Bayesian Data Analysis-英文文献.pdf

Statistical Science 2009, Vol. 24, No. 2, 176–178 DOI: 10.1214/09-STS284D Main article DOI: 10.1214/09-STS284 © Institute of Mathematical Statistics, 2009 Bayes, Jeffreys, Prior Distributions and the Philosophy of Statistics1 Andrew Gelman I actually own a copy of Harold Jeffreys’s Theory generally, for the principle of constructing noninfor- of Probability but have only read small bits of it, most mative, or minimally informative, or objective, or recently over a decade ago to confirm that, indeed, Jef- reference prior distributions from the likelihood (see, freys was not too proud to use a classical chi-squared for example, Kass and Wasserman, 1996). But it can p-value when he wanted to check the misfit of a model notoriously difficult to choose among noninformative to data (Gelman, Meng and Stern, 2006). I do, how- priors; and, even more importantly, seemingly nonin- ever, feel that it is important to understand where our formative distributions can sometimes have strong and probability models come from, and I welcome the op- undesirable implications, as I have found in my own portunity to use the present article by Robert, Chopin experience (Gelman, 1996, 2006). As a result I have and Rousseau as a platform for further discussion of become a convert to the cause of weakly informative foundational issues.2 priors , which attempt to let the data speak while being In this brief discussion I will argue the following: strong enough to exclude various “unphysical” possi- (1) in thinking about prior distributions