Provider profiling is used to gauge the level of care that is being offered across providers. The extensive repercussions of provider profiling on the health care system emphasizes the need to ensure that the statistical tools and methods used to generate these estimates of provider performance are appropriately capturing the actual quality of care at each provider. The objective of this dissertation is to propose improvements to current profiling methods that enable patients, providers, and policymakers to better estimate and understand the level of care in a healthcare system. Statistical methods in missing data and causal inference are used to accomplish this objective. For the first aim, methods for imputing race/ethnicity in an incomplete dataset based on multiple imputation are developed and implemented. This is valuable because patient-level racial and ethnic information is needed to identify racial/ethnic disparities in health care systems and may be used for provider profiling in risk-adjustment models that include sociodemographic variables. The second aim focuses on a novel approach for profiling providers that uses Bayesian mixture models to identify similar clusters of providers based on the admission characteristics of the patients they treat. Upon implementing this approach and grouping providers, within-group assessments of provider performance can be made. Because provider profiling can also be thought of as a comparison of multiple interventions or treatments, the third aim develops a flexible multiple imputation method for the estimation of causal effects in the multiple treatment setting that can be applied to estimate a wide range of provider performance metrics. As evidenced through the extensive simulation studies and results presented in this dissertation, the accomplishment of these aims provides the tools needed to better capture quality of care.
