Bayesian variable selection for binary outcomes in high-dimensional genomic studies using non-local priors.

Area of Science:

• Genomics
• Statistical Genetics
• Computational Biology

Background:

• Genomic technologies generate massive datasets requiring advanced statistical and computational methods.
• Variable selection for binary response prediction is crucial in genomic data analysis.
• Existing penalized likelihood methods have limitations in handling ultrahigh-dimensional genomic data.

Purpose of the Study:

• To propose a novel Bayesian method, iMOMLogit, for selecting explanatory variables in genomic data.
• To improve model identification and reduce prediction error compared to existing methods.
• To provide a computationally efficient algorithm for ultrahigh-dimensional genomic data analysis.

Main Methods:

• Utilized a Bayesian approach with a mixture of non-local prior densities and point masses.
• Developed a novel method for setting prior hyperparameters based on total variation distance.
• Implemented an efficient computational algorithm for ultrahigh-dimensional settings ([Formula: see text]).

Main Results:

• iMOMLogit demonstrated improved performance in simulation studies for model identification and error reduction.
• Application to genomic datasets yielded high-accuracy predictions with significantly fewer variables than competing methods.
• The developed algorithm effectively identifies the highest posterior probability model with provided diagnostics.

Conclusions:

• iMOMLogit offers a powerful and efficient Bayesian solution for variable selection in genomic data analysis.
• The method achieves superior prediction accuracy and parsimony in variable selection.
• The computational framework supports the analysis of ultrahigh-dimensional genomic datasets.