試題差異功能（DIF）的偵測已是目前測驗理論發展的重要課題。傳統上，以試題 反應理論來分析 DIF 的作法多建立在比較兩個團體的試題參數差異上。 這種方法的先決條 件為共變數矩陣估計的準確性，可是事實上它的精確估計非常困難。再者，這種作法並不直 接估計 DIF 的大小， 因此無法深入評估各個試題的 DIF 狀況。 本研究改進了 Thissen， Steinberg 和 Gerrard （ 1986 ）的作法，直接估計 DIF 參數。這種做法建立在多向度隨 機係數多項洛基模式（ Adms，Wilson，＆ Wang，1997 ）。 電腦模擬研究的結果發現所有 參數（含 DIF 參數）的回復性相當好。我們分別以試題參數差異法、DIF 參數 z 檢定法、 概率比法等三種方法分析了性向測驗中的語文分測驗。不論就理論上還是實際上，均以概率 比法的效果最佳。

　　　　　Differential item functioning(DIF) analysis has been a major issue in test development. DIF analysis with item response theory are usually based on differences in item parameters between two groups. This approach assumes that accurate estimates of the covariance matrix are available. However, it has been shown that they are extremely difficult to compute. In addition, this approach does not directly estimate DIF, which makes the evaluation of DIF difficult. In this paper, we elaborate the work proposed by Thissen, Steinberg, and Gerrard (1986) and directly estimate DIF parameters. This approach is made possible by the multidimensional random coefficients multinomial logit model (Adams, Wilson, & Wang, 1997). Results of the simulation study show that all the parameters, including DIF parameters, were recovered very well. A real data set of a verbal subscale from an aptitude test was analyzed in three ways: item parameter difference, DIF parameter z test, and likelihood-ratio. The likelihood-ratio approach gives best results in terms of both theoretical and practical advantages.