This paper compares several methods of exposure control. including: 1. MI, always selecting the maximum information item; 2. 1p, using only the difficulty for selection; 3. MM, 5-4-3-2-1 method; 4. RA, always random form the 5 most informative items; 5. SH, set the control parameter k to each item via iterative simulation first, controlling the exposure of each item under a pre-set parameter, like 0.4; 6. Rk, when an item's exposure rate exceed the k, then this item could not become the candidate for the next selection, then select item based on MI; 7. PR, adding an random component to the information part, increasing the effect of information as the test progress; 8. PRk, like Rk, but in step 2 select item based on PR function, but not MI.
The exposure rate and precision are trade off. MI is most precise but poor in controlling the exposure rate. Using random selection, the exposure is well controlled, but precision is poor.
What concern for the performance of the methods includes: exposure control (maximum rate, balancing the rate of each item), estimation precision, the efficiency of test (for variable-length test), and the number of items not used in the bank (as constructing an item bank is time and effort consuming).
Procedure Advantages Disadvantages
Completely at random (CO, control condition) Select item from the bank randomly The exposure rate is best and average across the bank. No item not used. Precision is low, efficiency of the test is low. Require most item in variable-length test.
Maximum information MI Always selecting the maximum information item Precision is best, as the items used are the most informative. Short length in variable-test. Exposure rate is high, out of control. Many items not used in the bank.
One parameter method 1P using only the difficulty for selection Encourage to use items with low discrimination, decrease the usage of highly discriminating items. Could use all items in the bank. Not consider other item characteristics, which is important for item selection. Could not control the maximum rate. Precision is low, like CO.
McBride and Martin MM 5-4-3-2-1 Better control on exposure than MI. Similar with MI.
Randomesque method RA always random from the 5 most informative items Better control on exposure than MI and MM. Similar with MI.
Sympson and Hetter SH set the control parameter k to each item via iterative simulation first, controlling the exposure of each item under a pre-set parameter, like 0.4. High exposure items in the simulation would receive a low k. Well control on the exposure, precision is good. High burden to determine the k for each item. Many item not used
Restricted maximum information Rk when an item's exposure rate exceed the k, then this item could not become the candidate for the next selection, then select item based on MI Direct control on the maximum exposure rate. Using the MI for selection, could not balancing the usage of each item, leaving many items not used in the bank. Participants may receive items of different qualities, resulting in better estimation precision for some participants.
Progressive method PR adding an random component to the information part, increasing the effect of information as the test progress Increase the minimum exposure, that is, reduce the number of items not used. Not a serious decrease in test precision. Could not directly control the maximum exposure rate.
Progressive and Restricted PRk like Rk, but in step 2 select item based on PR function, but not MI Control maximum rate. Unused items are small. Precision is good. Overcome the shortcoming of Rk and PR, and provide the best overall results.