Importance of Capturing Risk Adjustment Data for Veterans Using Multiple Healthcare Systems

Examine whether incorporating non-Veterans Health Administration (VA) and social determinants of health (SDOH) variables into the Gagne comorbidity measure improves model performance given VA enrollees' increased use of care outside VA.

Three-year observational study (10/1/2019-9/30/2022) of two Veteran cohorts who received care in VA or through VA-purchased care ("CC"): (1) those receiving total hip/knee arthroplasties (THA/TKA) and (2) those with serious mental illness (SMI) who had behavioral health (BH) hospitalizations. Outcomes included 30-day, all-cause readmissions. We ran descriptive statistics to examine the extent to which each data source (CC, Medicare, and Medicaid) identified additional Gagne comorbidities and readmissions. We fit logistic regression models to examine the impact of adding comorbidities, outcomes, and SDOH first to VA outcomes, and then after each new data source was added. We examined model discrimination and calibration to compare predicted risk with observed outcomes.

We used administrative/claims data from VA's Corporate Data Warehouse and the Centers for Medicare and Medicaid Services (CMS). Our sample included 75,684 THA/TKA and 152,427 BH index admissions.

The greatest increase in the number of comorbidities and readmissions came from adding CC data (e.g., percent increases in comorbidities and readmissions were > 30% for both cohorts; THA/TKA, readmissions increased by 153%). Adding new data sources and SDOH variables had little effect on model discrimination; changes in c-statistics were 0.01-0.03. Both models calibrated well (e.g., mean predicted vs. observed readmissions in the first and tenth deciles for THA/TKA were 0.014 vs. 0.012, and 0.090 vs. 0.075, respectively).

Adding CC data had the greatest impact on identifying additional comorbidities and outcomes, reflecting Veterans' increased use of CC. Incorporating non-VA and SDOH variables into risk-adjusted models had minimal impact on model discrimination (individual-level predictions) and model calibration (mean predicted vs. mean observed readmissions).