Model Validation for Survival Analysis by Smoothed Predictive Likelihood.

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Assessing predictive performance is vital for survival model selection and evaluation.
• The predictive log-likelihood is a standard measure, but problematic for semiparametric/nonparametric models due to step-function survival curves.
• Existing solutions like Verweij's predictive partial likelihood are restricted to Cox models.

Purpose of the Study:

• To propose a novel, broadly applicable method for evaluating predictive performance in general survival models.
• To address the limitations of existing methods when dealing with step-function survival curves.
• To demonstrate the utility of the new method in model selection and tuning.

Main Methods:

• Nearest-neighbor kernel smoothing applied to survival model predictions.
• Development of a generalized predictive likelihood measure.
• Comparative analysis with existing methods in Cox and other survival models.

Main Results:

• The proposed kernel smoothing method provides a viable alternative for predictive likelihood in general survival models.
• The new method demonstrates competitive performance in the Cox model setting.
• The approach is applicable to testing for frailty terms and optimizing smoothness in penalized additive hazards models.

Conclusions:

• A novel kernel smoothing approach enhances predictive performance assessment in diverse survival models.
• This method broadens the applicability beyond Cox models, offering flexibility in model evaluation.
• The technique facilitates model selection, parameter tuning, and assessment of complex survival model features.