2026 Seminars

Seminar Announcement

Indrila Ganguly
Postdoctoral Research Fellow
BBE Fred Hutchinson Cancer Center

"Flexible Sequential Validation Strategies for Assessing Incremental Value of Biomarkers"

Wednesday, March 25, 2026, 1:30pm-2:20pm (CT)
E0134 Howe Russell (HRKE)

Early detection studies in oncology often rely on panels of established biomarkers, yet these markers frequently show limited diagnostic performance. This motivates the need for statistical frameworks that can assess whether a new biomarker provides meaningful improvement when added to an existing panel. A major practical challenge, however, is the scarcity of high-quality biospecimens, particularly when many candidate biomarkers must be evaluated rigorously.

In this talk, I present a design and analysis framework that addresses this challenge by optimizing specimen allocation and the evaluation of incremental diagnostic performance across multiple biomarker candidates. The approach has two key components. First, we develop a robust hypothesis testing procedure within a group-sequential framework, enabling early stopping for promising or uninformative candidates. Second, we integrate this with a rotational sampling scheme to ensure efficient use of limited specimens. I will discuss how we define and quantify incremental value, outline the statistical properties of the proposed test, and explain why the inference remains valid even when risk models are mis specified.

Overall, this framework provides a practical and efficient strategy for improving the validation of early detection biomarkers when specimen resources are limited.

Seminar Announcement

Ruyu Zhou
Statistics PhD Candidate
University of Notre Dame

"Differential Privacy and Statistics: Privatized Inference and the Inherent Privacy of Sampling"

Wednesday, March 18, 2026, 1:30pm-2:20pm (CT)
E0134 Howe Russell (HRKE)

Privacy-preserving data analysis has become a central challenge in modern statistics, with Differential Privacy (DP) emerging as the gold standard for protecting individual-level information.

In this talk, I will present two projects at the intersection of DP and statistics. First, focusing on privatized inference, I will introduce a general framework for privacy-preserving statistical inference that constructs privatized interval estimators via consistent, privatized posterior quantile estimation. I theoretically establish mean-squared error consistency for the proposed estimators and demonstrate improved privacy-utility tradeoffs through extensive empirical experiments.

Second, focusing on the inherent privacy guarantees provided by posterior sampling, I develop a unified Rényi divergence framework to quantify the DP guarantees achieved "for free" when releasing a single posterior sample.

These theoretical results substantially tighten existing conservative bounds and broaden the class of Bayesian models for which inherent DP guarantees can be rigorously characterized.