Missing data considerations for patient reported outcome measures in randomized controlled trials

Patient-reported outcome (PRO) measures are widely used in clinical trials to capture patients’ perspectives on their health and treatment. However, missing data remains a major challenge and can lead to biased or incomplete results. Understanding how different methods for handling missing data perform under realistic conditions is essential for decision-making. We conducted a simulation study informed by a review of 87 PRO instruments, modeling both simple and complex questionnaire structures in a hypothetical randomized trial. Under both the missing at random (MAR) and missing not at random (MNAR) assumptions, four commonly used approaches were evaluated: multiple imputation (MI) at the item level, MI at the total score level, item mean, and mixed model for repeated measurements (MMRM). To make the study more realistic, we also ran simulations using actual trial data from the COU-AA-302 study, focusing on the Functional Assessment of Cancer Therapy–Prostate (FACT-P) questionnaire. The simulation results supported the use of the “Half Rule” under MAR. When the number of missing items did not exceed half of the total items, item mean imputation showed reasonable estimation with low root mean squared error, high power, and low type 1 error. MI on item scores consistently performed well, but inflated type 1 error under certain MNAR scenarios. As the missing visit rate increased, the advantage of item level imputation diminished, i.e., RMSE approached and eventually exceeded that of MI on the total score. In the trial data analysis, given the high missing visit rate in the collected FACT-P data, MI on total score level performed not worse than MI on item score level. No single method is optimal for all situations. Item-level imputation performs well when missing data are minimal and occur at random, but it becomes less reliable when data is missing not at random. For complex PROs or studies with many missing visits, imputing at the total score level with the “Half Rule” offers a practical alternative. Strategies should match the missingness pattern, supported by simulations and sensitivity analyses to ensure robust results.