Catalytic Priors

• Activity Number: 629
• Type: Invited
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Nathan Stein* and Samuel Kou and Donald B. Rubin
• Companies: University of Pennsylvania and Harvard University and Harvard University
• Keywords: prior distribution ; multilevel model ; empirical Bayes

Abstract

A strategy is proposed for building data-dependent prior distributions that are designed to stabilize one model, the target model, by encouraging it not to stray too far from the predictive distribution of another simpler model. The simpler model can be specified and estimated before fitting the target model, leading to straightforward two-stage estimation procedures. The framework is applied to generalized linear models. In simulations and an analysis of protein activity data, the proposed priors are competitive with and can offer improved performance over other state-of-the-art methods often used as default data analytic tools. The priors are derived from three different theoretical perspectives, unifying and extending several previous proposals for prior construction.