Evaluation of Diagnosis-Based Risk Adjustment Among Specific Subgroups: Can Existing Adjusters be Improved by Simple Modifications?

Abstract

Managed care organization use risk adjustment systems to allocate resources and evaluate provider performance. Managers of healthcare organizations need statistical methods to determine whether a particular risk adjustment system can be applied successfully to their organization's unique population and setting, and, if not, whether simple modifications can be effective. We demonstrate methods that can be used to evaluate risk adjustment systems in populations and in subgroups within those populations. We evaluate the use of two diagnosis-based risk adjustment systems, Adjusted Clinical Groups (ACGs) and Diagnostic Cost Groups (DCGs), to explain healthcare utilization within three subgroups of veterans who used Department of Veteran Affairs (VA) healthcare services: homeless individuals, individuals with post-traumatic stress disorder (PTSD), and individuals with spinal cord disorders (SCD). ACG and DCG models are modified to better predict mean level of use for each subgroup and explain the variation in use within the group by adding indicators for each of the three conditions. Predictive ratios (PRs) and R-squares are presented within each of the subgroups for base and revised models. Both models performed well for PTSD (PRs = 0.90 and 0.95, DCG and ACG, respectively), while the DCG model fit better for SCD (PRs = 0.93 and 0.72, respectively); both models underpredicted substantially among the homeless (PRs ∼ 0.67). Adding indicators for each subgroup forces perfect prediction of mean use within subgroups and substantially improved discrimination within groups. Overall R-squares moderately improved when indicators were added.