Model selection based on logistic regression in a highly correlated candidate gene region

Area of Science:

• Genetics
• Statistical genetics
• Bioinformatics

Background:

• Identifying causal genetic variants is crucial for understanding disease mechanisms.
• Dense single-nucleotide polymorphism (SNP) panels present challenges due to high linkage disequilibrium (LD) and multicollinearity.
• Standard statistical methods can be hampered by the high dimensionality of genetic data.

Purpose of the Study:

• To develop and evaluate methods for identifying causal variants from dense SNP data.
• To address challenges posed by linkage disequilibrium (LD) in variable selection.
• To compare the performance of different logistic regression-based variable selection strategies.

Main Methods:

• Investigated three variable selection methods based on logistic regression.
• Applied penalties using Akaike's Information Criterion (AIC) and the LASSO penalty.
• Implemented a Bayesian variable-selection logistic regression model.
• Utilized simulated dense SNP data, including the causal DR/C locus.

Main Results:

• Evaluated model selection performance using average prediction error across nine replicates.
• Demonstrated the application of methods using simulated dense SNP data.
• Compared the performance of AIC, LASSO, and Bayesian selection methods.

Conclusions:

• The newly developed Bayesian variable selection method performs well for identifying causal variants in dense SNP data.
• The Bayesian approach effectively handles multicollinearity arising from high LD.
• This method shows promise for genetic association studies with high-density genotyping data.