Columnwise Neural Imputation for Incomplete Ordinal Psychometric Data

This paper introduces columnwise neural imputation (COLNI), a novel neural network-based approach for imputing missing ordinal responses in psychological and educational assessments. COLNI operates in a columnwise manner, using observed responses to other items to predict missing values for each target item. Its neural network architecture incorporates a Compensator layer to accommodate incomplete input patterns and a consistent rank logits (CORAL) output layer to preserve the ordinal structure of responses. By combining dropout with logit-space nearest-neighbor sampling, COLNI generates multiple plausible completed datasets to account for imputation uncertainty. A comprehensive simulation study evaluated its performance under various missingness mechanisms, missingness rates, and latent dimensionalities. Results demonstrated that COLNI outperformed conventional imputation methods in recovering key psychometric quantities, including item means, inter-item correlations, and person and item parameters under the multidimensional graded response model. An empirical study using data from the Short Dark Triad test further demonstrated its effectiveness in practical multidimensional assessment settings.

Recommended citation: Zhang, L., Jeon, M., & Chen, P. (2026). Columnwise neural imputation for incomplete ordinal psychometric data. Psychological Methods. Advance online publication. https://doi.org/10.1037/met0000867
Download Paper