Comparison of statistical methods for the analysis of patient-reported outcomes in randomised controlled trials: A simulation study

Patient-reported outcomes (PROs) that aim to measure patients' subjective attitudes towards their health or health-related conditions in various fields have been increasingly used in randomised controlled trials (RCTs). PRO data is likely to be bounded, discrete, and skewed. Although various statistical methods are available for the analysis of PROs in RCT settings, there is no consensus on what statistical methods are the most appropriate for use. This study aims to use simulation methods to compare the performance (in terms of bias, empirical standard error, coverage of the confidence interval, Type I error, and power) of three different statistical methods, multiple linear regression (MLR), Tobit regression (Tobit), and median regression (Median), to estimate a range of predefined treatment effects for a PRO in a two-arm balanced RCT. We assumed there was an underlying latent continuous outcome that the PRO was measuring, but the actual scores observed were equally spaced and discrete. This study found that MLR was associated with little bias of the estimated treatment effect, small standard errors, and appropriate coverage of the confidence interval under most scenarios. Tobit performed worse than MLR for analysing PROs with a small number of levels, but it had better performance when analysing PROs with more discrete values. Median showed extremely large bias and errors, associated with low power and coverage for most scenarios especially when the number of possible discrete values was small. We recommend MLR as a simple and universal statistical method for the analysis of PROs in RCT settings.