Fairness constraint in Structural Econometrics and Application to fair estimation using Instrumental Variables

Samuele Centorrino, Jean-Pierre Florens, Jean-Michel LoubesarXiv 2022 · arXiv

Introduction

In the rapidly evolving field of algorithmic fairness, most research assumes a direct and observable relationship between inputs and outputs. However, in econometrics and social sciences, data often suffers from endogeneity—a condition where explanatory variables are correlated with unobserved factors (error terms). Standard fairness interventions, such as simply removing sensitive attributes, often fail in these contexts because bias "leaks" through these hidden causal correlations. This paper bridges the gap by integrating fairness constraints directly into structural econometric models, specifically focusing on Instrumental Variables (IV) estimation.

Theoretical Framework: Fairness as a Linear Operator

The authors formalize fairness by treating it as a linear operator $F$ acting on a functional space $\mathcal{E}$. In this framework, a model $\phi$ is considered "fair" if it belongs to the null space of the operator, denoted as $\mathcal{N}(F)$. This allows for the mathematical enforcement of common fairness definitions, such as Statistical Parity or Irrelevance in Prediction, within the rigorous bounds of structural equations.

Approaches to Fair Estimation

The paper explores three distinct methodologies for achieving fairness in IV models:

  1. Projection onto Fairness: This two-step approach involves first calculating a standard, potentially biased IV estimator and then projecting that solution onto the fair subspace $\mathcal{N}(F)$.
  2. Direct Fair Solution: Instead of post-hoc projection, this method solves the structural IV functional equation directly within the restricted "fair" subspace. The authors demonstrate that this is not mathematically equivalent to the projection method, as it accounts for the geometry of the IV operator $K$ from the outset.
  3. Approximate Fairness (Penalization): Recognizing that strict fairness may significantly degrade predictive accuracy, the authors introduce a Tikhonov-style regularization. By adding a penalty term $\rho \|F\phi\|^2$ to the objective function, researchers can tune the trade-off between accuracy and equity.
Geometric representation of the fair estimation problem in functional spaces.

As shown in the figure above, the relationship between the unconstrained solution $\phi$ in space $\mathcal{E}$ and the fair solution $\phi_F$ in the null space $\mathcal{N}(F)$ is mediated by the operator $K$, which maps these functions to the observation space $\mathcal{G}$. The distinction between projecting the result in $\mathcal{G}$ versus solving within the constrained subspace in $\mathcal{E}$ is a central contribution of the work.

Key Insights and Results

A critical finding of the research is the formal quantification of the "Price of Fairness." This represents the loss in accuracy incurred when moving from an unconstrained model to a fair one, essentially measured as the distance between the true underlying function and its best fair representation.

Furthermore, the paper provides formal proofs for the consistency and convergence rates of these estimators. For the approximate fairness model, the convergence rate is determined by a combination of the smoothing parameter $\alpha$ and the fairness penalty $\rho$. Simulation results using varying coefficient models demonstrate that the penalized estimator effectively navigates the middle ground, gradually shifting predicted outcome distributions until group parity is achieved.

Conclusion and Future Directions

This work provides a robust mathematical foundation for fairness in structural econometrics. By accounting for endogeneity, it offers a more reliable path for equitable decision-making in complex social systems. Future research may extend this framework to handle non-linear fairness constraints and individual fairness definitions, which remain challenging within the current linear inverse problem paradigm.

Summary updated

Abstract

A supervised machine learning algorithm determines a model from a learning sample that will be used to predict new observations. To this end, it aggregates individual characteristics of the observations of the learning sample. But this information aggregation does not consider any potential selection on unobservables and any status-quo biases which may be contained in the training sample. The latter bias has raised concerns around the so-called \textit{fairness} of machine learning algorithms, especially towards disadvantaged groups. In this chapter, we review the issue of fairness in machine learning through the lenses of structural econometrics models in which the unknown index is the solution of a functional equation and issues of endogeneity are explicitly accounted for. We model fairness as a linear operator whose null space contains the set of strictly {\it fair} indexes. A {\it fair} solution is obtained by projecting the unconstrained index into the null space of this operator or by directly finding the closest solution of the functional equation into this null space. We also acknowledge that policymakers may incur a cost when moving away from the status quo. Achieving \textit{approximate fairness} is obtained by introducing a fairness penalty in the learning procedure and balancing more or less heavily the influence between the status quo and a full fair solution.
BibTeX
@misc{centorrino2022fairness,
  author = {Samuele Centorrino and Jean-Pierre Florens and Jean-Michel Loubes},
  title = {{Fairness constraint in Structural Econometrics and Application to fair estimation using Instrumental Variables}},
  year = {2022},
  eprint = {2202.08977},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.08977},
}
Figure