Risk-Adjustment Models for Healthcare Provider Profiling

Abstract

End-Stage Renal Disease (ESRD) is a severe illness that is difficult to treat, and because dialysis facilities and transplant centers have a substantial influence on ESRD patient outcomes, stakeholders routinely evaluate these providers to ensure that they deliver adequate care. Fair and accurate evaluations depend on appropriate statistical methods that can adjust for confounding risk factors beyond the providers' control, such as high-risk patient populations. In this dissertation, we propose new methods to assess risk-adjustment model performance and to account for confounding effects due to unmeasured risk factors and observed cluster-level variables.

In Chapter II, we develop a Concordance Index (C-Index) to assess risk-adjustment models under left-truncated and right-censored data, a data type that commonly arises from the healthcare claims of prevalent ESRD patients. We show that conventional C-Index estimators have limiting values which depend on the distribution of the left-truncation times, and therefore produce biased estimates of the target concordance probability. The proposed C-Index, based on inverse-probability weighting, has a limiting value that is free of both the left-truncation and right-censoring distributions, and often outperforms conventional methods in terms of bias, mean-squared error, and coverage probability. We use this method to assess the risk-discrimination ability of a mortality model for ESRD patients.

In Chapter III, we propose an individualized empirical null standardization method to account for overdispersion in healthcare quality measures due to unobserved risk factors. In practice, quality measures are inevitably impacted by incomplete risk adjustment, which causes the theoretical null distributions to become misspecified and increases the probability of falsely-flagging providers for low-quality care. To overcome these limitations, we derive robust estimators of the empirical null distributions by invoking truncated-normal densities and modeling the null variance as a function of effective sample size. The proposed solution only depends on public summary statistics, circumventing common data-sharing limitations in transplant research. Motivated by urgent policy considerations in the assessments of kidney transplant centers, we build a composite scoring system, based on our empirical null methodology, to balance two key aspects of healthcare in the evaluations.

In Chapter IV, we develop a robust privacy-preserving model to estimate the effects of observable cluster-level (e.g. provider-level or geographic region-level) confounding variables, and we extend the empirical null methods from Chapter III to a Pseudo-Bayesian framework. Cluster-level confounding factors, such as geographic disparities in donor organ availability and COVID-19-related disruptions, introduce bias in transplant center evaluations, and considering that national transplant datasets are extremely large, often contain outlying providers, and are not easily shared due to patient privacy concerns, conventional models for patient-level risk-adjustment are not well-suited to estimate cluster-level confounding effect parameters. Simulations show that our proposed estimators, which do not require patient-level model fitting, are accurate and robust to outliers, and our Pseudo-Bayesian flagging method has better Frequentist properties than existing approaches. We apply these methods to adjust the Transplant Rate Ratio for geographic disparities in donor organ availability.