Epidemiology & Biostatistics PhD trainee Chenyu Liu presenting.
Overview: Logistic regression is a fundamental tool for modeling binary outcomes in biomedical studies, where the goal is to obtain interpretable odds ratios that support decision-making. Predictors can induce separation—wherein a single variable or linear
combination perfectly or nearly perfectly classifies cases and controls—causing maximum likelihood estimates to diverge. To address this problem, we develop a comprehensive detection-and-estimation framework. We develop a pre-hoc diagnosis to detect separation and identify problematic predictors or combinations, with a Separation Index that quantifies severity of separation. A key theoretical contribution is that under separation, the coefficient vector's direction remains identifiable, so pairwise ratios of nonzero coefficients are well-defined even when individual magnitudes diverge. Leveraging this, we propose a Bayesian estimator with multivariate exponential power prior that penalizes coefficients. For fair comparison, we construct a simulation framework that stably generates separation cases with ground truth. Simulations show our diagnostics discover more problematic data than existing methods, and our estimator controls bias better than widely used alternatives. In HIV-risk data, our framework outperforms alternatives by producing finite and sign-preserving odds ratios.
If unable to attend in person in Biomedical Research Building room 105, you may join via Zoom at
Meeting ID: 958 2937 2435
Passcode: 087450