OBJECTIVES: To estimate the effectiveness of statin+fibrate combination-therapy versus statin-monotherapy on cardiovascular disease(CVD) occurrence in subjects with type II diabetes in a managed care setting using appropriate econometric models dealing with selection bias in nonlinear settings. METHODS: Combination-therapy-group and monotherapy-group were identified among subjects with type II diabetes with two-years intake period(7/1/2002-6/30/2004) and three-years follow-up using administrative claims from a US health plan covering four million lives. Outcomes measure was CVD-occurrence. A univariate-probit model was developed to evaluate adjusted CVD-risk difference between groups. To control for selection bias, we used propensity score(PS) and instrumental variable(IV) method. To deal with nonlinear outcomes, we built two-stage-probit model with IV method using two-stage-residual-inclusion estimation. We used physician prescribing preference as the instrument. To test the validity of the instrument, we tested for the correlation between the instrument and treatment indicator using standard t-test. To check whether it is valid to exclude the instrument from the main equation, Wald-test was performed. Stock-and-Yogo test was used to check the weak instrument issue. To test the endogeneity of treatment indicator, we performed Hausman-test. RESULTS: Adjusting for age, gender, prior-CVD, CVD-related pharmacy-costs, Elixhauser-comorbidity, and diabetes with complication, combination-therapy-group experienced 9.1% less CVD compared with monotherapy-group at the mean of covariates(P=0.008). The results from probit and PS model were similar. For the IV model, specification-tests indicated that the validity of the instrument was satisfied. However, Hausman-test implied that treatment-indicator was not endogenous(p=0.172). CONCLUSIONS: To deal with nonlinearity issues when using IV method, we employed two-stage residual-inclusion estimation. Since we failed to identify selection bias which may be due to untestable assumptions and treatment effect heterogeneity, a univariate-probit model or PS model was used to draw conclusions. In diabetics after adjusting for known baseline differences, CVD-risk was significantly lower among subjects with statin+fibrate combination-therapy compared with those with statin-monotherapy.