A0658
Title: A robust minimum distance estimation of Cox PH models
Authors: Jingjing Wu - University of Calgary (Canada) [presenting]
Abstract: Cox proportional hazard (PH) models are simple but frequently used in survival analysis. The unknown coefficient parameters in a Cox PH model are usually estimated using the partial maximum likelihood estimation (PMLE) introduced in a past study. Nevertheless, PMLE is generally non-robust against model misspecification and outlying observations. When data is contaminated, PMLE produces inaccurate estimates with a large bias. A robust distance-based method is proposed instead, specifically, a minimum Hellinger distance estimation (MHDE). Both discrete and continuous covariates are considered, and different types of covariates are accommodated by introducing different versions of MHDE. Through an extensive simulation study, the finite-sample performance of the proposed MHDEs is examined and compared with the PMLE. Numerical results show that the proposed MHDEs are competitive with the PMLE under the true model, while they outperform the PMLE when outliers are present, which testifies to the robustness property of the MHDEs. The applications of the proposed MHDE are also demonstrated in real data analysis.
