Criticality of predictors in multiple regression.

Area of Science:

• Statistics
• Regression Analysis
• Computational Statistics

Background:

• Traditional methods for assessing predictor importance in multiple regression have limitations.
• Existing indices may not fully capture a predictor's role in identifying the best-fitting model.

Purpose of the Study:

• To introduce a new method for comparing predictors in multiple regression models.
• To develop a measure of predictor criticality distinct from traditional importance indices.

Main Methods:

• Utilizes the bootstrapping (resampling with replacement) procedure to generate numerous data samples.
• Fits all 2^p subset regression models for each sample and identifies the best subset model.
• Defines predictor criticality based on the probabilities of subsets including that predictor being the best.

Main Results:

• Generates the probability distribution for the best subset model across all possible subsets.
• Quantifies predictor criticality as a function of probabilities associated with models containing the predictor.
• Demonstrates the procedure's applicability to various regression models and goodness-of-fit measures.

Conclusions:

• The proposed predictor criticality measure offers advantages over traditional importance indices.
• Predictors included in a higher number of probable best-fitting models are identified as critical.
• The method provides a robust approach for understanding predictor roles in multiple regression and enhancing predictive modeling.