Cross-validation choices can quietly reshape which variables Lasso selects in Cox models

Small, routine decisions in statistical modelling can change which variables end up in a survival model, according to a simulation study published in BMC Medical Research Methodology. The study examined Lasso-penalized Cox proportional hazards models fitted to low-dimensional real-world medical data and found that the number of cross-validation folds, the use of stratification, and the choice between Lasso and elastic net all influenced how consistently variables were selected from one analysis run to the next.

In survival analysis, it is often combined with cross-validation to choose the penalty parameter. The researchers note that this step introduces randomness, because the folds used in cross-validation depend on how the data are randomly split, and different splits can lead to different penalty values and therefore different sets of selected predictors.

The team, led by Bartosz Dobies and Dörte Huscher, drew on data from the Berlin Initiative Study in older adults. They worked with a subcohort of 930 participants who had measurements of a novel biomarker, and they assessed two time-dependent outcomes: all-cause mortality, with 882 participants analysed, and a cardiovascular composite endpoint, with 638 participants. In each case, Lasso was applied to identify prognostic variables among 20 medically preselected predictors, a low-dimensional setting in which the number of candidate variables is far smaller than the number of observations.

In their simulations, the researchers systematically varied the number of cross-validation folds, implemented stratification by covariates that violated the proportional hazards assumption, and adjusted the mixing parameter of the elastic net, a related method that blends the Lasso penalty with a ridge penalty. They evaluated stability by examining how much the number of selected predictors varied, and they assessed predictive performance using the C-index.

The number of cross-validation folds had a substantial influence on model size and on selection stability across random seeds. Higher fold numbers generally produced more consistent results, with 10-fold cross-validation performing best for the mortality outcome and 15-fold cross-validation performing best for the cardiovascular composite endpoint. In other words, the same data, analysed with the same method but different fold counts, could yield different sets of selected variables.

Stratification also mattered. When the researchers stratified by covariates that violated the proportional hazards assumption, stability improved in some cases. This suggests that accounting for known violations of the model’s assumptions during cross-validation can reduce the variability in which variables are selected, although the benefit was not uniform across all settings examined.

The elastic net did not consistently outperform Lasso in terms of stability or predictive accuracy in this setting. Mixing parameters below 0.6 led to greater variability in the selected variables.

The broader message of the study concerns how routine modelling choices, including default cross-validation settings, may shape variable selection stability in clinical survival analyses. Higher fold numbers and appropriate stratification can improve stability, but the authors note that these improvements come at a computational cost. Because the study was based on a single real-world dataset with preselected predictors and two outcomes, the findings describe the behaviour of these methods in this specific low-dimensional setting rather than establishing universal rules, and the authors frame their conclusions as a caution about default settings rather than a prescription for a single best workflow.

Subject of Research: Stability of variable selection in Lasso-penalized Cox proportional hazards models under different cross-validation designs

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

Article References: Dobies, B., & Huscher, D. (2026). Investigating the instability of variable selection in Lasso-penalized Cox proportional hazards models: a simulation study on low-dimensional real-world medical data. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03019-4

Image Credits: AI Generated

DOI: 10.1186/s12874-026-03019-4

Keywords: Lasso, Cox proportional hazards model, cross-validation, elastic net, variable selection, survival analysis, simulation study, model stability, biostatistics, Berlin Initiative Study, C-index, clinical prediction