Matching One Sample According to Two Criteria in Observational Studies

Area of Science:

• Statistics
• Biostatistics
• Health Services Research

Background:

• Multivariate matching aims to create comparable treated and control groups in observational studies.
• Traditional methods struggle to achieve both covariate distribution balance and individual pair homogeneity when dealing with many covariates.
• Existing techniques often require separate approaches for balancing covariate distributions and forming homogeneous matched pairs.

Purpose of the Study:

• To introduce a novel polynomial-time multivariate matching method that achieves both group covariate distribution similarity and individual pair homogeneity.
• To generalize and improve upon existing multivariate matching techniques.
• To apply the method to a real-world study on nursing care and sepsis mortality in the Medicare population.

Main Methods:

• A new method utilizing minimum cost flow optimization on a tripartite graph is proposed.
• The tripartite graph structure allows simultaneous optimization for balancing covariate distributions (right side) and matching close pairs (left side).
• The method generalizes existing approaches and can minimize the earthmover distance between marginal distributions.

Main Results:

• The proposed method successfully addresses the dual goals of multivariate matching, which are often in conflict with traditional methods.
• The approach is computationally efficient (polynomial-time).
• The method was applied to a study investigating the association between nursing quality and sepsis mortality in Medicare beneficiaries.

Conclusions:

• The new tripartite graph-based minimum cost flow method offers a unified and effective approach to multivariate matching.
• This method enhances the ability to construct robust treated and control groups for observational research.
• The match2C package in R is available for implementing this advanced matching technique.