As pointed in the abstract, many statistical criteria for model comparison were introduced in this study, and their performances were evaluated under different conditions. A scenario was considered that, for most longitudinal studies, missing responses are commonly observed, and the missingness result in unbalanced repeated measurements. By manipulating types of mean and covariance structures, number of repeated measurements, sample size, population distribution shape, and estimation methods, listed criteria were computed and compared. The results matched the findings of previous studies: none of the criteria is the superiority under all conditions.
It is noticeable that on page 25 a dropout model was introduced to generate MAR responses. The model can be easily extended to generate MNAR responses in simulated sequential data.