Adjusted score functions for monotone likelihood in the Cox regression model.

Area of Science:

• Biostatistics
• Survival Analysis
• Medical Data Analysis

Background:

• Standard Cox model inference relies on maximizing the partial likelihood function.
• Monotone partial likelihood is a frequent issue in health science studies, especially with censored survival data and categorical covariates.
• This problem often arises when a categorical covariate level exclusively contains censored observations.

Purpose of the Study:

• To address the problem of monotone partial likelihood in Cox regression models.
• To propose and evaluate an alternative bias reduction method for improved Cox model inference.
• To compare the performance of the proposed median bias reduction approach against existing mean bias reduction methods.

Main Methods:

• Adapted an adjusted score function for median bias reduction, building on recent work by Kenne Pagui et al.
• Investigated the invariance properties of the proposed method under reparameterizations, crucial for hazard ratio interpretation.
• Conducted numerical simulations to assess the statistical properties and performance of the new method.
• Applied the methods to a real-world melanoma skin dataset for practical illustration and comparison.

Main Results:

• The proposed median bias reduction method effectively prevents infinite estimates, a common issue with monotone likelihood.
• The method demonstrates invariance under componentwise reparameterizations, simplifying hazard ratio interpretation.
• Numerical studies indicate superior inference properties compared to existing mean bias reduction techniques.
• The melanoma dataset analysis provided a practical basis for comparing the methods.

Conclusions:

• The proposed median bias reduction approach offers a robust solution to monotone partial likelihood in Cox models.
• This method enhances the reliability of statistical inference in survival analysis, particularly in challenging datasets.
• The approach provides a valuable alternative for researchers encountering estimation problems in Cox regression.