Investigating the instability of variable selection in Lasso-penalized Cox proportional hazards models: a simulation study on low-dimensional real-world medical data

Background This study examined the stability of Lasso-penalized Cox proportional hazards models in low-dimensional clinical data. While Lasso is a popular method for simultaneous variable selection and regularization, its performance is known to vary across cross-validation (CV) runs due to random fold assignments. We aim to quantify the effects of the number of folds, stratification procedures, and elastic net mixing parameters on variable selection consistency. Methods We analyzed real-world data from the Berlin Initiative Study in older adults, focusing on a subcohort with measurements of a novel biomarker ( n = 930) and assessing two time-dependent outcomes: all-cause mortality ( n = 882) and a cardiovascular composite endpoint (CE, n = 638). We applied Lasso to identify prognostic variables among 20 medically preselected predictors. In simulation studies, we systematically varied the number of CV folds, implemented stratification by nonproportional covariates, and adjusted the elastic net’s mixing parameter. We evaluated model stability by examining variability in the number of selected predictors and predictive performance via the C-index. Results In our datasets, the number of CV folds substantially influenced model size and selection stability across random seeds, with higher fold numbers generally yielding more consistent results (optimal: 10-fold for mortality, 15 for CE). Stratifying by covariates, which violated the proportional hazards assumption, enhanced stability in some cases. The elastic net did not consistently outperform Lasso in terms of stability or predictive accuracy; mixing parameters below 0.6 led to greater variability. Conclusions The stability of Lasso-penalized Cox models is strongly affected by CV design. Higher fold numbers and appropriate stratification can improve stability, but at a computational cost. Elastic net offers no clear benefit in this setting. These findings highlight how routine modelling choices, including default cross-validation settings, may influence variable selection stability in low-dimensional clinical survival analyses.