonicity: A new behavioral condition, D-monotonicity, is introduced as a key restriction for DRUM, generalizing the Weak Axiom of Stochastic Revealed Preference to a dynamic setting.
  • Power of the Test: Monte Carlo simulations demonstrate that the proposed statistical test for DRUM has good size and power properties in finite samples, suggesting that the empirical findings are not due to a lack of statistical power.

Limitations and Future Work

  • Uniqueness Assumption: The paper notes that the characterization of DRUM can be complex when uniqueness fails (i.e., when multiple preference profiles are consistent with the data). This is an area where DRUM imposes additional restrictions beyond static models.
  • Computational Complexity: While progress is made, the computation of certain matrices (like $A_T$) can grow exponentially with time periods ($T$), posing computational challenges for very long time horizons.
  • Parametric vs. Nonparametric: The paper focuses on nonparametric approaches. Future work could explore parametric versions of DRUM or integrate DRUM with existing parametric models.
  • Endogenous Expenditure: The paper relaxes the exogeneity of budgets to incorporate income uncertainty and consumption smoothing, suggesting avenues for further modeling of dynamic decision-making.
  • Relationship with DDC: The paper notes that Dynamic Discrete Choice (DDC) models are nested by DRUM in their setup, but a full comparison is limited due to differences in assumptions (e.g., logit shocks in DDC). Further exploration of this relationship is suggested.

Summary updated

Abstract

The (static) utility maximization model of Afriat (1967), which is the standard in analysing choice behavior, is under scrutiny. We propose the Dynamic Random Utility Model (DRUM) that is more flexible than the framework of Afriat (1967) and more informative than the static Random Utility Model (RUM) framework of McFadden and Richter (1990). Under DRUM, each decision-maker randomly draws a utility function in each period and maximizes it subject to a menu. DRUM allows for unrestricted time correlation and cross-section heterogeneity in preferences. We characterize DRUM for situations when panel data on choices and menus are available. DRUM is linked to a finite mixture of deterministic behaviors that can be represented as a product of static rationalizable behaviors. This link allows us to convert the characterizations of the static RUM to its dynamic form. In an application, we find that although the static utility maximization model fails to explain population behavior, DRUM can explain it.
BibTeX
@misc{kashaev2023dynamic,
  author = {Nail Kashaev and Victor H. Aguiar and Martin Plávala and Charles Gauthier},
  title = {{Dynamic and Stochastic Rational Behavior}},
  year = {2023},
  eprint = {2302.04417},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2302.04417},
}
Figure