Modelling extremes using approximate Bayesian Computation

Robert J. Erhardt, Scott A. SissonarXiv 2014 · DOI · arXiv

This paper, a chapter from a handbook on extreme value analysis, provides a comprehensive guide to using Approximate Bayesian Computation (ABC) to model rare, high-impact events. It demonstrates how to perform rigorous statistical inference when the "likelihood function"—the mathematical heart of most statistical models—is too complex to actually calculate.

TL;DR

The authors demonstrate how to use simulation-based "likelihood-free" inference (ABC) to fit complex models for extreme events like 100-year floods or material failures. By replacing impossible calculations with smart simulations and summary statistics, they provide a path to model spatial and structural extremes that were previously computationally out of reach.

What Problem Does This Paper Address?

In Extreme Value Theory (EVT), we care about the "tails" of a distribution—the rare events that cause disasters. However, as we move from simple models to complex ones (like those involving spatial dependence or hidden "latent" variables), the math breaks down.

Specifically, the likelihood function (the probability of seeing your data given a set of parameters) often suffers from a "combinatorial explosion." In high-dimensional spatial models, calculating this function can be computationally prohibitive or mathematically impossible (no "closed-form" expression exists). This makes standard Bayesian tools like Markov Chain Monte Carlo (MCMC) unusable. The authors address this gap by showing how ABC can bypass the likelihood entirely.

Methods

The core idea of ABC is "inference via simulation." Instead of calculating a likelihood, you:

  1. Pick a candidate parameter value from a prior.
  2. Simulate a "fake" dataset from the model using those parameters.
  3. Compare the simulated data to your real observed data.
  4. If they are "close" enough, keep the parameter; otherwise, toss it.

Key technical contributions discussed include: * Summary Statistics: To avoid the "curse of dimensionality" (where it's impossible to match raw datasets exactly), the authors use low-dimensional summaries like quantiles, L-moments, or Maximum Likelihood Estimates (MLEs). * Regression Adjustment: A sophisticated post-processing step. It uses local-linear regression to "correct" the accepted parameters based on the gap between simulated and observed summary statistics, drastically improving accuracy. * Distance Metrics: The use of Mahalanobis distance to account for different scales and correlations between summary statistics.

Datasets

The paper validates these methods using three distinct scenarios:

  1. Vargas Rainfall (Venezuela): Annual maximum daily rainfall (1951–1999). This includes the 1999 "Vargas Tragedy" debris flows. It serves as a "gold standard" test case where the true likelihood is known, allowing the authors to prove ABC works.
  2. Steel Inclusions (Stereology): Data on 112 microscopic particles (inclusions) in steel. The goal is to predict the largest inclusion in a 3D block based on 2D slices—a classic problem where assuming particles are "spheres" is easy but wrong, while "ellipsoids" is realistic but mathematically difficult.
  3. Midwestern US Temperatures: Annual maximum daily temperatures from 39 locations (1895–2009). This is used to fit a "max-stable process," a complex spatial model used by the insurance industry for weather derivatives.

Key Insights and Ideas

  • The Curse of Dimensionality is Real: The authors show that using the full raw dataset as a "summary" fails even with only 49 data points. ABC requires dimension reduction to be effective.
  • Sufficiency vs. Dimensionality: A summary statistic that is "sufficient" (contains all info) but high-dimensional is often worse in ABC than a non-sufficient but low-dimensional one.
  • Regression Adjustment is a "Free Lunch": Applying regression adjustment often makes the ABC approximation indistinguishable from the "true" posterior, even when the initial simulation match was mediocre.
  • Model Sensitivity: In the steel example, switching from a spherical to an ellipsoidal model (made possible by ABC) significantly changed the estimated rate of inclusions, proving that "tractable but wrong" models can lead to dangerous engineering conclusions.

Key Results

  • Accuracy: In the rainfall example, ABC using 3-dimensional summary statistics (MLEs) perfectly reconstructed the true posterior distribution.
  • Spatial Modeling: For the 39-location temperature dataset, ABC successfully identified a "crescent-shaped" relationship between spatial range and smoothness parameters, allowing for the calculation of 95% credible intervals for spatial correlation.
  • Efficiency: The authors found that using a larger "tolerance" (accepting more simulations) combined with regression adjustment is often more computationally efficient than trying to achieve a "perfect" match with a tiny tolerance.

Limitations and Future Work

  • Simulation Speed: ABC relies on the ability to simulate data quickly. If the model takes minutes to run a single simulation, generating the required millions of samples becomes a bottleneck.
  • The "Art" of Summary Statistics: There is still no universal "best" way to pick summary statistics. While the paper suggests using MLEs or score functions, this remains a major area of active research.
  • Evolution of Tractability: The authors wisely note that what is "intractable" today might have a closed-form solution tomorrow as mathematical proofs advance, but ABC remains the essential "bridge" for current researchers.

Summary updated

Abstract

By the nature of their construction, many statistical models for extremes result in likelihood functions that are computationally prohibitive to evaluate. This is consequently problematic for the purposes of likelihood-based inference. With a focus on the Bayesian framework, this chapter examines the use of approximate Bayesian computation (ABC) techniques for the fitting and analysis of statistical models for extremes. After introducing the ideas behind ABC algorithms and methods, we demonstrate their application to extremal models in stereology and spatial extremes.
BibTeX
@misc{erhardt2014modelling,
  author = {Robert J. Erhardt and Scott A. Sisson},
  title = {{Modelling extremes using approximate Bayesian Computation}},
  year = {2014},
  eprint = {1411.1451},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1411.1451},
}
Figure