High Dimensional Variable Selection with Error Control.

Area of Science:

• Genomics
• Biostatistics
• Bioinformatics

Background:

• Iterative Sure Independence Screening (ISIS) is common for high-throughput data but computationally intensive and prone to high false discovery rates (FDR).
• Controlling FDR is crucial for reliable variable selection in complex datasets.
• Existing methods may struggle with the balance between computational cost and accuracy.

Purpose of the Study:

• To propose and evaluate a novel screening method using FDR to reduce dimensionality.
• To control FDR while applying variable selection methods like LASSO, SCAD, and MCP.
• To improve the efficiency and reliability of variable selection in high-dimensional data analysis.

Main Methods:

• A new screening approach utilizing FDR was developed.
• The proposed FDR screening was integrated with LASSO, SCAD, and MCP variable selection methods.
• The methods were tested via simulations and applied to prostate cancer data predicting metastasis.

Main Results:

• Simulations confirmed that the proposed screenings controlled FDR and yielded high Area Under the Receiver Operating Characteristic Curve (AUROC) scores.
• In the prostate cancer data, LASSO and MCP achieved AUROC scores of 0.746 and 0.764, respectively.
• The integrated methods demonstrated effective FDR control and strong predictive performance.

Conclusions:

• Variable selection methods combined with sequential FDR and ISIS screening effectively control FDR in final models.
• The proposed approach offers a computationally efficient and reliable alternative for high-dimensional data analysis.
• This method shows promise for improving biomarker discovery and prediction accuracy in complex diseases.