Masked Autoencoder Transformer for Missing Data Imputation of PISA

This study introduces a scale-down transformer model to address the challenge of missing responses in educational assessments for psychometric evaluation. Traditional estimation methods for Item Response Theory (IRT) models are frequently computationally inefficient in generating estimates for a higher number of dimensions. The challenge becomes more pronounced when dealing with missing responses. We propose a Masked Autoencoder Transformer model for discrete input (DiTMAE) to impute missing answers for OECD-PISA response data. The model learns context information from the unmasked parts and reconstructs it using a decoder. For evaluation purposes, we estimate item and person parameters from two different approaches, (i) an adapted Variational Autoencoder that incorporates the Item Response Theory (IRT) method; (ii) the traditional statistical tool, Joint Maximum Likelihood (JML), that can produce estimates in occurrence of missing values.