Description
- Abstract:
- Incorporating historical data can substantially improve statistical efficiency, but standard power prior methods typically assume that all covariates are fully observed across data sources. In many practical settings, however, key covariates are systematically missing in historical data, which may induce biased inference if external information is naively borrowed. In this thesis, we study Bayesian inference under covariate missingness within the power prior framework. We first consider settings in which a covariate is unobserved in the historical dataset but no additional inclusion mechanism is present, and derive a marginal likelihood formulation that enables valid borrowing based on partially observed records. We then extend the framework to settings with covariate-dependent historical inclusion. When selection depends on the missing covariate, we propose an importance-sampling weighted power prior to correct for the induced distributional shift. Simulation studies and a real-data application demonstrate that the proposed methods improve efficiency while maintaining robustness under a range of sensitivity scenarios.
- Notes:
- Thesis (Sc. M.)--Brown University, 2026
Citation
Zhao, Zinan,
"Robust Bayesian Inference with Power Prior under Covariate Missingness"
(2026).
Biostatistics Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:unqaebkp/
Relations
Collection:
-
Biostatistics Theses and Dissertations
Theses and Dissertations for the Biostatistics department....