Comment: On Focusing, Soft and Strong Revision of Choquet Capacities and Their Role in Statistics

Thomas Augustin, Georg SchollmeyerStatistical 2021 · DOI

Abstract

In this commentary on the work of Ruobin Gong and Xiao-Li Meng (GM), we explore the power of imprecise probabilities in addressing statistical paradoxes. By discussing generalized uncertainty quantification and various conditioning rules, we characterize three major updating rules—Generalized Bayes, Dempster’s, and Geometric—as envelopes of specific sets of conditional probabilities. This perspective allows for the introduction of "soft revision" rules that act as a compromise between conservative and aggressive updating. Focusing primarily on Choquet capacities of order 2, we discuss their role as statistical models and the potential for imprecise probabilities to manage the multidimensional nature of uncertainty.

Introduction

In classical statistical inference, uncertainty is typically represented by a single probability distribution. However, in many real-world scenarios where information is scarce, conflicting, or vague, a single distribution may be insufficient. This leads to the framework of imprecise probabilities, where uncertainty is represented by a set of possible distributions, known as a credal set.

In their landmark paper, Gong and Meng (GM) demonstrated that different rules for updating these credal sets can resolve long-standing statistical paradoxes, such as Simpson’s Paradox and Dilation. This note extends their work by providing a unified mathematical characterization of these updating rules and proposing a family of "soft revision" rules to balance the trade-offs inherent in belief updating.

Unifying Conditioning Rules as Envelopes

The primary methodological contribution of this work is the characterization of the three major conditioning rules in imprecise probability—the Generalized Bayes Rule, Dempster’s Rule, and the Geometric Rule—as envelopes of certain sets of conditional probabilities.

Working within the framework of Choquet capacities of order 2 (also known as 2-monotone lower probabilities), we demonstrate that these rules are not disconnected procedures but can be expressed through a unified mathematical language. Specifically, they can be viewed as the infimum (minimum) or supremum (maximum) of a ratio of probabilities over specific subsets of the credal set. This "envelope view" clarifies the relationship between different modes of statistical reasoning.

Focusing vs. Belief Revision

A critical conceptual distinction in this framework is the difference between Focusing and Belief Revision:

  1. Focusing (Generalized Bayes): This approach applies existing knowledge to a specific event without altering the underlying model. It is inherently conservative; if the prior knowledge is "vacuous" (totally unknown), the posterior remains vacuous.
  2. Belief Revision (Dempster’s and Geometric Rules): This involves actually changing the model based on new evidence. Dempster’s rule, for instance, acts similarly to a maximum likelihood update. While powerful, it carries a risk of "overfitting" by discarding all distributions in the credal set except those that maximize the likelihood of the observed data.

The "Soft Revision" Compromise

To address the extremism of these rules—where one may learn nothing and the other may learn too much from limited data—we propose a "parameter of revision" ($\alpha$). By utilizing likelihood cuts ($\alpha$-cuts), researchers can define a spectrum of conditioning rules:

  • At $\alpha = 0$, the rule collapses to the conservative Generalized Bayes Rule (Focusing).
  • At $\alpha = 1$, the rule reaches the more aggressive Strong Revision (Dempster’s or Geometric rules).

This family of soft revision rules allows for a pragmatic middle ground, though we acknowledge a theoretical tension: these rules do not always satisfy the strict "coherence" requirements of Walley’s theory of imprecise probabilities. However, their utility in robust statistics may outweigh these formal limitations.

Statistical Relevance and 2-Monotonicity

While Choquet capacities of order 2 are a specific subset of imprecise probability models, they are highly relevant for robust statistics. They encompass a rich class of models, including $p$-boxes and $\epsilon$-contamination models (where a central distribution is assumed but allowed a small percentage of error). These models are essential for distinguishing between variability (randomness that can be averaged out) and indeterminacy (fundamental gaps in knowledge that persist regardless of sample size).

Conclusion

As statistics enters the era of Big Data, the need for tools that can handle the multidimensional nature of uncertainty becomes more acute. Following the "Law of Decreasing Credibility"—which suggests that the more assumptions we add to force a precise answer, the less credible that answer becomes—imprecise probabilities offer a more honest, set-valued approach to what we truly know. By unifying conditioning rules and introducing flexible revision parameters, we can better navigate the transition from prior ignorance to informed statistical inference.

Summary updated

Abstract

We congratulate Ruobin Gong and Xiao-Li Meng on their thought-provoking paper demonstrating the power of imprecise probabilities in statistics. In particular, Gong and Meng clarify important statistical paradoxes by discussing them in the framework of generalized uncertainty quantification and different conditioning rules used for updating. In this note, we characterize all three conditioning rules as envelopes of certain sets of conditional probabilities. This view also suggests some generalizations that can be seen as compromise rules. Similar to Gong and Meng, our derivations mainly focus on Choquet capacities of order 2, and so we also briefly discuss in general their role as statistical models. We conclude with some general remarks on the potential of imprecise probabilities to cope with the multidimensional nature of uncertainty.
BibTeX
@inproceedings{augustin2021comment,
  author = {Thomas Augustin and Georg Schollmeyer},
  title = {{Comment: On Focusing, Soft and Strong Revision of Choquet Capacities and Their Role in Statistics}},
  booktitle = {Statistical},
  year = {2021},
  doi = {10.1214/21-sts765d},
}
Figure