|
Abstract:
|
Healthcare policymakers, particularly those testing innovations at the Centers for Medicare and Medicaid Services, have always been aware that policy interventions do not affect all individuals equally. This has led to an increased interest in subgroup-specific impact estimates, so policymakers can understand under what settings interventions are most effective. Traditional approaches to subgroup analysis estimate a different model for each subgroup of interest, leading to increased noise, spurious statistical significance (type-I errors), and exaggerated significant estimates (type-M errors). We present a hierarchical Bayesian model that estimates all subgroup-specific impacts simultaneously, borrowing strength across subgroups as appropriate. We demonstrate how to set up hierarchical priors that are suitable to reflect the structure of the data and the effects of interest. We also address several practical challenges that arise when the approach is applied to data from large healthcare policy evaluations, which routinely include millions of records. Our approach is motivated by and applied to the evaluation of a large CMS Model Test of a healthcare service delivery intervention.
|
