Bayes Reading Group: Jonathan Bradley (Florida State)

Friday Nov 22, 2024, 11:00 AM → 12:00 PM America/New_York

3rd Floor Classroom (162 Fifth Avenue )

Discussion Lead: Jonathan Bradley (Florida State)

Topic: The Data Subset Approach: Scaling Bayesian Models by Removing Assumptions on the Data Generating Mechanism

Link: https://www.tandfonline.com/doi/full/10.1080/10618600.2021.1923518

Abstract:  The "data subset approach” is a new Bayesian method that assumes a parametric model for a low-dimensional training dataset and assumes the remaining holdout data follows its true non-parametric data-generating mechanism. The split into training and holdout is treated as a parameter in the Bayesian model. The model is specified in a way that leads to a posterior sampling strategy that re-samples a smaller subset of the data at each iteration of the sampler. Theoretical properties of the data subset model will be discussed including: propriety, partial sufficiency, and semi-parametric properties. Several illustrations will be discussed including simulations, an analysis of the U.S. Census Bureau’s Public Use Micro-Sample (PUMS), and an analysis of datasets from NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) remote sensing instrument.