Techniques to produce and evaluate realistic multivariate synthetic data.

Area of Science:

• Data Science
• Statistical Modeling

Background:

• Reproducible data modeling necessitates adequate sample sizes.
• Small sample sizes can hinder robust model evaluation and validation.
• Existing methods struggle with data scarcity in complex modeling scenarios.

Purpose of the Study:

• To evaluate a synthetic data generation technique for addressing small sample size issues in data modeling.
• To determine if synthetic data generated via kernel density estimation (KDE) is statistically similar to original samples.
• To assess the utility of this approach for datasets with latent multivariate normal characteristics.

Main Methods:

• Investigated three samples (n=667) with 10 input variables (X).
• Augmented sample size in X using univariate kernel density estimation (KDE).
• Transformed variables to T, approximated probability density functions as normal, and generated synthetic data.

Main Results:

• Generated synthetic data in Y and X by reversing the transformation steps.
• All samples approximated multivariate normality in Y, enabling synthetic data generation.
• Probability density function and covariance comparisons confirmed similarity between original and synthetic samples.

Conclusions:

• The evaluated synthetic data generation technique effectively addresses small sample size problems for data with latent normal characteristics.
• This method enhances reproducibility and facilitates model evaluation in data-scarce situations.
• Further research is needed to fully elucidate the properties of the latent class.