Fitting and Cross-Validating Cox Models to Censored Big Data With Missing Values Using Extensions of Partial Least...

Area of Science:

• Biostatistics
• Computational Statistics
• Big Data Analytics

Background:

• Fitting Cox models in big data settings with missing values is computationally challenging.
• Existing analytic tools struggle with the volume, intensity, and complexity of big data.
• Partial Least Squares (PLS) regression extensions offer potential for high-dimensional Cox models, including those with missing data.

Purpose of the Study:

• To evaluate the efficacy of standard cross-validation (CV) criteria for PLS extensions of Cox models in big data.
• To identify and validate new CV criteria that perform reliably with these advanced models.
• To enhance the performance of PLS-based Cox models through improved hyperparameter selection.

Main Methods:

• Simulation studies using three distinct data generation algorithms.
• Evaluation of over a dozen cross-validation criteria, including AUC and prediction error-based methods.
• Development and application of sparse group extensions for PLS Cox models and a new performance metric (integrated R Schmid Score weighted).

Main Results:

• Standard cross-validation schemes (naive and van Houwelingen) demonstrate failure when applied to PLS extensions of Cox models.
• Several newly evaluated cross-validation criteria effectively select an appropriate number of components.
• Benchmark reanalysis using the validated CV criteria shows enhanced performance of PLS-based Cox models.

Conclusions:

• Traditional cross-validation methods are unreliable for hyperparameter tuning in PLS-extended Cox models for big data.
• The proposed cross-validation criteria offer a robust solution for selecting hyperparameters, leading to improved model performance.
• New sparse group extensions and performance metrics further advance the utility of PLS for complex survival data analysis.