A transformation of Bayesian statistics:Computation, prediction, and rationality

Abstract

Bayesian approaches have long been a small minority group in scientific practice, but quickly acquired a high level of popularity since the 1990s. This paper shall describe and analyze this turn. I argue that the success of Bayesian approaches hinges on computational methods that make a class of models predictive that would otherwise lack practical relevance. Philosophically, however, this orientation toward prediction comes at a price. The new computational approaches change Bayesian rationality in an important way. Namely, they undercut the interpretation of priors, turning them from an expression of beliefs held prior to new evidence into an adjustable parameter that can be manipulated flexibly by computational machinery. Thus, in the case of Bayes, one can see a coevolution of computing technology, an exploratory–iterative mode of prediction, and the conception of rationality.