Statistical modeling of federated data through sufficient statistics

Abstract

Statistical modeling on individual-level data is indispensable in health research. However, regulations for accessing individual-level data for health research have become more stringent as technology has advanced drastically in recent years. Consequently, individual-level data may not be obtained in a timely manner, and research progress is at risk of being hampered. This thesis addresses the data sharing challenges faced by data providers and data analysts. We propose a framework that enables data analysts to perform statistical inference at the individual level without having access to personal health data. Only summary statistics are shared once, from which pseudo-data are generated. These pseudo-data are used to replace the actual data when estimating generalized linear mixed models (GLMM). Scalability issues are addressed by generating compressed pseudo-data with associated frequency weights. Communication and resource efficiency as well as wide applicability distinguish our proposed framework from existing approaches in the literature.