A straightforward extension was presented that a multilevel sampling structure of examinees was incorporated into the Rasch testlet model, so that a total 4-level model was formed. In both analyses, four models, including the Rasch model, the Rasch model with multilevel structure, the Rasch testlet model, and the proposed model, were fitted to generated or real data and subsequently their results were compared. The results consisted with the old conclusion summarized on page 94: fixed-effect model parameter estimates are the best linear unbiased estimators and the random-effect model parameters are the best linear unbiased predictors.
--------------
1. The N in Table 1 and 2 actually denoted the number of items.
2. Some details in respect of simulation design were skipped. Readers who unfamiliar with the process of conducting a simulation have to guess what they did by themselves.