Abstract

In the big data era, it is of crucial importance to have statistical methods that can robustly integrate multi-source auxiliary information to enhance main analysis. Herein I will illustrate in both genomics and observational studies that why information integration is essential, and how to equip integration schemes with robust properties:

1.     Cell type deconvolution in bulk tissue RNA sequencing (RNA-seq) data is an important step towards the understanding of cell type composition variation between disease conditions. I develop a Multi-Robust Deconvolution (MultiRD) algorithm, which enables the integration of external expression data from multiple sources, to infer cell type proportions from the target bulk RNA-seq data. More importantly, MultiRD is a robust deconvolution method even with incorrect auxiliary information.

2.     Many biomedical studies will collect a number of secondary measurements, in addition to the main trait of interest. In practice, these auxiliary variables are often served as secondary outcomes in separate analyses, thus not directly applicable to the main study. I propose a Multiple information Borrowing (MinBo) framework to synthesize multiple secondary outcomes into the main study, where the misspecification of any secondary models will have little impact on estimation consistency but still preserve considerable efficiency gain in the main model.