Beyond Bayes rule: Simulation experiments for principled data science

Monday, February 17, 2020 - 2:00pm to 3:00pm


Richard Hahn
Associate Professor
Arizona State University


Bayesian modeling has many attractive virtues and often leads to estimators that work exceptionally well in practice. But sophisticated Bayesian models have notable drawbacks as well: theoretical intractability, steep computational burden, and poor calibration under model misspecification. Through a detailed case study of my own work on Bayesian causal forests (BCF) and Accelerated Bayesian Additive Regression Trees (XBART) I describe an approach that blends Bayesian modeling with efficient non-Bayesian algorithms and evaluates the resulting estimators via prior predictive simulations. The resulting data analysis approach might be characterized as that of a committed (nonparametric) Bayesian who, in an effort to prevent the perfect from being the enemy of the good, has renounced allegiance to Bayes rule. More concretely, this talk is about the wide practical application of Monte Carlo evaluations of Bayes risk


Event Type