Testing many constraints in possibly irregular models using incomplete U-statistics

Nils Sturma, Mathias Drton, Dennis Y.c. LeungJournal of the Royal Statistical Society Series B (Statistical Methodology) 2024 · DOI

TL;DR

This paper introduces a new statistical testing framework for complex hypotheses with many constraints, even when the underlying data distribution has "irregularities." The method uses incomplete U-statistics and Gaussian multiplier bootstrap to provide reliable results, outperforming traditional methods in challenging scenarios.

What Problem Does This Paper Address?

Statistical hypothesis testing often involves checking if a parameter falls within a specific set (e.g., equality or inequality constraints). This becomes particularly difficult when:

  1. Many Constraints: The number of constraints is large, potentially exceeding the number of data samples.
  2. Irregularities: The null hypothesis might have "singularities" or "irregular points" where standard statistical approximations break down, leading to invalid tests. This is common in polynomial hypotheses.

Traditional methods like likelihood ratio tests or Wald tests struggle with these issues, often failing to control error rates or provide accurate results. The authors aim to develop a robust testing methodology that handles both high-dimensional constraints and irregular settings.

Methods

The core of the proposed methodology lies in two key components:

  1. Incomplete U-Statistics: Instead of using "complete" U-statistics (which average over all possible combinations of data points, becoming computationally expensive with many constraints), the authors use "incomplete" U-statistics. These are formed by randomly sampling a subset of combinations. This approach offers computational advantages and is crucial for handling high-dimensional settings.
  2. Gaussian Multiplier Bootstrap: To derive critical values for the test statistic (which is based on the maximum of studentized incomplete U-statistics), the paper employs a Gaussian multiplier bootstrap. This technique approximates the distribution of the test statistic using random Gaussian variables, allowing for data-dependent critical values.

A key theoretical contribution is showing that this bootstrap approximation remains valid even under "mixed degeneracy," a scenario where some components of the U-statistic might be degenerate (have zero variance) while others are not. This is crucial for handling irregular hypotheses.

Datasets

The paper does not introduce new datasets. Instead, it uses simulated data to evaluate the performance of its proposed methodology. Specifically, it generates data from:

  • Gaussian latent tree models: Used to test goodness-of-fit for these models, which involve a large number of polynomial constraints.
  • Two-factor analysis models: Used to test minors, another application involving polynomial constraints.

The simulations are designed to specifically test the method's performance in both "regular" and "irregular" settings, often by choosing parameters close to algebraic singularities.

Key Insights and Ideas

  • Incomplete U-statistics as a Solution: The paper demonstrates that incomplete U-statistics are not just computationally efficient but also theoretically sound for testing complex hypotheses, especially in high dimensions.
  • Handling Irregularities is Crucial: The authors highlight that standard methods fail dramatically when dealing with irregular hypotheses. Their proposed method's ability to control type I error in these scenarios is a significant advancement.
  • Mixed Degeneracy is Key: The theoretical breakthrough lies in proving the validity of the bootstrap approximation under "mixed degeneracy," a more general condition than previously studied, which directly addresses the challenges posed by irregular hypotheses.
  • Bootstrap for Robustness: The Gaussian multiplier bootstrap provides a practical way to obtain critical values that are robust to the underlying data distribution and the presence of irregularities.
  • Polynomial Constraints are General: The methodology is shown to be broadly applicable to polynomial constraints, a common and important class of hypotheses in statistics.

Key Results

  • Asymptotically Valid Tests: The proposed testing methodology asymptotically controls the type I error rate, even in settings with many constraints and irregular hypotheses.
  • Improved Performance in Irregular Settings: Simulations show that the proposed test significantly outperforms traditional likelihood ratio (LR) tests in irregular scenarios, where LR tests fail to control the type I error.
  • High-Dimensional Capabilities: The method is effective even when the number of constraints (p) is much larger than the sample size (n).
  • Practical Guidance: The paper suggests choosing the computational budget parameter (N) for incomplete U-statistics to be of the same order as the sample size (N = O(n)) as a good balance between robustness and efficiency.

Limitations and Future Work

  • Computational Budget Trade-off: While a higher computational budget (N) generally leads to better power, it can also make the bootstrap approximation less reliable when the true parameter is very close to an irregular point. The paper suggests N=2n as a practical compromise.
  • Strict Inequality Constraints: The current power analysis for strict inequality constraints could be improved. The authors suggest a two-step approach (Romano et al., 2014; Bai et al., 2021) for future work.
  • Generalization to Other Models: While the paper focuses on Gaussian latent tree models and two-factor analysis models, extending the methodology to other complex statistical models with polynomial constraints is a natural next step.
  • Optimality of Bounds: The theoretical bounds on the approximation error are not always optimal, and further research in high-dimensional statistics could refine these.

Summary updated

Abstract

Abstract We consider the problem of testing a null hypothesis defined by equality and inequality constraints on a statistical parameter. Testing such hypotheses can be challenging because the number of relevant constraints may be on the same order or even larger than the number of observed samples. Moreover, standard distributional approximations may be invalid due to irregularities in the null hypothesis. We propose a general testing methodology that aims to circumvent these difficulties. The constraints are estimated by incomplete U-statistics, and we derive critical values by Gaussian multiplier bootstrap. We show that the bootstrap approximation of incomplete U-statistics is valid for kernels that we call mixed degenerate when the number of combinations used to compute the incomplete U-statistic is of the same order as the sample size. It follows that our test controls type I error even in irregular settings. Furthermore, the bootstrap approximation covers high-dimensional settings making our testing strategy applicable for problems with many constraints. The methodology is applicable, in particular, when the constraints to be tested are polynomials in U-estimable parameters. As an application, we consider goodness-of-fit tests of latent-tree models for multivariate data.
BibTeX
@article{sturma2024testing,
  author = {Nils Sturma and Mathias Drton and Dennis Y.C. Leung},
  title = {{Testing many constraints in possibly irregular models using incomplete <i>U</i>-statistics}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2024},
  doi = {10.1093/jrsssb/qkae022},
}
Figure