where xi is the binary response of subject n to item i; is the ability of subject n; and ai, bi, and ci are, respectively, the discrimination, difficulty, and pseudo-guessing parameters of item i. The 2PL model can be obtained from Equation 1 by fixing ci to 0; the 1PL model can be obtained by additionally fixing ai to 1. In this study, results of IRT-based DIF detection methods were compared and the 2PL IRT method was used to estimate ability and item parameters. 1.1 IRT-based DIF methods.
The basic idea of LRT is that item parameters should be invariant across different subgroups. In order to test item parameter invariance, likelihood of a compact model in which the parameters are constrained to be the same and an augmented model in which all variables of interest are allowed to vary between the subgroups are compared. The significance of this comparison is tested by means of the usual likelihood ratio test. Based on the selected IRT model, not only the item difficulties (1PL model), but also discriminations (2PL model), and pseudo-guessing parameters (3PL model) are allowed vary between the groups. The main idea is to compare the likelihood of two models and choose the model which has the largest likelihood. The LRT test statistic is defined as