OBJECTIVES: To investigate statin treatment adherence, and predictors, in Taiwan incident diabetes patients. METHODS: From the Taiwan Longitudinal Cohort of Diabetes Patients (LHDB) database, we identified a cohort of incident diabetes patients aged above 40 years from year 2003—2011. The first date patients received statin after diabetes diagnosis was defined as the index date. Time varying statin treatment adherence was measured quarterly (in 3-month windows) by medication possession ratio (MPR), and patients were classified into adherent (MPR>=80%), non-adherent (MPR<80%) and treatment-stop groups (MPR=0%). We used multinomial logistic regression models with repeated measures to identify potential predictors of non-adherence or treatment-stop. RESULTS: There were 273,461 statin initiators after diabetes was first diagnosed, and 62.5% of them were adherent in the first 3-month window, then decreased gradually to about 20% at the end of first year and then stabilized during subsequent years. Results of multinomial logistic regression models (R-square: 0.3838) found baseline covariates, including age, female, cardiovascular and renal diseases were only marginal associated with non-adherence or stop, with point estimates either 10% increase or decrease. Stronger predictors for non-adherence/stop were found among time-varying covariates: more frequent lipid level tests (OR for non-adherence: 0.81; 95%CI: 0.81-0.82), treated at a higher tier of medical facilities (Medical center vs. clinics, OR: 0.64; 95%CI: 0.63-064) and lived in higher-urbanization cities were associated with lower odds for non-adherence/stop. CONCLUSIONS: The long-term statin adherence was suboptimal in Taiwan diabetes patients. With the incorporate of time-varying covariates, the efficiency of adherence prediction models were largely improved. The results of our study could be applied for future studies with long-term follow-up needed for adherence adjustment.