Risk adjustment in studies using administrative data.

Abstract

Administrative data are increasingly used to evaluate service use and cost of treatments in "real world" settings. However, the degree to which administrative data can be used to risk-adjust for differences between nonrandomized groups receiving different treatments has not been studied. This study used data from a large randomized trial to determine whether administrative data describing previous utilization and cost are as good as clinical data in predicting future resource utilization and cost. Clinical data (symptoms, quality of life, pharmacological side effects) and administrative data (inpatient utilization and total cost) were obtained from a large randomized clinical trial comparing clozapine and haloperidol. The combination of administrative measures prior to randomization was the best predictor in a multiple regression model of service use in the year after randomization (adjusted r2 = 0.27-0.31). Models using administrative data alone predicted 24 to 28 percent of the variance of service use and costs, while clinical data predicted only 6 to 7 percent of the variation in the same dependent variables. In this analysis, administrative data about previous utilization and costs were far more successful than clinical data alone at predicting future utilization and costs and are therefore likely to be better risk adjusters in studies with nonrandomized comparison groups.