Variable selection and estimation in generalized linear models with the seamless L₀ penalty.

Area of Science:

• Statistics
• Computational Biology
• Genetics

Background:

• Variable selection is crucial in generalized linear models (GLMs) for identifying relevant predictors.
• Existing methods may struggle with high-dimensional data and weak signals.
• The L0 penalty offers desirable sparsity but is computationally challenging.

Purpose of the Study:

• To propose a novel penalized likelihood approach, seamless L0 (SELO), for variable selection and estimation in GLMs.
• To develop an efficient algorithm for fitting SELO-GLM models.
• To evaluate the performance of the SELO-GLM/BIC procedure in terms of model selection consistency and finite sample performance.

Main Methods:

• The seamless L0 (SELO) penalty, a smooth approximation to the L0 penalty, is utilized.
• An efficient algorithm is developed for fitting the SELO-GLM.
• A Bayesian Information Criterion (BIC) is proposed for tuning parameter selection.
• Simulation studies are conducted to assess finite sample performance.

Main Results:

• The SELO-GLM procedure demonstrates the oracle property even with a diverging number of variables.
• The SELO-GLM/BIC procedure consistently selects the true model under regularity conditions.
• Simulation results indicate superior finite sample performance compared to existing methods, particularly for large, high-dimensional datasets with weak signals.

Conclusions:

• The proposed SELO-GLM/BIC method provides an effective and efficient approach for variable selection in GLMs.
• This method shows promise for analyzing complex genetic data, such as identifying single nucleotide polymorphisms (SNPs) associated with breast cancer risk.