Resampling-based tests for Lasso in genome-wide association studies.

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Genome-wide association studies (GWAS) typically use univariate regression to test millions of genetic variants against phenotypes.
• Lasso penalized regression offers an alternative for jointly modeling multiple genetic variants and phenotypes.
• Inference on individual Lasso coefficients, particularly in high-dimensional genetic data, remains a challenge.

Purpose of the Study:

• To evaluate six methods for inference on Lasso coefficients in high-dimensional genetic settings.
• To identify robust and efficient methods for hypothesis testing within penalized regression models for GWAS.

Main Methods:

• Six inference methods were evaluated: Lasso-Ayers, Lasso-PL, Lasso-AL, Lasso-RB, Lasso-MRB, and Lasso-PT.
• Methods were compared using simulations and an application to the Minnesota Center for Twins and Family Study.
• Specific focus on type-1-error control and power in the presence of numerous null predictors.

Main Results:

• Lasso-RB, Lasso-MRB, and Lasso-PT were found to be unreliable for inference with increasing numbers of null predictors in finite samples.
• Lasso-PL and Lasso-AL demonstrated sustained speed and power for inference, even in high-dimensional scenarios.
• The permutation selection procedure (Lasso-PL) and analytic selection method (Lasso-AL) proved effective.

Conclusions:

• Lasso-PL and Lasso-AL are recommended as fast and powerful alternatives to univariate analysis in GWAS.
• These methods provide viable solutions for conducting statistical inference with Lasso regression in high-dimensional genetic data.
• The findings support the use of Lasso-PL and Lasso-AL for robust genetic association analysis.