This presentation is part of: Poster Session - Public Health; Methodological Advances
COMBINING PROPENSITY SCORE WITH CLASSIFICATION AND REGRESSION TREE TO EVALUATE THE EFFECT OF INSULIN TREATMENT FOR TYPE II DIABETES
PATRICK THIEBAUD, PHD, University of Southern California, Parmaceutical economics and Policy, SAN DIEGO, CA, Michael B Nichol, PhD, University of Southern California, Pharmaceutical Economics and Policy, Los Angeles, CA, and BIMAL PATEL, PHARM.D, MEDIMPACT, INC., SAN DIEGO, CA.
Purpose: To estimate the effect of insulin treatment on the health care utilization of type II diabetes patients by developing a parsimonious propensity score-based model. Method: The data consists of the medical and pharmaceutical insurance claims for 379 type II diabetes patients. The estimation procedure uses a classification and regression tree (C&RT) analysis to determine the optimal set of explanatory variables for the propensity score weighted least square regression. The propensity score is calculated with logistic regression. This approach reduces the number of predictors to a small, manageable number without relying on any assumptions about the distributions of predictors or interactions between predictors. In addition, C&RT can determine a split value for continuous predictors, like age. Models relying on the logistic regression stepwise selection method are contrasted with models built with C&RT, and the results regarding the effect of insulin treatment are compared. Results: Logistic regression models created with C&RT have a higher sensitivity and specificity than parametrically constructed models, for the same number of predictors. Both specification methods produce similar results, namely reductions in total cost and drug cost. The C&RT-based models, however, show a larger effect for insulin treatment on health care utilization, particularly by patients over the age of 53. This may indicate that C&RT-based models can better correct for selection bias. Conclusion: Combining the C&RT and propensity score methods produces more accurate and parsimonious models than parametric model selection alone.
