Item Response Theory Model or Test Bias.

摘要

A multidimensional non-parametric IRT model of test bias is presented, providing an explanation of how individually-biased items can combine through a test score to produce test bias. The claim is thus that bias, though expressed at the item level, should be studied at the test level. The model postulates an intended-to-be-measured target ability and nuisance determinants whose differing ability distributions across examinee group cause bias. Multiple nuisance determinants can produce item bias cancellation, resulting in little or no test bias. Detection of test bias requires a valid subtest, whose items measure only target ability. A long-test viewpoint of bias is also developed.