Vector quantile regression beyond correct specification

Guillaume Carlier, Alfred Galichon, Victor ChernozhukovarXiv 2016 · DOI · arXiv

Abstract

This paper studies vector quantile regression (VQR), which is a way to model the dependence of a random vector of interest with respect to a vector of explanatory variables so to capture the whole conditional distribution, and not only the conditional mean. The problem of vector quantile regression is formulated as an optimal transport problem subject to an additional mean-independence condition. This paper provides a new set of results on VQR beyond the case with correct specification which had been the focus of previous work. First, we show that even under misspecification, the VQR problem still has a solution which provides a general representation of the conditional dependence between random vectors. Second, we provide a detailed comparison with the classical approach of Koenker and Bassett in the case when the dependent variable is univariate and we show that in that case, VQR is equivalent to classical quantile regression with an additional monotonicity constraint.

Introduction

Standard linear regression is a foundational tool in econometrics and statistics, primarily used to estimate the conditional mean of a distribution. However, mean-based models often fail to capture the behavior of the "tails"—the extreme values that are often of the greatest interest in fields like finance or risk management. Quantile Regression (QR), introduced by Koenker and Bassett, addresses this for scalar variables by modeling specific percentiles of the conditional distribution.

When the dependent variable is a vector (multivariate), the problem becomes significantly more complex. Unlike the real line, higher-dimensional spaces have no natural, canonical ordering, making the concept of a "quantile" difficult to define. Vector Quantile Regression (VQR) was developed to solve this by utilizing the mathematics of Optimal Transport (OT). While initial VQR research assumed "correct specification"—the idea that the model perfectly matches the underlying data generation process—this paper extends the theory to the more realistic case of misspecified models.

Methodology: Optimal Transport and Duality

The authors frame VQR as a correlation maximization problem. The goal is to find a random vector $U$ (typically uniformly distributed on a unit cube) that is maximally correlated with the dependent variable $Y$, while ensuring that the explanatory variables $X$ are mean-independent of $U$ (i.e., $E[X|U] = E[X]$).

To achieve this, the paper relies on several key mathematical frameworks:

  1. Brenier’s Theorem: This theorem from OT theory suggests that the optimal mapping between distributions is the gradient of a convex function. In VQR, this gradient serves as the multivariate analogue of a monotone quantile function.
  2. Variational Approach: The problem is treated as a search for a potential function that optimizes the relationship between the vectors.
  3. Convex Duality: To address misspecification, the authors utilize Kantorovich duality. By applying Komlos’ Theorem, they prove that a solution to the dual problem exists even when the linear model is an imperfect fit for the data. This allows the VQR solution to be interpreted as a "convex envelope"—the best possible convex approximation of the actual relationship.

Key Insights

Robustness to Misspecification

The primary contribution of this work is the proof that VQR does not "break" when the model is misspecified. Instead of failing, the framework provides a general representation of the dependence between $X$ and $Y$. It effectively finds the closest monotone representation of the data, ensuring that the model remains theoretically grounded and interpretable even in the presence of non-linearities that the modeler did not account for.

Solving the Quantile Crossing Problem

In classical univariate quantile regression, a common issue is "quantile crossing," where estimated quantile lines (e.g., the 10th and 20th percentiles) intersect. This is logically inconsistent, as it implies a value can be in a lower percentile and a higher percentile simultaneously. The authors demonstrate that VQR inherently prevents this by imposing a global monotonicity constraint, ensuring that the estimated quantiles are always consistent.

The Univariate Connection

The paper establishes a formal bridge between modern multivariate OT theory and classical 1D methods. The authors prove (Theorem 4.9) that when the dependent variable is a scalar ($d=1$), VQR is mathematically equivalent to the Koenker and Bassett method, provided that a monotonicity constraint is enforced. This result validates VQR as a true generalization of classical quantile regression.

Conclusion and Future Directions

This research provides the necessary theoretical foundation for applying Vector Quantile Regression to real-world datasets where model misspecification is a certainty. By extending Kantorovich duality to include mean-independence constraints, the authors have ensured that VQR is computationally and mathematically viable.

However, limitations remain. The current proofs assume bounded support for the variables, and the computational cost of solving the dual problems via linear programming can be high for very large datasets. Future work may focus on developing high-speed algorithms for high-dimensional data and extending these proofs to heavy-tailed distributions. Ultimately, this work paves the way for more robust econometric tools that can capture the full complexity of multivariate relationships without being hindered by the rigid assumptions of correct specification.

Summary updated

Abstract

This paper studies vector quantile regression (VQR), which is a way to model the dependence of a random vector of interest with respect to a vector of explanatory variables so to capture the whole conditional distribution, and not only the conditional mean. The problem of vector quantile regression is formulated as an optimal transport problem subject to an additional mean-independence condition. This paper provides a new set of results on VQR beyond the case with correct specification which had been the focus of previous work. First, we show that even under misspecification, the VQR problem still has a solution which provides a general representation of the conditional dependence between random vectors. Second, we provide a detailed comparison with the classical approach of Koenker and Bassett in the case when the dependent variable is univariate and we show that in that case, VQR is equivalent to classical quantile regression with an additional monotonicity constraint.
BibTeX
@misc{carlier2016vector,
  author = {Guillaume Carlier and Alfred Galichon and Victor Chernozhukov},
  title = {{Vector quantile regression beyond correct specification}},
  year = {2016},
  eprint = {1610.06833},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1610.06833},
}
Figure