Partial likelihood estimation of isotonic proportional hazards models

Area of Science:

• Biostatistics
• Survival Analysis
• Statistical Modeling

Background:

• Semiparametric proportional hazards models are crucial for analyzing survival data.
• Estimating monotone covariate effects with unspecified baseline hazards presents computational challenges.
• Existing methods like nonparametric maximum likelihood and iterative quadratic programming have limitations in terms of computational intensity and numerical stability.

Purpose of the Study:

• To develop a computationally efficient and theoretically justified method for estimating monotone covariate effects in semiparametric proportional hazards models.
• To address the limitations of existing estimation techniques, particularly for right-censored data.
• To provide a stable and practical alternative for analyzing complex survival data, including time-dependent covariates.

Main Methods:

• Partial likelihood estimation is employed for the semiparametric proportional hazards model.
• A novel pseudo-iterative convex minorant algorithm is proposed, leveraging pool-adjacent-violators techniques.
• The algorithm is extended to handle time-dependent covariates and provides a separate estimator for the baseline hazard function.

Main Results:

• The proposed pseudo-iterative convex minorant algorithm demonstrates significant computational speed improvements (orders of magnitude) compared to iterative quadratic programming and iterative convex minorant algorithms.
• The new algorithm exhibits enhanced computational stability and theoretical justification.
• Simulation studies indicate moderate reductions in bias and variance of the estimators.
• Analysis of HIV prevention study data highlights the practical utility of the isotonic methodology.

Conclusions:

• The pseudo-iterative convex minorant algorithm offers a computationally stable and efficient solution for semiparametric proportional hazards models with monotone covariate effects.
• This methodology is valuable for analyzing complex survival data, including nonlinear covariate effects and time-dependent covariates.
• The approach has practical implications for various fields, such as medical research and epidemiology, as demonstrated by the HIV study analysis.