Likelihood-Based Inference for Max-Stable Processes

Simone A. Padoan, Mathieu Ribatet, Scott A. Sisson

TL;DR

This paper introduces a practical, likelihood-based framework for analyzing spatial extremes using max-stable processes. By employing a composite likelihood approach, it overcomes the computational challenges of traditional methods, enabling simultaneous modeling of marginal and dependence structures for spatial extreme events.

What Problem Does This Paper Address?

Modeling extreme events in spatial data, such as record-breaking rainfall or wind speeds, is crucial for understanding and predicting their impact. Max-stable processes are a theoretically sound tool for this, but their practical application has been hindered by the intractability of their multivariate density functions. This makes standard likelihood-based inference difficult, forcing researchers to rely on approximations or less flexible methods. This paper aims to bridge this gap by developing a computationally feasible and flexible inferential framework for max-stable processes.

Methods

The core of this paper's contribution lies in its use of composite likelihood methods. Instead of relying on the full, intractable multivariate density, composite likelihoods construct a likelihood function from a set of simpler, marginal or conditional densities. Specifically, the authors propose a pairwise composite likelihood approach. This means they build the likelihood by considering all possible pairs of spatial locations. This strategy leverages the fact that while the full multivariate density is complex, bivariate (two-location) densities for max-stable processes are often analytically tractable. The paper then details how to estimate model parameters, including both marginal GEV (Generalized Extreme Value) parameters and dependence parameters, simultaneously within this composite likelihood framework. They also discuss model selection using criteria like the Composite Likelihood Information Criterion (CLIC).

Datasets

The paper uses a dataset of U.S. precipitation data. This dataset consists of daily rainfall records from 46 gauging stations over a period of 91 years. The authors do not introduce new benchmark datasets but rather apply their developed methodology to this existing real-world data.

Key Insights and Ideas

  • Composite Likelihood as a Practical Solution: The intractability of max-stable process densities is a major bottleneck. Composite likelihoods offer a computationally viable workaround, making these powerful models accessible for real-world spatial extreme analysis.
  • Joint Modeling of Margins and Dependence: A key advantage of the proposed method is its ability to estimate marginal (e.g., GEV) and dependence parameters simultaneously. This unified approach is more efficient and potentially more accurate than separate estimation procedures.
  • Extremal Coefficients as a Measure of Dependence: The paper highlights the importance of extremal coefficients as a practical measure of dependence strength between spatial locations. These coefficients provide insights into how many independent variables effectively contribute to the extreme behavior.
  • Gaussian Extreme Value Process as a Workhorse: The study focuses on the Gaussian extreme value process, a specific type of max-stable process, demonstrating its utility and the effectiveness of the composite likelihood approach for this class.

Key Results

  • The proposed composite likelihood method is shown to be reliable and versatile, allowing for simultaneous modeling of marginal and dependence parameters at a moderate computational cost.
  • Simulation studies indicate that the composite likelihood estimator is accurate and exhibits good behavior in finite samples, with minimal bias.
  • The analysis of U.S. precipitation data reveals spatial variations in extremal dependence, with stronger dependence observed in mountainous regions and a clear indication of anisotropy (directional dependence), potentially linked to geographical features like the Appalachian Mountains.
  • Model selection using CLIC identified a preferred model for the U.S. precipitation data, demonstrating the practical utility of the framework for choosing among competing spatial extreme models.

Limitations and Future Work

  • Asymptotic Normality Issues: In cases of very strong dependence, the asymptotic normality of the composite maximum likelihood estimator (MCLE) can fail, affecting the reliability of standard error estimates. This is noted in the simulation results.
  • Model Selection Performance: While the composite likelihood ratio tests show good performance, the CLIC criterion demonstrated "poor performance" in model selection tests, reaching a rejection rate of only 20% under the true model.
  • Future Work: The authors suggest that extending this framework to incorporate threshold excess models and limiting Poisson characterization of extremes would be beneficial, allowing for the incorporation of more data into the modeling process. Further exploration of non-parametric approaches for spatial extremes is also mentioned as a promising avenue.

Summary updated

Abstract

The last decade has seen max-stable processes emerge as a common tool for the statistical modeling of spatial extremes. However, their application is complicated due to the unavailability of the multivariate density function, and so likelihood-based methods remain far from providing a complete and flexible framework for inference. In this article we develop inferentially practical, likelihood-based methods for fitting max-stable processes derived from a composite-likelihood approach. The procedure is sufficiently reliable and versatile to permit the simultaneous modeling of marginal and dependence parameters in the spatial context at a moderate computational cost. The utility of this methodology is examined via simulation, and illustrated by the analysis of United States precipitation extremes.
BibTeX
@article{padoan2010likelihood,
  author = {Simone A. Padoan and Mathieu Ribatet and Scott A. Sisson},
  title = {{Likelihood-Based Inference for Max-Stable Processes}},
  journal = {Journal of the American Statistical Association},
  year = {2010},
  doi = {10.1198/jasa.2009.tm08577},
}
Figure