Distributed Quasi-Poisson regression algorithm for modeling multi-site count outcomes in distributed data networks.

Area of Science:

• Computational Statistics
• Health Informatics
• Epidemiology

Background:

• Real-world data from multiple institutions improve generalizability but raise privacy concerns for patient-level data sharing.
• Distributed regression methods are needed for multi-institutional analyses, particularly for count outcomes.
• Count data often exhibit overdispersion, requiring specialized statistical models.

Purpose of the Study:

• To introduce a novel, privacy-preserving computational method for distributed quasi-Poisson regression.
• To enable analysis of count outcomes across multiple institutions without sharing patient-level data.
• To address limitations in current distributed regression methods for overdispersed count data.

Main Methods:

• Developed a one-shot distributed algorithm for quasi-Poisson regression (ODAP) using a surrogate likelihood approach.
• ODAP requires only aggregate data sharing and minimal communication rounds (at most three).
• Evaluated ODAP via simulations and two proof-of-concept real-world data analyses, comparing it to meta-analysis and pooled regression.

Main Results:

• ODAP demonstrated negligible error relative to pooled regression estimates across simulated scenarios.
• Meta-analysis estimates showed increased variability with higher outcome heterogeneity, unlike ODAP.
• In real-world data, ODAP estimates closely matched pooled estimates, outperforming meta-analysis in certain scenarios.

Conclusions:

• ODAP provides a communication-efficient, privacy-preserving solution for distributed quasi-Poisson regression on multi-institutional count data.
• The method yields estimates comparable to pooled regression and superior to meta-analysis, especially with low counts and high heterogeneity.
• ODAP enhances the ability to study rare outcomes and improve generalizability in multi-site observational studies.