插補法在檢測試題差異功能的效果

The Effects of Imputation Methods on the Detection of Differential Item Functioning

本研究為一模擬研究，旨在探討不同遺失比率、插補法對偵測差別試題功能（DIF）的影響，藉以了解插補法在回復資料完整性後，是否有助於DIF 試題的偵測。研究中比較四種處理遺失值的方法（整筆刪除、零插補、雙向插補、多重插補），並以兩種DIF 偵測程序（MH、SIBTEST）進行檢驗，在不同DIF 強度及遺失比率下，檢視DIF偵測的型一誤差率與檢定力，藉以評估插補法在DIF偵測的效益。研究結果顯示，整筆刪除會降低DIF 偵測之型一誤差率與檢定力，DIF 強度可提高DIF偵測之型一誤差與檢定力，MH與SIBTEST的偵測結果具一致性，不同插補方法對回復DIF偵測之型一誤差與檢定力有不同效果，零插補與多重插補在中度DIF 有較佳的回復性，在重度DIF 則是在少量遺失比率的回復性較佳，雙向插補的表現較不穩定。

The study is intended to investigate the impact of missing data rate and imputation method on the detection of DIF. The study compares the performances of four imputation methods (listwise deletion， zero imputation， two-way imputation， and multiple imputation w/NORM) and two DIF detection methods (Mantel-Haenszel and SIBTEST) under three missing rates (10%， 20%， and 30%) and two DIF magnitudes (0.5 and 0.8)， using Type I error and the statistical power of DIF detection as comparison criteria. Several conditions are set as follows: a sample size of 1，000 used for both the focal and reference groups; missing completely at random set for the missing data mechanism; and uniform DIF throughout the study. Findings show that the listwise deletion method results in lower Type I error rate as well as power. When the DIF magnitude is medium， Type I error and power can be boosted， or namely recovered， by imputing the missing data with zero imputation and multiple imputation. In short， each imputation method has different efficacy. Furthermore， there is no much difference between the two targeted DIF detection methods， Mantel-Haenszel and SIBTEST， in terms of Type I error rate and power.