Fully Synthetic Data for Complex Surveys.

Area of Science:

• Statistics
• Data Science
• Survey Methodology

Background:

• Releasing confidential survey data for public use is challenging.
• Statistical agencies need methods to generate privacy-preserving synthetic data.
• Existing methods may not fully account for complex survey designs.

Purpose of the Study:

• To propose a novel approach for generating fully synthetic data from surveys with complex sampling designs.
• To ensure the synthetic data accurately reflects the original data's properties for analysis.
• To facilitate variance estimation for analyses using the synthetic data.

Main Methods:

• Generate pseudo-populations using weighted finite population Bayesian bootstrap to handle survey weights.
• Draw simple random samples from pseudo-populations to estimate synthesis models.
• Employ multiple imputation with two data generation strategies for variance estimation.
• Develop and present multiple imputation combining rules for each strategy.

Main Results:

• The proposed method generates fully synthetic data suitable for public release from complex surveys.
• Simulation studies demonstrate the repeated sampling properties of the multiple imputation combining rules.
• Comparisons show favorable results against synthetic data generation based on pseudo-likelihood methods.
• The approach was successfully applied to American Community Survey data.

Conclusions:

• The proposed approach provides a statistically sound method for creating fully synthetic survey data.
• This facilitates the wider dissemination of valuable data while protecting respondent confidentiality.
• The method is robust and adaptable for various complex survey designs.