Variable Selection via Knockoffs in Missing Data Settings with Categorical Predictors.

Area of Science:

• Statistics
• Machine Learning
• Educational Measurement

Background:

• Large-scale assessment data often contain numerous variables with missing values, posing challenges for predictor selection.
• Traditional variable selection methods, like knockoffs, may not effectively handle missing data or unordered categorical predictors.

Purpose of the Study:

• To extend the knockoffs method for predictor selection to accommodate datasets with missing values.
• To develop a flexible and effective framework for variable selection in complex, real-world datasets.

Main Methods:

• A preliminary multiple imputation (MI) phase to address missing values.
• Application of a knockoff filter to each imputed dataset for variable selection.
• Evaluation through simulation studies and application to INVALSI large-scale assessment data.

Main Results:

• The proposed method demonstrated satisfactory performance in simulation studies.
• The approach yielded effective results when applied to Italian grade 5 student test score data.
• The method successfully handled datasets with numerous unordered categorical predictors and missing values.

Conclusions:

• Implementing the knockoffs method within a multiple imputation framework is a feasible, flexible, and effective strategy for variable selection.
• This integrated approach addresses key limitations of traditional methods in handling missing data and complex predictor types.