[Paper Review] Quantile Estimation of A general Single-Index Model

Efang Kong, Yingcun Xia|ArXiv.org|Mar 17, 2008

Statistical Methods and Inference12 references3 citations

TL;DR

This paper proposes a novel estimation method for the index parameter in a general single-index quantile regression model, using iterative local polynomial fitting with kernel-weighted M-estimation. The method achieves uniform strong consistency and asymptotic normality under weak dependence assumptions, enabling robust inference in conditional heteroscedastic models like ARCH(p).

ABSTRACT

The single-index model is one of the most popular semiparametric models in Econometrics. In this paper, we define a quantile regression single-index model, which includes the single-index structure for conditional mean and for conditional variance.

Motivation & Objective

• To develop a robust estimation procedure for the index parameter in a general single-index quantile regression model, particularly for models with conditional heteroscedasticity.
• To address the curse of dimensionality in nonparametric quantile regression by employing a single-index structure that reduces dimensionality.
• To establish uniform strong consistency and asymptotic normality of the proposed estimator under weak dependence (strongly mixing) conditions.
• To extend existing mean regression methods to quantile regression, enabling more complete statistical analysis of stochastic relationships.
• To provide a computationally feasible algorithm based on iterative linear quantile regression subproblems for practical implementation.

Proposed method

• Estimate the index parameter θ₀ via iterative minimization of a kernel-weighted sum of check loss functions over pairwise differences of covariates.
• Use local polynomial fitting with kernel K(θᵀXᵢⱼ/h) to weight observations based on proximity to a given point, where Xᵢⱼ = Xᵢ − Xⱼ.
• Decompose the optimization into two alternating steps: (1) for fixed θ, estimate local intercepts aⱼ and slopes bⱼ via linear quantile regression; (2) update θ by solving a new quantile regression problem on transformed variables.
• Standardize the final estimate as ǁθ̂ǁ⁻¹θ̂ to ensure unit norm, preserving the direction of the index vector.
• Leverage existing efficient algorithms for linear quantile regression (e.g., from Koenker, 2005) to solve subproblems efficiently.
• Establish theoretical properties using stochastic equicontinuity and U-statistic theory, under assumptions of strong mixing and bounded density.

Experimental results

Research questions

• RQ1 Can a consistent and asymptotically normal estimator be constructed for the index parameter in a general single-index quantile model under weak dependence?
• RQ2 How can the curse of dimensionality in nonparametric quantile regression be mitigated using a single-index structure?
• RQ3 What is the theoretical justification for the iterative kernel-weighted M-estimation procedure in terms of uniform strong consistency?
• RQ4 How does the proposed method perform in models with conditional heteroscedasticity, such as ARCH(p) processes?
• RQ5 What are the regularity conditions under which the estimator achieves uniform convergence and asymptotic normality?

Key findings

• The proposed estimator achieves uniform strong consistency for the index parameter θ₀ under weak dependence (strongly mixing) conditions.
• The estimator is asymptotically normal, with convergence rate n⁻¹/², under regularity conditions including bounded density and smoothness of the conditional quantile function.
• The method successfully handles conditional heteroscedasticity, as demonstrated in the ARCH(p) model where mₜ(x) = g(θ₀ᵀX)Qₜ(ε), with Qₜ(ε) the τ-th quantile of the error.
• Theoretical justification relies on showing that the Hessian matrix (S₂ + θ₀θ₀ᵀ)⁻¹(Ω₀ + θ₀θ₀ᵀ) has all eigenvalues less than 1 and positive definite, ensuring convergence.
• The iterative algorithm converges to a unique solution under the assumption that θ₀ is the only eigenvector of S₂ and Ω₀ corresponding to eigenvalue 0.
• Numerical studies confirm the method’s robustness and finite-sample performance, particularly in median regression and ARCH-type models.