Good-enough risk adjustment models for physician payment and performance assessment

The Affordable Care Act moved diagnosis-based risk adjustment to center stage, both by embracing risk adjustment for health insurance exchanges and by encouraging payment innovation and improved performance measurement. Almost all envisioned reforms, including Affordable Care Organizations (ACOs), Patient Centered Medical Homes (PCMHs), and Pay-for-Performance, simply assume the existence of adequate risk adjustment models.

We have previously developed and evaluated risk adjustment models for paying primary care practices and others for assessing their performance (Ash and Ellis, 2012; Ellis and Ash, 2012; Vats et al. 2013). All models were developed with the health IT firm Verisk Health and use its DxCG-HCC classifications. One health plan has used our Primary Care Activity Level (PCAL) model to pay PCMH practices since 2009; Massachusetts Medicaid (MassHealth) will use PCAL models for payment in 2015.

First, we present the rationale for calculating bundled payment models for primary care, using existing (“good enough”) claims data. We contrast the principles used to develop these models with principles we used previously in developing CMS-HCC models for Medicare Advantage risk adjustment. We also discuss: using OLS versus nonlinear estimators, overfitting, topcoding, weighting, lags, adding new socioeconomic variables, and PCP assignment.

We used 2006-2007 MarketScan Commercial Claims and Encounter data to estimate payment models and assign patients to primary care practitioners (PCPs). To study the effects on PCP payments for their panels required developing plausible PCP “pseudo-panels.” For example, we randomly assigned the 29% of patients who saw no PCP over two years to practices in their county of residence, in proportion to the numbers of assignable patients in these practices. We developed the model on 17.4 million patients and evaluated it among the 1.67 million patients who could be assigned to 436 practices with 500 to 5000 patients each.

We examined nine service bundles – reflecting a range of narrow to broad expectations for practice responsibilities –  of potential use in bundled PCP payment, evaluating each for: predictiveness, implied financial risk to PCPs, and payment stability over time. Three alternative weighted sums of services were developed to approximate the demands on PCPs for managing patients’ primary care needs.  Among the practice-assigned patients, our preferred PCAL model using 653 parameters achieves an R² of 67% in average spending at the individual level and 82% at the Pseudo-PCP level.

Most outcomes are affected by patient risk. Thus, we developed models to predict both traditional (e.g., total cost, ED visits) and novel (e.g., prescriptions for antibiotics of concern) outcomes for risk adjusted performance assessment.

Diagnosis-based risk adjustment models can substantially improve predictions relative to demographic models, and can explain over 50 percent of practice-level variation in concurrent data; high accuracy is important for maintaining appropriate payments and performance assessments for PCP practices with 500-5000 patients.

Our goal is to provide practical details to encourage others to develop and evaluate the additional good-enough models needed to support ACA-envisioned payment and quality assessment reforms.