OBJECTIVES: To compare two statistical approaches for analyzing longitudinal data with missing observations: 1)imputation using Last Observation Carried Forward method (LOCF) and 2)Mixed-effects Model Repeated Measures method (MMRM) to analyze the change from baseline in health-related quality of life (HRQoL) by medication adherence level.  METHODS: HRQoL via SF-12 Health Survey and medication adherence via a 5-level categorical response was measured monthly for one-year for 184 patients in a U.S. multiple sclerosis observational study. HRQoL was summarized in two continuous variables: Physical Component Score (PCS-12) and Mental Component Score (MCS-12). Categorically collected medication adherence was converted to numeric values and average compliance was calculated over a 1-year period then categorized into two groups:  ≥ 90% (GT90) or < 90% (LT90) compliant. For validity of compliance, patients who had completed at least 6 measurements during 1-year on compliance question were included. LOCF used the last available change from baseline to impute the missing values for early drop-out. MMRM is a likelihood-based approach which models all actual observations jointly, with no attempt at imputation for missing values. RESULTS: A total of 131 patients were included in this analysis. The 12-month change from baseline in PCS-12 comparing patients with GT90 compliance vs. LT90 compliance using MMRM was 0.86 (p=0.277) and using LOCF was 1.30 (p=0.339).  For MCS-12, the improvement among patients with GT90 compliance over LT90 compliance using MMRM was 2.04, while the corresponding improvement using LOCF was 1.97.  For MCS-12, only the MMRM method produced statistically significant improvements (p-values: LOCF=0.234, MMRM=0.026). CONCLUSIONS: MMRM and LOCF yielded not only different results but also different statistical significance in the 12-month change from baseline in MCS-12. Since the approach to estimate and model is different between two methods, the pattern and shape of data must be investigated to find the right method to produce valid estimates.