A Simple Information Criterion for Variable Selection in High-Dimensional Regression.

Area of Science:

• Statistics
• Bioinformatics
• Pharmacovigilance

Background:

• High-dimensional regression, common in genomics and drug studies, requires selecting sparse regressors.
• Existing criteria like AIC and BIC are too liberal due to unaddressed multiple testing in variable selection.

Purpose of the Study:

• To propose a new information criterion, the extended AIC (EAIC), for robust sparse model selection in high-dimensional regressions.
• To ensure asymptotic Family-Wise Error Rate (FWER) control in variable selection.

Main Methods:

• Developed the extended AIC (EAIC) formula incorporating log-likelihood, model size, total candidate regressors, and FWER target.
• Evaluated EAIC against AIC, BIC, mBIC, mAIC, and EBIC using LASSO in simulations across linear and logistic regression settings.
• Applied EAIC to detect adverse drug reaction signals in pharmacovigilance data.

Main Results:

• EAIC demonstrated effective FWER control across diverse regression settings, unlike AIC and BIC which showed numerous false positives.
• Simulation studies confirmed EAIC's superior variable selection performance compared to other criteria.
• The method proved effective in identifying potential adverse drug reactions from real-world pharmacovigilance data.

Conclusions:

• The extended AIC (EAIC) provides a statistically sound method for sparse model selection in high-dimensional regression.
• EAIC offers improved accuracy and control over false positives, crucial for applications like genomic analysis and pharmacovigilance.
• EAIC is a valuable tool for automated signal detection and reliable variable selection in complex datasets.