AnM-Estimator of Spatial Tail Dependence

J.h.j. Einmahl, Anna Kiriliouk, Andrea Krajina, Johan SegersJournal of the Royal Statistical Society Series B (Statistical Methodology) 2015 · DOI

TL;DR

This paper introduces a new statistical method, a rank-based M-estimator, for modeling the extreme dependence structure of high-dimensional spatial data. This approach is more computationally feasible than previous methods, particularly for large datasets, and is demonstrated to be effective on real-world wind speed data.

What Problem Does This Paper Address?

Modeling extreme events, such as floods, droughts, or extreme wind speeds, often requires understanding how multiple related variables behave simultaneously during these extreme occurrences. This is known as modeling "spatial tail dependence." While max-stable processes are a standard tool for this, existing methods for estimating their parameters become computationally intractable in high dimensions (many locations or variables). This paper addresses this challenge by proposing a new estimator that focuses on pairwise relationships, making it scalable to high-dimensional spatial data.

Methods

The core of the paper is a novel rank-based M-estimator for spatial tail dependence. Instead of directly modeling the extreme values, the method uses the ranks of the data. This has two key advantages: it bypasses the need to model the marginal distributions (which can be difficult in high dimensions) and makes the estimator invariant to monotonic transformations of the data.

The proposed estimator is constructed by minimizing a distance between empirical and parametric representations of bivariate tail dependence functions. To improve efficiency, a data-driven weight matrix is introduced, which is optimized to minimize the asymptotic variance of the estimator. The authors provide theoretical guarantees for the estimator, proving its consistency and asymptotic normality under mild assumptions.

Datasets

The paper uses a dataset of daily maximal wind gust speeds from 35 weather stations in the Netherlands, collected by the Royal Netherlands Meteorological Institute (KNMI). For their analysis, they selected 22 inland stations and aggregated daily maxima to three-day maxima, restricting the observation period to summer months to minimize temporal dependence and handle missing data.

Key Insights and Ideas

  • Scalability through Pairwise Analysis: The key innovation is reducing the high-dimensional problem to analyzing pairs of variables. This makes the method computationally feasible for scenarios with many locations or variables, a significant hurdle for previous approaches.
  • Rank-Based Robustness: Using ranks makes the estimator robust to the specific marginal distributions of the data, which is a major advantage when these distributions are unknown or difficult to model.
  • Data-Driven Optimization: The introduction of a data-driven weight matrix allows the estimator to adapt to the specific dependence structure of the data, leading to improved efficiency.
  • Isotropy Suffices for Dutch Wind Data: Surprisingly, the analysis of Dutch wind speed data suggests that a simpler, isotropic Brown-Resnick model is sufficient, despite the common assumption of anisotropy in spatial tail dependence for wind speeds.

Key Results

  • The proposed rank-based M-estimator is shown to be consistent and asymptotically normal.
  • Simulation studies demonstrate that the pairwise approach, especially when using only neighboring pairs of locations, often leads to better performance (lower bias and RMSE) than using all possible pairs, particularly for the Smith and Brown-Resnick models.
  • The optimal weight matrix, derived from the data, generally leads to improved estimation accuracy compared to using an identity weight matrix.
  • The application to Dutch wind speed data shows that the proposed method can effectively model spatial tail dependence, and the analysis suggests that an isotropic Brown-Resnick model is adequate for this specific dataset.

Limitations and Future Work

  • Computational Cost of Optimal Weight Matrix: While the pairwise approach is scalable, calculating the optimal weight matrix can still be computationally intensive, especially when the number of pairs is large. The paper suggests a two-step procedure to address this.
  • Assumption of Continuous Distributions: The method requires continuous marginal distributions. The authors note that care must be taken with data exhibiting yearly seasonality or trends, and the method is not directly applicable to count data.
  • Future Work: The paper mentions that the proofs of some corollaries are straightforward extensions of existing work, suggesting potential for further theoretical development. The authors also note that guidelines for practical use are available in the spatialTailDep R package.

Summary updated

Abstract

Summary Tail dependence models for distributions attracted to a max-stable law are fitted by using observations above a high threshold. To cope with spatial, high dimensional data, a rank-based M-estimator is proposed relying on bivariate margins only. A data-driven weight matrix is used to minimize the asymptotic variance. Empirical process arguments show that the estimator is consistent and asymptotically normal. Its finite sample performance is assessed in simulation experiments involving popular max-stable processes perturbed with additive noise. An analysis of wind speed data from the Netherlands illustrates the method.
BibTeX
@article{einmahl2015i,
  author = {J.H.J. Einmahl and Anna Kiriliouk and Andrea Krajina and Johan Segers},
  title = {{An<i>M</i>-Estimator of Spatial Tail Dependence}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2015},
  doi = {10.1111/rssb.12114},
}
Figure