Transformers deep learning models for missing data imputation: an application of the ReMasker model on a psychometric...

Area of Science:

• Psychometrics
• Data Science
• Machine Learning

Background:

• Missing data is a significant challenge in psychometric research, potentially compromising study reliability and validity.
• Traditional imputation methods (e.g., mean imputation, Expectation-Maximization) often fail to meet the assumptions of psychological data, leading to biased results.
• Developing robust methods for handling missing data is crucial for accurate psychometric analysis.

Purpose of the Study:

• To evaluate the effectiveness of transformer-based deep learning for imputing missing data in psychometric research.
• To compare a novel transformer model, ReMasker, against conventional and other machine learning imputation techniques.
• To assess the performance of different imputation methods using a real-world psychometric dataset.

Main Methods:

• A masking autoencoding transformer model (ReMasker) was developed and compared with mean/median imputation, Expectation-Maximization (EM), K-nearest neighbors (KNN), MissForest, and Artificial Neural Networks (ANN).
• A psychometric dataset from the COVID distress repository was utilized for the evaluation.
• Imputation performance was quantified using the Root Mean Squared Error (RMSE) between original and imputed data matrices.

Main Results:

• Transformer-based models, specifically ReMasker, demonstrated superior performance in data reconstruction compared to conventional imputation methods.
• Machine learning approaches, including ReMasker, consistently outperformed traditional techniques across all tested scenarios.
• ReMasker achieved the lowest reconstruction error, indicating more accurate data imputation.

Conclusions:

• Transformer-based deep learning models offer a robust and effective solution for addressing missing data in psychometric research.
• The superior performance of ReMasker highlights the potential of advanced AI techniques to enhance data integrity and the generalizability of findings in psychological studies.
• These findings advocate for the adoption of advanced imputation methods to improve the quality of psychometric research outcomes.