Assessing New Predictors with Biased Data Augmented with Summary Statistics

Statistical Seminar

Organizer：

吴宇楠

Speaker：

张洪教授中国科学技术大学管理学院统计与金融系

Time：

Fri., 16:00-17:00, June 5, 2026

Venue：

C654, Shuangqing Complex Building A

Title：

Assessing New Predictors with Biased Data Augmented with Summary Statistics

Abstract:

In many real-world applications, evaluating the added value of new risk predictors is hampered by the lack of individual-level data in the target population—only summary statistics may be available. However, individual-level data often exist in a related source population, which may have different data distributions. In this talk, I will present a new statistical method that integrates these two sources to enable unbiased risk prediction and evaluation of new predictors in the target population. Under the mild assumption that the effects of new predictors are shared across populations, our approach constructs an unbiased risk model and further derives biascorrected estimators of predictive performance, including the AUC with and without the new predictors, as well as their difference, with valid inference. Simulation results demonstrate that the proposed method provides accurate and robust AUC estimates despite population heterogeneity. Finally, I will illustrate its practical utility through an application to the UK Biobank, where we confirm that the time between consecutive R waves on an electrocardiogram offers significant added predictive value for 10year acute myocardial infarction risk.