Analysis of longitudinal Patient-Reported Outcomes with informative and non-informative dropout: Comparison of CTT and Rasch-based methods

Patient-reported outcomes (PRO) are more and more used in health sciences to evaluate concepts such as health-related quality of life. These outcomes cannot be directly observed and are often referred to a latent variable. Two psychometric theories exist for the analysis of PRO: the classical test theory, the most common used in practice and the item response theory with its most used model, the Rasch model. In many studies, PRO are collected longitudinally in order to study the evolution of the outcome through time. Missing data are frequently encountered in longitudinal studies and can be potentially informative. This study aimed at comparing Classical Test Theory (CTT) and Rasch-based approaches to analyze longitudinal PRO collected from a scale validated with a Rasch model and studying the impact of dropout, informative or not, on both approaches. Data with informative dropout have shown estimation bias and have to be analyzed with more appropriate methods. For complete data and data with non-informative dropout, a method of analysis based on the Rasch model may be preferred for the analysis of longitudinal PRO collected from a scale validated with a Rasch model due to the generally observed slight gain of power and the psychometric properties of the model.