Dynamic Modeling and Equilibria in Fair Decision Making

Joshua Williams, J. Zico KolterarXiv 2019 · OpenReview

TL;DR

This paper introduces a novel continuous model for population dynamics in fair decision-making, using the Beta distribution to represent repayment probabilities. It demonstrates that while unconstrained optimal policies can lead to the "Matthew Effect" (widening inequality), common fairness constraints can ensure group convergence, even in the face of estimation errors, under specific conditions.

What Problem Does This Paper Address?

Automated decision-making systems, while aiming for efficiency, can inadvertently perpetuate or even exacerbate societal inequalities. Traditional fairness constraints (like demographic parity or equality of opportunity) often focus on immediate outcomes and can be incompatible with each other. Furthermore, these constraints are typically evaluated in static settings, failing to capture the long-term, dynamic impact of decisions on different population groups. This paper addresses the critical gap of understanding how fairness constraints play out over time in a continuous population state, particularly when dealing with imperfect information and the potential for "Matthew Effect" scenarios where advantages and disadvantages compound.

Methods

The authors propose a novel approach by modeling the population's propensity to repay loans using a continuous Beta distribution. This allows for a more nuanced representation of individual differences compared to discrete models. They then introduce a population dynamics model where loan granting decisions influence future repayment probabilities. Key to their method is:

  • Beta Distribution for Population State: Representing the distribution of repayment probabilities within a group using the Beta distribution, parameterized by mean ($\mu$) and shape ($c$).
  • Dynamic Update Function: A mathematical model describing how the mean repayment probability ($\mu$) of a group evolves over time based on loan granting decisions and their outcomes. This function incorporates parameters for loan repayment success, loan denial impact, and a discount factor for future rewards.
  • Fairness Constraints as Parameterizations: Showing how common fairness notions (demographic parity, equality of opportunity, etc.) can be expressed as specific parameterizations of a general class of thresholding policies within the Beta distribution framework.
  • Analysis of Equilibria: Investigating the long-term stable states (equilibria) of these dynamic systems under various policy constraints, including unconstrained optimal policies and fair policies.
  • Modeling Misestimation: Extending the model to incorporate uncertainty in estimated repayment probabilities, analyzing how different levels of misestimation affect group convergence.

Datasets

The paper uses a publicly available dataset of TransUnion TransRisk scores and associated credit data. This data is processed to derive histograms of repayment probabilities for different racial groups (Black, Non-Hispanic White, Hispanic, Asian), which are then used to fit Beta distributions for simulation.

Key Insights and Ideas

  • The "Matthew Effect" is Real and Dynamic: Even with "optimal" unconstrained policies, the system can naturally lead to widening inequality, where initially advantaged groups see their well-being increase while disadvantaged groups stagnate or decline. This is a dynamic, compounding effect, not just a static snapshot.
  • Fairness Constraints Can Be Powerful Equalizers: Counterintuitively, many common fairness constraints, when applied within this dynamic model, force different groups to converge to the same equilibrium distribution. This suggests that fairness interventions, when properly designed, can mitigate long-term inequality.
  • Misestimation is a Major Hurdle for Fairness: When repayment probabilities are inaccurately estimated (a common issue, especially for minority groups), even fair policies may fail to achieve group convergence. Differing levels of misestimation between groups can prevent them from reaching parity, a stark contrast to some prior models.
  • Continuous Modeling Offers Nuance: Using a continuous Beta distribution provides a richer representation of population states than discrete models, allowing for a more accurate capture of dynamic effects and the impact of various fairness policies.

Key Results

  • Unconstrained Optimal Policies: Lead to "Matthew Effect" bifurcations, where group means diverge significantly.
  • Fair Policies (under correct estimation): Ensure convergence of group means to a common equilibrium point, suggesting that fairness constraints can lead to long-term parity.
  • Fair Policies (under misestimation): If groups have different levels of misestimation, convergence is generally not achieved, even with fair policies. This highlights the critical importance of accurate estimation for achieving equitable outcomes.
  • Simulations: Real-world credit data simulations demonstrate these dynamics, showing how unconstrained policies lead to divergence, while demographic parity and equality of opportunity policies lead to convergence.

Limitations and Future Work

  • Model Specificity: The authors acknowledge that their model, while intuitive, is a simplification and may not perfectly capture all real-world complexities of loan granting or other decision-making scenarios. The dynamics are based on specific assumptions about how loans impact future well-being.
  • Generalizability of Dynamics: Future work could explore alternative population dynamics models beyond the Beta distribution and the specific update function used.
  • Stereotypes and Misestimation: While the paper analyzes misestimation, a deeper exploration of how stereotypes statistically affect outcomes and how to mitigate these effects is suggested as an important avenue.
  • Defining "Fairness": The paper highlights that "fair treatment" is not universally defined, and future work could explore the impact of more abstract notions of fairness.
  • Data Requirements: The model's reliance on accurate estimation of group distributions underscores the need for robust data collection and modeling techniques, especially for underrepresented groups.

Summary updated

Abstract

Recent studies on fairness in automated decision making systems have both investigated the potential future impact of these decisions on the population at large, and emphasized that imposing ''typical'' fairness constraints such as demographic parity or equality of opportunity does not guarantee a benefit to disadvantaged groups. However, these previous studies have focused on either simple one-step cost/benefit criteria, or on discrete underlying state spaces. In this work, we first propose a natural continuous representation of population state, governed by the Beta distribution, using a loan granting setting as a running example. Next, we apply a model of population dynamics under lending decisions, and show that when conditional payback probabilities are estimated correctly 1) ``optimal'' behavior by lenders can lead to ''Matthew Effect'' bifurcations (i.e., ''the rich get richer and the poor get poorer''), but that 2) many common fairness constraints on the allowable policies cause groups to converge to the same equilibrium point. Last, we contrast our results in the case of misspecified conditional probability estimates with prior work, and show that for this model, different levels of group misestimation guarantees that even fair policies lead to bifurcations. We illustrate some of the modeling conclusions on real data from credit scoring.
BibTeX
@misc{williams2019dynamic,
  author = {Joshua Williams and J. Zico Kolter},
  title = {{Dynamic Modeling and Equilibria in Fair Decision Making}},
  year = {2019},
  eprint = {1911.06837},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.06837},
}
Figure