Does Accounting for Community‐Level SDOH Matter for Readmission Policy? Insights from New York City

Abstract

Research ObjectiveTo inform how the measurement of community‐level social determinants of health (SDOH) at different levels of geography may affect the specification of the relationship between SDOH and hospital readmissions, as well as limitations and misclassifications that may occur at each level.Study DesignUsing a unique inpatient data set from New York State, including a subset of New York City residents and their geographic location geocoded to a block group, census tract, the NYC‐specific Neighborhood Tabulation Area (NTA), and Public Use Microdata Area (PUMA), we compare models predicting all‐cause unplanned hospital readmission rates within 30 days after inpatient discharge for three common conditions (acute myocardial infarction or AMI, heart failure or HF, and pneumonia or PN) using SDOH composite scores from variables collected at each geographic level with the baseline readmission model used by the Centers for Medicare & Medicaid Services (CMS). These models are compared on goodness of fit, statistical significance of the variables added to the baseline model, and the impact on hospitals treating high‐risk patients. SDOH data included geographic‐level variables for education, income, housing, English‐speaking, housing value, and employment.Population StudiedAll‐age and all‐payer residents of New York City that had an inpatient admission to a hospital in New York State for AMI (n = 17 994), HF (n = 53 998), or PN (n = 69 539) between January 1, 2013 and November 30, 2016.Principal FindingsCMS’s baseline readmission model performed worse than all four SDOH‐augmented models in goodness of fit. The added SDOH variables reached the highest levels of statistical significance in the model using the lowest level block group SDOH data, which also tended to have the highest levels of goodness of fit. Among all models, the model with a block group level SDOH composite showed the greatest measured performance improvement in readmission rate for those hospitals with the highest proportion of patients with SDOH high rates, compared to CMS’s current model.We also find that lower geographic levels, which have more numerous, smaller areas, have more widely distributed SDOH rates around the NYC average. As a result, when measuring SDOH at low geographic levels, more patients exhibited high‐risk levels for individual variables, compared to measuring at high geographic levels.ConclusionsNot only does including SDOH data improve model performance with statistically significant model variables and lead to changes in the distribution of hospital measurement, as compared to CMS’s readmission model, but smaller geographic composite variables tended to have even larger effects than those at larger geographies.Implications for Policy or PracticeThe absence of granular SDOH data means stakeholders face trade‐offs between precision and accessibility: While data at a higher geographic level might be less representative of the patient’s risk level, data at a more granular lever may be harder to obtain and measure accurately. This research suggests that SDOH data at more granular geographic levels provide better predictions of hospital readmission.