A general, prediction error-based criterion for selecting model complexity for high-dimensional survival models.

Area of Science:

• Biostatistics
• Bioinformatics
• Computational Biology

Background:

• High-dimensional data in survival analysis presents challenges in model complexity selection.
• Overfitting is a significant concern, necessitating robust model selection criteria.
• Predictive partial log-likelihood (PLL) via cross-validation is a common but not universally applicable method.

Purpose of the Study:

• To introduce and evaluate a relative integrated prediction error curve (IPEC) as an alternative criterion for selecting model complexity in predictive survival models.
• To compare the performance of IPEC with traditional PLL-based methods, particularly in scenarios where PLL is not available.
• To assess the impact of using bootstrap resampling versus cross-validation for estimating model performance.

Main Methods:

• A simulation study was conducted using microarray survival data to mimic real-world high-dimensional scenarios.
• The performance of model selection using bootstrap-estimated IPEC was compared against cross-validation-estimated PLL.
• Bootstrap resampling was also explored as an alternative to cross-validation for estimating PLL.

Main Results:

• Bootstrap estimation of PLL generally yielded similar prediction performance compared to cross-validation estimates.
• Model selection using bootstrap-estimated IPEC performed comparably to cross-validation-estimated PLL.
• The proposed IPEC method proved effective in a real-world microarray survival dataset from diffuse large-B-cell lymphoma patients.

Conclusions:

• The relative integrated prediction error curve (IPEC) is a viable and stable alternative criterion for model complexity selection in high-dimensional survival analysis.
• IPEC is particularly advantageous when predictive partial log-likelihood (PLL) is not available or difficult to compute.
• Bootstrap resampling provides a reliable alternative for estimating performance metrics like IPEC and PLL, ensuring robust model selection.