Perfect Separation Project

Photo by rawpixel on Unsplash

Logistic regression is widely used for binary outcomes but struggles with perfect separation, causing inflated odds ratios and unstable predictions. We propose a pre-hoc method to detect both direct and latent perfect separation and show that although latent separation leads to infinite estimates, the ratio of coefficients converges to a true value. Incorporating this into a Bayesian Power Prior framework corrects inflated estimates and improves convergence. Simulations and real-world examples show our method outperforms the Firth correction, enhancing model accuracy and interpretability. A user-friendly R package is to be available at https://github.com/bioscinema.

Liangliang(Lyon) Zhang
Liangliang(Lyon) Zhang
Assistant Professor

Dr. Zhang’s research interests center around Bayesian inference and prediction, high dimensional models, and complex structured data, such as brain imaging and metagenomic data.