Seminar by Lu Xia

Please Note: This seminar will be given in person.

Reliable High-dimensional Inference Beyond Linear Regression: Challenges and Opportunities

Modern technologies have made it possible to collect a large amount of information in biomedical studies, where the number of covariates is comparable to or even larger than the sample size. Statistical methods for reliable inference on regression parameters in the presence of high-dimensional covariates are warranted. While much of the existing literature focuses on linear regression, other widely adopted models for analysis of binary, count, time-to-event or correlated data in biomedical research pose additional challenges in both theoretical development and empirical performance.

In this talk, I will first introduce a projection-based approach for inference on linear combinations of regression parameters in generalized estimating equations, under the “large p, small n” regime, to analyze correlated data. Then, I will present a de-biased lasso approach for drawing inference on stratified Cox models under the “large n, diverging p” regime. The proposed methods are shown to enjoy more reliable empirical performance, especially in estimation bias and confidence interval coverage, than their competitors, and are applied to analyses of large omics and the Scientific Registry of Transplant Recipients data, respectively. Finally, I will briefly discuss my work on other areas and how these advances together lay a foundation for my future research.