Being Bayesian in the 2020s: opportunities and challenges in the practice of modern applied Bayesian statistics

Joshua J. Bon, Adam Bretherton, Katie Buchhorn, Susanna Cramb, Christopher Drovandi, Conor Hassan, Adrianne L. Jenner, Helen J. Mayfield, James McGree, Kerrie Mengersen, Aiden Price, Robert Salomone, Edgar Santos–Fernández, Julie Vercelloni, Xiaoyu Wang

TL;DR

Bayesian statistics has become a powerful and integral tool for statisticians and data scientists. This paper explores six key areas where Bayesian methods are evolving: intelligent data collection, leveraging new data sources, federated analysis, handling implicit models, transferring knowledge between models, and developing practical software products. These advancements highlight the growing maturity and future potential of Bayesian approaches in tackling complex real-world problems.

What Problem Does This Paper Address?

The paper addresses the ongoing evolution and application of Bayesian statistics in the 2020s. While Bayesian methods have a strong theoretical foundation and have seen widespread adoption, there are still significant opportunities and challenges in their practical application. The authors aim to identify and discuss six key areas where innovation is happening, pushing the boundaries of what's possible with Bayesian inference in modern data science. These areas represent the frontier of applied Bayesian statistics, moving beyond established methods to tackle new data types, computational constraints, and complex modeling scenarios.

Methods

The paper doesn't focus on a single method but rather surveys six distinct areas, each with its own set of techniques:

  • Intelligent Data Collection: Explores Bayesian optimal experimental design, focusing on maximizing information gain for inferential goals. This involves optimizing data collection strategies rather than simply collecting all available data. Techniques include maximizing expected utility functions and using approximations like Laplace approximation and variational Bayes for computationally efficient design.
  • New Data Sources: Discusses the application of Bayesian modeling to novel data types like virtual reality (VR) elicited data and citizen science data. Bayesian hierarchical models are used to handle the complexities and potential biases of these sources, aiming to increase trust and integrate them into mainstream analyses.
  • Federated Analysis: Addresses the challenge of analyzing data distributed across multiple custodians while preserving privacy. The paper discusses approaches for fitting models without centralizing data, highlighting the complexities of horizontal and vertical federated learning settings.
  • Inference for Implicit Models: Focuses on models where the likelihood function is intractable but the model can be simulated. Approximate Bayesian Computation (ABC) and its variants, like Sequential Monte Carlo ABC (SMC-ABC) and Sequential Neural Likelihood (SNL), are presented as key methods for performing Bayesian inference in these scenarios.
  • Model Transfer: Explores how information from a "source" domain with sufficient data can be used to improve inference in a "target" domain with limited data. Techniques discussed include naive updating, power priors, and more advanced methods like transfer learning for neural networks and transfer Gaussian Processes.
  • Purposeful Software Products: Showcases how Bayesian methods are being embedded into practical tools for decision-making. Examples include Bayesian networks for risk-benefit calculators (e.g., COVID-19 vaccine assessment) and spatio-temporal Bayesian models for environmental monitoring (e.g., coral reef health).

Datasets

The paper describes several case studies that utilize real-world data:

  • Coral Reef Monitoring: Data from underwater images of coral reefs, specifically focusing on the Great Barrier Reef.
  • Antarctic Aesthetics: Elicitation data from subjects immersed in a VR environment, assessing aesthetic values of Antarctic landing sites.
  • Citizen Science Data: Image classification data from citizen scientists, used for estimating coral cover.
  • Tumour Growth: Published in vivo tumour growth measurements for ovarian cancer, used to calibrate agent-based models.
  • COVID-19 Vaccine Risk-Benefit Calculator (CoRiCAL): Utilizes Australian data and international data to inform risk-benefit assessments.
  • ReefCloud: Leverages a large dataset of one million reef images for coral cover prediction.
  • Australian Cancer Atlas: Uses Australian statistical area level 2 (SA2) data for cancer incidence and survival estimates.

No new benchmarks are explicitly introduced, but the case studies often involve real-world data that serves as evaluation sets.

Key Insights and Ideas

  • Bayesian methods are democratizing complex modeling: The development of user-friendly software and accessible methods is putting powerful Bayesian tools into the hands of a wider range of researchers and practitioners.
  • "Intelligent" data collection is crucial: Instead of brute-force data collection, Bayesian optimal design allows for more efficient and informative data gathering, especially in resource-constrained environments.
  • New data sources require new modeling approaches: VR and citizen science data, while rich, come with unique challenges related to quality and bias. Bayesian hierarchical models offer a robust framework for handling these.
  • Privacy is a growing concern, and federated learning is a key solution: The ability to perform analysis without centralizing sensitive data is becoming increasingly important, and Bayesian approaches are being adapted for this.
  • Implicit models are becoming mainstream: The ability to simulate complex models, even without an explicit likelihood, opens up new avenues for realistic modeling, with ABC methods leading the way.
  • Knowledge transfer is key for data-scarce problems: Model transfer techniques allow us to leverage existing knowledge, accelerating learning and improving inference in situations with limited data.
  • Bayesian tools are moving from theory to impactful applications: The development of tools like risk calculators and environmental monitoring platforms demonstrates the tangible benefits of applied Bayesian statistics.

Key Results

The paper presents a survey of opportunities and challenges, rather than a single set of quantitative results. However, the case studies highlight the successful application of Bayesian methods:

  • Intelligent Data Collection: The coral reef monitoring case study demonstrates how Bayesian design can provide flexible sampling windows that balance information gain with practical constraints.
  • New Data Sources: Bayesian models are shown to effectively incorporate VR-elicited data and citizen science observations, improving the reliability of estimates for aesthetic values and coral cover.
  • Federated Analysis: While the paper outlines the challenges, it points to ongoing research aiming to develop methods for privacy-preserving analysis of spatially dependent latent variables.
  • Inference for Implicit Models: The tumour growth case study shows that SMC-ABC and SNL can effectively calibrate complex agent-based models, producing accurate posterior predictive distributions for real-world cancer data. SMC-ABC was found to be more robust to model misspecification, though SNL was computationally faster.
  • Model Transfer: The paper discusses methods that can improve inference by borrowing information from source domains, though specific quantitative results are not detailed for this section.
  • Purposeful Products: The CoRiCAL tool demonstrates how Bayesian networks can be used to create accessible and informative risk-benefit calculators for public health decisions. ReefCloud showcases the use of spatio-temporal Bayesian models for large-scale environmental monitoring, providing accessible insights into coral reef health. The Australian Cancer Atlas highlights the power of Bayesian spatial models for detailed regional health analysis.

Limitations and Future Work

  • Computational Cost: Many advanced Bayesian methods, particularly those involving MCMC or ABC, can be computationally intensive. The paper notes that research is ongoing to improve efficiency.
  • Model Misspecification: The performance of some methods, like SNL, can degrade if the simulator does not accurately reflect the real-world data. Robustness to misspecification remains an area for further research.
  • Privacy Guarantees in Federated Learning: While federated learning aims for privacy, ensuring strong, mathematically guaranteed privacy (e.g., differential privacy) in complex Bayesian models is an active area of research.
  • Learning Optimal Transfer Parameters: For model transfer, determining the optimal way to transfer information (e.g., learning the value of $\phi$ in power priors) and how to incorporate uncertainty in this transfer is still an open question.
  • Generalizing Model-Agnostic Methods: While many transfer learning methods are specific to neural networks, developing general, model-agnostic Bayesian approaches is a future goal.
  • Developing Robust Designs: The paper acknowledges that Bayesian designs can rely on assumptions about the data model, and future work aims to develop more robust designs that are less sensitive to these assumptions.
  • Addressing Complex Hierarchical Models in Federated Settings: The paper highlights that current federated Bayesian methods are limited, especially for models with hierarchical structures and latent variables, indicating a significant area for future development.

Summary updated

Abstract

Building on a strong foundation of philosophy, theory, methods and computation over the past three decades, Bayesian approaches are now an integral part of the toolkit for most statisticians and data scientists. Whether they are dedicated Bayesians or opportunistic users, applied professionals can now reap many of the benefits afforded by the Bayesian paradigm. In this paper, we touch on six modern opportunities and challenges in applied Bayesian statistics: intelligent data collection, new data sources, federated analysis, inference for implicit models, model transfer and purposeful software products. This article is part of the theme issue 'Bayesian inference: challenges, perspectives, and prospects'.
BibTeX
@inproceedings{bon2023being,
  author = {Joshua J. Bon and Adam Bretherton and Katie Buchhorn and Susanna Cramb and Christopher Drovandi and Conor Hassan and Adrianne L. Jenner and Helen J. Mayfield and James McGree and Kerrie Mengersen and Aiden Price and Robert Salomone and Edgar Santos–Fernández and Julie Vercelloni and Xiaoyu Wang},
  title = {{Being Bayesian in the 2020s: opportunities and challenges in the practice of modern applied Bayesian statistics}},
  booktitle = {Philosophical},
  year = {2023},
  doi = {10.1098/rsta.2022.0156},
}
Figure