Combining distributed regression and propensity scores: a doubly privacy-protecting analytic method for multicenter...

Area of Science:

• Biostatistics
• Health Informatics
• Epidemiology

Background:

• Sharing detailed individual-level data in multicenter studies presents significant privacy and logistical challenges.
• Analytic methods utilizing summary-level data offer a potential solution for privacy-preserving multivariable analyses.

Purpose of the Study:

• To assess the feasibility and validity of multivariable-adjusted distributed linear regression.
• To evaluate the combination of distributed linear regression with propensity scores in a large distributed data network.

Main Methods:

• Compared percent total weight loss 1-year post-surgery between Roux-en-Y gastric bypass and sleeve gastrectomy using data from 43,110 patients across 36 health systems.
• Employed distributed linear regression, which uses only summary-level data (sums of squares and cross products matrix), to fit three regression models.
• Adjusted for baseline variables using individual covariates, propensity score deciles, or both, comparing results to analyses using pooled individual-level data.

Main Results:

• Distributed linear regression yielded results identical to pooled individual-level data analyses across all variables and models.
• The maximum numerical difference in parameter estimates or standard errors was negligible (3×10^-11).

Conclusions:

• Distributed linear regression is a feasible and valid method for analyzing continuous outcomes in multicenter studies.
• Integrating distributed regression with propensity score modeling improves privacy protection and analytical flexibility.