Neural Network Copulas for Generating Synthetic Test Data Preserving Psychometric Properties.

Area of Science:

• Psychometrics
• Artificial Intelligence
• Data Science

Background:

• Sharing item response data in intelligence research is limited by privacy concerns.
• Traditional simulation methods struggle with complex response dependencies in cognitive assessments.

Purpose of the Study:

• To propose a novel neural network copula (NNC) framework for generating synthetic dichotomous item response data.
• To ensure the generated data preserves essential psychometric properties while safeguarding examinee privacy.

Main Methods:

• Utilized a deep autoencoder and kernel density estimation to model marginal probabilities and dependence structures separately.
• Developed a framework to handle the discrete nature of binary item response data with minimal distributional assumptions.

Main Results:

• NNC-based synthetic data demonstrated high correspondence with empirical data across various facets.
• Reproduces total score distributions and inter-item correlations accurately.
• Yielded consistent estimates for item characteristic curve parameters, item fit statistics, and test information functions.

Conclusions:

• The proposed NNC framework offers a robust method for generating privacy-preserving synthetic item response data.
• Demonstrated algorithmic stability and inferential precision through Monte Carlo replications.
• Facilitates data sharing in intelligence research without compromising sensitive examinee information or psychometric integrity.