Biostatistics Seminar Series: Di Shu, PhD

Abstract: Biomedical research often requires large and diverse samples to generate valid and generalizable evidence. It is common to analyze data from multiple sources to increase statistical power and generalizability. For example, the Sentinel System is a national electronic system funded by the U.S. Food and Drug Administration to monitor the safety of approved medical products using data from 18 health plans and delivery systems. In distributed data networks like the Sentinel System, privacy considerations often make it challenging or impossible to pool individual-level data across data-contributing sites. It is therefore imperative to develop valid statistical methods that leverage information from multiple sources while minimizing the sharing of sensitive individual-level data. In this talk, I will introduce one-step privacy-protecting methods to estimate the overall and site-specific hazard ratios using inverse probability weighted Cox models in distributed data network studies. The proposed methods only require sharing of summary-level risk-set tables to produce results identical to what would be obtained from the corresponding individual-level data analyses. To provide protection against misspecification of the propensity score model, we will further propose a weighted estimation method rooted in empirical likelihood theory which allows each site to simultaneously postulate a set of propensity score models. The resulting overall and site-specific hazard ratio estimators are multiply robust in that they are guaranteed to be consistent when each site includes a correctly specified site-specific propensity score model in its set of candidate models. I will justify these methods theoretically, illustrate their use, and demonstrate their statistical performance using both simulated and real-world data. Finally, I will discuss a few future research topics.