Prediction Modeling With Many Correlated and Zero-Inflated Predictors: Assessing the Nonnegative Garrote Approach.

Area of Science:

• Biostatistics
• Computational Biology
• Proteomics

Background:

• Mass spectrometry data presents challenges for prediction modeling due to correlated, zero-inflated features.
• Resource-intensive experiments necessitate feature reduction for concise and predictive models.

Purpose of the Study:

• To establish and examine regularized regression approaches for zero-inflated and correlated predictors.
• To introduce and evaluate a novel two-stage regularized regression approach, ridge-garrote.

Main Methods:

• Developed a novel two-stage regularized regression approach (ridge-garrote) using ridge and nonnegative garrotte estimators.
• Compared ridge-garrote with one-stage (ridge, lasso) and other two-stage (lasso-ridge, ridge-lasso) methods.
• Assessed predictive performance and predictor selection via simulation and a kidney function prediction case study using peptidomic data.

Main Results:

• Ridge-garrote consistently selected more parsimonious models than competitors in simulations.
• Lasso-ridge offered higher predictive accuracy but with high variability in selected predictors.
• Ridge-lasso showed slightly better accuracy than ridge-garrote but selected more noise predictors.

Conclusions:

• Ridge-garrote offers practical utility for selecting parsimonious predictor sets with minimal compromise in predictive accuracy.
• Ridge is suitable when variable selection is not a priority.
• The choice of method depends on the balance between predictive accuracy and model parsimony.