OBJECTIVES: Missing data occur in clinical trials and can have a negative impact on the results. In this study, we (1) outlined concepts related to missing patient reported outcomes (PRO) data, including patterns of missingness, missing data types (missing completely at random [MCAR], missing at random [MAR], missing not at random [MNAR]), and commonly used analytic strategies that address missingness (multiple imputation [MI], mixed model repeated measures [MMRM]), and (2) investigated the potential impact of missing data on PRO results in clinical trial settings.

METHODS: A simulation study was conducted to empirically evaluate the impact of missing PRO data in clinical trials. The simulation aligned with a hypothetical clinical trial where a PRO instrument was captured for N=500 subjects (n=250 per treatment group) at 3 visits (baseline, 3 months, 6 months) and explored various types (MCAR, MAR, MNAR) and rates (0%, 20%, 40%) of missingness. Both MI and MMRM were used.

RESULTS: Simulation results showed that when the missing PRO data were MCAR, descriptive statistics and model-based within and between-treatment group estimates were unbiased. However, when the PRO data were MAR, unbiased descriptive statistics were only obtained using MI, while unbiased model-based estimates were obtained using MI and MMRM. Findings also demonstrated that, regardless of the analytic strategy used (MI, MMRM), biases arose when missing data were MNAR. Supplemental simulation analyses suggested that capturing the reason(s) for missingness and integrating this information into MMRM as a covariate may help to reduce bias.

CONCLUSIONS: Using MMRM and MI, researchers can handle missing data that are MCAR or MAR, but these approaches may not adequately address missing data that are MNAR. Future work could focus on designing a method for capturing information regarding reasons for missing data and developing analytic strategies that can leverage these insights to accurately characterize treatment effects.