OBJECTIVES: To empirically develop and validate the RxCost Model, a prospective and concurrent risk assessment model that uses claims-based diagnostic information to predict future pharmacy expenditures for a US commercial population. Additionally, we sought to empirically develop, validate, and compare the Mixed RxCost (MRxCost) Model to explore the gain in predictive power associated with adding drug information to the RxCost Model. Prescription cost risk assessment models can be used to profile physician practices or control for comorbidity burden in economic studies. METHODS: A retrospective longitudinal cohort study using MEDSTAT MarketScan claims data (1998 - 2000) for ambulatory persons who were continuously enrolled for at least 13 months and were 18 to 64 years old was used. A training sample consisting of over 1.3 million lives was utilized to develop the models. Model coefficients were developed from AHRQ clinical classification software, clinical expert panel, and stepwise OLS regression to screen noise variables. A random holdout sample of 218,383 was utilized to validate the models and to compare the performance of each model. Measure of discrimination (R-squared), predictive ratios, and discrimination for hypothetical physician groups were computed and compared to each other as well as to a Demographic-only model and the proprietary DCG-HCC model. RESULTS: The R-square value for the prospective RxCost, the MRxCost Model and the DCG-HCC using the validation sample was 0.22, 0.34 and 0.16, respectively and was 0.34 for the concurrent RxCost model. The RxCost model's predictive ratio's varied between 0.93 and 1.05 for clinical subgroups and ranged from 1.03 to 1.04 across hypothetical physician patient groups of size 10 to 500. CONCLUSIONS: The RxCost Model was successfully developed and it outperformed the DCG-HCC model in terms of R-square after re-calibrating the DCG-HCC model. The MRxCost Model also proved that supplementing drug information can improve discriminatory power.