This paper studied several termination methods in CAT, mainly four kinds. In order to test their effects, they were combined into 14 conditions. Using simulation study, four item banks were constructed; each bank was used to test the termination methods in 14 conditions. So there are 4 x 14 tests total.
Four terminated methods:
1. Fixed-length
2. Standard error
3. Minimum information
4. Convergence of ability estimation.
Four item banks:
1. 500 items, a ~ lognormal (-0.1,0.3^2), b ~ U(-4,4), c ~ mixture of uniform distributions, (0.1, 0.225)
2. 500 items, a ~ lognormal (-0.1,0.3^2), 300 items’ b ~ U(-4,4), 200 items’ b ~ N(0,1), c ~ mixture of uniform distributions, (0.1, 0.225)
3. 100 items, a ~ lognormal (-0.1,0.3^2), b ~ U(-4,4), c ~ mixture of uniform distributions, (0.1, 0.225)
4. 100 items, a ~ lognormal (-0.1,0.3^2), 50 items’ b ~ U(-4,4), 50 items’ b ~ N(0,1), c ~ mixture of uniform distributions, (0.1, 0.225)
Item selection: Fisher’s information, maximum information at each estimation.
Ability estimation: MLE
Termination rules: combined into 14 kinds.
Results showed that:
1. More items, better ability estimation
2. Too short test could not give good estimation for low ability
3. Minimum number of items 15/20
4. Minimum information method would result in too many items administered
5. Better would be to use one or more variable termination criteria in combination with a minimum number of items constraint
Should control item exposure and content balance in future study.