Improving risk adjustment indices with drug exposure in administrative data

Abstract

Risk adjustment is essential in any study comparing patients' outcomes such as mortality and effectiveness of care. Medicaid programs would also benefit from cost risk adjustment models, as they have been moving away from a traditional fee-for-service payment system toward a capitated managed care system. Very little research, however, has been published on risk adjustment specific to Medicaid populations. Most risk adjustment methods have been based on ICD-9-CM diagnosis codes, which present with some limitations in coding comorbidities in the context of longitudinal studies. Therefore, there exist opportunities to complement code-based measures with another source of comorbidity information. In this research, we developed and independently validated Medicaid-specific prospective cost and mortality risk adjustment models based on ICD-9-CM codes, drug exposure, and combined information. We modeled mortality and cost outcomes for three populations: ambulatory Medicaid recipients, patients with a first stroke event, and patients with an initial diagnosis of Alzheimer's dementia or related dementias (AD/D). Prospective models developed on the GA Medicaid population were validated by panels of clinicians, re-estimated, 'frozen', and tested on the independent population of North Carolina Medicaid recipients. Either drug classes or ICD-9-CM codes can characterize the comorbidity burden of ambulatory, AD/D, and stroke patient populations independently, but used in conjunction with a hierarchical classification, the two sources of information increased the sensitivity to disease burden. Our prospective mortality risk adjustment models provide a tool to Medicaid programs and health service researchers to initially stratify or otherwise control for varying levels of disease severity and comorbid illnesses. A long-term goal for our prospective cost risk adjustment models is to forecast resources commensurate with actual needs of a large segment of the Medicaid population or for patient cohorts that will exact an increasing toll on Medicaid resources. However, further refinements (re-calibration) and independent testing of our disease-specific cost models may be needed before they can accurately predict future levels of resource needs in Medicaid cohorts, whereas the combined ambulatory cost model achieved good external predictive power. Drug exposure represents a new venue of information that will help enhance the quality and performance of health service research studies.