Dr. Jin Zhou Professor-in-Residence of Biostatistics at UCLA, will be speaking about "Beyond Average Treatment Effects: Borrowing Strength from External Predictions in All of Us" on Wednesday, October 14th, 2025 from 3:30 - 4:30pm in HSSB 1173.
Abstract:
Treatment decisions depend on both the magnitude of benefit and the uncertainty around it. Although large observational cohorts provide valuable evidence, precision can remain limited when estimating benefit for patients with similar baseline risk. In this talk, I will discuss how additional health records and external prediction models can improve the precision of treatment-effect estimates. We apply prediction-powered conditional inference to time-to-event outcomes, considering effects in the overall population, within subgroups, and among patients with similar baseline risk. Using the NIH All of Us Research Program, I will illustrate this approach through comparisons of diabetes medications and three-year kidney risk. Preliminary analyses suggest that GLP-1 receptor agonists are associated with lower three-year kidney risk than DPP-4 inhibitors or sulfonylureas. Incorporating auxiliary predictions from additional All of Us records improves the precision of the estimated treatment benefits across baseline-risk neighborhoods. I will conclude with ongoing work using synthetic Veterans Affairs data to build external prediction models, with attention to differences between the source and study populations.