Comparison of different approaches in handling missing data in longitudinal multiple-item patient-reported outcomes: a simulation study.

Background: Patient-reported outcomes (PROs) are important clinical outcomes widely used as primary and secondary endpoints in clinical studies. However, PRO data often suffers from missing values for various reasons, which pose challenges to data analysis. This simulation study aimed to compare the performance of existing state-of-the-art approaches in handling missing PRO data.

Methods: Using a real and complete multiple-item PRO dataset, we generated various missing scenarios with different missing rates, mechanisms, and patterns. The performances of eight methods were compared, including a mixed model for repeated measures (MMRM) with and without imputation at the item level, multiple imputation by chained equations (MICE) at the composite score and item levels, and three control-based pattern mixture models (PPMs) and the last observation carried forward (LOCF) imputation at the item level.

Results: We found that the bias (i.e., deviation of the estimated from the true value) in the treatment effect estimates increased, and the statistical power diminished as the missing rate increased, especially for monotonic missing data. Item-level imputation led to a smaller bias and less reduction in power than composite score-level imputation. Except for cases under missing-not-at-random mechanisms (MNAR) and with a high proportion of patients' entire questionnaire missing, MMRM imputation at the item level demonstrated the lowest bias and highest power, followed by MICE imputation at the item level. The PPM methods were superior to the other methods under MNAR mechanisms.

Conclusions: PPMs imputation at the item level was preferable for MNAR, whereas MMRM and MICE imputation at the item level were better for other scenarios. These findings provide valuable insight for selecting appropriate methods for handling missing PRO data.