A1344
Title: A leakage-free nested cross-validation approach for adaptive penalized Cox models
Authors: Gul Inan - Koc University (Turkey) [presenting]
Izzet Goksel - Istanbul Technical University (Turkey)
Abstract: Adaptive penalized Cox models, including the adaptive Lasso and adaptive Elastic Net, are widely used for variable selection and risk prediction in high-dimensional survival analysis. However, standard implementations often estimate adaptive weights using the full dataset prior to Cross-Validation, introducing data leakage. This leads to optimistically biased risk estimates and overly complex, unstable models. A leakage-free Nested Cross-Validation (Nested CV) framework for calibrating adaptive Cox models is proposed. In this approach, all data-dependent steps, including initial estimation, adaptive weight construction, and hyper-parameter tuning, are performed strictly within training folds. This ensures proper separation between training and validation data, yielding statistically valid estimates of generalization performance and selection stability. The framework unifies adaptive Cox Lasso and adaptive Elastic Net formulations. Extensive simulations across varying dimensionality, correlation structures, signal strengths, and censoring levels show that conventional approaches systematically underestimate prediction error and inflate model complexity. In contrast, the proposed Nested CV framework provides approximately unbiased risk estimates, improved sparsity recovery, and greater selection stability. A real-data application further demonstrates its practical advantages for reliable and interpretable survival modeling.
