Statistics with Imprecise Probabilities—A Short Survey

Thomas AugustinSpringerBriefs 2021 · DOI · OpenAlex

Abstract

This chapter aims at surveying and highlighting in an introductory way some challenges and big opportunities a paradigmatic shift to imprecise probabilities could induce in statistical modelling. Working with an informal understanding of imprecise probabilities, we discuss the concepts of model imprecision and data imprecision as the two main types of imprecision in statistical modelling. Then we provide a short survey of some major developments, methodological questions and applications of imprecise probabilistic models under model imprecision in the context of different inference schools and summarize some recent developments in the area of data imprecision.

Introduction: Moving Beyond Precise Probabilities

Traditional statistical theory is largely founded on the assumption of "precise" probability, where every event is assigned a single, exact numerical value $p(A)$. While this framework has served science well, it often fails to capture the inherent ambiguity of real-world modeling. Thomas Augustin's survey explores the shift toward "imprecise probabilities" (IP)—a paradigm that replaces single distributions with sets of distributions to more faithfully represent uncertainty.

The Two Pillars of Imprecision

The need for IP arises from two primary sources:

1. Model Imprecision

In many scenarios, a researcher lacks sufficient information to justify a single "correct" prior or sampling model. Forcing a precise choice, such as a specific Normal distribution, can lead to overconfident conclusions. A classic example in robust statistics is the similarity between a Normal and a Cauchy distribution; while they may look nearly identical, the sample mean behaves predictably for the former but fails entirely for the latter. IP provides a superstructure to handle these "tiny deviations" that can lead to significant errors.

2. Data Imprecision

Real-world data is rarely "clean." It can be ontic (inherently vague, such as an undecided voter) or epistemic (precise in reality but observed poorly, such as censored survival times). Traditional methods often bridge these gaps with strong, untestable assumptions (e.g., "Missing at Random"). IP allows for a more honest treatment of such data without forcing it into a precise box.

Methodological Frameworks

The IP toolbox offers several frameworks for handling uncertainty:

  • Generalized Bayesian Inference: Instead of a single prior, a set of priors is used. The Generalized Bayes Rule (GBR) then produces a set of posteriors.
  • Near-Ignorance Models: These models allow for a state of "not knowing" without the arbitrary assumptions of "non-informative" precise priors. The Imprecise Dirichlet Model (IDM) is a prominent example for categorical data.
  • Nonparametric Predictive Inference (NPI): This approach avoids priors entirely, using exchangeability to provide lower and upper probabilities for future observations.
  • Robust Statistics: IP links closely with frequentist robust statistics, using neighborhood models to ensure procedures remain stable even if data deviates slightly from the assumed model.

Key Conceptual Insights

Prior-Data Conflict

In standard Bayesianism, more data typically narrows the posterior. In the IP framework, if data strongly contradicts the prior, the set of posteriors can actually expand. This serves as a built-in "alert system," signaling that the model's assumptions and the observed data are in conflict.

Manski’s Law of Decreasing Credibility

The survey highlights the principle that the credibility of an inference decreases as the strength of the underlying assumptions increases. By relaxing assumptions, IP allows researchers to identify which conclusions remain robust across a wider range of possibilities.

Applications and Future Outlook

IP has found success in several specialized fields:

  • Engineering and Reliability: Used in Uncertainty Quantification (UQ) where failure data is scarce.
  • Survey Research: Providing more realistic intervals for pre-election polls by properly modeling "undecided" responses.
  • Categorical Data: The IDM has become a standard for learning from small samples.

Despite its potential, IP remains on the periphery of mainstream statistics due to computational complexity and a lack of integration with modern machine learning. The author argues for a shift from using IP "defensively" (for robustness checks) to "actively" (as a tool for incorporating vague expert knowledge). Future progress depends on developing better simulation methodologies, such as Imprecise Monte Carlo, to handle the increased complexity of working with sets of distributions.

Summary updated

Abstract

Abstract This chapter aims at surveying and highlighting in an introductory way some challenges and big opportunities a paradigmatic shift to imprecise probabilities could induce in statistical modelling. Working with an informal understanding of imprecise probabilities, we discuss the concepts of model imprecision and data imprecision as the two main types of imprecision in statistical modelling. Then we provide a short survey of some major developments, methodological questions and applications of imprecise probabilistic models under model imprecision in the context of different inference schools and summarize some recent developments in the area of data imprecision.
BibTeX
@inproceedings{augustin2021statistics,
  author = {Thomas Augustin},
  title = {{Statistics with Imprecise Probabilities—A Short Survey}},
  booktitle = {SpringerBriefs},
  year = {2021},
  doi = {10.1007/978-3-030-83640-5\_5},
}
Figure