Building generalized linear models with ultrahigh dimensional features: A sequentially conditional approach.

Area of Science:

• Statistics
• Bioinformatics
• Genomics

Background:

• Marginal screening methods can miss important variables.
• Existing conditional screening methods often rely on a fixed initial conditioning set, risking false positives/negatives.
• Current screening approaches frequently require tuning parameters and additional modeling steps.

Purpose of the Study:

• To develop an improved conditional screening approach that dynamically updates the conditioning set.
• To provide a method that avoids the need for pre-selected conditioning sets and tuning parameters.
• To enhance variable selection accuracy in statistical modeling.

Main Methods:

• A sequential conditioning approach with an iterative selection process was proposed.
• Theoretical properties were established within the framework of generalized linear models.
• An extended Bayesian information criterion was utilized as a stopping rule.

Main Results:

• The proposed method dynamically updates the conditioning set, improving selection accuracy.
• The approach successfully identifies marginally weak but conditionally important variables.
• The method eliminates the need for tuning or threshold parameters, simplifying model building.

Conclusions:

• The sequential conditioning approach offers a robust alternative to traditional screening methods.
• This method enhances the reliability of variable selection in statistical and genomic analyses.
• The approach was validated through simulations and a clinical study on multiple myeloma treatment response.