Distributed Statistical Analyses: A Scoping Review and Examples of Operational Frameworks Adapted to Health...

Area of Science:

• Health analytics
• Statistical inference
• Distributed data analysis

Background:

• Learning health systems require multi-organizational data, but ethical concerns limit traditional data pooling.
• Distributed algorithms offer alternatives but may not fit all health data frameworks.
• Heterogeneous and unevenly distributed data pose challenges for standard statistical methods.

Purpose of the Study:

• To overview statistical inference methods for horizontally partitioned data.
• To describe and assess generalized linear model (GLM) methods for health data.
• To adapt existing methods for practical use in heterogeneous health settings.

Main Methods:

• A scoping review identified 6 GLM approaches for horizontally partitioned data.
• Statistical theory adapted methods for uneven sample sizes and heterogeneous distributions.
• Workflows and algorithms were developed to detail information-sharing needs.

Main Results:

• Six standard GLM analysis approaches were identified, assuming data homogeneity.
• New statistical procedures were derived to handle data heterogeneity and uneven sample sizes across nodes.
• Developed workflows and algorithms clarify operational complexities and information exchange.

Conclusions:

• This work provides adapted methods for analyzing heterogeneous, horizontally partitioned health data.
• Clarified workflows and exchanged quantities enhance the usability of these methods in health analytics.
• Further research is needed on the confidentiality of shared summary statistics.