Abstract
This paper establishes a uniform error rate for the Bahadur representation of local polynomial quantile regression estimators. Unlike previous results that focus on fixed data-generating processes (DGPs), our findings hold uniformly over a range of quantiles, evaluation points, and a broad class of probability distributions. This "triple uniformity" is essential for robust hypothesis testing, particularly in the context of monotonicity tests where the null hypothesis is a complex composite. We further extend these results to bootstrap estimators and demonstrate the practical utility of our theoretical framework through Monte Carlo experiments.
Introduction
Quantile regression is a powerful tool for understanding the conditional distribution of a response variable beyond its mean. While local polynomial estimation offers a flexible, nonparametric approach to this task, performing reliable inference—such as testing for monotonicity—requires a deep understanding of the estimator's asymptotic behavior.
Traditional "pointwise" asymptotics, which assume a fixed DGP, often fail to provide accurate approximations in finite samples. This is especially problematic when the true function lies near the boundary of the null hypothesis (e.g., a function that is nearly flat). In such cases, standard theory can lead to tests that over-reject or under-reject. This paper addresses this gap by developing a uniform asymptotic theory that ensures reliability across a wide range of potential distributions.
The Bahadur Representation and Triple Uniformity
The core of our theoretical contribution is a refined Bahadur representation. A Bahadur representation "linearizes" a complex, non-linear estimator, expressing it as a sum of independent random variables plus a small remainder term. If this remainder is proven to be uniformly small, researchers can use well-understood linear theory to perform inference.
We establish that this error rate holds uniformly over a "triple" of parameters:
- Quantiles ($\tau$): The specific percentiles being estimated.
- Evaluation Points ($x$): The values of the regressors.
- Probability Distributions ($P$): The underlying data-generating process.
By employing convexity arguments from Pollard (1991) and maximal inequalities from Massart (2007), we achieve a tighter error bound than previous benchmarks. For a one-dimensional regressor, our established rate is $O_P[\sqrt{\log n} / (n^{1/4} h^{1/4})]$, which improves upon earlier literature by removing an extra $\log^{1/4} n$ factor. These refinements are vital for the theoretical validity of subsequent statistical tests.
Bootstrap Validity
In many practical scenarios, the asymptotic distribution of a test statistic is too complex for direct use. Researchers frequently turn to the bootstrap to calculate critical values. We prove that the same uniform error rates apply to bootstrap local polynomial estimators. This provides the necessary theoretical justification for using bootstrap-based inference in nonparametric quantile regression.
Application: Testing Monotonicity
To illustrate the utility of these results, we apply them to the problem of testing whether a quantile regression function is monotonic (e.g., always increasing). This is a challenging problem because the behavior of the test depends heavily on the "contact set"—the regions where the function's derivative is zero.
Pointwise theory struggles to handle functions that are flat or nearly flat. Our uniform asymptotics bridge this gap, ensuring that the monotonicity test remains "honest." This means the test maintains its nominal size (Type I error control) even when the underlying function is at the very edge of the null hypothesis.
Monte Carlo Experiments
We evaluated the performance of the proposed monotonicity test using synthetic data with a sample size of $n=200$.
- Size Control: Under the null hypothesis of a flat function ($Y=U$), the test's rejection rates were very close to nominal levels. For instance, at a 5% significance level, the test rejected the null approximately 5% of the time, confirming the reliability of the uniform theory.
- Power: The test demonstrated high power against various non-monotonic alternatives (e.g., quadratic curves and "wiggly" functions), often achieving rejection rates between 90% and 100%.
- Bandwidth Sensitivity: The experiments highlighted that power can be sensitive to the choice of bandwidth ($h$), a common characteristic in nonparametric statistics that warrants careful consideration in practice.
Conclusion and Future Work
This work provides a robust theoretical foundation for nonparametric quantile regression. By establishing triple uniformity, we enable more reliable statistical inference for complex properties like monotonicity. While this paper focuses on fixed bandwidth sequences and scalar regressors, future research could extend these results to allow for data-driven bandwidth selection and explore the performance of these tests in higher-dimensional settings where the "curse of dimensionality" is more pronounced.
Summary updated