Using item response theory as a methodology to impute categorical missing values.

Area of Science:

• Data Science
• Statistics
• Machine Learning

Background:

• Missing data is a common challenge in datasets, limiting statistical inference and model development.
• Existing imputation techniques have varying impacts on downstream analyses, including clinical score calculation and model testing.

Purpose of the Study:

• To evaluate an Item Response Theory (IRT) based approach for categorical data imputation.
• To compare the IRT method against established machine learning imputation techniques such as k-nearest neighbors (kNN), multiple imputed chained equations (MICE), and DataWig.

Main Methods:

• The IRT imputation method was applied to three datasets with ordinal, nominal, and binary categories.
• Datasets were manipulated to vary missing data proportions and systematization.
• Performance was assessed by accuracy in value reproduction and predictive performance post-imputation.

Main Results:

• The IRT approach demonstrated strong performance in categorical imputation.
• It outperformed several current multiple imputation methods under various conditions.
• The method showed particular promise in reproducing missing values and enhancing predictive accuracy.

Conclusions:

• Item Response Theory for categorical imputation provides a theoretically sound and effective alternative to existing methods.
• Its probabilistic approach to filling missing values offers advantages for data quality and subsequent analyses.
• The IRT method is a viable option for researchers dealing with missing categorical data.