A SEMIPARAMETRIC MULTIPLE IMPUTATION APPROACH TO FULLY SYNTHETIC DATA FOR COMPLEX SURVEYS

Area of Science:

• Statistics
• Data Science
• Survey Methodology

Background:

• Data synthesis is crucial for mitigating data disclosure risks in statistical analysis.
• Generating fully synthetic data from large, complex surveys presents modeling and application challenges.
• Existing methods may struggle with sophisticated analyses and data utility.

Purpose of the Study:

• To extend two-stage imputation for simultaneous missing value imputation and fully synthetic data generation.
• To develop a new combining rule for valid statistical inferences from synthetic data.
• To evaluate a novel semiparametric approach for generating synthetic data for skewed continuous and sparse binary variables.

Main Methods:

• Adapted two semiparametric missing data imputation models for synthetic data generation.
• Developed a new combining rule for statistical inference.
• Evaluated the approach using simulated and real longitudinal data (Health and Retirement Study).
• Compared the proposed method with parametric regression (IVEware) and nonparametric (synthpop) approaches.

Main Results:

• The proposed approach maintains high data utility across various descriptive and model-based statistics.
• It demonstrates superior performance compared to existing methods, particularly for complex analyses like factor analysis.
• Effective for both skewed continuous and sparse binary variables.

Conclusions:

• The extended two-stage imputation method offers an effective strategy for generating high-utility synthetic data from complex surveys.
• This approach enhances data privacy while preserving analytical integrity.
• It provides a robust alternative to existing data synthesis techniques for advanced statistical modeling.