Registries, network databases, and the electronic health record commonly impose a hierarchical structure in which patients are clustered in ways that may be relevant to analysis. Although propensity score methods are popular for confounding adjustment, how to appropriately extend these methods to clustered or multilevel data is unclear.
