The use of vector bootstrapping to improve variable selection precision in Lasso models.

Area of Science:

• Statistical Genetics
• Bioinformatics
• Computational Biology

Background:

• The Lasso (Least Absolute Shrinkage and Selection Operator) is a key shrinkage regression technique for variable selection in statistical genetics.
• K-fold cross-validation is commonly employed for fitting Lasso models, often supplemented by bootstrap confidence intervals to enhance selection precision.

Purpose of the Study:

• To systematically investigate the impact of nesting cross-validation within bootstrapping on Lasso variable selection precision (VSP).
• To compare the precision of nested cross-validation with standard Lasso methods and software defaults.

Main Methods:

• Simulation studies using genomic data under polygenic and typical Genome-Wide Association Study (GWAS) effect size models.
• Comparison of Lasso VSP with and without nested cross-validation, alongside bootstrapping and default software settings.
• Empirical data analysis using SNPs and SNP-SNP interactions from a GWAS of borderline personality symptoms.

Main Results:

• Nested cross-validation demonstrated superior variable selection precision at small effect sizes.
• No significant advantage of nesting was observed at larger effect sizes.
• The default Lasso approach selected low-reliability SNPs and interactions, which were excluded by the bootstrapping method in the empirical example.

Conclusions:

• Nesting cross-validation within bootstrapping offers enhanced precision for Lasso variable selection, particularly in scenarios with small effect sizes.
• The proposed nested approach provides more reliable SNP and interaction selection compared to default Lasso methods, as evidenced in empirical GWAS data.