A systematic review and evaluation of statistical methods for group variable selection.

Area of Science:

• Statistics
• Bioinformatics

Background:

• Variable selection is crucial for high-dimensional datasets, especially those with inherent group structures.
• Regularized regression techniques are widely used for feature selection in statistical modeling.

Approach:

• A systematic literature review following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines identified 14 R-implemented variable selection methods.
• Methods were classified into knowledge-driven (group-level, bi-level) and data-driven (two-step, collinearity-tolerant) approaches based on group structure handling.
• Performance was evaluated via a simulation study comparing identified methods.

Key Points:

• Group-level selection methods (e.g., group minimax concave penalty) are effective for identifying relevant variable groups.
• Bi-level selection methods (e.g., group bridge) are superior for pinpointing individual predictive variables within groups.
• Two-step and collinearity-tolerant methods (e.g., elastic net) offer results without prior knowledge but are generally less performant than knowledge-driven approaches.

Conclusions:

• The choice of variable selection method depends on the research question and the availability of prior knowledge about group structures.
• Group-level methods are recommended when identifying entire predictive groups is the priority.
• Bi-level methods are suggested for scenarios requiring precise identification of individual important variables within groups, particularly when group homogeneity is uncertain.