Federated difference-in-differences with multiple time periods in DataSHIELD.

Area of Science:

• Econometrics
• Biostatistics
• Epidemiology

Background:

• Difference-in-differences (DID) is crucial for causal inference but hindered by privacy regulations in sensitive data analysis.
• Current methods face challenges with reduced sample sizes and statistical power due to consent requirements.
• Existing federated DID software is limited, restricting its application in privacy-sensitive research.

Purpose of the Study:

• To develop and validate a federated version of the Callaway and Sant'Anna difference-in-differences (CSDID) method.
• To integrate the federated CSDID approach into the DataSHIELD platform for secure, privacy-preserving analysis.
• To demonstrate the utility of federated DID for causal impact evaluation with sensitive, multi-site data.

Main Methods:

• Development of a federated CSDID algorithm compatible with DataSHIELD's privacy framework.
• Utilizing aggregated statistics instead of raw individual data for privacy preservation.
• Validation using simulated datasets and real-world data from a malaria intervention study in Mozambique.

Main Results:

• The federated CSDID approach successfully reproduced key estimates and standard errors from traditional DID analyses.
• Federated analysis demonstrated increased sample sizes and reduced estimation uncertainty compared to non-federated approaches.
• The method enabled causal impact evaluations even when direct sharing of treated or untreated group data was not feasible.

Conclusions:

• Federated learning offers a viable solution for conducting rigorous DID analyses on sensitive data, overcoming privacy barriers.
• The developed federated CSDID method enhances statistical power and analytical capabilities in multi-center or cross-border research.
• This approach facilitates broader and more effective evaluation of policy interventions and treatments globally.