|
Abstract:
|
The availability of national level or registry medical databases provides a tremendously rich data resource for patients, clinicians, or policy makers to make informed decision. Propensity score (PS) technique is a widely adopted to address heterogeneity or selection bias in many retrospective studies based on those database. This technique has been majorly applied to two group comparisons in practice. The general propensity score (GPS) theory developed Imbens (2000) further extends application to multiple group comparisons. Among several available PS approaches, such as stratification, inverse probability weighting, PS as covariates, and matching, the matching approach provides easier and more intuitive interpretation, but its extension to multiple group comparisons is limited 3 groups based on the latest literature. In this work, we proposed a new algorithm that breaks such limitation with minimum computation burden. Simulation studies are conducted to explore its optimal operation characteristics, as well as its relative performance comparing to other PS approaches when 3 or more groups are under comparison. A NCDB case study illustrates its usage and interpretation.
|
