The author considers the missing data, heterogeneous variances in each time point, normal or nonnormal data, and ML or REML in repeated measurement. It tries to give a overall condition of all possible repeated measurement models and use those information criteria to investigate their utility. Simulation studies were given and no real data were fit. Results show that no any information criterion outperforms than each other.
1. The paper uses so many sorts of information criteria to do model comparison. But I have no expectation which information criterion will be the better one in each simulated condition. When I am watching those tables, I do not know how to figure out.
2. No real data was given and fit. I have no idea what is the structure of data can be appropriate to the basic formula (1). What I saw is symbols, not a practical application.
3. I’m not familiar with the area of repeated measurement. As for data generation, I’m curious about the process of generation. Why does it work when we follow his procedure as listed in paper. That is, how can it generate the data of the repeated measurement.
4. As for table 3, how to calculate the statistical power? What is the procedure? Is it assumed that we know the true model and then use those information criteria to find out the true model. Finally calculate the percentage of correctness.