A1675
Title: Regularized variable selection with missing data via stacked multiple imputation
Authors: Kyoji Furukawa - Kurume University (Japan) [presenting]
Natsumi Kumano - Kurume University (Japan)
Abstract: Missing data are ubiquitous in statistical analyses and pose challenges for estimation and variable selection. While multiple imputation (MI) is widely used, its direct application to variable selection is problematic, as different models may be selected across imputed datasets, complicating interpretation and inference. Stacking approaches combine multiply imputed datasets into a single weighted dataset, enabling standard model selection without relying on Rubins rules. A stacked MI approach that incorporates the analysis model through model-based weighting has been proposed, thereby avoiding complex procedures to ensure compatibility between imputation and analysis models. However, appropriate methods for variable selection within this framework remain unclear, and their performance has not been well examined. The proposed method integrates this approach with regularized variable selection techniques, such as LASSO and elastic net. It uses model-based weights to account for imputation uncertainty, enabling efficient penalized estimation within a single framework while remaining effective when the number of covariates is large. Through simulations under various missing data scenarios, the method improves model selection accuracy, reduces bias, and maintains predictive performance compared with existing methods. An application to a clinical dataset further demonstrates its practical utility and computational efficiency.
