A Two-Step Variable Selection Strategy for Multiply Imputed Survival Data Using Penalized Cox Models | JoVE Visualize

Area of Science:

Statistics
Biostatistics
Data Science

Background:

Multiple imputation (MI) is a standard technique for addressing missing data in statistical analyses.
Applying penalized regression methods after MI presents challenges due to potential inconsistencies in variable selection across imputed datasets.
Developing robust variable selection strategies is crucial for reliable analysis of multiply imputed data, especially in survival modeling.

Purpose of the Study:

To propose and evaluate a novel two-step variable selection method for multiply imputed datasets with survival outcomes.
To compare the performance of the proposed method against alternative approaches, including stacked MI datasets with weighted penalized regression and group LASSO.
To investigate the impact of different penalization techniques and selection rules on variable selection stability and estimation accuracy.

Main Methods:

A two-step variable selection procedure involving LASSO or ALASSO on individual imputed datasets, followed by ridge regression and aggregation of selected variables based on inclusion frequency (any or d% of datasets).
Comparison with stacked MI datasets using weighted penalized regression and a group LASSO approach enforcing consistent selection.
Simulation studies using Cox models, evaluating performance metrics and employing various model tuning strategies (AIC, BIC, cross-validation, 1SE rule).

Main Results:

Performance varied significantly depending on the specific penalization method and selection rule employed.
Conservative approaches, such as ALASSO with BIC and a 50% inclusion frequency, demonstrated better control of false positives and improved calibration stability.
The grouped LASSO approach yielded comparable variable selection but was associated with slightly higher estimation errors.
No single method consistently outperformed all others across all simulated scenarios.

Conclusions:

The choice of penalized method and variable selection rule significantly impacts the analysis of multiply imputed survival data.
Practitioners must carefully consider the trade-offs between variable selection stability, estimation accuracy, and model calibration when selecting a method.
The proposed two-step method, particularly with conservative settings, offers a promising strategy for robust variable selection in multiply imputed survival data analysis.