Title: Interpreting complex models: Efficient, valid posterior inference for meaningful quantities
Authors: Jared Murray - Carnegie Mellon University (United States) [presenting]
Carlos Carvalho - The University of Texas at Austin (United States)
Abstract: A conceptually simple framework is proposed for making Bayesian inferences about interpretable models that summarize complex posterior distributions. This provides a vehicle for understanding large, complicated, and often nonparametric models. Our approach is able to map the output of state-of-the-art predictive tools onto scientifically meaningful quantities while maintaining valid posterior inference. It also provides a bridge between Bayesian methods and recently popular frequentist methods for post-selection inference. We illustrate the general approach in two important special cases: Summarizing high-dimensional linear regression models with lower-dimensional alternatives, and interrogating the fit of a nonparametric regression model (Bayesian additive regression trees).