A Framework for Generating Realistic Synthetic Tabular Data in a Randomized Controlled Trial Setting.

Area of Science:

• Health Informatics
• Biostatistics
• Data Science

Background:

• Realistic synthetic data generation is crucial for health research, aiding data sharing and privacy protection.
• Generating complex synthetic tabular data, particularly for randomized controlled trials (RCTs), remains a significant challenge.
• Current methods lack consensus on preserving multivariate data distributions for synthetic tabular RCT data.

Purpose of the Study:

• To compare strategies and techniques for generating realistic synthetic tabular data for randomized controlled trials (RCTs).
• To identify the most effective method for preserving underlying data distributions in synthetic RCT datasets.
• To address the need for reliable synthetic data in epidemiological and clinical studies.

Main Methods:

• Empirical comparison of several data generation strategies and three techniques (two machine learning, one statistical).
• Utilized an R-vine copula model for generating baseline variables.
• Employed regression models for post-treatment allocation variables, mimicking RCT outcomes.

Main Results:

• A sequential generation approach proved most effective for creating synthetic tabular RCT data.
• The R-vine copula model successfully generated realistic baseline variables.
• Subsequent regression models accurately captured post-treatment allocation variable features, including trial outcomes.

Conclusions:

• The proposed sequential generation strategy, combining R-vine copulas and regression models, is the optimal method for generating synthetic tabular RCT data.
• This approach effectively preserves realistic data features and multivariate distributions.
• The findings offer a robust solution for creating privacy-preserving synthetic data for clinical trials.