Fast and fair simultaneous confidence bands for functional parameters

Dominik Liebl, Matthew ReimherrJournal of the Royal Statistical Society Series B (Statistical Methodology) 2019

TL;DR

This paper introduces a novel method for constructing "fast and fair" simultaneous confidence bands for functional data analysis. These bands offer improved interpretability, computational efficiency, and the ability to handle incomplete (fragmentary) data, addressing key limitations of existing methods.

What Problem Does This Paper Address?

In functional data analysis (FDA), quantifying uncertainty with confidence bands is crucial but has lagged behind estimation and hypothesis testing. Existing methods often suffer from being computationally expensive, overly conservative, unable to handle incomplete data (fragmentary functions), or lacking local interpretability. This paper tackles the challenge of developing confidence bands that are simultaneously fast, statistically fair (controlling false positive rates across different parts of the data domain), and applicable to a wider range of FDA scenarios, including those with missing data.

Methods

The core of the methodology lies in a generalized Kac-Rice formula for constructing confidence bands. Instead of relying on computationally intensive resampling techniques, the authors leverage random process theory. Key innovations include:

  • Non-constant, Adaptive Critical Value Functions: Unlike traditional methods that use a single constant critical value, this approach employs adaptive critical values that can vary across the domain of the functional parameter. This allows for better control of false positive rates and improved interpretability.
  • Fairness Constraint: The bands are designed to satisfy a "false positive rate balance" constraint. This means the probability of incorrectly rejecting a null hypothesis is balanced across predefined partitions of the data's domain, addressing fairness concerns in statistical decision-making.
  • Handling Fragmentary Data: The method does not require estimating the full covariance function, only a narrow band along the diagonal. This makes it suitable for situations where functional data is incomplete or "fragmentary."
  • Theoretical Foundation: The work builds upon the expected Euler characteristic inequality and generalizes Kac-Rice formulas for elliptical processes, providing a robust theoretical framework.

Datasets

The paper demonstrates its methods using:

  • Simulated Data: Extensive simulations are conducted using various mean function scenarios (shift, scale, local) and covariance structures (smooth, rough, smooth-to-rough) to evaluate the performance of the proposed bands against existing methods.
  • Sports Biomechanics Data: Real-world data from a study on running shoes, involving torque curves measured at the ankle joint, is used to showcase the application of the bands for fully observed functional data.
  • Fragmentary Growth Curves Data: Data on bone mineral density acquisition curves from male and female participants is used to illustrate the method's effectiveness with fragmentary functional data.

Key Insights and Ideas

  • Fairness is Quantifiable and Achievable: The paper demonstrates that statistical fairness, specifically balanced false positive rates, can be formally incorporated into confidence band construction, moving beyond qualitative discussions.
  • Adaptive Bands Offer Superior Interpretability: By allowing critical values to adapt to the data's characteristics (e.g., roughness), the resulting bands provide more nuanced local interpretations, which is crucial for high-stakes decisions.
  • Fragmentary Data is No Longer a Roadblock: The ability to construct valid confidence bands without full covariance estimation significantly expands the applicability of FDA to real-world datasets where data is often incomplete.
  • Speed and Fairness Can Coexist: The proposed "Fast and Fair" bands achieve computational efficiency without sacrificing statistical rigor or fairness guarantees, a significant improvement over many existing methods.

Key Results

  • Computational Efficiency: The proposed bands are significantly faster to compute than resampling-based alternatives, with computation times often an order of magnitude lower.
  • Type-I Error Control: Simulations show that the proposed bands effectively control the nominal Type-I error rate across various scenarios and sample sizes, including challenging cases with rough or non-stationary covariance structures.
  • Fairness Guarantees: The bands demonstrably satisfy the false positive rate balance criterion, as verified through simulations.
  • Performance on Real Data: The methods are successfully applied to sports biomechanics and bone mineral density data, yielding interpretable results and demonstrating their practical utility.

Limitations and Future Work

  • Price of Fairness: While the proposed method incorporates fairness, imposing more partitions for fairness can lead to more conservative bands. The paper quantifies this "price of fairness."
  • Higher Dimensional Domains/Manifolds: The authors suggest that extending the approach to higher-dimensional function domains or manifolds is a promising avenue for future research, building on existing work for classic Kac-Rice formulas.
  • Relaxing Distributional Assumptions: Further work could explore relaxing the asymptotic distribution assumptions, particularly for estimates that are not tight in the strong topology (e.g., from scalar-on-function regression).
  • Other Selection Criteria: While fairness is the focus, the framework could be extended to optimize other criteria, such as minimizing average band width.

Summary updated

Abstract

Abstract Quantifying uncertainty using confidence regions is a central goal of statistical inference. Despite this, methodologies for confidence bands in functional data analysis are still underdeveloped compared to estimation and hypothesis testing. In this work, we present a new methodology for constructing simultaneous confidence bands for functional parameter estimates. Our bands possess a number of positive qualities: (1) they are not based on resampling and thus are fast to compute, (2) they are constructed under the fairness constraint of balanced false positive rates across partitions of the bands’ domain which facilitates the typical global, but also novel local interpretations, and (3) they do not require an estimate of the full covariance function and thus can be used in the case of fragmentary functional data. Simulations show the excellent finite-sample behaviour of our bands in comparison to existing alternatives. The practical use of our bands is demonstrated in two case studies on sports biomechanics and fragmentary growth curves.
BibTeX
@article{liebl2019fast,
  author = {Dominik Liebl and Matthew Reimherr},
  title = {{Fast and fair simultaneous confidence bands for functional parameters}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2019},
}
Figure