Objective Bayesian analysis occupies a distinctive position within the Bayesian paradigm. While mainstream Bayesian statistics embraces the role of subjective prior information, objective Bayesian methods seek to develop prior distributions that are in some sense "default" or "non-informative," allowing the data to speak with minimal prior influence. The O'Bayes conference series is the principal forum for researchers working on these methods and the foundational questions they raise.
History and Intellectual Context
The objective Bayesian tradition has deep roots, going back to Laplace's principle of insufficient reason and Harold Jeffreys's invariant prior distributions. The modern development of objective Bayes was driven by researchers such as José M. Bernardo, who introduced the concept of reference priors, and James Berger, who championed the use of objective Bayesian methods in scientific applications. The O'Bayes conference series was established to provide a dedicated forum for this distinctive approach.
The first O'Bayes workshop is held, bringing together researchers interested in objective, reference, and non-subjective Bayesian methods.
The conference becomes a regular biennial event, with meetings in locations including Branson (Missouri), Valencia, and Philadelphia.
O'Bayes expands its scope to include connections between objective Bayes and frequentist methodology, model selection, and high-dimensional inference.
In the O'Bayes context, "objective" does not mean that the analysis is free of assumptions. Rather, it refers to the goal of developing prior distributions that are determined by the model and the inferential problem at hand, rather than by a specific individual's beliefs. Reference priors, Jeffreys priors, maximum entropy priors, and probability matching priors are all examples of objective Bayesian approaches. These methods aim to provide default analyses that have good frequentist properties while retaining the coherence of the Bayesian framework.
Key Themes
The O'Bayes conference addresses a range of interconnected topics. Reference prior theory seeks to develop principled default priors for specific classes of models. Objective model selection addresses the use of Bayes factors and information criteria when priors are chosen to be non-informative. Matching priors aim to find priors that ensure posterior credible intervals have correct frequentist coverage. High-dimensional and complex model settings present new challenges for objective Bayesian methods, as the choice of non-informative priors in high dimensions can have a dramatic effect on inference.
Community and Impact
The O'Bayes community is relatively small but intellectually influential. The conference typically attracts 50-100 participants, creating an intimate and highly interactive setting for deep technical discussions. The proceedings and invited papers from O'Bayes have been published in special issues of leading journals, and the ideas developed at these meetings have had a broad impact on Bayesian practice, influencing the default prior choices in widely used software such as Stan and JAGS.
"The O'Bayes conference addresses one of the deepest questions in Bayesian statistics: can we do Bayesian inference without being Bayesian? The answer, developed over decades of research presented at these meetings, is nuanced, sophisticated, and profoundly useful."— James Berger