To administer a computerized adaptive test (CAT), the parameters of items in a pool should be obtained in previous from the calibration with a collected sample. In the past, the point estimates of parameters are treated as true values and directly imported into the CAT system, suggesting the standard errors are totally ignored. Recently the algorithm of computing the standard error of person ability was modified which considered the point estimates and standard errors of items simultaneously. Previous studies discussed the issue of calibration error mainly focus on fixed-length CAT, so the study implemented the same methodology on variable-length CAT. Three independent variables were manipulated: IRT models, termination criteria, and calibration sample sizes, respectively, and three dependent measures were observed at different ability levels: average test length, empirical bias of ability estimate, and the percentage of correctly classified examinees, respectively.
1. Sample size is the critical factor of determining the calibration error of items, which is mainly manipulated in this study. However, it is more likely that the inaccuracy of item parameter is due to ignoring possible impacts (e.g. testlet effect, DIF, and test-taker's low motivation) and fitting improper models.
2. The values of bs for the 3PLM in Table 2 are too large to be acceptable. Are they typos?