چكيده فارسي :
Most conventional differential item functioning (DIF) detection procedures assume that the test under study is unidimensional. Any dimension beyond that primary dimension is considered construct-irrelevant and DIF is said to occur when an item measures a secondary dimension and performance of the observed groups (e.g., males vs. females) of similar abilities differs significantly on the secondary dimension. However, items may be designed intentionally to measure more than one dimension and differences on the secondary dimension show strengths and weaknesses of the test takers rather than malfunctioning of the items. Application of conventional DIF detection procedures which match groups according to observed total scores or unidimensional latent ability measures identify these dimensions as nuisance (construct-irrelevant) and if performance of different observed groups differ on these dimensions on any given item, the respective item is flagged for DIF. Thus for multidimensional tests when the matching criterion is the total score or unidimensional ability measure, more items are likely to be flagged for DIF when they only measure a relevant aspect of a latent variable space. DIF within the framework of the diagnostic classification models (DCMs) reduces inflated estimates of DIF by using the attribute profile of the test takers as the matching criterion. The present study used the DCM DIF procedure (Hou, de la Torre, Nandakumar, 2014) to study gender DIF in the reading comprehension section of the university entrance examination (UEE). To this end, item response data of about 7000 test takers who took the UEE to seek admission into the English programs at BA level were selected. Further, the item-by-attribute relationships for the test were adopted from Ravand (in press) and the attribute profiles of the test takers were used as the matching variable. The results of the DIF analysis showed that item parameters in 8 out of 20 items were significantly different across males and females (hence flagged for DIF) as indicated by the P- values ( .05). Since it was suspected that the results were affected by the large sample size, effect sizes for all the items which were flagged for DIF were calculated. It was found that just one of these items showed moderate gender DIF. To compare the DCM DIF results with the conventional DIF detection procedures, Mantel-Haenszel DIF was also estimated. The results showed that two of the items exhibited DIF. As more items were flagged for DIF by the Mantel-Haenszel DIF method, the hypothesis that the DCM DIF may reduce the estimates of DIF was supported
