OBJECTIVES: Propensity score (PS) is widely used in observational study. It is usually estimated over entire study time period. However, pattern of medication use is changed over time. The PS which is estimated without taking into account time might be biased. Little is known about the effects of time on propensity score. This study aims to determine the differences of PS estimated by conventional PS and time-specific PS approaches and their effects on estimates of treatment effects.  METHODS: A retrospective database analysis at a University-affiliated hospital in Thailand was used. Patients who aged 18 years or older and were diagnosed as diabetes during July 2008 – June 2011 with receiving glucose lowering agent were included. All included patients were categorized into thiazolidinedione (TZDs) and non-TZDs groups. PS were estimated by conventional and time-specific approaches. In the time-specific approach, PS was separately estimated into 3 groups by calendar time. The pair t-test was used to compare the PS of both approaches. Patients were matched using caliper nearest neighbor matching with no replacement. The multivariate Cox proportional hazard model was used to determine the adjusted hazard ratio of cardiovascular hospitalizations of TZDs and non-TZDs in matched cohorts.  RESULTS: A total of 2165 patients were included in this study. Patients were on average 58.8 ± 12.7 years of age with 44.5% of male. The average conventional PS was 0.198, which was significantly lower than  the average time-specific PS of 0.209.For CVD-related hospitalization, the adjusted hazard ratio (HR) of conventional PS-matched cohort was 1.05 (95% confidence interval (CI); 0.38 – 2.89), while that of time specific PS-matched cohort was 1.12 (95%CI; 0.43 – 2.92). CONCLUSIONS: Propensity scores estimated by conventional and time-specific approaches were different. The different PS approaches could lead to different estimates of treatment effects.