In February 2016, Focus on Hospitals, owned and maintained by the Missouri Hospital Association (MHA), became the first U.S. website to report hospital readmission data that had been formally adjusted for patients’ sociodemographic status (SDS). This effort had two main goals: (1) give patients better information about quality of care at MHA’s more than 150 hospitals; and (2) prompt a national dialogue about risk adjustment for social determinants of health, specifically emphasizing what we call “the disease of poverty.” We would like to tell the story of our achievement in this area and a bit about the research and data analysis that made it possible. First, we start with some important context.
Our Response to a National Imperative
Recent policies and communiqués from the Centers for Medicare & Medicaid Services (CMS) embrace an “upstreamist” view of improving health outcomes for socially disadvantaged populations. For example, in fiscal year 2017, CMS will begin risk-adjusting the Medicare Advantage (MA) Star Rating incentive program for differences in dual eligibility and disability status among beneficiaries in various MA+ insurance plans.
Several CMS officials are on record supporting the concept of more effective risk adjustment. Dr. Cara James, who runs the CMS Office of Minority Health, recently said in an interview that as much as 80% of health disparities are driven by social determinants of health, and that structural barriers prevent the health care system from addressing these conditions effectively. In a separate interview, Dr. Karen DeSalvo, Acting Assistant Secretary for Health at the U.S. Department of Health & Human Services (the parent of CMS) stated that “your ZIP code is more important to your health than your genetic code.”
In this context of inequality, the Hospital Readmissions Reduction Program (HRRP) of the Patient Protection and Affordable Care Act imposes financial penalties on hospitals for excess readmissions by traditional Medicare patients for specific surgical procedures and health conditions. In addition to exacting steep penalties from hospitals (about $420 million in FY2016 and a projected $523 million in FY2017), the program affects hospitals’ reputations on the CMS Hospital Compare website. The HRRP performance assessments, generated at the Yale–New Haven Center for Outcomes Research and Evaluation (Yale–CORE), aim to adjust for both patient-level (age, sex, and relevant comorbidities) and provider-level risk. In August 2014, an expert panel convened by the National Quality Forum (NQF) recommended including social determinants of health, such as SDS, in risk-adjustment models that affect performance measures. The NQF followed suit and established a 2-year trial period to evaluate SDS risk adjustment in national quality reporting and incentive programs.
The MHA and its data company, the Hospital Industry Data Institute, have taken action. Specifically, we have used models developed by Yale–CORE to develop a new methodology for reporting 30-day risk-standardized readmission rates and ratios at MHA’s member hospitals.
Our Models and Their Results
We developed models that adjust for both clinical risk (building on work by CMS and Yale–CORE) and several SDS factors, including the socioeconomic status of the patient’s U.S. census tract, the patient’s Medicaid status, and community-level variables designed to account for differences in access to post-acute care, nutrition, and transportation. The 6 outcomes of interest were acute myocardial infarction (AMI), congestive heart failure (CHF), pneumonia, chronic obstructive pulmonary disease, total-hip and total-knee arthroplasty, and hospital-wide readmissions among adults during a 36-month period ending in May 2015.
Compared with the standard CMS–Yale models, our SDS-enriched models yielded significant relative reductions in the range of risk-standardized readmission ratios for each of the six outcomes, from a 35% reduction for the AMI cohort to an 80% reduction for the total-hip and total-knee replacement cohort. Overall, SDS enrichment best improved the 30-day readmission assessments at hospitals that served higher concentrations of Medicaid patients and higher-poverty communities.
The SDS-enriched models also made it easier to predict which patients — in all 6 cohorts — had higher risk for readmission (as measured by observed readmission rates). For example, evaluating the CHF cohort by estimated risk deciles showed that the CMS–Yale model for CHF explained just 18% of the variation in observed readmissions, compared with 63% of the variation explained by the SDS-enriched model. The single largest SDS-related improvement in predicting readmission in the CHF cohort was recorded by a safety-net hospital in Missouri’s impoverished southeastern region, consistent with the finding that the SDS-enriched models were most useful in hospitals that serve poor communities.
A common concern is that including SDS factors in risk-adjustment models may mask actual variation in quality, thereby imposing what some have called “the soft bigotry of low expectations.” However, for each of the 6 outcomes, our SDS-enriched models actually produced a greater number of hospitals with risk-adjusted ratios that were significantly higher or lower than expected, compared with what the CMS–Yale models produced. Clearly, SDS-enriched models are not explaining away all of the hospital variation.
The scatterplot focuses specifically on the CHF cohort: The CMS–Yale approach identified six hospitals with statistically significant assessments (5 higher than expected, 1 lower), whereas the SDS-enriched approach identified eight (5 higher than expected, 3 lower). Our findings suggest that accounting for SDS does not mask or excuse actual differences in quality among hospitals — it makes them more apparent by revealing meaningful differences. The data, we believe, have two implications for policy: (1) more-accurate risk adjustments and reporting on quality and (2) useful progress toward eliminating disparities in access and health outcomes.
Why Risk Models Must Change
Our data show that adjustment for nonclinical SDS factors makes a risk model better at predicting which patients will actually be readmitted within 30 days after discharge. SDS enrichment also significantly reduces the variation in estimated readmission performance assessments, compared with standard CMS–Yale model specifications that are used to penalize hospitals under HRRP rules. Furthermore, our SDS-enriched models include factors that hospitals cannot easily modify under the existing set of parameters. According to some CMS officials, those parameters create structural barriers to reducing disparities in community-based care.
In its January 2016 Guide to Preventing Readmissions Among Racially and Ethnically Diverse Medicare Beneficiaries, CMS openly acknowledged the higher readmission rates for racial and ethnic minorities, as well as patients with social complexities. The Guide makes clear that two hospitals of equal quality but unequal SDS mix will face different penalties under the HRRP. This disconnect between CMS’s rhetoric and policy has recently enticed Congress to insert itself into the debate.
Nevertheless, NQF’s 2-year trial period to assess the influence of social determinants on readmissions has yielded unanimous recommendations, by measure developers at Yale–CORE, for continued exclusion of these factors despite strong, statistically significant effects of the tested SDS variables. The CHF model, for example, was tested and recommended for endorsement with continued exclusion of SDS variables, even though those variables showed stronger risk prediction than did diabetes and malnutrition.
More recently, in their article “Accounting for Patients’ Socioeconomic Status Does Not Change Hospital Readmission Rates,” Yale–CORE researchers found that including SDS risk adjustments would reduce by 3% to 4% the number of safety-net hospitals that the HRRP penalizes. That may seem like a small reduction, but it was achieved by including only one variable indicative of median income by ZIP code — and it would result in fewer penalties for 65 hospitals that treat low-income patients and typically operate on razor-thin margins. For those hospitals, the title of the Yale–CORE researchers’ article does not seem apt: Accounting for patients’ socioeconomic status does change hospital readmission rates.
And in an independent study of neighborhood-level factors on health outcomes for CHF patients, Yale–CORE researchers found a significant association between readmission risk and the socioeconomic status of patients’ communities. Results from that study and from our analysis of the MHA’s SDS-enriched readmission models reveal the significant effect of social determinants and community variables on health outcomes. If hospitals’ ability to mediate these factors is limited, or if structural barriers prevent hospitals from addressing the upstream effects of social factors on health outcomes, it is common sense to account for them in determining penalties under the HRRP. Without such change, the disease of poverty will fester.
This article originally appeared in NEJM Catalyst on August 31, 2016.