Consistent Estimation of Generalized Linear Models with High Dimensional Predictors via Stepwise Regression

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Penalized regression methods like LASSO are common for predictive modeling but yield biased estimates.
• Ultrahigh-dimensional data requires variable screening before penalized regression can be applied.
• Existing methods struggle with bias and variable selection in ultrahigh-dimensional settings.

Purpose of the Study:

• To propose a novel stepwise procedure for fitting generalized linear models with ultrahigh-dimensional predictors.
• To develop a method that provides a final model with consistent estimates and controls for false positives and negatives.
• To offer a reliable tool for accurately gauging the effect size of risk factors in complex datasets.

Main Methods:

• A stepwise procedure for fitting generalized linear models.
• Variable screening and selection integrated into a stepwise approach.
• Development of methods for consistent estimation in ultrahigh-dimensional settings.

Main Results:

• The proposed procedure yields a final model with consistent estimates.
• The method effectively controls both false negatives and false positives.
• Simulations and clinical study applications demonstrate the procedure's utility.

Conclusions:

• The stepwise procedure offers a robust solution for predictive modeling with ultrahigh-dimensional data.
• Consistent estimates improve the accuracy of risk factor effect size assessment.
• This method enhances decision-making by providing reliable predictive models.