The use of risk adjustment in provider reimbursement arrangements has increased as alternative payment arrangements are becoming more widespread in health insurance. Risk adjustment has been used by Medicare Advantage and managed Medicaid plans to reimburse health plans for the unique risks and populations in their care. More recently, as carriers have transferred utilization risk to providers through alternative payment arrangements, such as global budgets and bundled payments, risk adjustment has been used to reflect a provider’s patient’s severity.
However, many risk adjustment methodologies were developed using a standard population representing a combination of adults and children. Adults comprise a larger proportion of the average population, and as a consequence, the disease states recognized in these methodologies were optimized with greater emphasis on adults. Since a chosen risk adjustment methodology should reflect the characteristics of the underlying patient population organizations such as children’s hospitals, pediatric provider groups, and health plans that enroll a large proportion of children, health organizations have begun to question these standard risk adjustment models. These groups argue that there are fundamental differences in clinical profiles, patient mix, treatment options, and patient management needs between the pediatric population and the general population.
In a recent Milliman research paper, we built a control model for a standard commercial population and a model optimized for a pediatric population from the Truven Health Analytics MarketScan® [1] database using the open source hierarchical condition categories (HCC) [2] system to compare results. We limited our focus to New England States [3] and developed a concurrent [4] risk adjustment model with 184 disease classifications based on the HCC system.
The R-Squared value [5] in our control model is 58% on the standard population. This is very similar to the reported R-Squared values for many commercially available concurrent risk adjusters [6]. However, if we remove the adults from this population, our model’s R-Squared reduces significantly to 45%.
To improve upon the control model’s R-Squared of 45% for pediatric-only populations we developed the pediatric-only model through an iterative process using only the pediatric population included in our MarketScan database sample:
- We began the modeling at the DxGroup level that underlies the HCCs. There are 784 DxGroups in the original HCC classification system.
- We modeled DxGroups with more than 30 patients separately and left those with less than 30 patients in their original HCCs.
- We created two-way and three-way disease interactions for inclusion in the model (e.g., diabetes and chronic obstructive pulmonary disorder (COPD) would be included as an additional explanatory variable, in addition to diabetes alone, and COPD alone). We calculated the sample size of each and retained only those that had at least 30 patients in a cell.
- We regrouped DxGroups and disease interaction terms with statistically insignificant coefficients (at a 5% significance threshold) with the other small-cell DxGroups in the same HCC and recalculated their coefficients (risk weights).
- We reset the coefficients of DxGroups and disease interaction terms with statistically significant but negative coefficients to zero. Negative coefficients often imply a confounding variable; if left in the model, they will produce spurious relationships among conditions. From a payment perspective, negative coefficients result in reduction in payment for diagnosing or treating a condition, which does not have face validity either.
- We repeated steps (4) and (5) until all variables left in the model had statistically significant and non-negative coefficients. This resulted in 570 DxGroups/HCC categories.
This pediatric risk adjustment model results in an R-Squared of 58% on pediatric populations, which is a significant improvement from the control model’s R-Squared of 45%. This increase in statistical fit will impact the financial results of organizations bearing financial risk for pediatric populations. For example, using the pediatric-only model on children in the data used to develop our model, results in a risk score that is approximately 1.5% higher than the control model developed for a standard population.
In alternative payment models which use risk adjustment to distribute payments to providers, models calibrated using a standard population could result in inequitable reimbursement to providers specializing in pediatric populations. As a result, these providers should carefully review the risk models used in any alternative payment arrangement before participation.