Bayesian Dynamic Data Borrowing Methodologies for Source-Specific Inference

Abstract

The rapid advancement of big data in various fields of healthcare has sparked a growing interest in making inferences about source-specific parameters. While technologies have provided a wealth of data for each person, the optimal utilization of multisource population data to enhance source-specific inference remains an open research question. Numerous existing methods have been developed to borrow information from multiple supplemental sources, such as individuals, in order to support individualized parameter inference in a primary source. Multisource exchangeability models (MEM) have been introduced as such a dynamic data borrowing approach, determining the amount of information to borrow based on the exchangeability between the primary and supplemental sources, where exchangeability is defined as the equality of a target parameter. Motivated by MEM, this thesis proposes three innovative Bayesian statistical data-borrowing methods to leverage complicated types of information. The first method is the data-driven MEM (dMEM), a two-stage approach that incorporates source selection and clustering to enable the inclusion of an arbitrary number of sources in individualized inference while ensuring computational efficiency and data effectiveness. The second method is the generalized Reinforced Borrowing Framework (RBF), which utilizes a distance-embedded prior leveraging not only data related to the target parameter but also different types of auxiliary information sources, to reinforce inference on the target parameter. The third method is the Synthetic Conditional Borrowing (SCB) approach, which generates synthetic sources from predictive models based on a range of covariates specific to the primary source. This method facilitates the inference of target parameters by incorporating a novel definition of conditional exchangeability that is not limited to strict homogeneity between sources but also considers the alignment in covariates. All three methods demonstrate robustness in various simulation settings and outperform previously proposed methods in both simulated and real-world scenarios.