MULTILAYER KNOCKOFF FILTER: CONTROLLED VARIABLE SELECTION AT MULTIPLE RESOLUTIONS.

Area of Science:

• Statistics
• Genetics
• Bioinformatics

Background:

• Variable selection is crucial for understanding complex outcomes, especially in genetics where groups of variables (genes) are of interest.
• Existing methods often focus on individual variables, potentially missing group-level significance or leading to false discoveries.
• Controlling false discovery rates (FDR) is essential for replicable scientific findings.

Purpose of the Study:

• To develop a method for selecting important individual variables and groups of variables simultaneously.
• To control the false discovery rate at both individual variable and group levels.
• To improve the reliability and interpretability of findings in high-dimensional data analysis.

Main Methods:

• Introduced the Multilayer Knockoff Filter (MKF), building upon knockoff constructions and multilayer testing frameworks.
• Developed theoretical guarantees for simultaneous FDR control at multiple resolutions (individual variables and groups).
• Utilized simulations to evaluate MKF's performance against existing methods regarding power and false discovery control.

Main Results:

• MKF effectively controls the false discovery rate at both individual variable and group levels.
• Simulations demonstrate that MKF achieves strong FDR control with minimal power loss compared to single-resolution methods.
• Application to a genetic dataset successfully reduced false gene discoveries while maintaining significant power.

Conclusions:

• The Multilayer Knockoff Filter (MKF) provides a robust framework for variable and group selection with guaranteed FDR control.
• MKF offers a powerful tool for analyzing complex datasets, particularly in genomics, by balancing discovery and reliability.
• This method enhances the chances of meaningful and replicable discoveries in scientific research.