Development of a New Risk-Adjustment Procedure for Primary Care Panels in the Veterans Health Administration

Background. The Veterans Health Administration (VHA) system provides comprehensive care for nearly 6 million Veterans. Primary Care is provided by multidisciplinary Patient Aligned Care Teams (PACTs), under a model similar to primary care medical homes outside the VHA. The expected panel size for each PACT is currently calculated using a formula that adjusts for available resources and is pro-rated by the provider full-time equivalents (FTEs) on a team. Panel sizes are then risk-adjusted using a statistical model that generates a Primary Care Intensity Score (PCIS) for each patient, representing the patient’s expected workload for PACTs. We sought to develop a parsimonious model that was less biased by accounting for endogeneity of patient demand for care and providers’ practice styles.

Methods. We randomly sampled a quarter of panels at each VHA facility (1,099 total panels) to construct an analytical sample of all patients assigned to these panels continuously between July 1, 2015 and June 30, 2017 (N=1,006,964 patients). We extracted data on demographic (age, sex, race/ethnicity, marital status), and clinical characteristics (conditions classified by the Hierarchical Condition Categories, HCCs and mental health conditions) at baseline (July 2015-June 2016), as well as Veterans’ observed primary care and acute-care utilization in the VHA during both the baseline and outcome period (July 2016-June 2017). The outcome variable of interest was the number of visits to primary care clinics. We estimated ordinary least squares (OLS) regression models and generalized linear models (GLM) with various error distributions and link functions. Model predictive ability was assessed based on the 5-fold cross-validated adjusted R-squared, root mean squared prediction error (RMSPE), and mean absolute error (MAE). We selected predictors using lasso regression, which shrinks some predictor coefficients to zero in order to achieve a parsimonious model and protect against overfitting. The models were used to predict individual risk scores for each empaneled patient. We also corrected for bias by adjusting for panel fixed effects in the model training phase while leaving out fixed effects coefficients in the score prediction phase.

Results. The model selected using the lasso procedure had 28 patient-level predictors and explained 37.8% of the variation (RMSPE 1.96, MAE 1.26), as opposed to 38.1% using the saturated models which included all 192 predictors (RMSPE 1.95, MAE 1.25). The most important predictor was the number of baseline primary care visits, which on its own predicted over 36.2% of the outcome variation. On the other hand, demographic and clinical characteristics had little predictive power. The panel fixed effects explained an additional 5 percentage points and produced moderate bias correction. The new score distribution led to 45% of the clinics to have their panel sizes adjusted upwards, 20% to have their panels unchanged, and 35% adjusted downwards. Panels were adjusted by between -16% (highest nanile of the distribution) to + 10% (lowest nanile).

Conclusion. We present a novel method of risk-adjustment that balances high predictive ability with bias correction for the endogeneity between supply and demand-side effects. Our preferred model is also parsimonious and prevents overfitting due to sample variability.