Development and validation of risk assessment models to predict pharmacy expenditures for both commercial and medicaid populations
Cantrell, Christopher Ron
MetadataShow full item record
Numerous risk assessment models have been developed to predict overall healthcare costs, utilization, mortality and hospital length of stay, however, there have been no models to date developed specifically to predict prescription expenditures. The objective of this research was to empirically develop a claims-based risk assessment model to predict prescription expenditures for both a commercial and Medicaid population. The models were developed using three years, 1998 through 2000, of MEDSTAT MarketScan data (commercial) and California Medicaid data (Medicaid). Both datasets are claims-based data that include medical and pharmacy claims and enrollment information in a linkable format. The MarketScan training sample used to develop the commercial models included over 1.3 million lives after the inclusion/exclusion criteria were satisfied. The California Medicaid (MediCal) training sample used to develop the Medicaid models included over 138 thousand lives after the inclusion/exclusion criteria were satisfied. A random sample of each dataset was used to validate the models. Ordinary Least Squares (OLS) was utilized to estimate the model coefficients. The primary model for this research, the Rx Cost Model (RxCost), is a diagnostic-based model that was empirically developed using diagnostic information. Another model, the Mixed Rx Cost Model (MRxCost), is a diagnostic and drug-based model that was developed to explore the gain in predictive power of supplementing the RxCost Model with drug information. The MarketScan validation sample was utilized to compare the performance of the models developed for the commercial population to each other as well as a Demographic-only model and the commercially available DCG-HCC model. The MediCal validation sample was used to compare the performance of the models developed for the Medicaid population to each other as well as a Demographic-only model, a Demographic and Medicaid eligibility model and the CDPS model. The R-square values for the commercial RxCost Model, the MRxCost Model and the DCG-HCC using the validation sample were 0.22, 0.34 and 0.16 respectively. The R-square values for the Medicaid RxCost Model, the MRxCost Model and the CDPS using the validation sample were 0.24, 0.30 and 0.04 respectively. The RxCost Model for both the commercial and Medicaid population performed better then the DCG-HCC and the CDPS in terms of R-square. The MRxCost Model for each population also performed well and resulted in a substantial gain of predictive power in terms of R-square.