An Exploration of the Robustness of Four Test Equating Models

This study examined how four commonly used test equating procedures (linear, equipercentile, Rasch Model, and three-parameter) would respond to situations in which the properties or the two tests being equated were different. Data for two tests plus an external anchor test were generated from a three parameter model in which mean test differences in difficulty, discrimination, and lower asymptote were manipulated. In each case two data sets were generated consisting of responses of 2,000 examinees to a 35 item test plus 15 item anchor test. Each test equating case was comprised of 4,000 examinees and 85 items. The robustness with respect to violations of assumptions was tested for the linear and Rasch equating methods. For equipercentile equating, the results showed how the method responded to various conditions for the three-parameter model, the study primarily tested LOGIST's simultaneous estimation procedure. Results indicated that equipercentile equating was very stable across the cases studied. Linear and Rasch model equating were very sensitive to violations of their models' assumptions. Results for the three-parameter model were disappointing. Paying close attention to test item properties was advised when selecting an equating method. (BS)