Kawano, S. (STAT) – Model Validation and Dependent Shrinkage Priors for Small Area Estimation

This dissertation addresses two questions in survey methodology: How can we produce granular, reliable estimates? And how can we validate them without ground truth? We develop three methods: (1) Gaussian data thinning enables out-of-sample model validation using only direct survey estimates and their sampling variances. We characterize its bias–variance tradeoff to guide model comparison. — (2) Spatially selected and dependent random effects use similarities between neighboring areas to inform both effect selection and effect values. — (3) Cascading tree spike-and-slab priors select random effects from coarse to fine resolutions, removing unnecessary branches of effects. — We evaluate these methods through simulations and applications to rent burden in the South Atlantic United States, employment in California, contraceptive use in Kenya, and industry payroll in New England.

Event Host: Sho Kawano, Ph.D. Candidate, Statistical Science 

Advisors: Paul Parker & Zehang Richard Li

Zoom: https://ucsc.zoom.us/j/97616141146?pwd=eEKbv8zyM5p3dPZ2s86oaO79vNs6SF.1

Passcode: 454321