OBJECTIVES: To compare the doubly robust estimation and propensity score methods in a diabetes-related comparative health economics study. METHODS: We used large US claims databases that covered 115 million patients between 2000 and 2011. Patients with type 2 diabetes with continuous health insurance were selected into the intensification cohort or non-intensification cohort depending on whether they intensified basal insulin with insulin aspart or remained on treatment with basal insulin without bolus insulin. The study measured the patients’ demographic and clinical characteristics, as well as their healthcare utilization and medical costs in a 6-month pre-baseline period and a 12-month post-baseline period. Doubly robust estimation and propensity score methods were used to examine the impact of intensification on direct medical costs per patient per month (PPPM). RESULTS: The study included 26,016 qualified patients with 2,860 in the intensification cohort and 23,156 in the non-intensification cohort. Baseline characteristics were similar between the two cohorts (mean age=58 years, percentage male=52%). Both cohorts also had similar prevalence rates of diabetes-related conditions, and similar utilization and medical costs in the pre-baseline period. The propensity score method indicated that the intensification cohort spent $297.4 PPPM less in the post-baseline period than the non-intensification cohort (p<0.05), while the doubly robust estimation method estimated a $399.0 PPPM cost reduction due to treatment intensification (p<0.05). The doubly robust estimation method was able to control all confounding factors also controlled by the propensity score method, as well as additional confounding factors not controlled by the propensity score method. CONCLUSIONS: Although both the propensity score and doubly robust estimation methods revealed cost savings due to treatment intensification with insulin aspart ($297.4 vs. $399.0 PPPM), the propensity score method was less robust against unobserved confounding factors than the doubly robust estimation method.