Broken adaptive ridge method for variable selection in generalized partly linear models with application to the...

Area of Science:

• Statistics
• Biostatistics
• Computational Biology

Background:

• Generalized linear models (GLMs) are widely used but assume specific covariate relationships.
• High-dimensional data, common in genomics, pose challenges for traditional statistical modeling.
• Generalized partly linear models (GPLMs) offer flexibility by including non-parametric components.

Purpose of the Study:

• To develop a novel statistical method for simultaneous variable selection and parameter estimation in GPLMs with high-dimensional covariates.
• To address limitations of existing methods in handling complex covariate effects and large datasets.
• To apply the new method to real-world data for identifying factors associated with coronary artery disease.

Main Methods:

• Developed the broken adaptive ridge (BAR) estimator, approximating L0-penalized regression.
• Utilized iterative reweighted squared L2-penalized regression for implementation.
• Employed Bernstein polynomials as a sieve space for approximating non-parametric functions.
• Leveraged existing R packages for practical implementation.

Main Results:

• The BAR estimator demonstrated superior performance compared to other penalty-based variable selection methods in extensive simulation studies.
• Application to the CATHGEN coronary artery disease dataset yielded novel findings regarding genetic and non-genetic covariates.
• The method effectively handles high-dimensional covariates and simultaneous estimation and selection.

Conclusions:

• The BAR estimator is an effective and flexible statistical tool for analyzing high-dimensional data within the GPLM framework.
• The method provides a robust approach for variable selection and parameter estimation, outperforming existing techniques.
• This approach offers valuable insights for complex disease studies, such as coronary artery disease, by integrating diverse covariate types.