Cross-site imputation can recover missing variables in federated multicenter studies

Area of Science:

• Epidemiology
• Biostatistics
• Observational Research

Background:

• Multisite studies often face challenges with missing key variables at certain sites.
• Pooling data across sites can be logistically or legally infeasible.
• Existing imputation methods may not be suitable when data pooling is restricted.

Purpose of the Study:

• To introduce a novel multiple imputation method called cross-site imputation.
• To enable the recovery of missing variables across study sites without individual-level data pooling.
• To address data limitations in multisite observational research.

Main Methods:

• Cross-site imputation utilizes predicted regression coefficients and variances from sites with observed data.
• It imputes missing variables at sites lacking recorded data.
• The method was illustrated using Swedish hospital data to recover missing confounders.

Main Results:

• Cross-site imputation effectively recovered systematically missing confounding variables.
• Imputation was successful independently at sites where data were initially missing.
• The method facilitated the inclusion of all hospitals in the final, fully adjusted analysis.

Conclusions:

• Cross-site imputation presents a practical solution for handling missing variables in multisite studies.
• This method is valuable given the growing reliance on multisite research designs.
• It offers a viable alternative when data pooling is not an option.