报告时间：2025年09月22日 （周一）上午09:00-10:00

报告地点：腾讯会议号：220265323

报告人：李元章教授 乔治华盛顿大学

摘要：Identifying significant factors from high-dimensional datasets remains a critical challenge in biomedical research, particularly when the outcome of interest is binary. We propose a novel approach using Gradient Orthogonal Basis (GOB) decomposition to efficiently reduce dimensionality and select informative variables in logistic regression models. The method decomposes the factor space into gradient-based orthogonal directions, capturing directions with maximal discriminatory power while controlling for noise and correlation structures. Model fitting proceeds via conditional logistic regression and generalized estimating equations (GEE), allowing for flexible handling of correlation and clustering. Variable selection is guided through statistical tests including Wald, Score, AIC, and QIC, alongside interaction assessments and sum-based statistics for robustness.

We apply our method to biomarker datasets involving schizophrenia and bipolar disorder, where it demonstrates improved power, interpretability, and consistency over traditional penalized or projection-based methods. Simulation studies further validate its effectiveness in scenarios with complex correlation patterns and moderate sample sizes. Our results highlight the GOB framework as a promising direction for interpretable and statistically rigorous modeling of high-dimensional binary outcome data.