Efron-type measures of prediction error for survival analysis

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Accurate prediction error estimation is crucial for developing and applying statistical models.
• Survival analysis with right-censored data presents unique challenges for error estimation.
• Existing resampling tools need adaptation for complex models in censored survival data.

Purpose of the Study:

• To adapt Efron and Tibshirani's resampling tools for survival analysis with right-censored event times.
• To enable the assessment and comparison of flexible statistical models against less flexible ones.
• To provide a robust method for evaluating predictive accuracy in survival data.

Main Methods:

• Adaptation of resampling techniques (e.g., bootstrapping) for right-censored survival data.
• Application of methods to evaluate flexible predictive rules including artificial neural networks, classification and regression trees, and regression splines.
• Comparative analysis of model performance on the same dataset.

Main Results:

• The adapted resampling methods effectively assess prediction error for flexible models in survival analysis.
• Flexible models (neural nets, trees, splines) can be reliably compared to less flexible models.
• The methodology is demonstrated as effective using real-world breast cancer trial data.

Conclusions:

• The adapted resampling approach provides a valuable tool for statistical model development in survival analysis.
• This method facilitates rigorous comparison of diverse modeling strategies, including advanced machine learning techniques.
• The findings support the use of these techniques for improving predictive accuracy in clinical trial data analysis.