UC Berkeley

There are deep and important philosophical differences between Bayesian and frequentist approaches to quantifying uncertainty. However, some practitioners choose between these approaches primarily on the basis of convenience. For instance, the ability to incorporate parameter constraints is sometimes cited as a reason to use Bayesian methods. This reflects two misunderstandings: First, frequentist methods can indeed incorporate constraints on parameter values. Second, it ignores the crucial question of what the result of the analysis will mean. Bayesian and frequentist measures of uncertainty have similar sounding names but quite different meanings. For instance, Bayesian uncertainties typically involve expectations with respect to the posterior distribution of the parameter, holding the data fixed; frequentist uncertainties typically involve expectations with respect to the distribution of the data, holding the parameter fixed. Bayesian methods, including methods incorporating parameter constraints, require supplementing the constraints with a prior probability distribution for parameter values. This can cause frequentist and Bayesian estimates and their nominal uncertainties to differ substantially, even when the prior is “uninformative.” This paper gives simple examples where “uninformative” priors are, in fact, extremely informative, and sketches how to measure how much information the prior adds to the constraint. Bayesian methods can have good frequentist behavior, and a frequentist can use Bayesian methods and quantify the uncertainty by frequentist means'but absent a meaningful prior, Bayesian uncertainty measures lack meaning. The paper ends with brief reflections on practice.