OBJECTIVE: The objective of this study was to compare the ability of risk stratification models derived from administrative data to classify groups of patients for enrollment in a tailored chronic disease management program. SUBJECTS: This study included 19,548 Medicaid patients with chronic heart failure or diabetes in the Indiana Medicaid data warehouse during 2001 and 2002. MEASURES: To predict costs (total claims paid) in FY 2002, we considered candidate predictor variables available in FY 2001, including patient characteristics, the number and type of prescription medications, laboratory tests, pharmacy charges, and utilization of primary, specialty, inpatient, emergency department, nursing home, and home health care. METHODS: We built prospective models to identify patients with different levels of expenditure. Model fit was assessed using R statistics, whereas discrimination was assessed using the weighted kappa statistic, predictive ratios, and the area under the receiver operating characteristic curve. RESULTS: We found a simple least-squares regression model in which logged total charges in FY 2002 were regressed on the log of total charges in FY 2001, the number of prescriptions filled in FY 2001, and the FY 2001 eligibility category, performed as well as more complex models. This simple 3-parameter model had an R of 0.30 and, in terms in classification efficiency, had a sensitivity of 0.57, a specificity of 0.90, an area under the receiver operator curve of 0.80, and a weighted kappa statistic of 0.51. CONCLUSION: This simple model based on readily available administrative data stratified Medicaid members according to predicted future utilization as well as more complicated models.
OBJECTIVE: The objective of this study was to compare the ability of risk stratification models derived from administrative data to classify groups of patients for enrollment in a tailored chronic disease management program. SUBJECTS: This study included 19,548 Medicaid patients with chronic heart failure or diabetes in the Indiana Medicaid data warehouse during 2001 and 2002. MEASURES: To predict costs (total claims paid) in FY 2002, we considered candidate predictor variables available in FY 2001, including patient characteristics, the number and type of prescription medications, laboratory tests, pharmacy charges, and utilization of primary, specialty, inpatient, emergency department, nursing home, and home health care. METHODS: We built prospective models to identify patients with different levels of expenditure. Model fit was assessed using R statistics, whereas discrimination was assessed using the weighted kappa statistic, predictive ratios, and the area under the receiver operating characteristic curve. RESULTS: We found a simple least-squares regression model in which logged total charges in FY 2002 were regressed on the log of total charges in FY 2001, the number of prescriptions filled in FY 2001, and the FY 2001 eligibility category, performed as well as more complex models. This simple 3-parameter model had an R of 0.30 and, in terms in classification efficiency, had a sensitivity of 0.57, a specificity of 0.90, an area under the receiver operator curve of 0.80, and a weighted kappa statistic of 0.51. CONCLUSION: This simple model based on readily available administrative data stratified Medicaid members according to predicted future utilization as well as more complicated models.
Authors: Marc B Rosenman; Ann M Holmes; Ronald T Ackermann; Michael D Murray; Caroline Carney Doebbeling; Barry Katz; Jingjin Li; Alan Zillich; Victoria M Prescott; Stephen M Downs; Thomas S Inui Journal: Milbank Q Date: 2006 Impact factor: 4.911
Authors: Mary E Charlson; Martin T Wells; Balavenkatesh Kanna; Van Dunn; Walid Michelen Journal: BMC Health Serv Res Date: 2014-11-14 Impact factor: 2.655