Inference for Stereological Extremes

Paola Bortot, Stuart Coles, Scott A. SissonJournal of the American Statistical Association 2007 · DOI

TL;DR

This paper tackles the challenge of inferring the size of the largest imperfections (inclusions) in steel, which is crucial for quality control. They develop a new, computationally intensive but robust method for analyzing non-spherical inclusions, addressing limitations of previous models that assumed spherical shapes.

What Problem Does This Paper Address?

In the production of high-quality steel, microscopic imperfections called "inclusions" are unavoidable. The strength of the steel is significantly affected by the size of the largest inclusion. Therefore, accurately estimating the distribution of these largest inclusions is vital for quality control. The standard method for observing these inclusions involves taking planar slices of the steel and measuring the cross-sectional size of any inclusions that intersect the slice. This creates a stereological problem: how to infer the true 3D size of inclusions from 2D measurements. Previous work successfully addressed this for spherical inclusions using extreme value theory and Bayesian MCMC methods. However, the assumption of spherical inclusions is likely unrealistic. This paper aims to: 1) assess how much the previous methods are affected when inclusions are actually non-spherical, and 2) develop a new method that can handle non-spherical inclusions.

Methods

The core of the paper lies in extending the existing spherical inclusion model to a more general ellipsoidal inclusion model.

  • Spherical Model: This is the baseline, assuming inclusions are perfect spheres. It uses extreme value theory (specifically the Generalized Pareto Distribution for tail behavior) and a Bayesian hierarchical model with standard Markov Chain Monte Carlo (MCMC) for inference.
  • Ellipsoidal Model: This generalizes the spherical model by allowing inclusions to be ellipsoids. This requires more detailed specification of shape and orientation.
  • Likelihood-Free MCMC (LF-MCMC): A major challenge with the ellipsoidal model is that calculating the likelihood function (a key component of standard MCMC) is computationally infeasible. To overcome this, the authors adapt a recently developed "likelihood-free" MCMC algorithm. This algorithm replaces direct likelihood calculations with simulations from the model.
  • Augmented State Space MCMC (LFA): To further improve the efficiency and robustness of the LF-MCMC for continuous data (like inclusion sizes), they propose a modification using an "augmented state space." This is analogous to simulated tempering and involves introducing an extra parameter to help the algorithm explore the parameter space more effectively, especially when dealing with potentially complex posterior distributions.

Datasets

The paper uses two types of data:

  • Simulated Data: They generate data from the ellipsoidal inclusion model to test the robustness of the spherical model and to validate their new LF-MCMC algorithm.
  • Real Inclusion Data: They analyze a dataset of 112 cross-sectional inclusion diameters from a steel block, previously analyzed by Anderson and Coles (2002) using the spherical model.

Key Insights and Ideas

  • Shape Matters (But Maybe Not Always Critically): The spherical assumption can lead to biased estimates, particularly for the rate parameter ($\lambda$). However, for estimating the characteristic size ($v_c$), the bias might be less severe, especially if the inclusion size distributions are "heavy-tailed."
  • Likelihood-Free Inference is Necessary for Complex Shapes: When dealing with realistic, non-spherical shapes, standard MCMC breaks down due to the intractability of the likelihood. Likelihood-free methods, while computationally demanding, become essential.
  • Augmentation Improves MCMC Mixing: The proposed augmented state space approach (LFA) is crucial for improving the mixing (efficiency) of the likelihood-free MCMC, especially for continuous data, by allowing the algorithm to explore the parameter space more broadly.
  • Filtering is Key for Accuracy: The augmented method produces a chain that needs to be filtered based on a parameter controlling the "closeness" of simulated data to observed data. This filtering step is critical for obtaining accurate posterior inferences.

Key Results

  • Robustness Assessment: Simulations show that fitting a spherical model to ellipsoidal data leads to biased estimates, with the rate parameter ($\lambda$) being particularly underestimated.
  • Algorithm Validation: The proposed likelihood-free MCMC algorithm (LFA) is shown to be viable and accurate for fitting the ellipsoidal inclusion model, providing results comparable to standard MCMC when applied to simulated spherical data.
  • Real Data Analysis: When applied to real steel inclusion data, the ellipsoidal model (fitted using LFA) yields different parameter estimates compared to the spherical model. Notably, the characteristic size ($v_c$) estimates show differences, with the spherical model underestimating $v_c$ for small block sizes and overestimating it for large block sizes relative to the ellipsoidal model.
  • Computational Cost: The likelihood-free methods are significantly more computationally intensive than standard MCMC, requiring much longer simulation runs.

Limitations and Future Work

  • Computational Cost: The primary limitation acknowledged is the high computational cost of the likelihood-free MCMC algorithms. The authors suggest that using simpler summary statistics could improve efficiency at the cost of accuracy.
  • Shape Generalization: The current work is limited to ellipsoidal inclusions. The authors express interest in investigating models that allow for even more general shapes and departures from the Poisson assumption for inclusion distribution.
  • Choice of Metric and Prior: The performance of the LFA algorithm depends on the choice of the metric $p$ and the pseudo-prior $\pi(\epsilon)$. Careful selection is needed to balance mixing and accuracy.
  • Filtering Threshold: The choice of the filtering threshold $\epsilon_T$ is crucial for balancing bias and variability in the posterior estimates. The paper suggests a method for selecting this threshold retrospectively.

Summary updated

Abstract

In the production of clean steels, the occurrence of imperfections—so-called “inclusions”—is unavoidable. The strength of a clean steel block is largely dependent on the size of the largest imperfection that it contains, so inference on extreme inclusion size forms an important part of quality control. Sampling is generally done by measuring imperfections on planar slices, leading to an extreme value version of a standard stereological problem: how to make inference on large inclusions using only the sliced observations. Under the assumption that inclusions are spherical, this problem has been tackled previously using a combination of extreme value models, stereological calculations, a Bayesian hierarchical model, and standard Markov chain Monte Carlo (MCMC) techniques. Our objectives in this article are twofold: (1) to assess the robustness of such inferences with respect to the assumption of spherical inclusions, and (2) to develop an inference procedure that is valid for nonspherical inclusions. We investigate both of these aspects by extending the spherical family for inclusion shapes to a family of ellipsoids. We then address the issue of robustness by assessing the performance of the spherical model when fitted to measurements obtained from a simulation of ellipsoidal inclusions. The issue of inference is more difficult, because likelihood calculation is not feasible for the ellipsoidal model. To handle this aspect, we propose a modification to a recently developed likelihood-free MCMC algorithm. After verifying the viability and accuracy of the proposed algorithm through a simulation study, we analyze a real inclusion dataset, comparing the inference obtained under the ellipsoidal inclusion model with that previously obtained assuming spherical inclusions.
BibTeX
@article{bortot2007inference,
  author = {Paola Bortot and Stuart Coles and Scott A. Sisson},
  title = {{Inference for Stereological Extremes}},
  journal = {Journal of the American Statistical Association},
  year = {2007},
  doi = {10.1198/016214506000000988},
}
Figure