More Predictive Modeling of Total Healthcare Costs Using Pharmacy Claims Data: Adherence and Boosted Regression
- Presenter:
Chair: Joel Hay; Discussant: TBA Tue June 6, 2006 15:30-17:00 Room 225
OBJECTIVE A variety of disease state classification systems derived using claims data are used in risk stratification and predictive modeling applications. Algorithms based solely on pharmacy claims data have the advantages of timeliness, cleanliness, and availability, while still being robust to predicting prospective healthcare outcomes and costs relative to their integrated medical and pharmacy counterparts. Following on Powers et al. (2005), this study expands Pharmacy Health Dimensions (PHD) to include controls for patient adherence to drug therapy. The study also evaluates the use of boosted regression as an alternative to other econometric approaches for predicting commonly right-skewed and leptokurtotic healthcare cost data.
METHODS: Using 2003 and 2004 data from a large health plan (N=369,985), PHD is used to identify participants having any of five diseases: diabetes, congestive heart failure, asthma, hypercholesterolemia, and hypertension. A split-sample design is employed to train and validate the predictive models. For the base model, total healthcare costs in 2004 are estimated as a function of age, gender, 2003 pharmacy costs, and comorbidities using PHD. Subsequent models add two variables of patient adherence to drug therapy, a commonly used measure of compliance, Medication Possession Ratio (MPR), and a novel gauge of persistency, Maximum Gap in Therapy. It is hypothesized that MPR would be negatively related to future total healthcare costs, while Maximum Gap in Therapy would be positively associated with future total healthcare costs. Furthermore, the magnitude of these effects should vary substantially by disease state. Several econometric modeling techniques are explored included ordinary least squares (OLS), log-OLS, two-part modeling (with a generalized linear model second part), and boosted regression. Standard metrics of predictive model fit and accuracy are reported.
RESULTS: As expected, the inclusion of the two measures of adherence increased the proportion of explained variation in all five of the disease state models. Furthermore, MPR was significantly (p<0.10) and negatively related to future total healthcare costs in all models except for hypercholesterolemia and hypertension. Maximum Gap in Therapy was positively associated with future total healthcare costs in all model (p<0.10). As in the prior study (Powers et al., 2005), results from OLS regressions were comparable to those of log-OLS and two-part modeling approaches in terms of their validation R2, positive predictive value, and specificity. Boosted regressions, however, provided the most accurate predictive model of future total healthcare costs for several of the disease states.
CONCLUSION: Predictive modeling of future total healthcare costs using pharmacy claims only offers several advantages over using integrated medical and pharmacy data. The predictive power of basic models of demographic and disease classification can be enhanced by adding measures of patient compliance and persistency. The magnitude and significance of the relationship between adherence to drug therapy and future total healthcare costs varies substantially by health condition.
REFERENCE: Powers, C.A., C.M. Meyer, M.C. Roebuck, and B.Vaziri. 2005. Predictive Modeling of Total Healthcare Costs Using Pharmacy Claims Data: A Comparison of Alternative Econometric Cost Modeling Techniques. Medical Care 43(11): 1065-1072.
